Professional Documents
Culture Documents
Dictionary
of
Statistics
and Research
Methods
APA
Dictionary
of
Statistics
and Research
Methods
Sheldon Zedeck, PhD
Editor in Chief
The citation for this publication is American Psychological Association. (2014). APA dictionary
of statistics and research methods. Washington, DC: Author.
DOI: 10.1037/14336.000
1 dedicate this volume to the memory of my parents, Judith and Hyman; to
the support and love from my brother Morris and his wife Ellen; to my
mentors and colleagues who influenced my data analytic and research
perspectives, particularly Carol Vale, Robert M. Guion, Patricia Cain
Smith, Edwin E. Ghiselli, and Geoffrey Keppel; to my wife Marti, our
children, and their spousesCindy and Jason Singer, Jason and Stacey
Skura Zedeck, and Tracy; and to my grandchildren, Molly, Ella, Lilly,
Aidan, and Noah, all of whom have played a statistically and practically
significant role in my life.
Sheldon Zedeck
Preface ix
Editorial Staff xv
About the Editorial Board xvi
Quick Guide to Format xvii
Appendixes
Abbreviations and Acronyms 423
Entry Illustrations 428
Overview of Research Design 431
Considerations
Symbols 432
Preface
The APA Dictionary of Statistics and Research Methods builds on a strong core
of lexicographical reference works published in the past seven years by APA
Books. APA inaugurated this collection with the parent APA Dictionary of
Psychology (2006)the culmination of some ten years of research and
lexicographic activity. A resource of 25,000 entries, it won critical
endorsement from both the publishing and reference library communities.
Two derivative works followed almost immediately: the abridged APA
Concise Dictionary of Psychology (2008; available both in print and as an app
for iPhone, iPad, and Android), whose 10,000 entries were selected for a
more general readership, and the student version, the APA College Dictionary
of Psychology (2009), which features 5,000 entries that are essential for
psychology advanced placement high-school students and university-level
psychology majors alike.
A third derivative (and the first With a subdisciplinary focus) followed in
2012; the APA Dictionary of Clinical Psychology (available in print and via
Kindle). Its 11,000 entries directly address the needs of clinical
psychologists, whether they work in health and mental health clinics, in
independent and group practices, or as consultants to professionals in such
fields as medicine, law, social work, and consumer relations. The content
focuses on clinical training, clinical supervision, and the diagnosis,
treatment, and prevention of emotional and behavioral disorders, with
definitions describing the biological, developmental, social, and individual-
difference underpinnings of health and mental health.
A fourth derivative, again subdisciplinary in nature, was published in
2013: The APA Dictionary of Lifespan Developmental Psychology is specifically
tailored to offer scholars and students balanced coverage (in some 7,500
entries) in such core areas as developmental theory; genetics; and the
biosocial, cognitive, and psychosocial perspectives of development.
Entries span all the stages of life, from birth through childhood;
through adolescence; and through the early, middle, and late periods of
adulthood.
To some degree, editorial work on each of the derivatives noted above
resulted in changes to APA's overall collection of dictionary terms, whether
those changes are revisions or updates of entries that appeared in the parent
dictionary or whether they are entirely new entries added to offer a more
complete representation of the evolving lexicon of psychology.
The APA Dictionary of Statistics arid Research Methods, however, has been
something of a more consciously advanced enterprise in the development
of our series. Although partially derived from the original corpus of terms in
the parent dictionary, the work may more accurately be seen as APA's first
endeavor to create an entirely new reference in this subareaone that
stands on its own and one that reaches out beyond psychology in its
applicability to allied social, behavioral, and health sciences.
Future Perspectives
It is commonplace among those in the field of lexicography to assert that a
dictionaryany dictionary, but especially a first editionis a work-in-
progress, the only limitations upon the future of which are the publishing
company's continued interest in keeping the work alive and current and its
ability to locate and task the appropriate people to provide solid content
and stylistic expertise for revised and new editions.
We therefore issue this first edition of the APA Dictionary of Statistics and
Research Methods with an eye to these potential limitations and with the
recognition that, although all parties have collectively done the best they
couldgiven the temporal imperatives of scholarly publishingthere is
always room for correction and improvement. We are, in fact, proud of our
work and send it out into the world confident that it is a very strong start,
that it offers greater and better coverage than the competition of which we
are aware, and that we fully intend to keep the work a living project in
future generations.
To assist us in this task, we invite you to contact APA Reference at
books@apa.org to note errors of omission, inaccuracies, infelicities of
phrasing, new vocabulary, and omitted senses. As always, we welcome your
thoughtful appraisal and suggestions.
Acknowledgments
The development and production of a dictionary requires an extraordinary
amount of effort and collaboration with many participants. Our
experiences as the editor and as publisher of this dictionary benefitted
immensely from an incredible group of colleagues and staff who devoted
many hours in guidance, consultation, demonstrated patience, fortitude,
and energy. It is difficult to express our true appreciation to the many
participants in a short space, but, to start, we want to acknowledge the
associate editors: Shelley Blozis, Lisa Harlow, and Abigail Panter. This team
contributed invaluable wisdom to the generation of the plan for the
dictionary, spent endless hours reviewing potential entries, and spent even
more hours writing and revising definitions. We also want to acknowledge
the contribution of the consulting editors, Dan Bauer and Jay Myung, for
their efforts in reviewing entries for possible inclusion in the dictionary. On
the project genesis and development side, we owe gratitude and thanks to
Ted Baroody (Director, Reference, APA Books) and to Trish Mathis (Senior
Reference Development Editor, APA Books), who worked with the editorial
team from day one, helping to generate the structure for the dictionary,
reviewed drafts on process and procedure, and provided wise feedback on
any issue requested throughout the project; they performed their tasks with
grace and support that resulted in the final product being a personally
rewarding partnership. Finally, we thank our consummately professional
U.K. editorial and production team at Market House Books. We want to
thank all of these participants for their support, encouragement, and
wisdom.
Sheldon Zedeck, PhD
Editor in Chief
Editor in Chief
Sheldon Zedeck, PhD
Associate Editors
Lisa L. Harlow, PhD
Shelley A. Blozis, PhD
A. T. Panter, PhD :
Consulting Editors
Daniel J. Bauer, PhD
Jay Myung, PhD
Akaike's i n f o r m a t i o n c r i t e r i o n
AIC) a SUMMARY STATISTIC used in com- Abbreviation
Cross-reference paring the relative IGOODNESS OF FIT| of
two or models for a given set of data,
while taking into account the number
of parameters in each model. The model
with the lowest AIC is considered the
best among all models specified. |[Hirot-
sugu Akaike (1927-2009), Japanese stat-
Etymology istician]
autocorrelation In] the situation in Part-of-speech label
which values of a variable measure over
time are correlated with other values of
the same series separated from them by a
specific interval. This often occurs with
economic or demographic data. Auto-
correlations are generally assumed to be
linear relationships and may be presented
Hidden entries graphically in an autocorrelogram| (or
I correlogram)|or formulaically in an auto-
correlation function (ACF). Also called] se-
Alternative name rial correlation.!See TIME-SERIES ANALY-
SIS.
I 1
ability level
the person weighed one pound more which a random sample is chosen from a
than last week. If the absolute value of larger group of items and used to make a
the weight difference was taken, it would decision about the quality of the items
simply reveal a difference of 1 pound, in that group. Acceptance sampling is
without any indication of whether it was often used in commerce to test the qual-
plus or minus. Also called modulus. ity of merchandise in a batch: A random
absolute zero a value on a measure- sample from the batch is inspected and
ment scale that denotes the complete the results used to determine whether
absence of the measured characteristic. the batch as a whole meets desired stan-
A RATIO SCALE has an absolute zero, dards or whether it fails to meet stan-
whereas an INTERVAL SCALE does not. dards and should be rejected as
Also called true zero. ; defective. Also called lot acceptance
sampling.
absorbing state a condition from
which there is no possibility of transi- accessible adj. in a MARKOV CHAIN, de-
tioning to another condition. In a se- scribing a state / that there is a possibility
quence of events such as a MARKOV of reaching from another state i in some
CHAIN, for example, a state is absorbing number of steps. accessibility n.
if there is zero probability of leaving that accidental s a m p l i n g see CONVE-
state once it has been obtained. Com- NIENCE SAMPLING.
pare TRANSIENT STATE.
accelerated f a i l u r e time model in accrual rate the rate at which some-
SURVIVAL ANALYSIS, a model in which thing accumulates. For example, the ac-
the response variable is a known trans- crual rate of volunteers to participate in
formation of the time until the event of a research study may be documented
interest occurs. When this transforma- and used to help determine how best to
tion is the LOGARITHM, the model is re- recruit individuals for similar studies in
ferred to more specifically as a censored the future.
linear regression model. In aging re- accuracy n. 1. lack of error or BIAS in a
search, for example, an investigator measure; the more accurate the measure,
might use an accelerated failure time the closer the measurement is to the
model to evaluate whether a specific in- TRUE SCORE for an individual. 2. a mea-
tervention increases lifespan. sure of performance on a task, usually
accelerated longitudinal design defined as the proportion of correct re-
see COHORT-SEQUENTIAL DESIGN. sponses. accurate adj.
acceleration n. in mathematics and sta- ACE abbreviation for ALTERNATING CON-
tistics, the rate of change in the SLOPE,of DITIONAL EXPECTATION.
a function or the rate of change in one
ACES abbreviation for ACTIVE CONTROL
variable as a function of an increase in a
EQUIVALENCE STUDY.
second variable. Compare DECELERATION.
acceptance region the range of values ACF abbreviation for AUTOCORRELATION
for a test statistic that leads to accep- function.
tance of the NULL HYPOTHESIS, such that acquiescent response set see YEA-
the ALTERNATIVE HYPOTHESIS Is rejected SAYING.
as a valid explanation for observed data.
Compare CRITICAL REGION. a c t i o n research socially useful and
theoretically meaningful research devel-
acceptance-rejection method see
oped and carried out in response to a so-
REJECTION METHOD. ;
cial issue or problem, results of which
acceptance s a m p l i n g a process in are applied to improve the situation
active control equivalence study
ING; that is, an alternative hypothesis two or more groups on one or more IN-
posits meaningful differences or relation- DEPENDENT VARIABLES. In other words,
ships between the variables under inves- it is a statistical method of studying the
tigation; responses of different groups to a DE-
alternative hypothesis distribu- PENDENT VARIABLE that adjusts for the
t i o n a theoretical set of plausible values influence of a variable that is not being
of a characteristic under certain assurnp- investigated but nonetheless is related to
tions that is compared to its correspond- the dependent variable and thus may in-
ing NULL DISTRIBUTION in the process of fluence the study results. An analysis of
conducting a POWER ANALYSIS. covariance is appropriate in two types of
cases: (a) when experimental groups are
alternative test f o r m see ALTERNATE suspected to differ on a background-
FORM. correlated variable in addition to the dif-
A M L abbreviation for ASYMMETRIC MAX- ferences attributed to the experimental
IMUM LIKELIHOOD.
treatment (i.e., the analysis corrects for
chance differences between groups that
analogue observation a response re- arise when participants are assigned ran-
corded from a participant in an ANA- domly to the treatment groups) and (b)
LOGUE STUDY designed to induce a where adjustment on a covariate can in-
particular behavior in a controlled envi- crease the precision of the experiment
ronment, such as a laboratory or clinic. (i.e., reduce the ERROR TERM). For exam-
Compare NATURALISTIC OBSERVATION; ple, suppose a researcher analyzes
SELF-MONITORING OBSERVATION. whether there is a difference in learning
among three types of instructionin-
analogue sample a group of individu-
class lecture, online lecture, and text-
als selected for inclusion in a study who
book only. He or she divides a random
possess the symptoms of a particular dis-
selection of adult students into three
order but who have not sought treat-
groups, implements the different instmc-
ment for the disorder. For example, a
tion types, and administers the same test
researcher may use college students with
to all parficipants to determine how much
mild problems resembling those seen in
they learned. If the researcher knows
outpatient clinical settings and expose
each participant's educational back-
them to specific therapy techniques to
ground, he or she could use an analysis
study their responses.
of covariance to adjust the treatment ef-
analogue study a study intended to in- fect (test score) according to educational
duce a particular behavior under con- level, which would reduce the observed
trolled environmental conditions that variation between the three groups
resemble or approximate the real-worM caused by variation in education levels
situation of interest. Examples include rather than by the instruction itself.
the use of hypnosis, drugs, and sensory
deprivation in a laboratory setting to in- analysis of covariance structures a
duce brief periods of abnormal behavior method of examining the relationships
that simulate those of psychopathologi- among a set of variables with regard to
cal conditions. Also called analogue how each one varies according to varia-
design; analogue experiment; ana- tion in the others. It is similar to STRUC-
logue research. TURAL EQUATION MODELING but disting-
uished by its emphasis on simultaneous
analysis of covariance (ANCOVA) an variation (i.e., covariance). Also called
extension of the ANALYSIS OF VARIANCE covariance structure analysis.
that adjusts for the influence of a CO-
VARIATE in testing whether there is a sig- analysis o f unweighted means see
nificant difference between means of UNWEIGHTED MEANS ANALYSIS.
analysis of variance
analysis o f variance (ANOVA) a sta- anchor test a set of test items used as a
tistical method of studying the variation reference point in comparing ALTERNATE
in responses of two or more groups on a FORMS of a test. One alternate form is ad-
DEPENDENT VARIABLE. A N O V A S test for ministered to one group of participants,
significant differences among the mean another is administered to a different
response values of the groups and can be group, and the items comprising the an-
used to isolate both the joint INTERAC- chor test are administered to both groups.
TION EFFECTS and the separate MAIN EF- Scores on each alternate form are then
FECTS of INDEPENDENT VARIABLES. compared with scores on the anchor test.
analysis o f weighted means a A N C O V A acronym for ANALYSIS OF CO-
method of comparing samples of un- VARIANCE.
equal sizes in which the data for each A n d e r s e n - G i l l model an extension of
group are averaged in a manner that the cox REGRESSION ANALYSIS model
takes into account the differing number that is used in SURVIVAL ANALYSIS for data
of observations contributing to the cal- that show LEFT CENSORING, time-varying
culation. See WEIGHTED AVERAGE. COVARIATES, recurrent events, and dis-
analysis u n i t see UNIT OF ANALYSIS. continuous intervals of risk. [Per Kragh
Andersen (1952- ), Danish statistician;
a n a l y t i c approach any method based Richard D. Gill (1951- ), Brifish-born
on breaking down a complex process Dutch mathematician]
into its parts so as to better understand
A n d e r s o n - D a r l i n g test see DARLING
the whole. For example, a researcher
TEST. [Theodore W. Anderson (1918- ),
studying cognition in children might
U.S. mathematician; Donald A. Darling
identify such component skills as prob-
(1915- ), U.S. mathematician]
lem solving, reasoning, imagination,
and memory; examine each in turn; and Andrews plot a method for displaying
then detemine how they interact. Com- multidimensional data using only two
pare S Y N T H E T I C A P P R O A C H . dimensions in which one curve is gener-
ated for each row of data, as in the hypo-
a n a l y t i c i n d u c t i o n a QUALITATIVE RE- thetical illustration overleaL Here f(t) is
SEARCH strategy for developing and test- a function derived for an observation
ing a theory in which the researcher (e.g., occupation type) that is based on
tentatively defines a phenomenon, cre- multiple variables (e.g., education re-
ates a hypothesis to explain it, and ex- quired, starring salary, promotion oppor-
amines a single specific occurrence of tuniries). The different values of t indi-
the phenomenon in order to confirm or cate the similarity of the three occupa-
refute the hypothesis. If the hypothesis tions shown. [David F. Andrews, Cana-
is confirmed, additional cases are exam- dian statistician]
ined until a sufficient degree of certainty
about the correctness of the hypothesis a n d r u l e see MULTIPLICATION RULE.
is obtained and the study may be con- anecdotal method an investigational
cluded. If the hypothesis is not con- technique in which informal verbal re-
firmed, the phenomenon is redefined or ports of incidents casually observed are
the hypothesis revised so as to accom- accepted as useful information. The an-
modate the findings. ecdotal method is scientifically inade-
quate but can offer clues as to areas of
anchor n. a number or descriptive qual- investigation that warrant more system-
ity used as a point of reference for mak- atic, controlled research.
ing a subjective judgment. For instance,
a study participant may be asked to rate angular t r a n s f o r m a t i o n see ARC
his or her health on a scale that uses an SINE TRANSFORMATION.
upper anchor labeled perfect health. a n i m a l rights the belief that animals
10
Anscombe residual
n n \ \ r
0.2 0.4 0.6 0.8 1.0
Andrews plot
should be treated with respect and be senting the most important DESCRIP-
free from exploitation and abuse by hu- TIVE STATISTICS obtained from an ANAL-
mans. Animal research is monitored in YSIS OF VARIANCE (ANOVA). It typically
many universities and other organiza- provides the DEGREES OF FREEDOM, the
fions by an animal care committee, whose SUM OF SQUARES, the MEAN SQUARE, the
purpose is to ensure the humane care F RATIO, and the SIGNIFICANCE LEVEL as-
and use of animals by assuring compli- sociated with each effect studied, as in
ance with federal and state regulations the example below.
and by supporting veterinary oversight
and continuing education and training. Ansari-Bradley test a NONPARA-
METRIC procedure for determining the
a n o n y m i t y n. a principle of research equivalence of two samples without re-
ethics stating that the identity of a study quiring that the variables of interest
participant should remain unknown. have a NORMAL DISTRIBUTION. Rather, it
Relatedly, CONFIDENTIALITY applies to assumes the middle values of the distri-
situations in which a participant's iden- butions are equal and evaluates whether
tity is known but should not be dis- the DISPERSIONS are equal, whether the
closed by the researcher. scores within each sample are independ-
A N O V A acronym for ANALYSIS OF VARI- ent, and whether the shapes of the dis-
ANCE.
tributions are identical.
Source df SS MS F P
Lecture topic (L) 2 1,200 600 21.18 < .001
Presentation method (?) 1 2,200 2,200 77.66 <.001
L X P interaction 2 700 350 12.35 <.001
Error 60 1,700 28.33
of error or unexplained variance in a re- a data point may fall. For example,
sponse variable that does not require the in displaying ages from 1 to 100, a re-
values of the response variable to follow searcher may use intervals such as 1-10,
a NORMAL DISTRIBUTION. It is used with 11-20, and so forth. The values deter-
GENERALIZED LINEAR MODELS. [Francis J. mining the intervals (e.g., 1 and 10) are
Anscombe (1918-2001), British statisti- the apparent limits.
cian]
a p p l i e d research studies conducted to
antecedent variable any variable that solve real-world problems, as opposed to
precedes a response variable. For ex- studies that are carried out to develop a
ample, in REGRESSION ANALYSIS it is the theory or to extend basic knowledge. Ex-
INDEPENDENT VARIABLE, in a MEDI- amples include ACTION RESEARCH and
ATIONAL PROCESS it is the intervening EVALUATION RESEARCH. Compare BASIC
variable, and in PATH ANALYSIS it is the RESEARCH.
variable beginning a chain of causal links.
Compare CONSEQUENT VARIABLE. applied science research conducted to
serve a practical human purpose rather
antedependence n. a relationship than with the purpose of extending
among multiple values of a response knowledge for its own sake. Compare
variable observed over time such that BASIC SCIENCE.
each value depends upon one or more
previous values. See AUTOCORRELATION. a p p l i e d statistics the use of statistical
methods and procedures to understand
A - o p t i m a l design a research design in data in psychology, sociology, econom-
which the set of INDEPENDENT VARI- ics, and other disciplines. Compare THE-
ABLES causing the least average variance ORETICAL STATISTICS.
in a DEPENDENT VARIABLE is chosen
from a larger pool of possible variables apprehensive-subject role behavior
for predicting values of that variable. See displayed by a research participant who
also D-OPTIMAL DESIGN; E-GPTIMAL DE- dislikes being evaluated or is otherwise
SIGN.
anxious about the study but nonetheless
tries to do well and convey a positive
a posteriori denoting conclusions de- impression. Compare FAITHFUL-SUBJECT
rived from observations or other mani- ROLE; GOOD-SUBJECT ROLE; NEGATIVIS-
fest occurrences: reasoning from facts. TIC-SUBJECT ROLE.
Compare A PRIORI. [Latin, "from the lat-
ter"] approximate bootstrap c o n f i -
dence method a procedure in which
a posteriori comparison see POST many subsets of an observed sample
HOC COMPARISON. data set are used to produce a range of
apparatus n. any instrument or equip- plausible values for a population charac-
ment used in an experiment or other re- teristic, such as the MEAN or VARIANCE.
search. It is less computationally involved than
similar BOOTSTRAPPING methods.
apparent error rate the inaccuracy
that can be observed when applying a approximation n. the process of ob-
REGRESSION EQUATION hypothesizing taining a value that is at least close to the
associations among a set of variables to a desired or actual value. For example, one
set of actual data points. An estimate of might round a measurement to the near-
how well the model fits the data, the ap- est decimal place for ease of subsequent
parent error rate often underestimates calculations. The degree of inaccuracy
the true error rate. inherent to this process is known as ap-
proximation error.
apparent l i m i t the lower or upper
bound of a CLASS INTERVAL within which a p r i o r i denoting conclusions derived
arithmetic mean
13
arithmetic progression
15
ATI
17
Bb
b a c k g r o u n d variable see SUBJECT neously summarizing the data for both
VARIABLE. and showing the degree of relationship
between them. It includes a filled square
back-to-back stem-and-leaf plot a or other shape indicating the MEDIAN of
variation of a STEM-AND-LEAF PLOT in the scores, a bag or shaded polygon en-
which the trailing digits (leaves) from circling the middle 50% of scores, a loop
two data sets are displayed on either side or differently shaded polygon encircling
of the same central column of initial val- the remaining data points, and a fence
ues (stem). The leaf values from one data or curve marking the edge of the loop
set are given on the right-hand side beyond which any OUTLIERS are dis-
while those of the second group are shown played. The DISPERSION of the data
on the left-hand side. For example, con- points is indicated by the size of the bag,
sider the following hypothetical sets of while the shape and orientation of the
test scores obtained by participants on bag indicate SKEWNESS and correlation,
two different occasions: respectively. A generic example is given
79 79 83 83 86 87 92 99 opposite.
and
Balaam's design an experimental ap-
80 81 82 83 87 87 90 99 proach in which participants are divided
The back-to-back stem-and-leaf plot for into four groups and experience differ-
these values is ent conditions or treatments (A and B)
|7|99 in one of four possible sequences: AA,
012377|8|3367 BB, AB, or BA. More specifically, in one
09 I 9 I 2 9 sequence participants receive the first
treatment over two periods, in a second
back-translation see TRANSLATION sequence participants receive the second
AND BACK-TRANSLATION. treatment over two periods, in a third
b a c k w a r d e l i m i n a t i o n a technique sequence participants receive the first
used in REGRESSION ANALYSIS in which treatment in the first period and the sec-
the goal is to forecast an outcome or re- ond treatment in the second period, and
sponse variable according to a subset of in the fourth sequence participants re-
predictor variables narrowed down from ceive the second treatment in the first
a large initial set. In background elimina- period and the first treatment in the
tion, all available predictors are included second period. A type of CROSSOVER DE-
originally and then examined one at a SIGN, Balaam's design is used in studies
time, with any predictors that do not in which a treatment CARRYOVER EFFECT
contribute in a statistically meaningful is anticipated. [L. N. Balaam, Australian
manner systematically dropped until a biometrician]
predetermined criterion is reached. Also
called backward deletion; back- balanced design any research design
ward selection; stepdown selec- in which the number of observations or
tion. See also STEPWISE REGRESSION. measurements obtained in each experi-
mental condition is equal. For example,
bagplot n. an extension of a BOX-AND- a health researcher interested in exercise
WHiSKER PLOT to two Variables, simulta- and depression would be using a bal-
ls
balanced replication
0)
.Q
.2
"T" T "T"
20 40 50
Variable A
bagplot
19
balanced scale
or treatments are randomly assigned to them are used to display data for vari-
participants and administered the same ables defined by qualities or categories.
number of times at each location. 2. an For example, to show the political affili-
occasional synonym for BALANCED RE- ations of Americans, the different parties
PEATED REPLICATION. would be listed on the x-ax'is, and the
height of the bar rising above each party
balanced scale a test or survey in which,
would represent the number or propor-
for each possible response, there is a re-
tion of people in that category. Also
sponse that means the opposite. A rating
called bar chart; bar diagram. See
scale with the four alternatives very poor,
also HISTOGRAM.
poor, good, and very good is an example, as
is a set of survey questions in which half
of the questions characterize a particular
trait (e.g., perceived stress level) in one
direction (e.g., low) and the other half
characterize the trait in the opposite di-
rection (e.g., high).
balancing score a score used for MATCH-
ING participants in a treatment group
and a comparison group on some vari-
able, with the goal of forming groups Democrat Republican Independent No Affiliation
21
Bayes factor
22
Bentler-Bonett index
23
Bernoulli distribution
parative FIT INDEX (CFI) in being normed. take on only one of two possible values
Also called normed f i t index (NFl). indicating either success or failure.
[Peter M. Bentier, U.S. quantitative psy- Usually a value of 0 is used to denote
chologist; Douglas G. Bonett, U.S. psy- a failure and a value of 1 is used to de-
chometrician] note a success. See BERNOULLI PRO-
CESS; BERNOULLI DISTRIBUTION. [JacqUCS
B e r n o u l l i d i s t r i b u t i o n a theoretical
distribution of the number of trials re- Bernoulli]
quired before the first success is obtained Bessel's correction an adjustment
in a BERNOULLI PROCESS. Such a distri- used to reduce bias when calculating the
bution is defined by two values: 0 and 1. VARIANCE and S T A N D A R D DEVIATION in
Usually a value of 0 is used to denote a a sample. Specifically, it is the use of n -
failure (i.e., the item of interest does not 1 instead of n in the denominator of the
occur) and a value of 1 is used to denote computational formulas used to com-
a success (i.e., the item of interest does pute these values, where n is the number
occur). On this basis, the likeliness of a of observations in the sample. [Friedrich
success is denoted as p and the likeliness Wilhelm Bessel (1784-1846), German-
of a failure is denoted asq = l - p . For ex- born Russian astronomer and mathema-
ample, a single toss of a coin has a Ber- tician]
noulli distribution with p = 0.5 (where 0
= heads and 1 = tails). A Bernoulli distri- best estimator the statistic obtained
bution is a special case of a BINOMIAL from a sample data set that over re-
DISTRIBUTION. [Jacques Bernoulli (1654- peated sampling exhibits the least
1705), Swiss mathematician and scien- amount of variance relative to other sta-
tist] tistics in estimating the corresponding
value in the larger population.
B e r n o u l l i process a sequence of unre-
lated events whose outcome values can best f i t the theoretical pattern that best
only be equal to 0 or 1 (denoting either accounts for the relafionships among
failure or success) with each outcome variables in a data set. For example, a RE-
having the same probability. For exam- GRESSION EQUATION having the best fit
ple, the results from tossing a coin a to sample data is the one that minimizes
given number of times when the condi- differences between observed and pre-
tions for each toss of the coin are iden- dicted values. On a SCATTERPLOT, a line
tical could be described as a Bernoulli of best fit provides a visual depiction of
process. In other words, a Bernoulli pro- this pattern, allowing extrapolation to
cess is a sequence of BERNOULLI TRIALS. values not part of the original data set. It
[Jacques Bernoulli] is important to note that the best-fitting
pattern is not necessarily the one that
Bernoulli's theorem see LAW OF LARGE generated the observed data, as other
NUMBERS. [Jacques Bernoulli] patterns not considered may provide a
B e r n o u l l i t r i a l a single experiment in superior fit.
which the only two possible outcomes
best linear unbiased estimator
are success or failure. Usually a value of 0
(BLUE) the LINEAR EQUATION that most
is used to denote a failure (i.e., the item
accurately characterizes the relationship
of interest does not occur) and a value of
between an outcome variable and a set
1 is used to denote a success (i.e., the
of predictor variables when applied to a
item does occur). A sequence of Ber-
sample data set. Thus, the quantity ob-
noulli trials is known as a BERNOULLI
tained from a BLUE on average equals
PROCESS. See also BERNOULLI DISTRIBU-
the quantity in the larger population
TION. Oacques Bernoulli]
from which it is derived and over re-
B e r n o u l l i variable a variable that can peated sampling has the least amount of
between-groups variance
25
between-subjects analysis of variance
the confidence interval for that subset, BIC abbreviation for BAYESIAN INFORMA-
return the values to the larger pool, and TION CRITERION.
then repeat the process numerous times. b i m o d a l d i s t r i b u t i o n a set of scores
He or she would then calculate a single with two peaks or MODES around which
overall confidence interval from the set values tend to cluster, such that the
of intervals obtained. frequencies at first increase and then de-
biased estimator a value obtained crease around each peak, as in the fol-
from sample data that consistently un- lowing hypothetical depiction.
der- or overestimates the true quantity
in the larger population of interest. In
other words, a biased estimator is one
whose value on average differs from the
value of the PARAMETER it purports to
represent. Also called biased statistic.
Compare UNBIASED ESTIMATOR.
biased s a m p l i n g selecting individuals
or other study units from a population
in such a manner that the resulting sam- As shown, when graphing the heights
ple is not representative of the popula- of a sample of adolescents one would ob-
tion. Compare UNBIASED SAMPLING. See tain a bimodal distribution if most peo-
also SAMPLING BIAS. ple were either 5'7" or 5'9" tall. See also
MULTIMODAL DISTRIBUTION; UNIMODAL
biased statistic see BIASED ESTIMATOR.
DISTRIBUTION.
bias-variance t r a d e o f f the situation b i n a r y t r i a l a single experiment in
in which increasing the complexity of a which there are only two possible out-
model to better account for the observed comes. For example, consider a behav-
relationships among a set of variables ioral researcher observing aggression in
reduces the precision of any estimated a group of schoolchildren during recess.
values subsequently derived from that If he or she makes a record every 1 min-
model. That is, the model becomes so
ute, then each such period may be con-
specific to the particular sample from sidered a binary trial, with aggression
which it was derived that it no longer is present or aggression absent as its only two
an accurate representation of the larger possible outcomes.
population of interest. For example, a
REGRESSION EQUATION for predicting a b i n a r y variable see DICHOTOMOUS
person's income may include two vari- VARIABLE.
ables (education level and career field) b i n o m i a l n. an algebraic expression
that only partially account for variance consisting of two terms, such as (x + y)^
in salary but nonetheless are applicable or 2x + 4y. See BINOMIAL EXPANSION; BI-
to a broad group of individuals beyond NOMIAL THEOREM. See also POLYNOMIAL.
those studied. Adding further variables
to the equation (e.g., area of residence, b i n o m i a l confidence i n t e r v a l a
parents' education levels, prior experi- CONFIDENCE INTERVAL for estimates of
ence, age, sex, ethnicity) may ensure the proportion of a population that ex-
that all variance is fully accounted for, hibits a particular characteristic or out-
but only in the limited set of individuals come, as determined from a sample of
studied. Researchers therefore tend to that population. A BINOMIAL DISTRIBU-
seek a balance, opting for either greater TION is assumed. For example, one could
intricacy or greater certainty as circum- sample the pass rate for students in sev-
stances warrant. eral different states on a nationally ad-
27
binomial distribution
ministered mathematics exam and then the categorical proportions in the ob-
use the information to compute a bino- tained data differ significantly from
mial confidence interval indicating a their proportions in the population
range of values for the proportion of all from which they are believed to derive.
U.S. students who will pass the exam. See also BINOMIAL PROBABILITY.
b i n o m i a l d i s t r i b u t i o n the PROBABIL- b i n o m i a l theorem the mathematical
ITY DISTRIBUTION for each possible se- rule that allows one to reformulate an al-
quence of outcomes on a variable that gebraic expression of two termsfor ex-
has only two possible outcomes, with ample, (x + y)^ or (x + yf'as a longer
the likeliness of obtaining each outcome series of sums. See BINOMIAL EXPANSION.
remaining constant. For example, a bi-
b i n o m i a l variable a variable for
nomial distribution of the results of try-
which there are only two outcomes, typ-
ing to predict the outcomes of 10 coin
ically designated 0 and 1, with each
tosses would display the probability of
possible outcome having a specified
observing each possible set of results, in-
likeliness of being observed. Flipping a
cluding one success and nine failures,
coin is an example, as it may land on ei-
two successes and eight failures, and so
ther heads or tails. See also BINOMIAL
forth. This distribution is often denoted
DISTRIBUTION.
by b(,9).
b i o g r a p h i c a l method the systematic
binomial expansion the EXPANSION
use of personal historiesgathered
of an algebraic expression consisting of
through such means as interviews, focus
two terms into a larger series of sums.
groups, observations, and individual re-
For example, a possible binomial expan-
flections and other narrativesin psy-
sion of (x + y)^ is (x + y)(x + y) or x^ + xy +
chological research and analysis. This
yx + / .
method emphasizes the placement of
b i n o m i a l p r o b a b i l i t y within a series the individual within the context of so-
of independent observations, the proba- cial connections, historical events, and
bility of observing a particular combina- life experiences.
tion of outcomes for a variable that can
biostatistics n. the branch of statistics
assume only one of two values (e.g., 0 or concerned with collecfing and analyzing
1; pass or fail) where there is a fixed data that pertain to biological processes
likeliness of each value occurring. For or health characteristics, especially in
example, consider a group of 10 individ- medicine and epidemiology. Also called
uals selected from a population of per- biometrics; biometry. biostatis-
sons aged 40-45, with each selection tical adj. biostatistician n.
unrelated to any other selection. A bino-
mial probability could be calculated to b i p l o t n. a multivariate graphic that si-
determine the likeliness that, for exam- multaneously displays information
ple, seven of those individuals are mar- about variables and participants. Data
ried while three are not, given known about two or more variables usually are
information on the probability of per- denoted by arrows or vectors, and spe-
sons of that age being married. cific measurements for individual cases
usually are denoted by dots. The exam-
b i n o m i a l sign test see SIGN TEST. ple (opposite) shows the relationship
b i n o m i a l test a statistical procedure to among five task variables: the Boston
determine whether an observed data Naming Test, various letter and category
pattern for a variable that can have only fluency tests, an adapted Simon task, the
one of two values (often represented as 0 Stoop Color-Word Interference Test,
and 1) matches a theoretical or expected and the Sustained Attention to Response
pattern. In other words, it tests whether Task. The dimensions indicate the per-
28
bivariate normality
Simon \ Stroop
E
5
"I 1 i 1 1 1 1
Dimension 1 (33.1%)
biplot
29
bivariate outlier
vice versa. Bivariate normality is an ular aspects of the study from which the
important assumption in certain statisti- data derive.
cal procedures, such as CORRELATION
ANALYSIS. b l i n d review an evaluation of a manu-
script to assess its suitability for publica-
bivariate o u t l i e r a data point whose tion or of a grant proposal to assess its
combination of values on two variables suitability for funding by a person who
is extreme or unusual relative to the does not know the identity of the author
other combinations in a data set. For ex- or proposer.
ample, a student who obtains perfect
scores of 100 points on two different block n. 1. a group or subset of study par-
math tests would be a bivariate outlier if ticipants who share a certain characteris-
most other students had combination tic and are treated as a unit in an
scores between 75 and 90 points. experimental design. 2. a set of variables
entered as a single entity into a REGRES-
bivariate probability distribution SION ANALYSIS or similar statistical pro-
see BIVARIATE DISTRIBUTION. cedure.
black box a system in which both the
input and output are observable but the block design a type of research study in
processes that occur between them are which participants are divided into rela-
unknown or not observable. For exam- tively homogeneous subsets (blocks)
ple, in the relationship between leader- from which they are assigned to the ex-
ship ability (input) and on-the-job perimental or treatment conditions. For
performance as a project manager (out- example, in a simple block design to
put), the role of the organizational cli- evaluate the efficacy of several antide-
mate could be considered a black box if pressants, participants with similar pre-
it is not understood. test depression scores might be grouped
into homogeneous blocks and then as-
b l i n d adj. denoting a research procedure signed to receive different medications.
in which information about particular The purpose of a block design is to en-
aspects or protocols of a study is un- sure that a characteristic of the study
known to participants, investigators, or participants that is related to the target
both. A single blind (or single masked) outcome (i.e., a COVARIATE) is distrib-
procedure is one in which the study par- uted equally across treatment condi-
ticipants do not know the experimental tions. See COMPLETE BLOCK DESIGN; IN-
conditions or groups to which they have COMPLETE BLOCK DESIGN; RANDOMIZED
been assigned; a double blind (or double BLOCK DESIGN.
masked) procedure is one in which nei-
ther the study participants nor the re- b l o c k d i a g r a m a graphical representa-
searchers who interact with them know tion of a system in which blocks repre-
which people have been assigned to sent key parts and connecting lines
which groups; and a triple blind (or tri- indicate relationships between parts. For
ple masked) procedure is one in which example, a block diagram could be used
the study participants, researchers, and to display the decision rules one might
individuals evaluating the data are all follow when conducting statistical tests
unaware of the specific study conditions to compare data from groups of study
to which people were assigned. Blinding participants.
is undertaken to prevent conscious or
unconscious BIAS by eliminating knowl- b l o c k i n g n. the process of grouping re-
edge that could skew results. search participants into relatively ho-
mogeneous subsets on the basis of a
b l i n d analysis an evaluation of data in particular characteristic. Such blocking
which the analyst is not aware of partic- helps adjust for preexisting patterns of
30
box-and-whisker plot
but instead is applied to the independent ance, is a helpful preliminary step in de-
variables (predictors) involved. [George termining whether the required AS-
E. P. Box; Paul W. Tidwell] SUMPTIONS have been met for certain
statistical procedures, such as ANALYSIS
b r a c k e t i n g n. in estimation problems, OF VARIANCE. [Morton B. Brown, Cana-
the specification of a range or interval diah-born U.S. statistician; Alan B.
known to contain a target value. For ex- Forsythe, U.S. statistician]
ample, the probability of an event may i
be expressed asp < .05 or as .01 <p < .05. B r o w n - S p e a r m a n f o r m u l a see
SPEARMAN-BROWN PROPHECY FORMULA.
Bradley-Terry model a model used to I
1
create a scale of preferences from data B R R abbreviation for BALANCED RE-
obtained from PAIRWISE COMPARISONS PEATED REPLICATION.
of items. It is widely used to estimate the
probability of certain items being pre- b r u s h i n g n. an interactive data-analysis
ferred over others. For example, the technique in which a computer user
Bradley-Terry model might be applied manipulates several multidimensional
in determining whether a depressed in- SCATTERPLOTS in real time, using a
dividual will opt to receive one treat- mouse or other device to continuously
ment (A) rather than another treatment change display features and immedi-
(B). [Ralph A. Bradley (1923-2001), Ca- ately view the results.
nadian-born U.S. statistician; Milton E. bubble plot a graphical representation
Terry, U.S. statistician] of data that is similar to a SCATTERPLOT
b r e a k d o w n p o i n t the smallest num- butj includes an additional third vari-
ber or proportion of extreme observa- able whose values are represented by dif-
tions (OUTLIERS) that can be present in a ferently sized circles. For example, a
data set before an ESTIMATOR (e.g., the researcher might use a bubble plot to
sample mean) WiU assume arbittarily large show the relationship between income,
values. years of education, and age within a
particular profession. The first two vari-
Breslow-Day test a test for the homo- ables would be arrayed along the hori-
geneity of ODDS RATIOS across different zontal A;-axis and vertical y-axis, re-
levels of a variable (e.g., participant in- spectively, with circles placed in the
comes) within a data set. [Norman E. graph to indicate points of intersecting
Breslow (1941- ), U.S. biostatistician; values; the radius of each circle would
Nicholas E. Day (1939- ), British bio- correspond to values of age, with larger
statistician and epidemiologist] circles indicating greater age (see exam-
ple overleaf).
Breslow test a method used in SUR-
VIVAL ANALYSIS to compare the out- b u m p h u n t i n g an examination of the
comes of different groups of individuals. distribution of values in a data set to
More specifically, it evaluates the equal- identify any SPIKES or MODES; in graphic
ity of two or more SURVIVAL FUNCTIONS terms, it is a search for any significant
depicting mortality or failure across "bumps" on an otherwise smooth curve.
time. Also called Gehan's generalized Bump hunting is useful for finding por-
Wilcoxon test. [Norman E. Breslow] tions of data that are different enough to
potentially represent distinct subgroups
Brown-Forsythe test a statistical of cases.
method for determining whether there
is HOMOGENEITY OF VARIANCE across two b u r n - i n n. in the MARKOV CHAIN MONTE
or more populations. The test, which uses CARLO METHOD, an initial series of sam-
the absolute values of the deviations of pling runs whose values are discarded
scores from the MEDIAN to calculate vari- and not incorporated into subsequent
33
Burt matrix
bubble plot
34
Cc
C symbol for COEFFICIENT OF CONTIN- tends MULTIPLE REGRESSION, which re-
GENCY. lates several INDEPENDENT VARIABLES
with only one DEPENDENT VARIABLE. For
c a l i b r a t i o n n. the process of assigning
example, a researcher might use a ca-
values to a measuring device (instru-
nonical correlation analysis to investi-
ment, test, or scale) relative to a refer-
gate whether several measures of
ence standard. For example, it would be
attitude toward learning (e.g., confi-
useful to compare the scores on a new
dence, anxiety, interest, enjoyment) are
test of intelligence with those from an
related to several measures of achieve-
older, well-accepted test to ensure that
ment (e.g., grade point average, the
the new test scores provide comparable
number of math classes taken, the num-
ratings or values. To do so, a researcher
ber of memberships in honor societies).
might select a group of people (the cali-
bration sample), administer each indi- canonical correlation coefficient
vidual both the old and new tests, and (symbol: R^) an index of the magnitude
then assess the results. of the linear relationship between a lin-
ear combination of one set of variables
caliper m a t c h i n g see MATCHING.
and a linear combination of a different
canonical adj. in data analysis, describ- set of variables, with each linear combi-
ing procedures that utilize linear combi- nation made in such a way as to maxi-
nations of measured variables rather mize the strength of the relationship
than the actual individual variables. between the two variable sets. It ranges
in value from -1 to +1.
canonical analysis any of a class of
statistical procedures that assess the de- canonical discriminant analysis
gree of relationship between sets of vari- see DISCRIMINANT ANALYSIS.
ables via interpretation of a limited
number of linear combinations of spe- canonical discriminant function
cific values of those variables. The over- see DISCRIMINANT FUNCTION.
all objective of such procedures is to canonical l o a d i n g see DISCRIMINANT
reduce the DIMENSIONALITY of the data LOADING.
under investigation. Examples include
CANONICAL CORRELATION ANALYSIS, canonical variate the result of a
DISCRIMINANT ANALYSIS, MULTIPLE RE- weighted, linear combination of a set of
GRESSION, and MULTIVARIATE ANALYSIS variables, often used as a dependent or
OF VARIANCE, among others. independent variable in a CANONICAL
CORRELATION ANALYSIS.
canonical correlation analysis a
method of data analysis that provides a c a p i t a l i z a t i o n o n chance drawing a
measure of the strength of the relation- conclusion from data wholly or partly
ship between a linear combination of biased in a particular direction by
predictor variables and a linear combi- chance. A common example of capital-
nation of outcome measures. Essen- ization on chance is the presentation of
tially, it extends a basic CORRELATION all the significant results in a study with-
ANALYSIS, which relates only one vari- out considering the number of results
able with another variable, and it ex- examined.
35
capture-tag-recapture sampling
36
Cauchy-Schwarz inequality
37
causal analysis
vidual absolute values. For example, evaluative tool, strict conditions con-
consider the numbers -2 and +3, whose cerning the measurement of the vari-
product is -6, the absolute value of which ables must be met.
is +6; similarly, the pair of +2 and -3 also
yields a product of -6, which also has an causal variable see INDEPENDENT
VARIABLE.
absolute value of +6. In either case, the
absolute values of the individual num- causation n. the empirical relation be-
bers are 2 and 3, which yields the same tween two events, states, or variables
product of 6. The Cauchy-Schwarz in- such that change in one (the cause)
equality is fundamental in many areas of brings about change in the other (the ef-
statistical theory, including probability fect). causal adj.
theory. [Augustin-Louis Cauchy (1789-
1857), French mathematician; Hermann CDF abbreviation for CUMULATIVE DIS-
Amandus Schwarz (1843-1921), German TRIBUTION FUNCTION.
mathematician] c e i l i n g effect a situation in which the
causal analysis an attempt to draw de- majority of values obtained for a vari-
pendable inferences about cause-and- able approach the upper limit of the
effect relationships from research data. scale used in its measurement. For exam-
Encompassing a variety of methods (e.g., ple, a test whose items are too easy for
PATH ANALYSIS, STRUCTURAL EQUATION those taking it would show a ceiling ef-
MODELING), such analyses differ in the fect because most people would achieve
degree to which they are statistically or be close to the highest possible score.
complex and the degree to which causal In other words, the test scores would ex-
inferences from them are, in fact, justi- hibit SKEWNESS and have little VARI-
fied. ANCE, thus prohibiting meaningful
analysis of the results. Compare FLOOR
causal d i a g r a m see PATH DIAGRAM. EFFECT.
38
central limit theorem
take public transit, males who take pub- portion of the scale. A response mea-
lic transit). sured by a meter, for instance, may not
be recorded because it is too small to be
cell mean a mathematical average of
detected or is so large that it exceeds the
the numeric values associated with a meter's capability to record. In such a
specific combination of two or more situation, the minimum value is as-
characteristics as displayed in a statisti- signed to an undetectable observation,
cal table. For example, results from a and the maximum value is assigned to
study of response time under specific an excessive observafion.
combinations of several factors may rely
on a comparison of cell means to GROUP censored regression see TOBIT ANALY-
DIFFERENCES in rcsponse times between SIS.
experimental conditions.
censoring n. the situafion in which some
cell-means model an approach in observafions are missingfroma set of data
which a set of observed responses on a (see CENSORED DATA). Censoring is com-
DEPENDENT VARIABLE are analyzed at mon in studies of survival time, in which
each level or condition of an INDEPEN- the research often ends before the event
DENT VARIABLE in order to identify dif- of interest occurs for all study units. See
ferences in their average values. Each also LEFT CENSORING; RIGHT CENSORING.
such average is obtained by adding a
random error term to the typical re- census n. the complete count of an en-
sponse value for the larger population tire population. A census differs from
from which the observations derive, as most experimental studies, which use a
estimated from the sample data. For ex- SAMPLE from a population in hopes of
ample, a researcher examining the influ- generalizing from that observed subset
ence of sleep on exam performance to the larger group.
might use the cell-means model to deter- centile a shortened name for PERCENTILE.
mine whether there are any differences
in the average exam score obtained by centile reference chart a table used
participants who got 6 hours of sleep the to compare a measurement on an indi-
previous night, those who got 8 hours, vidual to values in a population. The
PERCENTILE value from the population
and those who got 10 hours.
may be used to judge whether the indi-
censored data a set of data in which vidual's score is atypical in that context.
some values are unknown because they Thus, a centile reference chart may be
are not observed or because they fall used in assessing a child's growth rela-
below the minimum or above the maxi- tive to her weight and height (see chart
mum value that can be measured by the overleaf from the U.S. Centers for Dis-
scale used. For example, in a study of ease Control).
survival rates of a group of people with a
central d i s t r i b u t i o n see NONCEN-
particular disease, censored data would
TRAL DISTRIBUTION.
be obtained if the deaths of some partici-
pants occurred after the study ended. See central l i m i t theorem (CLT) the sta-
also DOUBLY CENSORED DATA. tistical principle that the sum of inde-
pendent values from any distribution
censored linear regression model
will approach a NORMAL DISTRIBUTION
see ACCELERATED FAILURE TIME MODEL.
as the number of values in the distribu-
censored observation a score that is tion increases. In other words, the larger
missing from a data set, either because the sample size, the more closely the
the event of interest has not occurred by SAMPLING DISTRIBUTION approximates a
the end of the study period or because normal distribution. The central limit
the response falls into an unmeasurable theorem is used to justify certain data
39
central moment
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
analysis methods when the appropriate- point of a set of scores. The central ten-
ness of a method relies on an assump- dency of a sample data set, for instance,
tion of normality. may be estimated by a number of differ-
ent statisfics (e.g., MEAN, MEDIAN,
central moment a MOMENT that de- MODE). See also MEASURE OF CENTRAL
scribes the shape of a set of scores with TENDENCY.
regard to its deviation about the mean.
Four common central moments that de- central-tendency bias see END-
scribe a RANDOM VARIABLE are the MEAN AVERSION BIAS.
(the first central moment), the VARI-
ANCE (the second central moment), the central-tendency measure see MEA-
SKEWNESS (the third central moment), SURE OF CENTRAL TENDENCY.
and the KURTOSIS (the fourth central
moment). Also called moment about centroid n. 1. in geometry, the center
the mean. point of an object or area in multidi-
mensional space. It may be determined
central tendency the middle or center by calculating the intersection of all
40
change-point problem
straight lines or planes that divide the c h a i n graph a graph in which the di-
area into two equal parts. 2. the MEAN of rectional and nondirectional relation-
multivariate data. It may be obtained by ships among variables in a data set are
calculating the average of the VECTORS indicated by arrows and lines, respec-
for the set of variables. See also CEN- tively. Consider the following generic
TROID METHOD. example.
centroid f a c t o r a single variable iden-
tified through its CENTROID value as
summarizing responses on multiple
other variables. Centroid factors are ob-
tained by the CENTROID METHOD, a
data-reduction technique sometimes
used in FACTOR ANALYSIS.
chain-of-events data data regarding a
centroid method 1. a technique used series of events that have occurred to a
in FACTOR ANALYSIS to reduce multiple group of individuals in the same order
correlated variables to a smaller set of ex- but not necessarily at the same times.
planatory variables. It is similar toand
indeed was a precursor ofPRINCIPAL c h a i n p a t h model a rare synonym of
COMPONENTS ANALYSIS but involves MARKOV C H A I N .
simpler calculations. Generally, the cen- chance difference a difference be-
troid method is now used only in situa- tween two samples that arises from the
tions when computers are not available. nature of sampling itself, rather than
2. an approach to AGGLOMERATIVE CLUS- from any intervention by a researcher.
TERING in which CENTROIDS are calcu- For example, scores obtained from two
lated to determine similarity among sep- groups of research participants might
arate entities and combine them into vary because of personal factors affect-
higher order groups. Each individual is ing some individuals (e.g., financial wor-
assigned to the subset closest in centroid ries) and not because of the conditions
value, with the process continuing until of the study.
the maximum grouping of cases has been
obtained. chance v a r i a t i o n see RANDOM VARIA-
TION.
CERES p l o t combined conditional ex-
pectations and residuals plot: a graph change i n R^ a change in the COEFFI-
that is similar to a PARTIAL RESIDUAL CIENT OF MULTIPLE DETERMINATION (R^)
PLOT in displaying relationships among from one model to another. In proce-
the predictors of a DEPENDENT VARIABLE dures such as HIERARCHICAL REGRES-
but distinct in being appropriate for use SION and STEPWISE REGRESSION, in
with nonlinear entities. which predictor variables are individu-
ally added to or removed from a set, a
CF abbreviation for CUMULATIVE FRE- significant change in R^ indicates that
QUENCY. the predictor of interest has a meaning-
CFA abbreviation for CONFIRMATORY ful role in the explanation of the re-
FACTOR ANALYSIS. sponse variable.
CFI abbreviation for Bentier comparative change-point p r o b l e m in the obser-
FIT INDEX. vafion of TIME-SERIES data, the difficulty
in judging when exactly a meaningful
CGF abbreviation for CUMULANT GENER- change in the pattern of data occurs,
ATING FUNCTION.
given a certain degree of indeterminacy
CHAID abbreviation for CHI-SQUARE AU- or randomness in the behavior of the
TOMATIC INTERACTION DETECTOR. variable (see STOCHASTIC PROCESS).
41
change score
change score see DIFFERENCE SCORE. pletely unknown except for mean and
changing-criterion design an exper- variance. [Pafnuty Chebyshev (or
Tchebyshev; 1821-1894), Russian math-
imental approach in which an initial
ematician]
baseline phase is followed by implemen-
tation of a treatment delivered in a series checklist n. a list of items that are to be
of phases. Each phase has its own crite- observed, recorded, corrected, or other-
rion rate for the target behavior, and wise considered in some manner. See
once responding becomes stable an in- also BEHAVIOR CHECKLIST. .
cremental shift in the criterion occurs to C h e r n o f f faces a representation of data
implement another phase. For example, in the form of stylized faces, designed to
a researcher studying the effectiveness of take advantage of the ability of observers
money in reducing caffeine consump- to discern subtle changes in facial ex-
tion may use a changing-criterion de- pressions.
sign with four treatment phases, each
gradually decreasing the amount below
which participants must maintain their
caffeine intake. The changing-criterion
design is distinguished from other ap-
proaches (e.g., the A-B-A DESIGN) in that In the example above, features such as
it does not require the withdrawal of an head and eye eccentricity, eyebrow slant,
intervention, an extended baseline, or and mouth size represent different values
treatment implementation across multi- of the variables of interest. [Herman
ple behaviors or treatments. Chernoff, 20th-century U.S. statistician]
homogeneous with regard to the re- served data. Examples include the C H I -
sponse variable. The response and the SQUARE GOODNESS-OF-FIT TEST, the CHI-
predictor variables may be CATEGORICAL, SQUARE TEST FOR HOMOGENEITY, the
continuous, or any combination thereof. CHI-SQUARE TEST FOR INDEPENDENCE,
the CHI-SQUARE TEST FOR TREND, and
c h i - s q u a r e d i s t a n c e a measure of the
the CHI-SQUARE TEST FOR VARIANCE.
difference between two or more groups
When unqualified, however, this term
with regard to their respective average
usually refers to the test for independ-
values on a set of variables. It is similar to
ence. Also called chi-square proce-
EUCLIDEAN DISTANCE in purpose but
dure.
used with C A T E G O R I C A L D A T A .
c h i - s q u a r e test f o r h o m o g e n e i t y a
chi-square d i s t r i b u t i o n (^ flistri-
statistical method to evaluate an as-
b u t i o n ) a distribution of the sums of in-
sumption that some feature of interest is
dependent squared differences between
equal between two or more groups (see
the observed scores in a data set and the
HOMOGENEITY). The test is used to eval-
expected score for the set. If a random
uate whether the frequencies within two
sample is repeatedly drawn from a nor-
or more categories of a given N O M I N A L
mal population and measured on some
variable are equivalent in different pop-
variable and the obtained scores trans-
ulations, such as whether the same
formed via STANDARDIZATION, multi-
number of fifth graders and sixth graders
plied by themselves, and then added,
watch particular cartoons.
the result will be a chi-square distribu-
tion with DEGREES OF FREEDOM equal tO c h i - s q u a r e test f o r i n d e p e n d e n c e a
the size of the samples drawn. See CHI- statistical method to evaluate whether
SQUARE TEST. there is a relationship between two vari-
ables whose values are categories. For ex-
c h i - s q u a r e g o o d n e s s - o f - f i t test a
ample, it may be used to test whether sex
statistical method of assessing how well
(male vs. female) is independent of
a mathematical model or theoretical ex-
that is, not related to or associated
pectation fits a set of observed data. The
withhaving a household pet (yes vs.
test is used to evaluate hypotheses about
no). It compares observed data to the
the proportions of individuals within
data that would be expected in each cell
different categories of a given N O M I N A L
variable, such as whether there are dif- of a CONTINGENCY TABLE if the two vari-
ferent percentages of males and females ables were entirely independent. This
enrolled in various college majors. The discrepancy between observed and ex-
smaller the statistic obtained from this pected counts is then used to compute
test, the smaller the difference between the CHI-SQUARE statistic. Overall, larger
the proposed model and the obtained values of the statistic relative to the
data and thus the better the model fit. number of participants in the study (de-
grees of freedom) are more likely to be
c h i - s q u a r e p r o b a b i l i t y p l o t a graph statistically significant. Also called chi-
of the frequency with which particular square test of association.
CHI-SQUARE values are obtained in a
c h i - s q u a r e test f o r t r e n d a statistical
data set as compared to the frequency
method to evaluate change over time in
expected according to theory. See PROB-
a CATEGORICAL VARIABLE Or to deter-
ABILITY PLOT.
mine whether two or more trends in cat-
c h i - s q u a r e test any of various proce- egorical data differ from each other. For
dures that use a CHI-SQUARE DISTRIBU- example, it may be used to test whether
TION to evaluate whether there is a there is a difference in incidences of can-
relationship between variables or the de- cer severity (low, medium, or high) as
gree to which a theory fits a set of ob- people age.
43
chi-square test for variance
Xll
Activation
Unpleasant Deactivation (0-g-. aroused, activated) Pleasant Activation
(e.g-. frenzied, jittery) (e.g.. energetic, excited)
III
Pleasure
(e.g-, un/iappy, dissatisried) (e-g.. satisfied, pleased)
Vlll
Deactivated Displeasure Deactivated Pleasure
(e.g., sad, gloomy) (e.g-, serene, peaceful)
circumplex
classification table
45
classification tree
classification table
Predict
Medication early gain
consult
38/64 = 59.4%
No I I , , \ I Yes
p=.043 I
classification tree
46
clustered data
class l i m i t the uppermost and lower- cases involving specific procedures for
most values between which lie a range of diagnosis or treatment.
values constituting a segment or class of
a variable: the upper and lower bound- closed i n t e r v a l a range of values that
aries of a CLASS INTERVAL. includes both its endpoints. For exam-
ple, the closed interval [0,3] consists of
class size the width of a range of values all values not smaller than 0 and not
constituting a segment or class of a vari- larger than 3. Compare OPEN INTERVAL.
able: the width of a CLASS INTERVAL.
closed question see FIXED-ALTERNA-
c l i n i c a l adj. in medical diagnosis or TIVE QUESTION.
treatment, describing a decision based
on a clinician's personal opinion and CLT abbreviation for CENTRAL LIMIT THE-
OREM.
judgment, as opposed to one based on
statistical evidence. Compare ACTUAR- cluster analysis a method of
IAL. multivariate data analysis in which indi-
c l i n i c a l sample a sample made up of viduals or units are placed into distinct
individuals who have been given a for- subgroups based on their strong similar-
mal diagnosis of a disorder. ity with regard to specific attributes. For
example, one might use cluster analysis
c l i n i c a l significance the extent to to form groups of individual children on
which a study result is judged to be the basis of their levels of anxiety, ag-
meaningful in relation to the diagnosis gression, delinquency, and cognitive
or treatment of disorders. An example of difficulties so as to identify useful
a clinically significant result would be an typologies that could increase under-
outcome indicating that a new interven- standing of co-occurring mental disor-
tion strategy is effective in reducing ders and lead to more appropriate
symptoms of depression. See also PRAC- treatments for specific individuals.
TICAL SIGNIFICANCE; STATISTICAL SIG- There are several different forms of clus-
NIFICANCE. ter analysisincluding HIERARCHICAL
CLUSTERING, K-MEANS CLUSTERING, and
c l i n i c a l test a test or measurement
LATENT CLASS ANALYSISand each is ap-
made in a clinical or research context for
the purpose of diagnosis or treatment of
propriate for use with different types of
a disorder. data. Results of a cluster analysis often
are presented in a DENDROGRAM. Also
c l i n i c a l t r i a l a research study designed called clustering.
to compare a new treatment or drug
with an existing standard of care or cluster centroid the central or typical
other control condition (see CONTROL value in a set of values relating to multi-
GROUP). Trials generally are designed to
ple variables. For example, in a data set
answer scientific questions and to find that contains patient ages and sex, the
better ways to treat individuals who cluster centroid would be an individual
have a specific disease or disorder. Also of the average age and the most com-
called clinical study; therapeutic mon sex for that group. See also CEN-
trial. See also RANDOMIZED CLINICAL TROID METHOD.
TRIAL. clustered data a set of observations or
c l i n i c a l u t i l i t y research see EFFEC- scores that can be grouped into multiple
TIVENESS RESEARCH.
subsets (clusters), such that the items in
each subset are similar to one another
c l i n i c a l v a l i d a t i o n the process of ac- with respect to certain attributes and the
quiring evidence to support the accu- distinctions between subsets help ex-
racy of a theory by studying multiple plain the overall variation among the
clustering
values as a whole. See CLUSTER ANALY- have been rounded to the nearest
SIS. integer or survival times that are missing
clustering n. see CLUSTER ANALYSIS.
observations for some units.
49
coefficient of reliability
from two or more groups of different get taller). Collinearity leads to diffi-
ages (COHORTS). If, for instance, individ- culties in interpreting the unique influ-
uals ranging in age from 5 to 10 years are ences of the independent variables and
sampled and then the members of each requires the use of PARTIALING proce-
age group are studied for a 5-year period, dures to distinguish their separate ef-
the resulting data would span 15 years of fects. See also MULTICOLLINEARITY.
development. Such studies essentially
are a combination of a LONGITUDINAL c o l u m n m a r g i n a l a summary of the
DESIGN and a CROSS-SECTIONAL DESIGN. values across each vertical set of cells in a
Also called accelerated longitudinal table. For a table containing frequency
design. counts it is a sum of the number of
counts in each column, whereas for
cohort study see LONGITUDINAL DE- other types of tables it is the average
SIGN. value of all data observations within a
given column. Compare ROW MAR-
cold-deck i m p u t a t i o n one of several
GINAL.
methods of inserting values for missing
data (see IMPUTATION) in which missing c o l u m n sum of squares 1. in an
observations are replaced by values from ANALYSIS OF VARIANCE involving two in-
a source unrelated to the data set under dependent variables, the amount of vari-
consideration. Suppose, for example, ance among individuals that is
that a patient questionnaire was admin- associated with either variable, as de-
istered in a hospital and that five people rived from the values given in the rele-
failed to respond to an item. Substi- vant column of the data table. It is
tuting responses from a similar item on a obtained by determining the average of
survey conducted previously would be all observations in the column, calculat-
an example of cold-deck imputation. ing how much each score deviates from
Compare HOT-DECK IMPUTATION. that average, multiplying the resulting
collaborative evaluation see PAR- value by itself, and adding it to the simi-
TICIPATORY EVALUATION.
larly obtained values for all other indi-
viduals in the column. The calculated
collapsing n. the process of combining quantities for each column are then
multiple response options or categories used to compute the BETWEEN-GROUPS
to form a smaller number of responses or SUM OF SQUARES, which in turn is used
categories. For example, a researcher to compute an F RATIO. See also TOTAL
could collect demographic data for stu- SUM OF SQUARES. Compare ROW SUM OF
dents at a particular college by each year SQUARES. 2. in EXPLORATORY FACTOR
of matriculation (i.e., freshman, sopho- ANALYSIS, the variance accounted for by
more, junior, and senior) and then col- each factor across variables. It is deter-
lapse the information from four to two mined by raising the FACTOR LOADINGS
categories (e.g., lower division and for each column in a factor loading ma-
upper division). trix to the second power and then total-
collinearity n. in REGRESSION ANALY- ing the values.
SIS, the situation in which two INDE- c o l u m n vector a data matrix with a
PENDENT VARIABLES are so highly associ-
single column of values. In other words,
ated that one can be closely or perfectly
it has the dimensions r x l , where r refers
predicted by the other. For example,
to the number of rows and 1 denotes the
collinearity likely is present if a re-
single column. Compare ROW VECTOR.
searcher examines how height and age
contribute to children's weight, since c o m b i n a t i o n n. the selection of r ob-
the two predictors are highly interre- jects from among n objects without re-
lated (i.e., as children grow older they gard to the order in which the objects
51
commensurate variable
52
complete factorial design
53
complete-linkage clustering
having two dosages and the other (B) set involving categorical variables. For
having three dosages, a complete facto- example, a researcher who observed the
rial design would pair the dosages ad- frequencies of males and females by ma-
ministered to different individuals or triculation year at a particular college
groups of participants as follows: A j with might present the results in a component
BJ, A l with B2, A l with B3, A2 with Bj, A2 bar graph (see below). The categories of
with B2, and A2 with B3. Compare FRAC- matriculation (freshmen, sophomores,
TIONAL FACTORIAL DESIGN. juniors, seniors) would be shown along
the A:-axis, and the number of males and
c o m p l e t e - l i n k a g e c l u s t e r i n g in HI-
females would be given by differently
ERARCHICAL CLUSTERING, a method in
colored or patterned bars rising above
which the distance between two clusters
each category. Also called sectional
of items (e.g., people, objects) is com-
bar graph; segmented bar graph;
puted as the greatest distance between
stacked bar graph.
any two objects in the different clusters.
Also called f a r t h e s t n e i g h b o r . Com- c o m p o n e n t i a l a n a l y s i s 1. any analy-
pare AVERAGE-LINKAGE CLUSTERING; sis of data in which a process or system is
SINGLE-LINKAGE CLUSTERING. separated into a series of subprocesses or
c o m p l e t e n u l l h y p o t h e s i s a state- components. 2. a set of information-
ment regarding the outcome of an ex- processing and mathematical tech-
periment that totally (rather than niques that enables an investigator to
partially) contradicts the ALTERNATIVE decompose an individual's performance
HYPOTHESIS. on a cognitive task into the underlying
elementary cognitive processes.
c o m p l e x c o m p a r i s o n an evaluation
that involves comparing some combina- c o m p o n e n t - p l u s - r e s i d u a l p l o t see
tion of two or more groups against one PARTIAL RESIDUAL PLOT.
or more other groups. For example, a re-
c o m p o s i t e h y p o t h e s i s a statistical hy-
searcher investigating the influence of a
pothesis that is not specific about all rel-
new teaching style on test scores might
evant features of a population or that
examine whether the average score
does not give a single value for a charac-
across two classrooms of students differs
teristic of a population but allows for a
from that of a third classroom. Compare
range of acceptable values. For example,
SIMPLE COMPARISON. a statement that the average age of em-
c o m p o n e n t b a r g r a p h a BAR G R A P H ployees in academia exceeds 50 is a com-
in which stacked columns are used to posite hypothesis, as there are a variety
represent two or more aspects of a data of ages above that number that the aver-
Matriculation Year
component bar graph
54
compound symmetry
Female
a Male
College Major
compound bar graph
55
comprehensive process analysis
regression line
n confidence interval
n forming band
n I I T"
Higti
Level of Job Satisfaction
confidence band
57
confidence limit
ticular confidence level used is up to the fest) variables are examined in relation to
researcher but generally is 95% or 99%. all underlying (latent) variables, confir-
matory factor analysis imposes explicit
confidence limit either of the values theoretical restrictions so that observed
at the upper and lower ends of a CONFI- measures relate with some (often just
DENCE INTERVAL, which provide an es-
one) latent factors but not others.
timated range for the value of a
population PARAMETER. Also called con- c o n f i r m a t o r y research research con-
fidence bound. ducted with the aim of testing one or
more preexisting hypotheses. Compare
confidentiality n. a principle of ethics
EXPLORATORY RESEARCH.
requiring mental health care or medical
care providers to limit the disclosure of a confound n. in an experiment, an inde-
patient's identity, his or her condition pendent variable that is conceptually
or treatment, and any data entrusted to distinct but empirically inseparable
professionals during assessment, diag- from one or more other independent
nosis, and treatment. Similar protection variables. Confounding makes it impos-
is given to research participants and sur- sible to differentiate that variable's ef-
vey respondents against unauthorized fects in isolation from its effects in
access to information they reveal in con- conjunction with other variables. For
fidence. See INFORMED CONSENT. con- example, in a study of high-school stu-
fidential adj. dent achievement, the type of school
(e.g., private vs. public) a student at-
c o n f i g u r a t i o n a l frequency analy- tended prior to high school and his or
sis a nonparametric method of compar- her prior academic achievements in that
ing observed counts within the context are confounds. Also called con-
categories of a variable with the frequen- founder; confounding factor; con-
cies that are theoretically expected. It is founding variable.
an alternative to similar methods (e.g.,
CHI-SQUARE TESTS) that require a mini- confounded comparison a compari-
mum expected count within each cate- son of values obtained by different ex-
gory in order for the results to be valid. perimental groups on an outcome or
dependent variable when two or more
c o n f i r m a t i o n bias the tendency to
predictor or independent variables vary
gather evidence that confirms preexist-
simultaneously across the groups. In
ing expectations, typically by emphasiz-
such cases it is impossible to differenti-
ing or pursuing supporting evidence
ate the effects of the independent vari-
while dismissing or failing to seek con-
ables. For example, consider a researcher
tradictory evidence.
studying how material presentation for-
c o n f i r m a t o r y data analysis a statis- mat (lecture vs. computer) and teacher
tical analysis designed to address one or sex (male vs. female) affect student
more specific research questions, gener- learning. If the investigator were to ex-
ally with the aim of confirming precon- amine data for only two groups of stu-
ceived hypotheses. Compare EXPLORA- dents (those who had a male teacher and
TORY DATA ANALYSIS. lecture presentation format vs. those
who had a female teacher and computer
confirmatory factor analysis (CFA) presentation format), he or she would
any method of testing A PRIORI hypoth- have created a confounded comparison.
eses to the effect that the relationships Compare UNCONFOUNDED COMPARI-
among a set of observed variables are SON.
due to a particular set of unobserved
variables. Unlike EXPLORATORY FACTOR confounded effects the indivisible ef-
ANALYSIS, in which all measured (mani- fects of two or more predictor or inde-
58
consistency check
59
consistent estimator
60
contingency table
tors relating to the variable under study CONTINGENCY TABLE. 2. more generally,
are permitted to influence the collection the extent to which one event or out-
and interpretation of data about that come is dependent on another.
variable. contingency analysis an approach to
content analysis a procedure for as- measuring the association between two
signing codes to identify themes in writ- variables whose values are represented
ten or spoken records. A content analy- by unordered categories. For example, a
sis of a speech, for example, may involve researcher may use contingency analysis
a count of the number of times a particu- to determine whether children who are
lar behavior occurs or the number of undergoing a painful medical procedure
times a particular idea is mentioned. become less distressed when nurses reas-
sure them or whether nurses reassure
content-referenced test see CRITE- children more when they are distressed.
RION-REFERENCED TEST.
See COEFFICIENT OF CONTINGENCY.
content v a l i d i t y the extent to which a contingency coefficient see COEFFI-
test measures a representative sample of CIENT OF CONTINGENCY.
the subject matter or behavior under in-
vestigation. For example, if a test is de- contingency table a two-dimensional
signed to survey arithmetic skills at a table in which frequency values for cate-
third-grade level, content validity indi- gories of one variable are presented in
cates how well it represents the range of the rows and values for categories of a
arithmetic operations possible at that second variable are presented in the col-
level. Although researchers traditionally umns: Values that appear in the various
relied upon their own subjective impres- CELLS then represent the number or per-
sions or the judgments of experts as the centage of cases that fall into the two
primary means of determining content categories that intersect at this point.
validity, modern approaches involve the For example, the sex and geographical
use of EXPLORATORY FACTOR ANALYSIS
locations of a sample of individuals ap-
and other multivariate statistical proce- plying for a particular job may be dis-
dures. played in a contingency table.
contextual analysis any method of Sex
evaluating data that takes into account
the characteristics of the environment Location Female Male Total
in which the information was collected Los Angeles 19 14 33
and their influence upon study units.
New York 19 15 34
contextual effect the influence on an
outcome or response variable of the en- Seattle 15 10 25
vironment in which data are collected. Tampa 13 12 25
For example, a study participant may be-
have differently in a laboratory than in a Washington, DC 17 16 33
public setting. Total 83 67 150
contextual variable any variable that
describes the environment in which
Thus, the number of women from Los
data are collected, such as its location or
Angeles are given, the number of men
the degree of social interaction.
from Los Angeles are given, the number
contingency n. 1. the degree to which of women from New York City are given,
one CATEGORICAL VARIABLE is associ- the number of men from New York City
ated with another such variable, as rep- are given, and so on. Also called cross-
resented by the frequencies recorded in a classification table.
61
contingent probability
62
controlled sampling
63
controlled trial
64
correlational research
65
correlation coefficient
variables are simply observed without sider the hypothetical example below de-
any control over the setting in which picfing correlations among anhedonia,
those relationships occur or any manipu- hopelessness, low self-esteem, and suicidal
lation by the researcher, FIELD RESEARCH ideation in a sample of college students.
often takes this form. For example, con-
correlation ratio (symbol: r|) a mea-
sider a researcher assessing teaching
sure of the strength or degree of relation-
style. He or she could use a correlational
ship between two variables whose
approach by attending classes on a col-
association is nonlinear (i.e., it cannot
lege campus that are each taught in a dif-
be depicted graphically by a straight
ferent way (e.g., lecture, interactive,
line). Often referred to as eta, the corre-
computer aided) and noting any differ-
lation ratio is the nonlinear equivalent
ences in student learning that arise. Also
called correlational design; correla- of the CORRELATION COEFFICIENT.
tional method; correlational study. correlogram n. see AUTOCORRELATION.
Compare EXPERIMENTAL RESEARCH.
correspondence analysis a statistical
correlation coefficient a numerical technique in which information from a
index reflecting the degree of linear rela- two-way CONTINGENCY TABLE is trans-
tionship between two variables. It is formed into a graphical display. This can
scaled so that the value of +1 indicates a then be analyzed to understand the as-
perfect positive relationship (such that sociations between the various catego-
high scores on variable x are associated ries of the different variables involved.
with high scores on variable y), -1 indi-
cates a perfect negative relationship c o t w i n c o n t r o l see TWIN CONTROL.
(such that high scores on variable x are count data information about the fre-
associated with low scores on variable y, quency of an attribute or event. An ex-
or vice versa), and 0 indicates no rela- ample would be the number of times a
tionship. The most commonly used type third-grader acts out in class during a 30-
of correlation coefficient is the Pearson minute observation period.
PRODUCT-MOMENT CORRELATION COEF-
FICIENT.
counterbalancing n. arranging a se-
ries of experimental conditions or treat-
correlation m a t r i x a symmetrical ments in such a way as to minimize the
SQUARE MATRIX displaying the degree of influence of extraneous factors, such as
association between all possible pairs of practice or fatigue, on experimental re-
variables contained in a set. A CORRELA- sults. In other words, counterbalancing
TION COEFFICIENT between the ith and is an attempt to reduce or avoid CARRY-
/'th variables in a set of variables is dis- OVER EFFECTS and ORDER EFFECTS, A
played in the intersection of the ith row simple form of counterbalancing would
and the /th column of the matrix. Con- be to administer experimental condi-
Variable 1 2 3 4
1. Anhedonia
2. Hopelessness .82
correlation matrix
covariance matrix
tions in the order A-B to half of the par- pothesis on the basis of a significance
ticipants and in the order B-A to the test with having demonstrated a scien-
other half; a LATIN SQUARE would be a tifically important effect.
more complex form.
c o u n t i n g process a type of STOCHAS-
counterfactual n. an alternative out- TIC PROCESS used to describe or analyze
come: a consideration of what would data that represent an ordered series of
have been observed had something taken the times to multiple events, such as a
place that in fact did not. Counter- study of the time to the first event, time
factuals are used primarily in RUBIN'S to the second event, and so on.
CAUSAL MODEL. For example, consider a
covariance n. a scale-dependent mea-
researcher seeking to determine the ef-
sure of the relationship between two
fect of a new depression treatment as
variables such that corresponding pairs
compared to an existing standard treat-
of values of the variables are studied
ment. According to Rubin's model, the
with regard to their relative distance
investigator would need to determine
from their respective means. A positive
the average difference for all participants
covariance results when values of one
between (a) the outcome after adminis-
variable that lie above the mean of that
tering the new treatment and (b) the
variable tend to be paired with values of
outcome after administering the alterna-
the second variable "that also lie above
tive treatment. Because the same partici-
the mean of that variable. A negative
pants cannot simultaneously be given
covariance results when values of one
the new and old treatments, the situa-
variable that lie above the mean tend to
tion arises in which certain experimen-
be paired with values of the second vari-
tal outcomes cannot be observed and
able that lie below the mean. A co-
become counterfactuals that instead must
variance is equal to 0 when two variables
be estimated.
are independent, or unrelated to one an-
c o u n t e r n u l l value the magnitude of other.
an effect (such as a difference in averages
between two groups or a measure of the covariance analysis see ANALYSIS OF
COVARIANCE.
strength of the relationship between
two variables) that carries with it a level covariance m a t r i x a SQUARE MATRIX
of statistical support that is equivalent to that represents how variance in each
the statistical support attributed to the variable in a set is related to variance in
value specified under a competing hy- all other variables in the set. Consider
pothesis (i.e., the NULL HYPOTHESIS). the following hypothetical example for
The counternull value helps eliminate five variables pertaining to perceived
two common errors: (a) equating failure health. The covariances between pairs of
to reject a null hypothesis with the esti- variables are located at the intersection
mation of the effect size as equal to zero of the row and column that correspond
and (b) equating rejection of a null hy- to the two variables. The quanfifies along
Variable 1 2 3 4 5
covariance matrix
67
covariance model
the diagonal of the matrix are variances in the investigation. See DECEPTION RE-
rather than covariances. Also called <lis- SEARCH.
persion matrix; varianee-covari-
ance matrix. COVRATIO a numerical value corre-
sponding to a unit (e.g., individual, ani-
covariance model a mathematical mal) that indicates the effect of
specification of the pattern of relation- removing the unit from a REGRESSION
ships among a set of quantitative vari- ANALYSIS. More technically, it is a ratio
ables that expresses how variance in of the DETERMINANTS Of COVARIANCE
each variable may be related to variance MATRICES with and without a given ob-
in each of the other variables. servation. Values less than 1 indicate
that removal of the unit has little influ-
covariance structure analysis see
ence on the precision of the estimates
ANALYSIS OF COVARIANCE STRUCTURES. obtained from the analysis.
covariate n. a variable that exhibits
Cox-Mantel test see LOG-RANK TEST.
COVARIATION with a measured outcome
[David Cox (1924- ), British statisti-
or DEPENDENT VARIABLE: It is Often in-
cian; Nathan Mantel (1919-2002), U.S.
cluded in an analysis so that its effect
biostatistician]
may be taken into account when inter-
preting the effects of the INDEPENDENT Cox regression analysis a statistical
VARIABLES of interest. For example, co- technique used to build multivariate
variates are used in ANALYSES OF CO- models that relate one or more continu-
VARIANCE to statistically adjust groups ous or categorical variables to SURVIVAL
so that they are equivalent with regard TIMES, without requiring researchers to
to these variables; they may also be used specify in advance the form or nature of
in MULTIPLE REGRESSION to minimize such relationships. For example, one
error that may arise from omitting any might use Cox regression to determine
noncentral but potentially influential how likely it is that alcoholics who are
variables. Also called concomitant abstinent at three months and at six
variable. months will relapse. A key methodologi-
cal concept in Cox regression analysis is
covariation n. a relationship between
the hazard, that is, the immediate po-
two quantitative variables such that as
tential or "risk" of event occurrence. It is
one variable tends to increase (or de-
computed from the baseline hazard rate,
crease) in value, the corresponding val-
which estimates the overall risk of event
ues of the other variable tend to also
occurrence as a function of time; various
increase (or decrease). For example, if a
coefficients, which describe the relation-
person's weight consistently rises as he
ship between each predictor variable
or she grows older, then the two vari-
and the rate of event occurrence; and
ables would be exhibiting covariation.
each individual's values on the predictor
cover story a plausible but false state- variables. There are two types of Cox
ment about the purpose of a research regression: the simpler standard Cox
study that is given to participants to regression model (or [Cox's] propor-
avoid disclosing to them the true hy- tional hazards model) and a more com-
pothesis being investigated. Such decep- plex generalization known as the ex-
tion may be practiced when the par- tended Cox regression model (or time-
ticipants' behavior in the study is apt to dependent Cox regression model). The
be affected by knowledge of the experi- standard model requires that the PRO-
ment's true purpose. For ethical reasons, PORTIONAL HAZARDS ASSUMPTION be
the decepfion should not flagrantly vio- met and thus is used when the risk of
late the participants' right to know what event occurrence for the reference and
they will be getting into by taking part comparison groups remains constant
68
criterion contamination
relative to one another over all time the degree of association between two
points. The extended model is used variables that have two or more unor-
when the effect of particular variables dered response categories. More specifi-
on the occurrence of the event of inter- cally, it is an omnibus EFFECT SIZE that
est changes over time, such that one quantifies the overall association among
group has a higher risk of event occur- the rows and columns in a CONTIN-
rence at early time points but a lower GENCY TABLE. Also called Cramer's
risk at later time points. [David Cox] phi. [Harald Cramer]
Cox-Snell residual the discrepancy Cramer-von Mises goodness-of-fit
between observed and predicted values test a statistical procedure to evaluate
for a variable defined as the time to a how well the distribution of an observed
particular event (e.g., time to death). set of scores matches a theoretical distri-
Used in cox REGRESSION ANALYSIS, such bution (e.g., a NORMAL DISTRIBUTION).
RESIDUALS provide a means for evaluat- It examines the discrepancies between
ing how a theory about changes in SUR- obtained scores and their corresponding
VIVAL TIMESfitsan observed set of data. expected scores to derive a single statis-
[David Cox; E.Joyce Snell] tic (W^), the larger the value of which
the poorer the model fit. [Harald
Cox's proportional hazards model Cramer; Richard E. von Mises (1883-
see cox REGRESSION ANALYSIS. [David 1953), Russian-born U.S. mathematician
Cox] and engineer]
Cox-Stuart test f o r t r e n d a statisti-
credible interval in BAYESIAN meth-
cal procedure for evaluating whether a
ods, a range within which a particular
sequence of independent scores follows
value of a population characteristic has a
a systematic pattern, for example,
specified probability of falling. It is the
whether earlier observations have an
Bayesian equivalent of the CONFIDENCE
overall tendency to be larger or smaller
INTERVAL used in FREQUENTIST ap-
than later observations. [David Cox; Alan
proaches. Also called Bayesian confi-
Stuart (1922-1998), British statistician]
dence interval; credible region.
Cramer-Rao l o w e r b o u n d a lower
c r i t e r i o n n. a standard against which a
limit on the extent to which estimated
judgment, evaluation, or comparison
values of a population parameter may
can be made. For example, a well-
vary, given that the ESTIMATOR is unbi- validated test of creativity might be used
ased. For example, suppose a researcher as the criterion to develop new tests of
examines a sample to derive possible es- creativity.
timates of the incidence of a rare disor-
der in the larger population. He or she c r i t e r i o n analysis a method of analyz-
could obtain a Cramer-Rao lower bound ing responses to a test in which it is as-
and compare it to the variance of the es- sumed that some unobserved variable
timates obtained using a particular UN- underlies the target behavior or concept
BIASED ESTIMATOR. Such a comparison of interest and accounts for the pattern
indicates whether the estimator in ques- of relationships (CORRELATIONS) among
tion is efficient, that is, whether given the test items.
alternative estimators it is the one with
the lowest degree of variance. Also called c r i t e r i o n c o n t a m i n a t i o n a situation
m i n i m u m variance bound. [Harald in which a response measure (the crite-
Cramer (1893-1985), Swedish mathe- rion) is influenced by factors that are not
matician; Calyampudi R. Rao (1920- ), related to the concept being measured.
Indian-born U.S. statistician] Evidence of this may be observed
through correlations of the response
Cramer's V (symbol: V; a measure of measure with variables that are concep-
69
criterion cutoff
tually distinct from that measure. For ex- that serves as a standard against which
ample, performance discrepancies (in other scores may be judged in a CRI-
dollars sold) among insurance agents TERION-REFERENCED TEST. 2. in REGRES-
may arise not from any actual differ- SION ANALYSIS, a predicted score on an
ences in ability but rather from socioeco- attribute or variable.
nomic differences in territories assigned
to the salespeople. c r i t e r i o n v a l i d i t y an index of how
well a test correlates with an established
c r i t e r i o n c u t o f f see CUTOFF SCORE. standard of comparison (i.e., a CRI-
TERION). Criterion validity is divided into
c r i t e r i o n group a group tested for three types: PREDICTIVE VALIDITY, CON-
traits its members are already known to CURRENT VALIDITY, and RETROSPECTIVE
possess, usually for the purpose of dem- VALIDITY. For example, if a measure of
onstrating that responses to a test repre- criminal behavior is valid, then it should
sent the traits they were intended to be possible to use it to predict whether
represent. For example, a group of chil- an individual (a) will be arrested in the
dren with diagnosed visual disabilities future for a criminal violation, (b) is cur-
may be given a visual test to assess its va- rently breaking the law, and (c) has a
lidity as a means of evaluating the pres- previous criminal record. Also called
ence of visual disabilities. criterion-referenced validity; cri-
c r i t e r i o n index an index that mea- terion-related validity.
sures a specific quality on which a set of c r i t e r i o n variable see DEPENDENT
scores may be evaluated. Examples in- VARIABLE.
clude indices of RELIABILITY and VALID-
ITY. c r i t i c a l difference in comparisons of
means or other statistics obtained from
criterion-referenced test an exam two or more samples, the minimum dif-
from which decisions are made about an ference that is deemed necessary to
individual's absolute level of accom- judge a test result as having STATISTICAL
plishment (i.e., mastery or nonmastery) SIGNIFICANCE. The Critical difference
of the material covered in that exam ac- will vary according to the procedure
cording to some standard reference point. used.
For example, if a student obtains a score
of 70% on a reading exam and a passing c r i t i c a l experiment see CRUCIAL EX-
score is 65%, then he or she has done PERIMENT.
well. Also called content-referenced
test. See also DOMAIN-REFERENCED TEST; c r i t i c a l r a t i o the result of dividing a
NORM-REFERENCED TEST. particular quantity resulting from a sta-
tistical test by a measure of the error re-
c r i t e r i o n scaling a method of chang- lated to the test result. For example, a
ing the values of a predictor and a re- calculated sample mean may be divided
sponse variable to make the relationship by its corresponding STANDARD ERROR
between the two variables more closely to derive a critical ratio, which may then
resemble a straight line. In criterion scal- be used to evaluate the STATISTICAL SIG-
ing, one calculates the averages of the re- NIFICANCE of that mean. Thus, if one ob-
sponse variable for subsets of individuals tains a critical ratio of 2.0, the observed
who have similar values on the predictor mean difference is twice as large as that
variable, replaces the predictor values expected on the basis of sampling error.
with these subset averages, and then
c r i t i c a l region a range of values that
uses the substituted averages to predict
may be obtained from a statistical proce-
the response variable.
dure that would lead to rejecting a spe-
c r i t e r i o n score 1. a score on a variable cific claim about a population. More
70
crossed-factor design
72
cross-validation
73
cross-validation sample
initial values, and statistical methods sible values for a variable, the second
generally are concerned only with the column (labeled f, for frequency) lists
first three cumulants or moments of a the number of scores that occur at each
distribution, as well as the fourth mo- of the possible values given in the first
ment, KURTOSIS. column, and the third column (labeled
CF, f o r C U M U L A T I V E FREQUENCY) giveS
cumulant generating function the running total of each of the values in
(CGF) a formula for obtaining values the second column. For example, a
that describe the basic nature of a DISTRI- teacher administers a test and the stu-
BUTION. A cumulant generafing function dents' scores are 1 F, 2 Ds, 4 Cs, 3 Bs, and
is generally the logarithm of a MOMENT 2 As. In a cumulative frequency distribu-
GENERATING FUNCTION, producing val- tion, the first column (X) represents
ues related to the MEAN, VARIANCE, and exam scores, with F, D, C, B, and A listed
SKEWNESS of a set of variables. from the bottom to the top. In the sec-
cumulative distribution function ond column (/) are the values of 1, 2, 4,
(CDF) a formula that gives the PROBA- 3, and 2 to indicate 1 F, 2 Ds, 4 Cs, 3 Bs,
BILITY, from 0 to 1, that a variable will and 2 As. In the final column (CF) are
have a score less than or equal to a spe- running totals of the second column
cific value. When this is plotted on a from the bottom up, listing 1, 3, 7, 10,
graph, the vertical y-axis will indicate and 12 to indicate the summed total of
the probability value from 0 to 1 for each scores at each of the grades, with the
possible score of the variable listed along total number of scores listed at the top
the horizontal ;if-axis. This type of graph (i.e., there are 12 total scores in this cu-
usually shows a pattern with a curve that mulative frequency distribution).
rises from the lower left up to a peak at
the upper right, where for the last score Score Frequeticy Cumulative
the probability that a variable will have a (X) (/) frequency (CP)
value equal to or less than the highest A 12
possible score is 1. Also called distribu-
tion function. B 10
C 7
cumulative frequency (CF) a run-
ning total of how often specific values D 3
occur. Cumulative frequencies are used F I
in DESCRIPTIVE STATISTICS when listing
the number of participants who fall into
each of several categories of a variable This type of table is useful in DESCRIP-
that can be ordered from low to high. TIVE STATISTICS to depict the number of
For example, if test scores in a particular scores at or below each score level, and
classroom are 1 F, 2 Ds, 4 Cs, 3 Bs, and 2 to provide an organized display of data
As, the cumulative frequency is obtained that could also be graphed in a CUMULA-
by successively adding the number of TIVE FREQUENCY POLYGON. AISO Called
students at each score from an F to an A. cumulative distribution; cumula-
Thus, the cumulative frequency values tive frequency table.
from the lowest to the highest would be
1, 3, 7, 10, and 12 for F, D, C, B, and A, cumulative frequency polygon a
respectively. See also CUMULATIVE FRE- graphical representation of a CUMULA-
QUENCY DISTRIBUTION. TIVE FREQUENCY DISTRIBUTION, With
points along the vertical y-axis depicting
cumulative frequency flistribu- the number of scores falling at or below
t i o n a table with three columns where the particular values given along the
the first column (labeled X) lists the pos- horizontal .x-axis. The resulting points
75
cumulative hazard function
D C
Exam Score
are connected with straight lines, reveal- mulative odds ratio is an extension of
ing a pattern that tends to increase or re- the single ODDS RATIO, which is the value
main constant as one moves from the formed at just one time point.
lower left to the upper right portion of
the graph (see example above). Also cumulative outcome analysis a
called cumulative frequency curve; procedure for assessing the PROPORTION
cumulative frequency diagram; cu- of occurrence of an event, usually across
mulative frequency graph. time. This procedure can be performed
as part of SURVIVAL ANALYSIS, which
cumulative h a z a r d f u n c t i o n in a studies the rate of survival or success (or
SURVIVAL ANALYSIS, a mathematical de- conversely the rate of death or failure)
scription of the accumulating number of over a period of time. For example, a re-
failures or deaths over time: It can be in- searcher may want to assess patients
terpreted as giving the probability of who remain alive at one-year intervals
death or failure at a specific time, given across afive-yearperiod, based on an ini-
survival until that time. For example, a tial sample of 100 patients. Before the
cumulative hazard function might be start of the study, the proportion is
used in a research project to examine the equal to 1.0, as all of the patients that
mnning pattern of deaths due to cancer will be in the study are now alive. At the
over a five-year period. It is an extension end of the first year, if 85 patients are
of a HAZARD FUNCTION beyond a specific still alive, the proportion would be
individual or single occurrence. 85/100 or .85. At the end of the second
year, if 62 patients are still alive, the pro-
cumulative odds ratio a WEIGHTED portion would be 62/85 = .73; and so on.
AVERAGE of a Set of values formed by Cumulative outcome analysis may also
calculating the probability of an event be used in examining LIFE TABLES.
occurring in one group over the proba-
bility of the event not occurring in the cumulative percentage a mnning
same group, then dividing this by the total of the PERCENTAGE values occur-
probability of the same event occuring ring across a set of responses. The total
in a second group over the probability of will either remain the same or increase,
the event not occurring in this group. reaching the highest value of 100% after
For example, a medical researcher may totaling all of the previous percentages.
examine the probability of cancer occur- For example, if the percentage of fresh-
rence in a treatment group versus a con- men, sophomores, juniors, and seniors
trol group every month for six months. among all of the students at a college
The researcher then finds the average of were 40%, 25%, 20%, and 15%, respec-
these values, after weighting them for tively, the cumulative percentage values
the number of individuals involved in would be 40%, 65%, 85%, and 100%
each of the monthly calculations. A cu- when summing from the percentage of
cumulative relative frequency
Exam Score
tions from 0 to 1.0 would be listed verti- cure rate model a representation of
cally and the grades F, D, C, B, and A the PROPORTION of individuals who
would be listed horizontally. When plot- have survived an illness after treatment,
ted, the running total of the proportions which includes those who have survived
of each score that occurred would have a up to a certain point but have not been
height of .083 for the score of F, on up to cured. This proportion is important in
a height of 1.0 for a score of A. Also SURVIVAL ANALYSES of complex ill-
called cumulative proportion. nesses, such as cancer. For example, it
cumulative scale see GUTTMAN SCALE. might be useful to construct a cure rate
model that estimates the number of pa-
cumulative sum chart (CuSum tients who survive cancer and appear to
chart) a graph of a group of values that no longer have the disease as well as the
are calculated by setting an initial value number who still have cancer but might
of 0 and obtaining each subsequent be expected to still be alive at specific
value by adding the previous value to times after treatment.
the difference between the current score
and the mean of all of the scores. For ex- curse o f dimensionality the problem
ample, if there were three IQ scores of that the volume of a mathematical space
92, 103, and 105, the mean is equal to increases exponentially with every new
(92 + 103 + 105)/3 = 100. Then, the ini- facet that is added, such that it becomes
tial value of the cumulative sum chart increasingly difficult or even intractable
would be set to equal 0, and the next to study spaces as the number of the
value would equal 0 + (92 - 100) = -8. variables increases.
The next value would equal -8 + (103 - curve f i t t i n g any of various statistical
100) = -5; and the last value would equal techniques for obtaining a function that
-5 + (105 - 100) = 0. graphically represents a given set of data
as closely as possible, with minimal
error. A simple example of curve fitting
occurs in LINEAR REGRESSION, where a
straight line function is said to provide
the closest approximation to the ex-
pected values of the outcome with a spe-
cific predictor and have the fewest
discrepancies between the predicted val-
ues and the actual, observed values.
c u r v i l i n e a r adj. describing an associa-
tion between variables that does not
consistently follow an increasing or de-
creasing pattern but rather changes di-
As shown, the cumulative sum values
rection after a certain point (i.e., it
are given along the vertical y-axis and the
involves a curve in the set of data
IQ scores are given along the horizontal x-
points). For example, the relationship
axis. between anxiety and achievement often
cumulative test a type of test that as- has a curvilinear pattern of increasing
sesses all of the information taught up to achievement with increasing anxiety
the point that it is taken. For example, (i.e., motivation to study) up to a certain
cumulative tests in school include an in- point when there is so much anxiety
tegration of all of the material that a that achievement tends to decrease.
teacher covered over the entire term, Thus, individuals who are not at all anx-
rather than just the most recent infor- ious and those who are extremely anx-
mation that was taught. ious would both be expected to have
78
cycle plot
poor performance, whereas moderately SERIES of data for a specific variable. The
anxious individuals would be expected pattern can be evenly spaced (e.g., as with
to have reasonably high performance. seven-day weekly salesfigures)or irregu-
See also NONLINEAR. lar (e.g., as with shifts in sales due to
weather changes or economic factors).
curvilinear regression see NONLIN-
EAR REGRESSION. cycle plot a graph that highlights regu-
lar patterns in TIME-SERIES data by dis-
CuSum chart abbreviation for CUMU- playing a separate line for each phase
LATIVE SUM CHART.
(e.g., week, month) for every major point
cutoff score a value or criterion that is (e.g., day of the week) along the series.
held to mark the lowest point at which a For example, a professor may want to ex-
certain status or category is attained. For amine the number of student e-mails he
example, the cutoff score for passing a or she receives each week over a 14-week
course is often 60%. Similarly, the cutoff semester by charting the 14 points for
score for being considered overweight is specific numbers of e-mails received on
a body mass index of 25 to 29. Also Sundays in the first line of the graph, fol-
called cutoff point; criterion cut- lowed by the 14 points for the number of
off. e-mails received on Mondays in the sec-
ond line, and so forth. Whenfinished,the
Cuzick's t r e n d test an extension of graph displays the number of e-mails on
the WILCOXON RANK-SUM TEST: a the verfical y-axis and the days of the
nonparametric procedure used with OR- week on the horizontal x-axis, with the
DINAL DATA for estimating the pattern of specific patterns over the 14-week semes-
relationship between rankings across ter for each day clearly visible (see below).
three or more independent random
In this way, it would be more obvious
samples. For example, Cuzick's trend
on what days of the week and during
test could assess the similarity of sets of
which weeks of the semester (e.g., just
standardized achievement test scores for
before exams, after holidays) more time
applicants to private, Ivy League, and
would need to be designated for attend-
state colleges, [jack Cuzick, U.S. epidemi-
ing to student questions and feedback.
ologist and statistician]
A cycle plot is thus more revealing than
cycle n. a distinct pattern within a TIME a standard time series graph that, for
Day of Week
cycle plot
79
cyclic component
this example, would list all of the days of cyclic data a set of information, espe-
the week over each of the 14 weeks (i.e., cially a TIME SERIES, in which a recurring
7 X 14 = 98 points) along the x-axis, with distinct pattern is identifiable. For exam-
the latter making it difficult to perceive ple, a psychiatrist may note that patients
daily or weekly patterns as clearly. tend to require more visits or medica-
cyclic component the portion of a tion during regular time periods around
plot of TIME-SERIES data that shows a no- holidays or the anniversary of the death
ticeable, distinct pattern, presumably of a loved one.
due to some identifiable factor. For ex-
ample, it could be important to identify cyclic v a r i a t i o n a change in the pat-
the cyclic component in car sales due to tern of data across a TIME SERIES that is
long-term fluctuations in the economy, recurring and identifiable. For example,
in addition to taking into account a sea- sales, doctor visits, and student e-mails
sonal pattern showing more predictable may show regular patterns over time
increases at the end of the year and dur- that are useful to identify. See also CY-
ing holidays. CLIC COMPONENT.
80
Dd
d 1. see COHEN'S D. 2. see GLASS'S D. haviors such as frequency and amount
of exercise, number of calories con-
d' symbol for D PRIME.
sumed per day, number of cigarettes
D 1. symbol for DIFFERENCE SCORE. 2. see smoked per day, number of alcoholic
COOK'S DISTANCE. drinks per day, and so forth.
symbol for MAHALANOBIS DISTANCE. data analysis the process of applying
graphical, statistical, or quantitative
DAG abbreviation for DIRECTED ACYCLIC techniques to a set of observations or
GRAPH. measurements in order to summarize it
D'Agostino test 1. a statistical assess- or to find general patterns. For example,
ment of the degree of SKEWNESS (lopsid- a very basic data analysis would involve
edness) in a distribution. Its full name is calculating DESCRIPTIVE STATISTICS (e.g.,
the D'Agostino test for skewness. 2. (sym- MEAN, MEDIAN, MODE, STANDARD DEVI-
bol: K^) a statistical assessment of the ATION) and possibly graphing the obser-
degree to which a distribution departs vations with a HISTOGRAM or BAR GRAPH.
from a NORMAL DISTRIBUTION, in which
data augmentation (DA) the use of
skewness and KURTOSIS are expected to
algorithms to simulate data so as to esti-
be zero. Its full name is the D'Agostino-
mate functions that are difficult to cal-
Pearson omnibus test for normality.
culate directly from the existing data.
[Ralph B. D'Agostino (1940- ), U.S. bio-
Data augmentation has been used in
statistician]
MARKOV CHAIN MONTE CARLO METHODS
damped regression see RIDGE REGRES- and other approaches.
SION.
database n. a large, structured collec-
Danieli weights values that can be tion of information stored in retrievable
used to adjust the beginning, end, and form on a computer. For example, most
middle of a MOVING AVERAGE form of researchers keep a record of information
TIME SERIES, allowing a clearer view of collected from their studies on a com-
the nature of the trends in the data that puter to allow for easy access, manipula-
are distinct from RANDOM ERROR. tion, and analysis in testing relevant
D a r l i n g test an assessment, more for- HYPOTHESES.
mally known as the Anderson-Darling data capture the automatic collection
test, of whether a sample of data came of data, often via scanning or computer
from a particular probability distribu- devices that record information quickly
tion. It is considered a very powerful and unobtrusively. For example, infor-
test of whether data differ from a NOR- mation is retained about a person's on-
MAL DISTRIBUTION. [Donald A. Darling line purchases with a particular credit
(1915- ), U.S. mathematician] card, which in turn may be used by a
company to help determine that per-
data pl. n. (sing, datum) observations or
measurements, usually quantified and son's preferences and make suggestions
obtained in the course of research. For about other items he or she might be in-
example, a researcher may be interested terested in purchasing.
in collecting data on health-related be- data c o d i n g the process of putting in-
81
data collection
formation into understandable, usually tions, and other regularities that can be
Q U A N T I T A T I V E Or Q U A L I T A T I V E , f o r m S tO used for predictive purposes. Although a
allow analyses that can summarize the relatively new discipline, data mining
main themes that emerge or test rele- has become a widely utilized technique
vant HYPOTHESES. For example, a re- within commercial and scientific re-
searcher may collect information on a search. For example, retailers often use
20-item survey that contains questions data mining to predict the future buying
on a number of areas of substance use. trends of customers or design targeted
Based on this information, several quan- marketing strategies, whUe clinicians
titative variables can be formed assess- may use it to determine variables pre-
ing, for example, the amount of alcohol dicting hospitalization in psychological
use, the frequency of alcohol use, the disorders. Data mining incorporates
amount of hard drug use, and the fre- methods from statistics, logic, and artifi-
quency of hard drug use. cial intelligence.
data collection a systematic gathering data p o i n t a specific piece of informa-
of information for research or practical tion derived from a larger set of data. A
purposes. Examples include mail sur- data point is often formed by the inter-
veys, interviews, laboratory experi- section of two other pieces of informa-
ments, and psychological testing. tion, as in the intersection of a person's
score on one variable with his or her
data dredging the inappropriate prac- score on a second variable in a SCATTER-
tice of searching through large files of PLOT. For example, 10 people could be
information to try to confirm a precon- asked their average number of hours of
ceived HYPOTHESIS or belief without an sleep per night and the number of colds
adequate design that controls for possi- they tend to have in a year. Then, the
ble CONFOUNDS or alternate hypotheses. number of sleep hours could be recorded
Data dredging may involve selecting along the horizontal A;-axis and the
which parts of a large data set to retain in number of colds per year recorded along
order to get specific, desired results. An the vertical y-axis, resulting in 10 sepa-
extreme example might occur if a mar- rate data points falling, most likely, in a
keting researcher found that 91 out of negative pattern from the upper left por-
100 people surveyed were opposed to a tion to the bottom right portion of the
certain product and then chose to only scatterplot (i.e., the more hours of sleep,
focus on the last 10 people in order to the fewer colds).
state that 9 out of 10 people prefer this
product, when in fact there were only 9 data p o o l i n g combining the informa-
out of 100 who preferred the product. tion from two or more studies or
See DATA SNOOPING. substudies, for example, by averaging
STANDARD DEVIATIONS Or VARIANCES
data m a t r i x (symbol: X) an arrange- across groups to form a single value for
ment of data in a MATRIX, usually with use in a T TEST or ANALYSIS OF VARIANCE.
one row for each participant and one Although it can be helpful to synthesize
column for each piece of information information in this way, data pooling
gathered from the participants. A data can sometimes lead to misleading con-
matrix may be used in analyses to test clusions, as in SIMPSON'S PARADOX.
HYPOTHESES and examine patterns.
data record a set of information, often
data m i n i n g the automated (comput- relating to a single participant, that is
erized) examination of a large set of stored in a larger file or DATABASE. For
observations or measurements, particu- example, a researcher may want to focus
larly as collected in a complex database, on the specific information or data re-
in order to discover patterns, correla- cord gathered from the 10th participant,
82
decision function
which would most likely be found in the symptom as evidence that a treatment is
10th row of the larger file. working effectively. Compare ACCELER-
ATION.
data reduction the process of reducing
a set of measurements or variables into a deception by commission see AC-
smaller, more manageable, more reli- TIVE DECEPTION.
able, or better theoretically justified set
or form. For example, a researcher may deception by omission see PASSIVE
conduct a FACTOR ANALYSIS on a set of DECEPTION.
50 items on well-being and satisfaction
to determine whether the information deception research any study in which
could be summarized more efficiently participants are deliberately misled or
on UNDERLYING DIMENSIONS of rela- not informed about the purpose of the
tionship satisfaction, degree of meaning investigation in order to avoid the possi-
in life, job satisfaction, and general bility that responses may be given to
health. meet the perceived expectations of re-
searchers. For example, a social psychol-
data screening a procedure in which ogist may use a deception experiment in
one subjects a large set of information to which participants are randomly as-
preliminary review for any of a variety of signed to either of two scenarios that
reasons: to check for accuracy, to iden- each describe job applicants in identical
tify unusual patterns (e.g., OUTLIERS terms except that one is said to be a male
that are very different from the informa- and the other is said to be female. The re-
tion from most participants) or any searcher may then assess any gender bias
missing information, to determine in the participants by asking how likely
whether the information would meet a the applicant is to be hired, instead of
statistical ASSUMPTION (e.g., of normal- directly asking participants about their
ity), to reduce the data to more manage- attitudes toward gender. See ACTIVE DE-
able dimensions, and so forth. CEPTION; DOUBLE DECEPTION; PASSIVE
DECEPTION.
data set a collection of individual but re-
lated observations or measurements decile n. one of a series of values that di-
considered as a single entity. For exam- vide a statistical distribution into 10
ple, the entire range of scores obtained equal-sized parts. Thus, the first decile is
from a class of students taking a particu- the value below which lie 10% of cases,
lar test would constitute a data set. the second decile is the value below
data snooping 1. looking for unpre- which lie 20% of cases, and so on.
dicted, post hoc effects in a body of data. decision error a conclusion that a
2. examining data before an experiment study obtained positive results when in
has been completed, which can some- fact the research did not do so in 95% of
times result in erroneous or misleading similar studies or a conclusion that a
conclusions. See also DATA DREDGING. study was not effective when it actually
d e b r i e f i n g n. the process of giving par- produced effective results in most other
ticipants in a completed research project reported research. The former conclu-
a fuller explanation of the study in sion of a false positive is referred to as a
which they participated than was possi- TYPE I ERROR, whercas the latter conclu-
ble before or during the research. sion of a false negative is called a TYPE II
ERROR.
deceleration n. a decrease in speed of
movement or rate of change. For exam- decision f u n c t i o n a procedure or set
ple, a medical researcher may be inter- of procedures that determines the con-
ested in the deceleration of an illness or clusions to be drawn or actions to be
83
decision rule
taken on the basis of certain observed will occur with the probability equal to a
data. Also called decision rule. value designated as BETA (P); if it is re-
decision rule 1. in HYPOTHESIS TEST- jected, a correct decision has been made
ING, a formal statement of the set of val- with a probability equal to 1 - p, which is
labeled as POWER.
ues of the test statistic that will lead to
rejection of the NULL HYPOTHESIS that decision theory a broad class of pre-
there is no significant effect in the study sumptions in the quantitative, social,
being examined. For example, a com- and behavioral sciences that aim to ex-
mon decision rule is to reject the null plain the process and identify optimal
hypothesis when the value of a z TEST ways of arriving at conclusions in such a
stafisfic exceeds 1.96. 2. see DECISION way that prespecified criteria are met.
FUNCTION. decision tree a diagram that uses a tree-
decision table a table that sets out the like structure of branches and nodes to
key conditions and actions involved in delineate conditions, action choices,
coming to a conclusion. In statistics, a and the further conditions these choices
widely used decision table sets out the give rise to. A decision tree starts small
conditions of there being either a true and grows in size and complexity as it
NULL HYPOTHESIS or a false null hypoth- advances from left to right. For example,
esis, the actions of rejecting the null to decide which statistical analysis to
hypothesis or retaining the null hypoth- conduct in a given case, a diagram could
esis, and the corresponding PROBABIL- begin at the left with a single node ask-
ITY associated with each action. If the ing what the focus of the analysis is to
null hypothesis is true and it is rejected, be.
a TYPE I ERROR has occurred with a prob- A branch going off to the upper right
ability equal to ALPHA (a; often .05); if it could specify a node with the focus of
is retained, a correct decision has been obtaining mean differences between
made with a probability equal to 1 - a groups, and another branch going off to
(in this case, .95). If the null hypothesis the lower right could specify a node for
is false and it is retained, a TYPE II ERROR examining relationships among vari-
dedsion tree
84
defender's fallacy
85
definitional formula
definitional formula the formal ver- rather than randomly. This includes
bal definition of a statistical concept. For simple nonrandom techniques, such as
example, the definitional formula of JUDGMENT SAMPLING and CONVENIENCE
VARIANCE states that it is the mean SAMPLING. In more sophisticated meth-
squared difference between a score and ods, such as QUOTA SAMPLING and
the mean of all of the scores. This con- STRATIFIED SAMPLING, various Steps are
trasts with the COMPUTATIONAL FOR- taken with the aim of creating a sample
MULA, which is the equation used to that is representative of the larger popu-
calculate values for the concept. Also lation.
called conceptual formula; defini-
tion formula. delta (symbol: A) n. 1. a measure of the
change in a PARAMETER. For example,
d e f i n i t i o n a l v a l i d i t y the extent to AR'^ indicates how much of the change
which the methods or approaches used in R^ (the COEFFICIENT OF MULTIPLE DE-
by a researcher are consistent with the TERMINATION) was caused or explained
significance claimed (explicitly or im- by a given step (e.g., adding a specific
plicitly) for the research. See also EXTER- variable to an analysis). 2. see GLASS'S D.
NAL VALIDITY; INTERNAL VALIDITY; E M -
PIRICAL VALIDITY; FACE VALIDITY. delta method a procedure used to ar-
rive at the approximate PROBABILITY
degrees o f freedom (symbol: df) the DISTRIBUTION of a Variable that is ex-
number of elements that are allowed to pected to have a NORMAL DISTRIBUTION.
vary in a statistical calculation, or the For example, the delta method would
number of scores minus the number of allow a researcher to find the probability
mathematical restrictions. If the MEAN of having a low, medium, or large in-
of a set of scores is fixed, then the num- come if income was expected to be nor-
ber of degrees of freedom is one less than mally distributed in the population of
the number of scores. For example, if interest. The delta method is based on
four individuals have a mean IQ of 100, the CENTRAL LIMIT THEOREM. Also called
then there are three degrees of freedom, delta technique.
because knowing three of the IQs deter-
mines the fourth IQ. d e m a n d characteristics in an experi-
ment or research project, cues that may
dehoaxing n. DEBRIEFING participants influence or bias participants' behavior,
who have been involved in DECEPTION for example, by suggesting the outcome
RESEARCH to inform them that they or response that the experimenter ex-
were misled as part of the study. This pects or desires. Such cues can distort the
may involve desensitizing participants, findings of a study. See also EXPERI-
so that their self-image is not harmed by MENTER EFFECT.
having participated in a deceptive study.
Deming-Stephan a l g o r i t h m a pro-
deletion residual in a statistical proce- cedure for deriving a LOG-LINEAR ANALY-
dure, the difference that results from SIS or model from a CONTINGENCY
deleting data based on one of the partici- TABLE of frequencies for two CATEGORI-
pants in a sample. If this difference is CAL VARIABLES. For example, this proce-
large, there is some evidence that the re- dure might be used to examine whether
moved participant was an OUTLIER, who gender and college major are indepen-
was very different from the remaining dent (the NULL HYPOTHESIS expectation)
participants. Also called likelihood re- or related (the ALTERNATIVE HYPOTHESIS
sidual. See REGRESSION DIAGNOSTICS. expectation) in a sample of data from
deliberate s a m p l i n g any method of college students. [W. Edwards Deming
selecting individuals to participate in re- (1900-1993) and Frederick F. Stephan,
search on the basis of a specific plan, U.S. statisticians]
86
dependent samples
Agglomerative
2 3
Distance Measure
Divisive
dendrogram
87
dependent-samples analysis of variance
88
design effect
For example, a teacher may form a de- removing short-term cyclical variations
scriptive hypothesis that students will in a TIME SERIES to reveal the underly-
understand at least 70% of the course ing trend. For example, a marketing
material, without delving into the spe- researcher may want to investigate
cific reasons for such understanding. whether a particular advertising strategy
increased sales, after removing the fluc-
descriptive measure a variable that
tuations in the data that have to do with
measures something essential about a
the time of year (e.g., holiday sales, sum-
data SAMPLE, without making an infer-
mer spending). Compare DETRENDING.
ence as to the population PARAMETER
for that variable. For example, a teacher desensitizing n. see DEHOAXING.
may provide a descriptive measure to a
specific class that the average grade on design n. the format of a research study,
an exam was 75%, without inferring describing how it will be conducted and
that all students outside of the class also the data collected. For example, an EX-
would understand 75% of the exam ma- PERIMENTAL DESIGN involves an INDE-
terial. PENDENT VARIABLE and at least two
groups, a treatment or EXPERIMENTAL
descriptive research an empirical in- GROUP and a CONTROL GROUP, to which
vestigation designed to test prespecified participants are randomly assigned and
hypotheses or to provide an overview of then assessed on the DEPENDENT VARI-
existing conditions and, sometimes, re- ABLE. A variety of other design types
lationships, without manipulating vari- exist, including CORRELATIONAL RE-
ables or seeking to establish cause and SEARCH, QUASI-EXPERIMENTAL DESIGNS,
effect. For example, a survey undertaken LONGITUDINAL DESIGNS, NATURAL EX-
to ascertain the political party prefer- PERIMENTS, and OBSERVATIONAL STUD-
ences of a group of voters would be a de- IES, among others.
scriptive study because it is intended
simply to identify attitudes rather than designatory scale a system in which a
systematically infer or analyze influenc- number is assigned without that num-
ing factors. ber having an actual quantitative value.
For example, social security numbers use
d e s c r i p t i v e s t a t i s t i c s procedures for a designatory scale to uniquely identify
depicting the main aspects of sample each person, but a larger social security
data, without necessarily inferring to a number does not indicate anything dif-
larger population. Descriptive statistics ferent than a smaller social security
usually include the MEAN, MEDIAN, and number.
MODE to indicate CENTRAL TENDENCY,
as well as the RANGE and STANDARD DE- design effect an adjustment, used in
VIATION that reveal how widely spread GROUP-RANDOMIZED TRIALS Or HIERAR-
the scores are within the sample. De- CHICAL LINEAR MODELS, for the depen-
scriptive statistics could also include dence in the data when groups of similar
charts and graphs such as a FREQUENCY individuals (e.g., those in particular
DISTRIBUTION or HISTOGRAM, among classrooms or hospitals) are randomly
others. Compare INFERENTIAL STATIS- assigned to treatment conditions. The
TICS. formula for the design effect takes into
account the number of individuals in
descriptive survey a questionnaire de- each cluster or group and the degree of
signed to assess a specific sample on a dependence among scores (the INTRA-
given set of items without implying or CLASS CORRELATION). STANDARD ERRORS
inferring any causal link between vari-
are adjusted with the design effect to
ables.
correct the negative BIAS that occurs
deseasonalization n. the process of when using dependent data.
89
design matrix
design m a t r i x a grid of data whose ele- ing no difference from normal. See also
ments denote the presence or absence of NORMAL PROBABILITY PLOT.
each participant (row) in a treatment
(column) of an experimental study. detrending n. the practice of removing
a specific existing pattern from data,
determinant n. a value that represents often collected in a TIME SERIES, in order
the generalized variance in a MATRIX of to reveal other expected patterns or se-
JM numbers, with large values indicating quences in the data, DESEASONALIZ-
that the matrix numbers are very dissim- ATION is a similar process but removes
ilar and thus have more varied informa- the effect of seasonal shifts in the data.
tion. A determinant can be calculated by
multiplying the EIGENVALUES of a ma- developmental survey an assessment
trix by one another. For example, a 2 x 2 of information across time on a set of in-
CORRELATION MATRIX with two OR- dividuals. For example, a researcher may
THOGONAL (i.e., unrelated) variables wish to assess the lifestyle habits and at-
would have eigenvalues of 1 and 1, titudes of adolescents over a 10-year pe-
which would yield a determinant of 1 x riod.
1, or 1. In contrast, a correlation matrix deviance n. in statistics, a measure of
with two collinear or completely related the GOODNESS OF FIT between a smaller
variables would have eigenvalues of 1 HIERARCHICAL MODEL and a fuller model
and 0, yielding a determinant of 0. Thus, that has all of the same parameters plus
the determinant of an orthogonal set of more. The difference or deviance be-
variables is very high (i.e., 1), as the tween these models follows a CHI-
two variables provide very different in- SQUARE DISTRIBUTION, with the DE-
formation. In contrast, the determinant GREES OF FREEDOM equal to the number
of the correlation matrix with com- of parameters that are added by the
pletely collinear variables would be as fuller model. If the deviance reveals a
low as possible (i.e., 0), as the two vari- SIGNIFICANT DIFFERENCE, then the
ables are redundant. Also called gener- larger model is needed. If the deviance is
alized variance. not significant, then the smaller, more
determination coefficient see COEF- parsimonious model is retained as more
FICIENT OF DETERMINATION. appropriate.
deterministic model a mathematical deviance information criterion
function in which the outcome can be (DlC) an index that reveals the DEVI-
exactly established. In other words, the ANCE between HIERARCHICAL MODELS,
model explains all of the variance in a where one model has a larger number of
DEPENDENT VARIABLE and nO ERROR parameters than the other. It is similar to
TERM is needed. This contrasts with a AKAIKE'S INFORMATION CRITERION and
STOCHASTIC MODEL, from which a range the BAYESIAN INFORMATION CRITERION
of possible values may result. but usually easier to calculate. In all
three indices or criteria, a smaller value
deterministic process see STOCHAS- indicates a more acceptable model.
TIC PROCESS.
deviance residual an index of the con-
detrended n o r m a l plot a graph in tribution made by a single observation
which the vertical y-axis represents devi- to the DEVIANCE between two models.
ations from normal and the horizontal
Ai-axis indicates the range of values for deviant case analysis a procedure
an observation. When observations are used first to understand why specific ob-
normally distributed, the points of a servations or data points differ from
detrended normal plot will fall on a what is expected in a specific model and
straight horizontal line at zero, indicat- then to adjust the model to accommo-
90
diagonal matrix
date most of the data. Also called nega- number of observations or cases in the
tive case analysis. sample.
deviate n. in statistics, the extent to DFFITS in REGRESSION ANALYSIS, dif-
which a score differs from a specified ference in fits: an index of the influence
value, such as the MEAN. Thus, a nega- that a particular case (/) has upon the fit-
tive deviate from the mean indicates ted value y,. It is one of several indices
how much lower than average the score that are useful in diagnosing problems
is, while a positive deviate from the in regression analysis, and shows how
mean indicates how much higher than much a predicted y (y) changes when a
average the score is. particular case is excluded from the cal-
culation of the weights in the regression
deviation n. a significant departure or equation. One rule of thumb is to delete
difference. This conceptually broad term a case if its DFFITS absolute value is
has a variety of applications in psychol- greater than 2/V(p/n), where n is the
ogy and related fields but in statistics number of observations or cases in the
refers to the arithmetical difference be- sample and p is the number of predictor
tween one of a set of values and some variables plus 1.
fixed amount, generally the MEAN of the
set or the value predicted by a model. See diachronic adj. see SYNCHRONIC.
STANDARD DEVIATION.
diagnostics pl. n. procedures for evalu-
deviational f o r m u l a an equation ating how much a model differsfromex-
that computes the average difference of pected patterns and ASSUMPTIONS. For
each score in a set of data from the MEAN example, diagnostics may reveal how
score. In other words, it is the equation much a set of data departs from assump-
to calculate STANDARD DEVIATION, tions of NORMALITY, LINEARITY, and
which is equal to the square root of the HOMOGENEITY OF VARIANCE, possibly
VARIANCE. showing skewed, curvilinear, and un-
equal variances across levels of another
deviation score the difference between variable, respectively. See also REGRES-
an observation or value x and the MEAN SION DIAGNOSTICS.
value (i.e., x - mean) in a set of data. The
sum of the deviation scores for a given diagnostic test any examination or as-
data set will equal zero, as approxi- sessment measure that may help reveal
mately half of the values will be less than the nature and source of an individual's
the mean and half will be greater than physical, mental, or behavioral problems
the mean. Also called deviation value. or anomalies. In medical research, for
example, a diagnostic test would be ex-
d f symbol for DEGREES OF FREEDOM. pected to show SENSITIVITY (i.e., cor-
rectly identifying individuals with a
DFBETAS in REGRESSION ANALYSIS, certain illness) and SPECIFICITY (i.e., cor-
di/ferences in beta values: an index that rectly identifying those who do not have
describes the influence of the ith case a specific illness).
upon the estimates of REGRESSION COEF-
FICIENTS. It is one of several indices that d i a g o n a l m a t r i x a SQUARE MATRIX in
are useful in diagnosing problems in re- which all of the values in the lower and
gression analysis, showing how much a upper triangular sections are equal to
coefficient would change if a single zero. Probably the most common diago-
given case (e.g., an OUTLIER) were nal matrix is the IDENTITY MATRIX,
dropped from the data. One mle of which has values of one on the main di-
thumb is to delete a case if its DFBETAS agonal (upper left to lower right) and
value is greater than 2/V, where n is the values of zero elsewhere.
91
diagonals model
o 0
O B
b
~1 1 1 1 1 1
4 8 t2 16 20 24
Sum
93
differential scoring
scale, whereas a weight of 145.5 to 145.9 starting point. The following is a generic
may register as 145.5. Thus, the scale example.
and researchers using it can be said to
have a digit preference for weight re-
corded in decimals ending in 0 or 5.
digram-balanced adj. describing an
experimental design in which treat-
ments are arranged so as to avoid a SE-
QUENCE EFFECT. For example, a research It may be used to represent data from
study that had four possible treatment any of various statistical methodssuch
conditions (A, B, C, D) that were all as- as STRUCTURAL EQUATION MODELING,
signed to each participant would be PATH ANALYSIS, and REGRESSION ANALY-
digram-balanced if different individuals SISin which one variable predicts one
received the treatments in varying or- or more other variables but the reverse is
ders: ABCD, DCBA, ACBD, DBCA. In such not tme; that is, prediction is not al-
a case, the research design itself is known lowed in both directions (e.g., if x pre-
as a B A L A N C E D L A T I N S Q U A R E . AlSO Called dicts y, then y cannot also predict x).
diagram-balanced. See also COUNTER- Also called acyclic digraph.
BALANCING; GRAECO-LATIN SQUARE. flirected g r a p h any display in which
f l i g r a p h n. a shortened name for DI- points (nodes) of information are con-
RECTED GRAPH. nected by lines (edges) showing direc-
tion (i.e., the edges have arrows or the
dimension n. in statistics, a factor or nodes are numbered). Consider the fol-
component that is applied in measuring lowing generic example.
a set of variables. For example, some in-
telligence scales are considered to have
two important dimensions consisting of
verbal and nonverbal intelligence.
dimensionality n. the number of
DIMENSIONS applied in measuring a
constmct. In FACTOR ANALYSIS and
PRINCIPAL COMPONENTS ANALYSIS, it iS
important to assess the dimensionality
of a set of items on a scale in order to
form cohesive subscales that each de- Such graphs are used to represent a di-
scribe a similar set of items. rectional process (e.g.. Point A predicts
Point B and Point B predicts Point C).
dip test o f u n i m o d a l i t y a statistical Also called digraph. Compare UNDI-
procedure to assess whether there is RECTED GRAPH.
more than one peak or MODE in a set of
data. For example, a dip test of uni- d i r e c t i o n a l hypothesis a scientific
modality for a bell-shaped NORMAL DIS- prediction stating (a) that an effect will
TRIBUTION would be nonsignificant, as occur and (b) whether that effect will
such a distribution has only one peak. specifically increase or specifically de-
crease, depending on changes to the IN-
direct correlation see POSITIVE COR-
DEPENDENT VARIABLE. For example,
RELATION.
a directional hypothesis could predict
directed acyclic g r a p h (DAG) a DI- that depression scores will decrease fol-
RECTED GRAPH in which there is no way lowing a six-week intervention, or con-
of moving through all of the points versely that well-being will increase
(nodes) in a sequence that returns to the following a six-week intervention. Also
95
directionality problem
96
discourse analysis
may have been incorrectly entered into connected. For example, a sequence of
computer records. ORDINAL numbers is discontinuous be-
cause it does not have a smooth flow
disaggregation n. the process of break- from one value to the next. See also DIS-
ing down data into smaller units or sets CRETE. Compare CONTINUOUS.
of observations. For example, faculty sal-
ary data initially may show a signiflcant discontinuous f u n c t i o n a graph in
difference between men and women. which there is a sudden break or discon- W
After disaggregating the data into sepa- nection in the flow of points that are
rate levels (e.g., assistant, associate, full plotted. Compare CONTINUOUS FUNC-
professor), however, one may find that TION.
there are no significant differences in discontinuous variable a variable
salary among men and women at the as- that has distinct, DISCRETE values but no
sistant professor level but that there are precise numerical flow. For example,
differences at the full professor level. gender can be thought of as a discontin-
Thus, disaggregating the data reveals a uous variable with two possible values,
finer-tuned pattern, suggesting that rnale or female. In contrast, a CONTINU-
long-term faculty differential compensa- OUS VARIABLE involves numerically pre-
tion practices applied in the past have cise information, such as height, weight,
been maintained over time. Compare and miles per hour. Also called discrete
AGGREGATION. variable.
disattenuated correlation a CORRE- discordance n. the state or condition of
LATION COEFFICIENT that removcs the being at variance. For example, in TWIN
effects of MEASUREMENT ERROR, on the STUDIES discordance refers to dissimilar-
principle that the obtained correlation ity between a pair of twins with respect
between x and y is limited by the RELI- to a particular trait or disease. Compare
ABILITY ofeach of the variables; an unre- CONCORDANCE. discordant adj.
liable measure (e.g., x) cannot predict
another unreliable measure (e.g., y). It is discordant s i b l i n g study research
formed by dividing a correlation be- that involves family members who differ
tween X and y hy the square root of the on specific characteristics or traits. For
product of the respective reliabilities for example, researchers may collect data
X and y. For example, assume the correla- on families that have children with dif-
tion between x and y was .30, the reli- fering gender, eye color, or blood types.
ability of A; was .70, and the reliability of discourse analysis a method of study-
y was .60. The disattenuated or corrected ing verbal communications that extends
correlation would be beyond the single sentence to encom-
.30/V(.70x.60) = .46. pass conversations, narratives, and writ-
ten arguments. Discourse analysis is
Thus, removing the effect of unreliabil- particularly concerned with the ways in
ity in both X and y would result in a which a sequence of two or more sen-
medium-size correlation of .30 becom- tences can produce meanings that are
ing closer to a large (i.e., .50) correlation different from or additional to any
of .46. Compare ATTENUATION. found in the sentences considered sepa-
rately. An important source of such
disclosure o f deceptions in DECEP-
meanings is the "frame" or format ofthe
TION RESEARCH, a participant's revela-
discourse (news item, fairytale, joke,
tion of the true nature of the study, even
etc.) and a recognition of the various
though he or she was asked not to give
norms that this implies. The norms and
others any information about it.
expectations that govern conversation
discontinuous adj. intermittent or dis- are a major concern of discourse analy-
97
discovery research
98
disjoint sets
99
dismantling study
intersection of the groups is empty, and MEAN. The larger the dispersion mea-
they are mutually exclusive. An example sure, the more spread out the scores.
of disjoint sets would be the set of all
dispersion parameter an index of the
males and the set of all females.
SKEWNESS (lopsidedness) or KURTOSIS
d i s m a n t l i n g s t u d y research that ex- (peakedness) of a distribution. The NOR-
amines whether individual components MAL DISTRIBUTION thus does not need a
of a larger protocol can be just as effec- dispersion parameter because skewness
tive and with less effort and expense. and kurtosis are not present (i.e., equal
For example, a condition traditionally zero) in such a distribution. Also called
treated by several months of therapy in scale parameter.
addition to medications could be inves- d i s p e r s i o n test an assessment of
tigated via a dismantling study to assess whether two samples have a similar
whether just the therapy, just the medi- spread of values on a variable of interest.
cation, or perhaps a reduced time i n This is useful in determining whether
therapy plus some medication would be there is evidence for H O M O G E N E I T Y OF
equally as effective as the full interven- VARIANCE, which allows for pooling or
tion. averaging variances across samples.
d i s o r d i n a l i n t e r a c t i o n in a FACTO- d i s p l a y n. the presentation of stimuli to
RIAL ANALYSIS OF VARIANCE, an INTER- any of the senses.
ACTION EFFECT in which the direction
of influence of one of the I N D E P E N D E N T
d i s s i m i l a r i t y c o e f f i c i e n t an index
VARIABLES differs depending on the of difference or distance between two
objects, variables, or samples. It may
level of a second independent variable.
be based on C O R R E L A T I O N , with lower
A disordinal interaction is indicated by a
values indicating more difference or dis-
pattern of crossed lines when plotting
tance, or VARIANCE, with higher values
the two MAIN EFFECTS from the analysis.
indicating greater difference or distance.
For example, if a researcher found that a
A dissimilarity coefficient often is used
particular treatment for dementia im-
to assess the difference between vari-
proved functioning in individuals with
ables or entities in multivariate proce-
mild and moderate degrees of the dis- dures such as CLUSTER ANALYSIS,
ease but hindered functioning in those CORRESPONDENCE ANALYSIS, MULTIDI-
with a severe degree, then a disordinal MENSIONAL SCALING, and PRINCIPAL
interaction would be present. Also called COMPONENTS ANALYSIS. Also Called dis-
crossover i n t e r a c t i o n . Compare OR- tance measure. Compare SIMILARITY
DINAL INTERACTION.
COEFFICIENT.
d i s p e r s i o n n. the degree to which a set d i s s i m i l a r i t y m a t r i x a M A T R I X used
of scores deviate from the mean. Also to indicate the distance or difference be-
called spread. See also RANGE; STANDARD tween pairs of objects, variables, or sam-
DEVIATION; VARIANCE. ples. Also called distance m a t r i x . See
d i s p e r s i o n m a t r i x see C O V A R I A N C E also DISSIMILARITY COEFFICIENT. Com-
MATRIX. pare SIMILARITY MATRIX.
d i s t a l c a u s e see U L T I M A T E C A U S E .
d i s p e r s i o n m e a s u r e an index of vari-
ability, or how dispersed a set of values is d i s t a l c o n t r o l a C O N T R O L G R O U P for
for a given variable. Common measures an experiment or research project that is
of dispersion include the RANGE, which kept physically distant from the EXPERI-
is the highest minus the lowest score, M E N T A L GROUP. A distal control group
and the S T A N D A R D D E V I A T I O N , which is may permit a better experimental com-
the average distance of scores from the parison in that there is less chance of
divisive clustering
101
domain
tion. [Howard E. Doran, U.S. econo- pants, and it uses half of the participants
metrician] needed for a double BLIND between-
groups study in which separate groups
dose-ranging trial a clinical trial in
are randomly assigned to different treat-
which patients are randomly assigned to
ments. A possible problem with the
treatment groups to receive one of sev-
crossover facet of the design is that the
eral different dosages of a drug. A dose-
effects of an initial treatment may not
ranging trial provides information on
fully be gone by the time the second
how much medicine is needed for it to
treatment is administered, thus causing
be effective. Also called dose-ranging
potential bias in the interpretation of
study.
which treatment is causing the most ef-
dot p l o t see SCATTERPLOT. fective outcome. See also CROSSOVER
DESIGN.
double bar g r a p h see DUAL BAR
GRAPH. double bootstrapping a procedure in
which samples are randomly drawn,
double-barreled question an item with replacement, from an initial data
on a survey that simultaneously asks set and their PARAMETERS and STAN-
about two kinds of information instead DARD ERRORS estimated and averaged
of just one, thus making the responses across the set of samples. This is then fol-
difficult to interpret. An example of a lowed by another random sampling of
double-barreled question is "How often the data, again with replacement, after
do you use a condom after drinking alco- which parameter estimates and standard
hol?" If a respondent said very often, a errors are obtained a second time. Dou-
researcher does not know whether this ble bootstrapping usually is less biased
refers to the frequency of condom use, than using a single set of BOOTSTRAP-
the frequency of drinking alcohol, or PING samples; additional procedures
both. It would be best to ask two sepa- often are applied between the two boot-
rate questions to obtain this informa- strapping procedures in order to further
tion (e.g., "How often do you use a reduce BIAS.
condom?" and "How often do you drink
alcohol?"), with each question having double cross-validation a procedure
its own set of response choices (e.g., in which a sample is split into two por-
"Never" up to "Quite Often"). tions or subsamples and a statistical
model (e.g., a MULTIPLE REGRESSION
double b l i n d see BLIND. EQUATION) is estimated from the first
double-blind crossover t r i a l a type portion and verified with the second
of clinical trial in which participants re- portion. The process is then reversed: An
ceive two different treatments and nei- estimate of the model is obtained using
ther the researchers nor the participants the second subsample and verified using
are aware of which treatments are ad- the first subsample. If results from botfi
ministered to which individuals at what validations are similar, there is strong
time. Participants are randomly assigned evidence that the model is robust and
to a particular order of treatments generalizable across samples that are
(Treatment A followed by Treatment B, similar to those used in the two sets of
or Treatment B followed by Treatment analyses. See CROSS-VALIDATION.
A); between treatments they go through
a brief WASH-OUT PERIOD to eliminate double deception in DECEPTION RE-
the effect of the initial treatment before SEARCH, an apparent debriefing after the
the second treatment is administered. experiment, which is itself a further
This kind of design should neutralize po- ploy.
tential bias from researchers or partici- double-dummy technique a re-
103
double exponential distribution
search design used when two treatments initial start date of the study would have
are dissimilar enough for participants to to be substituted instead. Similarly, the
know which treatment they are receiv- quit date for those who had not yet quit
ing just by its form (e.g., a comparison of smoking by the conclusion of the study
tablet and capsule medication types). In also would not be known, and the end
the double-dummy technique, each par- date of the study would have to be sub-
ticipant is randomly assigned to at least stituted instead. Thus, both ends of the
one inactive PLACEBO (i.e., dummy) distribution of quit dates would be cen-
condition plus a regular treatment. sored. See CENSORED DATA.
Thus, with a double-dummy technique
participants are randomly assigned to doubly m u l t i v a r i a t e describing a de-
one of the following three conditions: sign in which two sets of multiple vari-
(1) Treatment A (e.g., active tablet) plus ables are assessed. For example, assume a
Placebo B (e.g., inactive capsule); (2) Pla- researcher evaluates several reading per-
cebo A (inactive tablet) plus Treatment B formance outcomes (e.g., number of
(active capsule); or (3) Placebo A plus books read, difficulty level of the books,
Placebo B, which serves as a CONTROL. vocabulary level) for a set of grade-
school students. Further, imagine the re-
double exponential d i s t r i b u t i o n a searcher assesses the students at the be-
graphed curve that rises very quickly to a ginning, middle, and end of the school
peak and then falls very quickly to the year. In assessing multiple reading out-
horizontal axis, forming a symmetrical comes at multiple time points, the re-
shape with the appearance of a witch's searcher would be using a doubly
hat. The pattern is similar to a normal multivariate design.
BELL CURVE except that it is not as full at
the peak and tails. Also called Laplace doubly stochastic 1. describing a
distribution. SQUARE MATRIX of ttonnegative num-
bers in which all of the rows sum to 1
double masked see BLIND. and all of the columns sum to 1. 2. de-
scribing a model in which observations
double s a m p l i n g a research process in of a RANDOM VARIABLE are modeled in
which data can be sampled a second two stages, in the second of which one
time if the results from a first statistical or more of the parameters used in the
test just missed significance due to low first stage are themselves treated as ran-
POWER. Because this process can result dom (STOCHASTIC) variables. Doubly
in a higher TYPE I ERROR rate, it should stochastic models are used to model pro-
be used only when needed to conserve cesses with a random probability pat-
costs and maximize power. tern, such as TIME SERIES.
doubly censored data data in which d p r i m e (symbol: d') a measure of an in-
exact measurements are not known for dividual's ability to detect signals; more
events at both ends of a distribution. For specifically, a measure of sensitivity or
example, assume a health researcher ex- discriminability derived from SIGNAL
amined the time at which a group of DETECTION THEORY that is unaffected by
individuals quit smoking cigarettes. Fur- response biases. It is the difference (in
ther, consider that there was a subgroup standard deviation units) between the
who had already quit smoking at the means of the noise and signal + noise
time of the initial assessment, as well as a distributions. A value ofrf'= 3 is close to
subgroup who had not yet quit smoking perfect performance; a value of d' = 0 is
at the end of the final assessment. Thus, chance ("guessing") performance.
the exact quit dates for those who quit
before the study began would not be d r i f t n. 1. a reduction in variation in ge-
known, and for purposes of analysis the netic traits that can occur when sam-
dual scaling
0)
o
C 40
Male
Pizza Choice
105
Duhem-Quine thesis
107
Ee
e 1. abbreviation for EXPONENT. 2. in university and then find the association
measurement theory, abbreviation for between these sets of means. This would
ERROR. provide information at a departmental,
instead of an individual, level.
E symbol for EXPECTED VALUE.
ecological f a l l a c y a mistaken conclu-
Eberhardt's statistic an index of ran- sion drawn about individuals based on
domness in a spatial distribution. If it is findings from groups to which they be-
significant, a researcher may conclude long. For example, if a university admin-
that the results are not due to random- istrator found that the ECOLOGICAL
ness but rather have a meaningful pat- CORRELATION between faculty salary
tern. and number of publications at the de-
ECM a l g o r i t h m abbreviafion for EXPEC- partmental level was strong and positive
TATION-CONDITIONAL MAXIMIZATION AL-
(e.g., r = .60), it would be an ecological
GORITHM. fallacy to assume that for any particular
faculty member the correlation would
ECME a l g o r i t h m abbreviation for EX- be the same.
PECTATION-CONDITIONAL MAXIMIZATION
EITHER ALGORITHM. ecological inference an assumption
about individuals or subsets of individu-
ecological assessment the gathering als on the basis of aggregate data about
of observations in various environments the larger group. For example, a re-
to examine whether individuals or enti- searcher may wish to make an ecological
ties behave differently depending on the inference about the education of women
surroundings. For example, a teacher at from ethnic minorities on the basis of
a child-development center may investi- data from a report that provides infor-
gate whether children demonstrate mation by gender or ethnicity but not
more or less problem behavior in various both together. Various strategies have
settings at the center. Findings from the been suggested for making vahd ecologi-
ecological assessment may offer ideas on cal inferences from summary statistics
how to intervene to improve the overall (i.e., inferences that do not commit the
behavior of the children in multiple set- ECOLOGICAL FALLACY).
tings.
ecological momentary assessment
ecological correlation a correlation (EMA) the process of examining the be-
that is calculated between group means havior of individuals at random, multi-
instead of between individual scores, ple time points to get a clearer picture of
thereby showing the relationship be- how they behave in various real-world
tween variables at the group level. For settings. For example, a researcher may
example, a university administrator may conduct an EMA on individuals who are
want to assess the relationship between trying to quit smoking by contacting
faculty salary and productivity as mea- them randomly throughout a week to
sured by the number of publications. An inquire about urges to smoke, mood,
ecological correlation would find the av- other people who are interacting with
erage salary and average number of pub- them, and other relevant circumstances
lications for each department at the that may be occurring at these times.
108
effective sample size
109
effect modifier
110
empirically keyed test
111
empirical method
differ. For example, on a test measuring that the mind at birth is like a blank
problem solving, the correct alternative sheet of paper. 2. the view that experi-
among the response choices would be mentation and systematic observation is
the one preferred by members of a CRI- the most important, if not the only,
TERION GROUP who were administered foundation of scientific knowledge and
the test previously. the means by which individuals evalu-
ate truth claims or the adequacy of theo-
empirical method any procedure for ries and models. 3. in philosophy, the
conducting an investigation that relies position that all linguistic expressions
upon experimentation and systematic that are not tautologous must be empiri-
observation rather than theoretical spec- cally verifiable if they are to be deemed
ulation. The term is sometimes used as a valid or meaningful. This principle was
vague synonym for SCIENTIFIC METHOD. essential to the philosophy of LOGICAL
empirical probability the number of POSITIVISM, empiricist adj., n.
times a specific event occurs over the
empowerment evaluation see PAR-
total number of trials in a process. For
TICIPATORY EVALUATION.
example, the THEORETICAL PROBABILITY
of getting a head whenflippinga coin is empty cell a category of a specified vari-
equal to .5 or V2, whereas the actual em- able or CROSS-TABULATION of variables
pirical probability obtained in an experi- in which there is no membership. For
ment may be different. Also called example, a marketing researcher might
experimental probability. be interested in color preferences (pink,
blue) for baby clothes by gender (girl,
empirical question a problem that
boy). He or she could collect data from a
can be solved by conducting an experi-
random sample of 100 people in a baby
ment or investigation, rather than by
clothes department and then create a
logic or theory. For example, "How
CONTINGENCY TABLE of the findings. For
many people own a white car?" is an em-
example, 40 people may prefer pink
pirical question. Determining its answer
clothes for girl babies, 10 may prefer
would require collecting a rather large,
blue clothes for girl babies, 50 may pre-
random sample that is representative of
fer blue clothes for boy babies, and none
the population of car owners and asking
prefer pink clothes for boy babies. The
each person in the sample the color of
CELL representing the intersection of
his or her car.
"pink" and "boy" is an empty cell
empirical test the test of a hypothesis because there is no one who endorses
by means of experiments or other sys- that combination of categories.
tematic observations.
empty set in SET THEORY, a collection of
e m p i r i c a l v a l i d i t y the degree to entities with no ELEMENTS or members.
which the accuracy of a test, model, or Also called null set.
other construct can be demonstrated
end-aversion bias the tendency for in-
through experimentation and system-
dividuals to avoid the extreme choices
atic observation (i.e., the accumulation
on a scale and instead select a choice in
of supporting research evidence) rather
the middle of the scale, closer to neutral.
than theory alone.
Such an approach results in a narrower
empiricism n. 1. an approach to EPISTE- range of responses that most likely will
MOLOGY holding that all knowledge of not be an accurate representation of the
matters of fact either arises from experi- variable being measured. For example,
ence or requires experience for its valida- a supervisor assessing employees would
tion. In particular, empiricism denies show end-aversion bias if his or her rat-
the possibility of innate ideas, arguing ings generally were around 4 on a 7-point
112
equal intervals
113
equality constraint
that, for example, the difference be- equiprobable adj. describing two or
tween agree and strongly agree is the same more different events that have the
as that between agree and strongly dis- same probability of occurring. For exam-
agree. ple, when flipping a fair coin, the occur-
equality constraint a restriction re- rence of a head or a tail is equiprobable.
quiring that two or more PARAMETERS equivalence n. a relationship between
have the same value. This is sometimes two or more items, scales, variables, or
used in procedures such as CONFIRMA- stimuli that permits one to replace an-
TORY FACTOR ANALYSIS, for example, other. See also EQUIVALENCE COEFFI-
when testing whether a FACTOR LOAD- CIENT; MEASUREMENT EQUIVALENCE.
ING is the same across different measures
of the same construct. Compare IN- equivalence coefficient a numerical
EQUALITY CONSTRAINT. index of the similarity of two different
assessments of the same measure, often
equality of variance see HOMOGENE- used as an indication of ALTERNATE-
ITY OF VARIANCE. FORMS RELIABILITY for slightly different
equal p r o b a b i l i t y of selection forms of the same test given on two dif-
method (EPSEM) a procedure for ran- ferent occasions. For example, there is
domly choosing from a larger popula- often a need to calculate an equivalence
tion, such that each person or entity has coefficient for standardized tests admin-
an equal chance of being chosen. See istered to a group of persons more than
RANDOM SAMPLING. once, to ensure that each form of the test
is measuring the same construct, such as
equal p r o b a b i l i t y s a m p l i n g see UN- verbal or quantitative ability.
EQUAL PROBABILITY SAMPLING.
equal w e i g h t i n g see WEIGHTING. equivalence paradox the situation in
which two different procedures produce
equated score the score distribution very similar outcomes, despite having
from Measure B transformed to match different initial assumptions or features.
the distribution of Measure A in one or In statistics, the term refers to the fact
more features. See TRANSFORMATION. that PRINCIPAL COMPONENTS ANALYSIS
equation n. a formal, usually brief, and FACTOR ANALYSIS often produce a
statement showing that two expressions similar pattern of dimensions and load-
are the same or equal. In mathematics or ings, although they each have a different
statistics, an equation is often a state- premise. Whereas principal components
ment with y on the left of an equal sign analysis mathematically redistributes all
and an expression for delineating or of the variance in the items to form di-
modeling y on the right side. For exam- mensions or components, factor analy-
ple, a commonly used equation for a sis more realistically separates out
straight line isy=mx-^-b, where m stands unique variance in the items before
for the slope of the line, x stands for dif- forming dimensions or factors using
ferent values along the X-AXIS, and b only the variance that is shared across
stands for the point where the line the variables.
crosses the Y-AXIS.
equivalent f o r m see ALTERNATE FORM.
equipercentile method a procedure equivalent-forms r e l i a b i l i t y see AL-
for showing how two measures are simi- TERNATE-FORMS RELIABILITY.
lar, such that a shared value of x on the
two measurements implies that the equivalent-groups design a study in
probability of a person drawn at random which the groups of participants are as-
having a score greater than x is the same sumed to be the same on all possible
for both measures. variables at the beginning of a study,
error of measurement
such that at the end of the study any dif- rate information (e.g., a wrong response,
ferences on the response measure can a mistaken belief). 2. in experimenta-
be attributed to the experimental ma- tion, any change in a DEPENDENT VARI-
nipulation. Although the best way to ABLE not attributable to the mani-
achieve an equivalent-groups design is pulation of an INDEPENDENT VARIABLE.
by randomly selecting a large and rele- 3. in statistics, a deviation of an ob-
vant sample from the population and served score from a true score, where
then randomly assigning participants true score is often defined by the mean
to either an EXPERIMENTAL GROUP or of the particular group or condition in
a CONTROL GROUP, sometimes this which the score being assessed for error
is not possible. In such situations, re- occurs, or from the score predicted by a
searchers can instead use MATCHING, model. Errors generally are categorized
with the expectation that the resulting as SYSTEMATIC ERROR Or RANDOM ERROR.
two participant groups then would be See also RESIDUAL.
approximately equivalent (i.e., come
from the same population) at the start of error bar an area on a graph used to de-
the study. pict how much uncertainty or slippage
For example, a reading researcher may there is around an estimated PARAMETER
want to implement a reading interven- value, such as the mean. An error bar
tion in one school and a control or PLA- usually is given as one STANDARD DEVIA-
CEBO condition in another similar TION above and below a point, but it also
school. Before analyzing results between can be a STANDARD ERROR (i.e., the stan-
the two schools, the researcher could ad- dard deviation divided by the square
minister an intelligence test to students root of the SAMPLE SIZE).
at both schools. Then, the researcher
error d i s t r i b u t i o n the pattern de-
could create similar sets by matching
scribing how scores can vary randomly
students with the same or highly similar
around a particular value, such as the
scores across the schools and conduct-
mean. It is often a NORMAL DISTRIBU-
ing the study with just those students.
TION with a mean of 0 and a STANDARD
See also MATCHED-PAIRS DESIGN. Com-
DEVIATION of 1.
pare NONEQUIVALENT-GROUPS DESIGN.
erf abbreviation for ERROR FUNCTION. error f u n c t i o n (erf) a roughly "S" or
half-mountain-shaped pattern on a
ergodicity n. a principle stating that the graph with values of a particular variable
average value of a variable over a set of (x) given along the horizontal axis and
individuals in a defined space or time, values from -1 to +1 given on the verti-
such as a SAMPLE, will be the same as the cal axis. Such a pattern indicates there
average across a long TIME SERIES of would be a high level of accuracy if using
points for a single individual. For exam- the mean to estimate a value for x, and
ple, if ergodicity held for a measure of conversely a large likelihood of error or
satisfaction in an organization, the aver- inaccuracy if using more extreme values.
age satisfaction score of all employees in
the organization would be the same as error mean square see MEAN
the average satisfaction score across a SQUARED ERROR.
one-year period for one employee. In re-
ality, ergodicity does not always hold, error o f expectation an error arising
thus giving rise to different streams because of a preconceived idea of the na-
of NOMOTHETIC and IDIOGRAPHIC re- ture of the stimulus to be presented or
search, which focus on the group or the the timing of the presentation.
individual, respectively.
error o f measurement see MEASURE-
error n. 1. a deviation from true or accu- MENT ERROR.
115
error of the first kind
117
etics
their natural habitat but also involving ered a Euler diagram. The following is a
experiments both in the field and in cap- generic depiction of a Euler diagram.
tivity. Ethology was developed by be-
Set1
havioral biologists in Europe and
is often associated with connotations
of innate or species-specific behavior
patterns. The term increasingly is used
to describe research involving obser-
vation and detailed descriptions of
human behavior as well. ethological
adj. ethologist n.
can accurately map what occurred at priate on different versions of a test and
various points in time. Event history methods or parameters that work well
data often are used in psychology, soci- under different conditions show ex-
ology, health sciences, education, and changeability.
engineering.
exclusive adj. completely separate and
event record a log containing the de- incompatible. For example, gender has
tails of a particular situation or occur- two exclusive categories of male and fe-
rence as assessed over a period of time. male. See also MUTUALLY EXCLUSIVE
For example, an event record may col- EVENTS.
lect information over numerous time
points throughout a week on the mood exclusive u n i o n in probability studies,
of a patient who is enrolled in a study to the combination of two or more events
reduce depression. that have no overlap. For example, the
set of all hearts and spades in a deck of
event s a m p l i n g a strategy commonly cards forms an exclusive union of two
used in DIRECT OBSERVATION that in- separate events, as a card cannot be both
volves noting and recording the occur- a heart and a spade at the same time. In
rence of a carefully specified behavior probability, the chance of two separate
whenever it is seen. For example, a re- events occurring together (i.e., an exclu-
searcher may record each episode of sive union) is equal to the probability of
apnea that occurs within a 9-hour pe- the first event plus the probability of the
riod overnight while a person sleeps. second event.
event space a subset of the set of all pos- exhaustive adj. complete or all encom-
sible outcomes in a situation (the SAM- passing. For example, a set of 13 hearts,
PLE SPACE). For example, the sample 13 spades, 13 diamonds, 13 clubs, and
space of working days for a particular or- two jokers forms an exhaustive set of all
ganization might comprise Sunday, possible playing cards that could be
Monday, Tuesday, Wednesday, Thurs- drawn.
day, Friday, and Saturday for all employ-
ees, but the event space for working days exogenous parameter a PARAMETER
might include only Monday through that occurs outside of and is not ex-
Friday for some employees. plained by a system or model but never-
theless has an effect on variables within
events p a r a d i g m a research design or that system. For example, poor socioeco-
therapeutic approach based on the inter- nomic status is sometimes seen as lead-
pretation of EVENT HISTORY DATA. ing to lower levels of education, which
exact i d e n t i f i c a t i o n see JUST-IDENTI- in turn could lead to lower income lev-
FIED MODEL. els. Thus, in a study of the effect of edu-
cation on income level socioeconomic
exact r e p l i c a t i o n see REPLICATION.
status might be considered an exoge-
exact test any statistical test in which nous parameter: It predicts a mediator
the probability of finding a result as ex- (education), which then predicts an
treme or more extreme than the one ob- endogenous or DEPENDENT VARIABLE
tained, given the NULL HYPOTHESIS, can (income). Compare ENDOGENOUS PARA-
be calculated precisely rather than ap- METER.
proximated. The FISHER EXACT TEST iS
exogenous variable an INDEPENDENT
an example.
VARIABLE whose value is determined by
exchangeability n. the ability to be factors outside the model or system
used in different circumstances or situa- under study. Also called exogenous
tions. For example, in measurement and factor. Compare ENDOGENOUS VARI-
statistics, items that are equally appro- ABLE.
119
exp
exp abbreviation for EXPONENTIAL FUNC- first variable plus the long-term average
TION. of the second variable.
expansion n. a description of a mathe- expectation-conditional maximi-
matical or statistical function via a series zation algorithm (ECM algorithm)
of sums. The Taylor expansion (see TAY- a procedure for finding M A X I M U M LIKE-
LOR SERIES) is an example. LIHOOD estimates of PARAMETERS in sta-
tistical models. Its computations are
expansion f a c t o r the amount by easier than those of traditional MAXI-
which scores can change over time or M U M LIKELIHOOD-EXPECTATION MAXI-
with differences in testing conditions. It MIZATION ESTIMATION methods.
is often important to consider an expan-
sion factor in ITEM RESPONSE THEORY expectation-conditional m a x i m i -
when predicting a pattern of scores. zation either a l g o r i t h m (ECME
algorithm) an iterative procedure for
expectancy n. see EXPECTATION. ex-
finding MAXIMUM LIKELIHOOD estimates
pectant adj.
of PARAMETERS in Statistical models that
expectancy chart a table or graph that alternates between performing an ex-
provides estimates of the future outcome pectation step and a maximization
for a specific variable. For example, in- step. It is an improved extension of the
dustrial and organizational psycholo- EXPECTATION-CONDITIONAL MAXIMIZA-
gists may use expectancy charts based TION ALGORITHM that is easier and faster
on test scores to determine the likeli- to compute.
hood of an individual's success in a job,
and insurance companies may use ex- expectation surface the space or
pectancy charts to predict how long a plane that contains a series of points
person will live based on such factors as that are the likely values of an outcome
health, current age, financial resources, given a set of predictor or explanatory
and occupation. variables or parameters.
120
experimental error
121
experimental ethics
react somewhat differently to the treat- tific study of behavior, motives, or cog-
ment, some more positively and some nition in a laboratory or other con-
more negatively, this is a random error trolled setting in order to predict,
that is not easily eradicated but that may explain, or influence behavior or other
be somewhat reduced (e.g., by randomly psychological phenomena. Experimen-
assigning a large set of individuals to tal psychology aims at establishing
EXPERIMENTAL GROUPS VS. CONTROL quantified relationships and explana-
GROUPS). tory theory through the analysis of re-
sponses under various controlled
experimental ethics see RESEARCH
conditions and the synthesis of ade-
ETHICS. quate theoretical accounts from the re-
experimental group a group of par- sults of these observations.
ticipants in a research study who are ex-
posed to a particular manipulation of experimental realism the extent to
the INDEPENDENT VARIABLE (i.e., a par- which a controlled study is meaningful
ticular treatment or TREATMENT LEVEL). and engaging to participants, eliciting
The responses of the experimental group responses that are spontaneous and nat-
are compared to the responses of a CON-
ural. See also MUNDANE REALISM.
TROL GROUP, other experimental groups, experimental r e p l i c a t i o n see REPLI-
or both. Also called treatment group. CATION.
experimental hypothesis a premise experimental research research uti-
that describes what a researcher in a sci- lizing randomized assignment of partici-
entific study hopes to demonstrate if pants to conditions and systematic
certain conditions are met, such as RAN- manipulation of variables with the ob-
DOM SELECTION of participants, RAN- jective of drawing causal inference. It
DOM ASSIGNMENT tO EXPERIMENTAL is generally conducted within a labora-
GROUPS or CONTROL GROUPS, and ma- tory or other controlled environment,
nipulation of an INDEPENDENT VARI- which in reducing the potential influ-
ABLE. ence of extraneous factors increases IN-
TERNAL VALIDITY but decreases EXTERNAL
experimental m a n i p u l a t i o n in an
VALIDITY. Compare CORRELATIONAL RE-
experiment, the manipulation of one or
SEARCH; FIELD RESEARCH; QUASI-EXPERI-
more INDEPENDENT VARIABLES in order
MENTAL RESEARCH.
to investigate their effect on a DEPEND-
ENT VARIABLE. An example would be the experimental series the trials admin-
assignment of a specific treatment or istered to an EXPERIMENTAL GROUP in 3
PLACEBO to participants in a research controlled research study, as opposed to
study in order to control possible CON- those administered to the CONTROL
FOUNDS and assess the effect of the treat- GROUP.
ment.
experimental treatment 1. In re-
experimental method a system of search, the conditions applied to one or
scientific investigation, usually based on more groups that are expected to cause
a design to be carried out under con- change in some outcome or DEPENDENT
trolled conditions, that is intended to VARIABLE. 2. an intervention or regimen
test a hypothesis and establish a causal that has shown some promise as a cure
relationship between independent and or ameliorative for a disease or condition
dependent variables. but is still being evaluated for efficacy,
experimental p r o b a b i l i t y see EM- safety, and acceptability.
PIRICAL PROBABILITY.
experimental u n i t the unit to which
experimental psychology the scien- an experimental manipulation is ap-
experimenter modeling effect
123
experimenter observer effect
124
exponential notation
125
exponential smoothing
by 10 raised to the negative of the num- group of participants who did not, yet
ber of zeros. For example, 5,000,000 the researcher has no control over which
written in exponential notation is 5 x specific people actually view the adver-
10^ and 0.0000005 written in exponen- tisement. Also called ex post facto
tial notation is 5 x 10"^. study.
exponential smoothing a method of ex post facto hypothesis a HYPOTHE-
weighting TIME-SERIES data using expo- SIS about what caused or brought about
nentially decreasing weights for older a condition that is made only after one
events. This reduces random fluctua- has examined information already col-
tions in the data so that one is better able lected about that condition.
to see the underlying trends in the series extended Cox regression model see
and gives recent data relatively greater
cox REGRESSION ANALYSIS.
importance in forecasting. Exponential
smoothing is widely used with financial external r e l i a b i l i t y the extent to
and economic data. which a measure is consistent when as-
sessed over time or across different indi-
exponential t r e n d a pattern in which viduals. External reliability calculated
data values increase over time propor- across time is referred to more specifi-
tional to a specific value. For example, cally as RETEST RELIABILITY; external re-
housing costs that increase annually by liability calculated across individuals is
a factor of 1.04 could be said to manifest referred to more specifically as INTER-
an exponential trend, such that one RATER RELIABILITY.
could use the trend to predict the future
value of any given house based on its external v a l i d i t y the extent to which
original cost. Thus, a house that cost the results of research or testing can be
$100,000 to purchase originally would generalized beyond the sample that gen-
cost an estimated $148,000 to purchase erated the results to other individuals.
10 years later, as determined by multi- Situations, and time periods. For exam-
plying the original house cost by 1.04, ple, if research has been conducted only
multiplying the resulting value by 1.04, with male participants, it cannot be as-
and then repeating the second multipli- sumed that similar results will apply to
cation step 10 times. female participants. The more special-
ized the sample, the less likely will it be
ex post facto data information about that the results are highly generalizable.
past occurrences. For example, a health Compare INTERNAL VALIDITY.
researcher could collect ex post facto
data on the number of hours per week a extraneous variable a measure that is
person believes he or she exercised over not under investigation in an experi-
the past month. ment but may potentially affect the out-
come or DEPENDENT VARIABLE and thus
ex post facto design research that ex- influence results. Such potential influ-
amines past occurrences in order to un- ence often requires that an extraneous
derstand a current state. Although this variable be controlled during research.
type of design involves both a DEPEND- See also CONFOUND.
ENT VARIABLE and an INDEPENDENT
VARIABLE, the investigator cannot ma-
extraneous variance variation in
nipulate the latter. For example, a re- scores that is' caused by CONFOUNDS,
searcher who is interested in deter- HIDDEN VARIABLES, or Other factors that
mining the effectiveness of a particular are not under investigation in a study.
television advertisement on consumer extrapolation n. the process of esti-
behavior may recruit a group of partici- mating or projecting unknown score
pants who saw the advertisement and a values on the basis of the known scores
eyeballing
obtained from a given sample. For exam- least values of a large set of random vari-
ple, a researcher might estimate how ables that form an approximately
well students will do on an achievement straight line from the lower left to the
test on the basis of their current perfor- upper right of a graph. The GUMBEL DIS-
mance, or estimate how well a similar TRIBUTION is an example.
group of students might perform on the eyeballing n. slang for a preliminary
same achievement test. casual look at research results, usually
extreme-value d i s t r i b u t i o n a lim- prior to a more formal analysis of the
ited DISTRIBUTION of the greatest or data. Also called eyeball test.
127
Ff
f 1. in an ANALYSIS OF VARIANCE or an pendent variable. 4. in FACTOR ANALY-
ANALYSIS OF COVARIANCE, an EFFECT SIZE SIS, an underlying, unobservable LATENT
index that represents the STANDARD DE- VARIABLE thought (together with other
VIATION of a set of group means divided factors) to be responsible for the interre-
by the average standard deviation of lations among a set of variables. 5. in
scores across the set of groups. Its value mathematics, a number that divides
ranges from 0 to infinity, with small, without remainder into another num-
medium, and large effect size values sug- ber.
gested as .10, .25, and .40, respectively.
2. in a table or distribution, symbol for f a c t o r analysis (FA) a broad family of
FREQUENCY. mathematical procedures for reducing a
set of intercorrelations among MANIFEST
F symbol for F RATIO. VARIABLES to a Smaller set of unobserved
FA abbreviation for FACTOR ANALYSIS. LATENT VARIABLES or factors. For exam-
ple, a number of tests of mechanical
facet n. 1. any component of an entity ability might be intercorrelated to en-
that itself may be assessed as an inde- able factor analysis to reduce them to a
pendent entity. For example, coopera- few factors, such as fine motor coordina-
tion and trust might be considered facets tion, speed, and attention. This tech-
of agreeableness in personality assess- nique is often used to examine the
ment. 2. in GENERALIZABILITY THEORY, common influences believed to give rise
an aspect of data that causes scores to to a set of observed measures (measure-
vary. For example, fluctuation in a set of ment structure) or to reduce a larger set
scores due to using different raters, dif- of measures to a smaller set of linear
ferent time points, or different items for composites for use in subsequent analy-
a particular test could be considered a sis (data reduction). See CONFIRMATORY
facet. 3. in FACTOR ANALYSIS, a primary FACTOR ANALYSIS; EXPLORATORY FAC-
or FIRST-ORDER FACTOR associated with TOR ANALYSIS; PRINCIPAL COMPONENTS
one or more SECOND-ORDER FACTORS. ANALYSIS.
face v a l i d i t y the apparent soundness f a c t o r i a l 1. adj. describing a design for
of a test or measure. The face validity of an ANALYSIS OF VARIANCE in which the
an instrument is the extent to which the effect of multiple INDEPENDENT VARI-
items or content of the test appear to be ABLES, each having two or more catego-
appropriate for measuring something, ries of response, is assessed. For example,
regardless of whether they actually are. a researcher may wish to examine the ef-
A test with face validity, however, may fect by gender (male vs. female) of a new
lack EMPIRICAL VALIDITY. drug treatment (drug vs. no drug). 2. n.
f a c t o r n. 1. anything that contributes to the value obtained when multiplying a
a result or has a causal relationship to a given integer by each of the positive in-
phenomenon, event, or action. 2. an tegers preceding it in value. Factorials
underlying influence that accounts in are indicated by a number followed by
part for variations in individual behav- an exclamation point. For example, 4! (4
ior. 3. in ANALYSIS OF VARIANCE and factorial) denotes 1 x 2 x 3 x 4 , which
other statistical procedures, an inde- equals 24.
factorial validity
129
factoring
Item
factor solution thatfitsthe data to a spe- show the number of additional studies
cific degree there will exist an infinite that would need to be added (e.g., from
number of equally good mathematical unpublished FILE-DRAWER ANALYSIS) in
solutions, each represented by a differ- order to make the effect size non-
ent FACTOR STRUCTURE MATRIX. ThuS, significant. The larger the value of fail-
rotation is required to obtain a solution safe JV, the higher the credibility of the
that is both mathematically viable and estimated meta-analytic effect size.
logically sound. See OBLIQUE ROTATION;
ORTHOGONAL ROTATION. f a i l u r e time see SURVIVAL TIME.
131
false positive
the variation of individual scores within CRITICAL VALUE of the F RATIO needed to
each group also increases or decreases. reject the NULL HYPOTHESIS. See F TEST.
The fan-spread hypothesis recognizes
feasibility study see PILOT STUDY.
that changes over time may get smaller
as individuals reach an ultimate goal. f i d e l i t y n. the degree of accuracy of a
For example, students in a program to measuring instrument or STATISTICAL
increase learning would not be expected MODEL. For example, a representation
to keep a constant rate of improvement derived from STRUCTURAL EQUATION
but rather to improve faster than a com- MODELING that depicts a pattern of rela-
parison group of students. tionships between health attitudes and
farthest neighbor see COMPLETE- behaviors could be said to have fidelity if
LINKAGE CLUSTERING.
it accurately explains the VARIATION
and COVARIATION in the data.
fatigue effect a decline in performance
on a prolonged or demanding research f i e l d n. somewhere other than a labora-
task that is generally attributed to the tory, library, or academic setting in
participant becoming tired or bored which experimental or NONEXPERI-
with the task. The fatigue effect is an im- MENTAL work is carried out or data col-
portant consideration when administer- lected.
ing a lengthy survey or test in which f i e l d experiment a study that is con-
participants' performance may worsen ducted outside the laboratory in a "real-
simply due to the challenges of an ex- world" setting. Participants are exposed
tended task. to one of two or more levels of an INDE-
FDA abbreviation for FUNCTIONAL DATA PENDENT VARIABLE and observed for
ANALYSIS. their reactions; they are likely to be un-
aware of the research. Such research
F d i s t r i b u t i o n a theoretical PROBABIL- often is conducted without RANDOM
ITY DISTRIBUTION widely used in the SELECTION or RANDOM ASSIGNMENT of
ANALYSIS OF VARIANCE, MULTIPLE RE- participants to conditions and no delib-
GRESSION, and other statistical tests of erate experimental manipulation of the
hypotheses about population variances. independent variable by the researcher.
It is the ratio of the variances of two in- See FIELD RESEARCH; QUASI-EXPERI-
dependent random variables each di- MENTAL RESEARCH.
vided by its DEGREES OF FREEDOM. In an
analysis of variance, for example, the F f i e l d notes notes on observations made
distribution is used to test the hypothe- in natural settings (i.e., the field) rather
sis that the variance between groups is than in laboratories. Field notes com-
significantly greater than the variance prise the data for subsequent analysis
within groups, thus demonstrating evi- in FIELD EXPERIMENTS and FIELD RE-
dence of some differences among the SEARCH.
means. Also called Fisher distribu-
f i e l d research studies conducted out-
tion; Fisher's F distribution; Fisher-
side the laboratory, in a "real-world" set-
Snedecor distribution; Snedecor's F
ting, which typically involve observing
distribution. See F RATIO; F TEST.
or interacting with participants in their
F d i s t r i b u t i o n table a table of values typical environments over an extended
listing the DEGREES OF FREEDOM for the period of time. Field research has the ad-
BETWEEN-GROUPS SUM OF SQUARES along vantages of ECOLOGICAL VALIDITY and
one side and the degrees of freedom for the opportunity to understand how and
the WITHIN-GROUPS S U M OF SQUARES why behavior occurs in a natural social
along another right-angle side, with val- environment; it has the disadvantages of
ues in the middle corresponding to the loss of environmental control and abil-
133
field survey
135
first passage time
called the first-order partial correlation John H. Halton, British-born U.S. com-
coefficient. Also called first-order par- puter scientist]
tial. See also PARTIAL CORRELATION.
Fisher F test see F TEST. [Sir Ronald
f i r s t passage time the amount of time Aylmer Fisher]
that elapses until the occurrence of a Fisher g test a statistical procedure for
random process in an entity or individ- analyzing whether there is any differ-
ual. For example, a medical researcher ence between the patterns of observed
may be interested in the first passage and expected CATEGORICAL DATA. The
time for a specific disease in a specific pa- Fisher g test is a LIKELIHOOD-RATIO TEST
tient. that is approximately distributed as a
CHI-SQUARE and can be used in the same
f i r s t q u a r t i l e see QUARTILE. situations. [Sir Ronald Aylmer Fisher]
Fisher-Behrens problem see BEHRENS- Fisher-Hayter m u l t i p l e compari-
FISHER PROBLEM. son test a statistical procedure for as-
sessing whether pairs of means are
Fisher d i s t r i b u t i o n see F DISTRIBU-
significantly different that is used after a
TION. [Sir Ronald Aylmer Fisher (1890-
significant F TEST. The Fisher-Hayter
1962), British statistician and geneticist]
multiple comparison test is a modifica-
Fisher exact test a statistical procedure tion of the FISHER LEAST SIGNIFICANT
to determine whether two CATEGORICAL DIFFERENCE TEST that controls for FAM-
variables are related. Appropriate for ILY-WISE ERROR RATE; it is relatively easy
small samples (a CHI-SQUARE TEST FOR to calculate. See POST HOC COMPARISON.
INDEPENDENCE is used with large sam- [Sir Ronald Aylmer Fisher; Anthony J.
ples), it examines the associations be- Hayter, British-born U.S. statistician]
tween the rows and columns of data in a Fisher least significant difference
fourfold (2 X 2) CONTINGENCY TABLE. For test (Fisher LSD test; LSD test) a sta-
example, the Fisher exact test could be tistical procedure to compare pairs of
used to assess whether there is a relation- means, conducted after an F TEST has re-
ship between gender (male or female) vealed that at least one pair of means is
and pizza topping preference (cheese or significantly different. The test calcu-
multitopping) in a group of 20 individu- lates the smallest value that would be
als. The test yields an exact p value statistically different from chance when
rather than a range of p values (e.g., p < subtracting one mean from another
.05). Also called Fisher-Irwin test; mean. If the absolute value of the actual
Fisher-Yates test. [Sir Ronald Aylmer difference between a pair of means is
Fisher; Joseph Oscar Irwin (1898-1982), larger than this least significant differ-
British statistician; Frank Yates (1902- ence (LSD), a researcher can reject a
1994), British statistician] NULL HYPOTHESIS that the means are
equal and conclude that they are signifi-
Fisher-Freeman-Halton test an ex- cantly different. Also called protected
tension of the FISHER EXACT TEST used to t test. [Sir Ronald Aylmer Fisher]
assess whether two variables, each hav-
ing two or more categories, are related in Fisher scoring method a procedure
a relatively small data sample. For exam- for finding a MAXIMUM LIKELIHOOD so-
ple, the Fisher-Freeman-Halton test lution when estimating PARAMETERS,
could be used to assess whether gender such that the resulting estimates are se-
and pizza topping preference are related lected to make the sample data have the
when the latter has several choices (e.g., highest probability of being drawn from
cheese, vegetable, meat, everything). [Sir a population with the given estimates.
Ronald Aylmer Fisher; G. H. Freeman; [Sir Ronald Aylmer Fisher]
136
fit index
Fisher's F d i s t r i b u t i o n see F DISTRI- SIS is tme (i.e., assuming the means are
BUTION. [Sir Ronald Aylmer Fisher] equal). Thus, the null hypothesis can be
rejected and a researcher can conclude
Fisher's i n f o r m a t i o n a variance or that there is evidence that the two
covariance, often of M A X I M U M LIKELI- means differ more than would occur by
HOOD estimates. When the variances chance. [Sir Ronald Aylmer Fisher]
and covariances are organized together
in a SQUARE MATRIX the resulting dis- Fisher's r to z t r a n s f o r m a t i o n a sta-
play is called Fisher's information ma- tistical procedure that converts a Pear-
trix. [Sir Ronald Aylmer Fisher] son PRODUCT-MOMENT CORRELATION
COEFFICIENT to a Standardized z SCORE
Fisher's least s i g n i f i c a n t d i f f e r - in order to assess whether the correla-
ence see FISHER LEAST SIGNIFICANT DIF- tion is statistically different from zero.
FERENCE TEST.
The test is useful in providing a nor-
Fisher's linear d i s c r i m i n a n t func- mally distributed statistic (called the
t i o n an additive combination of con- Fisher transformed value or Fisher's z)
tinuous variables that helps explain a that can be used in HYPOTHESIS TESTING
CATEGORICAL outcome. For example, an or in forming a CONFIDENCE INTERVAL.
organizational psychologist could ex- Also called Fisher transformation;
amine whether measures of perceived Fisher z transformation; z trans-
career influence, work respect, and insti- formation. [Sir Ronald Aylmer Fisher]
tutional climate distinguish among
three different departments within the Fisher-Yates test see FISHER EXACT
institution. A Fisher's linear discrimi- TEST. [Sir Ronald Aylmer Fisher; Frank
nant function analysis would reveal two Yates (1902-1994), British statistician]
linear discriminant functions in which f i s h i n g n. an informal term for search-
the discriminant coefficients associated ing unsystematically through a data set
with each predictor variable for each in order to look for any noticeable pat-
function indicate how strongly that terns of relationship among variables.
variable relates to the set of depart- More formal procedures called DATA
ments. See DISCRIMINANT FUNCTION. MINING have been developed using sta-
[Sir Ronald Aylmer Fisher] tistical and computer-science methods
to discern recognizable patterns in huge
Fisher-Snedecor d i s t r i b u t i o n see F
data sets. Fishing should not substitute
DISTRIBUTION. [Sir Ronald Aylmer Fisher;
for more rigorous methods based on
George W. Snedecor (1881-1974), U.S.
sound EXPERIMENTAL DESIGN. See also
statistician]
DATA DREDGING; DATA SNOOPING.
Fisher's randomization procedure
f i t n. the degree to which values pre-
a statistical test that compares the actual
dicted by a model correspond with em-
difference between sample means for
pirically observed values. For example,
two distinct groups, with a set of mean
in STRUCTURAL EQUATION MODELING, a
differences obtained by randomly split-
researcher may want to see how well
ting the combined data from the two
groups into all possible rearrangements his or her hypothesized model of the
of two arbitrary groups. The proportion relationships ambng a set of variables ac-
of mean differences from the various tuaUy fits the VARIATION and COVARI-
sets of randomly split data is viewed as a ATION in the data.
p VALUE, such that if the proportion is f i t index a quantitative measure of how
less than, say, .05, it indicates that there well a statistical model corresponds to
is a less than 5% chance of finding the VARIATION and COVARIATION in a
means as different or more so than the set of data. For example, in STRUCTURAL
two sample means if the NULL HYPOTHE- EQUATION MODELING, a well-recognized
137
fitted distribution
fit index is the Bentier comparative fit in LINEAR REGRESSION the fitted value is
index (CFI), which indicates how well a y', where y' = a + bx, in which a equals
proposed model fits the data compared the Y-INTERCEPT, b equals the unstan-
to a null model that just posits variances dardized REGRESSION COEFFICIENT, and
for the variables and no relationships A; is a predictor variable. Note that sub-
among the variables. An index value tracting a fitted value from an actual y
greater than .90 (or preferably .95) indi- score in linear regression will equal a FIT-
cates a model that explains the pattern TED RESIDUAL, which will be close to
of relationships in the data reasonably zero if the linear regression model is a
well. If the fit of a model is poor, then good match to the data.
the model needs to be respecified and
five-number s u m m a r y in EXPLOR-
then reanalyzed. See GOODNESS OF FIT.
ATORY DATA ANALYSIS, the Characteriza-
See also BENTLER-BONETT INDEX.
tion of a data set through the use of five
f i t t e d d i s t r i b u t i o n a known statistical summary statistics: the two extreme
pattern of data (e.g., a NORMAL DISTRI- scores, the upper and lower QUARTILES,
BUTION) that is compared to sample and the MEDIAN or middle quartile. For
data. To assess whether afitteddistribu- example, suppose the scores on a test are
tion is a good match for the data, one (a) 100, 93, 90, 82, 76, 72, 64, 61, 60, and
plots the sample data with a HISTOGRAM 47. The five-number summary from
or BAR GRAPH; (b) compares the sample highest to lowest numbers would be
graph with the known statistical pattern 100, 90, 74, 61, and 47that is, 100 is
(e.g., using a program designed for that the highest extreme score, 90 is the
purpose); and (c) assesses how closely highest quartile (the midpoint in the top
the sample data and statistical pattern half of the data), 74 is the middle point,
match using a GOODNESS-OF-FIT statistic 61 is the lowest quartile (the midpoint in
(e.g., CHI-SQUARE). If the fitted distribu- the bottom half of the data), and 47 is
tion describes the sample data fairly the lowest extreme score. These values
closely, the goodness-of-fit statistic provide the basic highlights of a set of
should be small and nonsignificant, in- data and can be used to create a BOX-
dicating there is little difference between AND-WHISKER PLOT.
the statistical pattern (e.g., normal dis- fixed-alternative question a test or
tribution) and the sample data pattern survey item in which several possible re-
(depicted in the initial histogram). sponses are given and participants
f i t t e d frequency see EXPECTED FRE- are asked to pick the correct response or
QUENCY. the one that most closely matches their
preference. An example of a fixed-
f i t t e d residual the value obtained alternative question is "Which of the
when subtracting an expected or pre- following most closely corresponds to
dicted value from the original data your age: 12 or younger, 13 to 19, 20 to
value. For example, in LINEAR REGRES- 39, 40 to 59, 60 to 79, or 80 or older?" A
SION the fitted residual equals y - y', fixed-alternative question is sometimes
where y' is the predicted value formed referred to as a closed question, al-
from filling actual x values into the re- though this can also refer to any inquiry
gression equation; and y' = a + bx, where requesting a short definite answer (e.g.,
a equals the Y-INTERCEPT, b equals the "How old are you?"). Also called fixed-
unstandardized REGRESSION COEFFI- choice question; forced-choice ques-
CIENT, and jit is a predictor variable. tion; multiple-choice question. Com-
pare FREE-RESPONSE QUESTION.
f i t t e d value the predicted value of a
particular score where a statistical model f i x e d effect an INDEPENDENT VARIABLE
is compared to actual data. For example, whose levels are specified by the re-
us
flowchart
searcher rather than randomly chosen for study rather than randomly chosen
within some level of permissible values. from a wide range of possible values. For
For example, a health researcher who' example, a researcher may wish to inves-
specifically chose to examine the effect tigate the effects of the available dosages
on weight loss of no exercise, one hour of a certain drug on symptom allevia-
of exercise, or three hours of exercise per tion. Fixed-effects models generally are
week would be treating time spent exer- intended to make inferences solely
cising as a fixed effect. In other words, all about the specific levels of the inde-
levels of interest are included in the de- pendent variables actually used i n the
sign and thus anyone wanting to repli- experiment. Compare MIXED-EFFECTS
cate the study would have to use the MODEL; RANDOM-EFFECTS MODEL.
same levels of exercise as in the original.
f i x e d f a c t o r see FIXED EFFECT.
As a rule, one should not generalize re-
sults of fixed-effect studies beyond the f i x e d p a r a m e t e r a specific value as-
specific levels or conditions used in the signed (as opposed to estimated) by a re-
experiment (see FIXED-EFFECTS F A L L A C Y ) . searcher when testing a statistical
Also called f i x e d factor. See FIXED- model. For example, in STRUCTURAL
EFFECTS MODEL; RANDOM-EFFECTS MODEL. E Q U A T I O N M O D E L I N G , researchers may
Compare R A N D O M EFFECT. use a fixed parameter of 1.0 for one of
the F A C T O R L O A D I N G S or variances of
fixed-effects analysis o f variance a
each LATENT VARIABLE in a model. Com-
statistical procedure to determine
pare FREE PARAMETER.
whether means for an outcome or D E -
PENDENT VARIABLE differ across a spe- f i x e d p o p u l a t i o n a specific range of
cific set of conditions (i.e., levels of the individuals or entities that constitute
INDEPENDENT VARIABLE) that have been the larger base of interest to a researcher.
selected by a researcher. This is con- See POPULATION.
trasted with a RANDOM-EFFECTS ANALY-
SIS OF VARIANCE, i n whlch the f i x e d v a r i a b l e a variable whose value is
conditions have been randomly selected specified by a researcher or otherwise
from a wide range of possible choices be- predetermined and not the result of
fore any mean differences are examined. chance. Compare R A N D O M VARIABLE.
139
fluctuation
Stepi
Jl
flowchart
Flowcharts also are used in PATH ANAL- test; Hartley f^^x test for homoge-
YSIS and STRUCTURAL EQUATION MOD- neity of variance; Hartley test;
ELING to illustrate the causal links Hartley test for homogeneity of
among the variables being studied. variance.
f l u c t u a t i o n n. 1. variation in size or focused comparison in research, the
value. 2. in TIME-SERIES ANALYSES, an ir- practice of putting the same set of care-
regularity that is not related to any trend fully chosen, strictly relevant questions
and is thus often removed via SMOOTH- to each participant in an experiment (or
ING. to the same participant on multiple oc-
Fmax d i s t r i b u t i o n in an ANALYSIS OF casions in a WITHIN-SUBJECTS DESIGN).
VARIANCE, a pattern of values used for This allows for systematic comparison of
assessing whether variances are homo- the answers given and produces more
geneous across the groups being studied. generalizable and replicable findings
See F,^x STATISTIC. than does asking questions that vary
from case to case.
Fmax statistic in an ANALYSIS OF VARI-
ANCE, a value formed from the ratio of focus group a small set of people, typi-
the largest variance over the smallest cally 8 to 12 in number, who share com-
variance from the set of groups being as- mon characteristics (e.g., working
sessed. The statistic is assessed for parents with 5- to 8-year-old children)
significance by looking up the CRITICAL that are relevant to the research ques-
VALUE in a statistical table for the rele- tion and who are selected to discuss a
vant number of groups and DEGREES OF topic of which they have personal expe-
FREEDOM. If the calculated F^^,^ statistic rience (e.g., their children's reading abil-
is larger than the critical ratio, then ities and school performance). A leader
there is evidence of HETEROGENEITY OF conducts the discussion and keeps it on
VARIANCE, which Violates one of the as- target while also encouraging free-
sumptions of the analysis and renders its flowing, open-ended debate. Originally
results invaUd. Otherwise, a researcher used in marketing to determine con-
can assume that there is not enough evi- sumer response to particular products,
dence for heterogeneity and the assump- focus groups are now used for determin-
tion of HOMOGENEITY OF VARIANCE Can ing typical reactions, adaptations, and
be reasonably maintained. solutions to any number of issues,
events, or topics and are associated par-
Fmax test the procedure used to obtain ticulariy with QUALITATIVE RESEARCH.
an F^^Ax STATISTIC and evaluate it for sig-
nificance. Also called Hartley F^^^ foldover design a type of FRACTIONAL
140
foreperiod
Participant 1
Participant 2
Participant 3
20 40
Time (months)
follow-up plot
141
forest plot
Author A, 1999
Author B, 2001
Author G, 2004
Author H, 2007
Odds Ratio
forest plot
142
fractional polynomial
143
frailty model
dratic, cubic, or further power when the study at any time without penalty. See
relationship is believed to be NONLIN- also FREEDOM FROM HARM; PARTICI-
EAR. For example, if a researcher expects PANTS' RIGHTS.
a relationship between achievement
motivation and performance that re- Freeman-Tukey test a GOODNESS-OF-
sembles an upside-down U (CURVI- FiT TEST used with data that have multi-
LINEAR) shape, a fractional polynomial
ple categories of counts or frequencies.
could be formed by squaring the [M. F. Freeman, U.S. statistician; John
achievement motivation score for each Wilder Tukey (1915-2000), U.S. statisti-
individual in order to predict achieve- cian]
ment more accurately as a nonlinear free parameter a value that is esti-
pattern. mated from data, usually in a modeling
procedure such as STRUCTURAL EQUA-
f r a i l t y model a form of SURVIVAL
TION MODELING. A free parameter con-
ANALYSIS that allows for HETEROGENE-
trasts with a FIXED PARAMETER, whlch is
ITY in the sample. It incorporates a RAN-
kept at a known, specific value. For ex-
DOM EFFECT to account for COVARIATES
ample, in CONFIRMATORY FACTOR ANAL-
across different subsets of participants in
YSIS, one FACTOR LOADING for each
the sample.
LATENT VARIABLE Is usually a fixed pa-
F ratio (symbol: F) in an ANALYSIS OF rameter of 1.0 and the other loadings are
VARIANCE or a MULTIVARIATE ANALYSIS free parameters that are estimated.
OF VARIANCE, the amount of EXPLAINED
free-response question a test or sur-
VARIANCE divided by the amount of
vey item that allows the respondent to
ERROR VARIANCE; that is, the ratio of be-
respond entirely as he or she pleases, as
tween-groups variance to within-group
opposed to a FIXED-ALTERNATIVE QUES-
variance. Its value determines whether
TION, in which the respondent must
or not to accept the NULL HYPOTHESIS
choose from several provided responses.
stating that there is no difference be-
tween the treatment and control condi- frequency (symbol: f ) n. the number
tions, with a large value indicating the of occurrences of a particular phenome-
presence of a significant effect. Ideally, a non, particularly a CATEGORICAL VARI-
researcher prefers to have rather small ABLE such as gender. For example, it is
variation within each group and maxi- often of interest to find the frequencies
mal variation between the groups in or counts of the men and women who
order to demonstrate significant group are participating in a research study.
differences. Also called F statistic; F Also called absolute frequency. See
value. also RELATIVE FREQUENCY.
freedom f r o m h a r m one of the basic frequency data information that rep-
rights of research participants that is en- resents the counts or number of occur-
sured by an INSTITUTIONAL REVIEW rences of particular response classes,
BOARD. Freedom from harm states that a usually for a CATEGORICAL VARIABLE.
research participant should not incur For example, a political psychologist
undue risk as a result of taking part in a may want to record frequency data on
study. See also FREEDOM TO WITHDRAW; the number of men and women, the
PARTICIPANTS' RIGHTS. number of different ethnic groups, and
the number of individuals who endorse
freedom to w i t h d r a w one of the a Democratic, Republican, or other po-
basic rights of research participants that litical party affiliation.
is ensured by an INSTITUTIONAL REVIEW
BOARD. Freedom to withdraw allows a frequency density the number of oc-
research participant to drop out of a currences of an event divided by the
144
frequentist inference
o o
o o o o o o o o o o o o o o o
o o o o o o
o oo_oo oo oooooooo_o o_
ooooo o
o o
o o o
ooo oo oo oo o
t - Cv[ CO lO CD h-" co" Oi O T-" CN CO Tt If) (D r^" (30" 05 O
Debt (dollars)
frequency polygon
145
Friedman two-way analysis of variance
siders any research study to be one of a F-to-enter for any of the remaining vari-
very large possible number of replica- ables was at least 4.0 and if so would add
tions. Also called classical inference. that variable into the analysis as well.
Compare BAYESIAN INFERENCE. When the researchers found an F-to-
enter for a variable less than 4.0, they
Friedman two-way analysis of would stop the process and examine
variance a NONPARAMETRIC TEST of only those variables that had been
whether two categorical INDEPENDENT added to that point. Compare F-TO-
VARIABLES have a consistent relation-
REMOVE.
ship with a DEPENDENT VARIABLE that
involves ORDINAL DATA. For example, a F-to-remove n. in model-building pro-
researcher might use the Friedman two- cedures such as BACKWARD ELIMINATION,
way analysis of variance to determine STEPWISE REGRESSION, Or backward STEP-
whether three different tests produce WISE DISCRIMINANT ANALYSIS, the Spe-
consistent outcomes when they are used cific ratio of variances needed to justify
to rank the performance of students un- keeping a variable as a predictor. Gen-
dergoing three different training pro- erally, an F-to-remove value around 4.0
grams. Also called Friedman rank is sufficient to allow a variable to be re-
test; Friedman test. [Milton Fried- tained in an analysis. For example, if re-
man] searchers want to assess which of a set of
six variables are the most important in
F statistic see F RATIO. predicting an outcome, they could con-
F test any of a class of statistical proce- duct a backward stepwise multiple re-
gression. All six variables initially would
dures, such as ANALYSIS OF VARIANCE Or
be included in the regression analysis,
MULTIPLE REGRESSION, that rely on the
providing an index of the amount of
assumption that the calculated statis-
variance explained by the set of six vari-
ticthe F RATIOfollows the F DISTRI-
ables. Then, the variable that had the
BUTION when the null hypothesis is
smallest F-to-remove value (i.e., less
true. F tests are tests of hypotheses about
than 4.0) would be removed from the
population variances or of whether RE-
analysis. The researchers would con-
GRESSION COEFFICIENTS are zero. Also
tinue by evaluating each of the remain-
called Fisher's F test; variance ratio
ing five variables and dropping any
test.
that had an F-to-remove less than 4.0,
F-to-enter n. in model-building proce- such that only the retained variables ul-
dures such as FORWARD SELECTION and timately are examined in the analysis.
forward STEPWISE DISCRIMINANT ANALY- Compare F-TO-ENTER.
SIS, the specific ratio of variances needed
f u g i t i v e literature see GRAY LITERA-
to justify adding a variable as a predic-
tor. Generally, an F-to-enter value TURE.
around 4.0 is sufficient to allow a vari- f u l l model a statistical representation
able to be included in an analysis. For that includes all of the variables of inter-
example, if researchers want to assess est, or all of the main PARAMETERS (e.g.,
which of a set of six variables are the REGRESSION COEFFICIENTS), among a set
most important in predicfing an out- of variables. For example, in STRUC-
come, they could conduct a STEPWISE TURAL EQUATION MODELING, a full
REGRESSION. The regression analysis model might include estimated regres-
would begin with no variables. Then, sion parameters between a set of INDE-
the variable with the largest F-to-enter PENDENT VARIABLES, MEDIATORS, and
that met the initial criterion (e.g., 4.0) DEPENDENT VARIABLES. Where there are
would be added as a predictor. The re- as many parameters estimated in a full
searchers would then assess whether the model as there are DEGREES OF FREEDOM.
funnel plot
it is said to be a SATURATED MODEL. Also ical equation that describes how values
called unrestricted model. of one variable are dependent on an-
other variable. The formula for a straight
f u l l - n o r m a l plot see NORMAL PROBA-
line, y = mx + b, describes a functional re-
BILITY PLOT.
lationship: The value y can be deter-
f u l l r a n k the quality of a MATRIX in mined by knowing the SLOPE, m, of
which each of the rows (or columns) are a line multiplied by a value, x, plus
independent of each other, such that the point where the line crosses the Y-
they do not provide redundant informa- AXIS, b.
tion. For example, a CORRELATION MA- f u n c t i o n a l variable a variable with
TRIX of three variables would be of fuU quantitative values that form a curve
rank if each of the three variables was when plotted on a graph.
unique, such that no variable com-
pletely overlapped with another vari- f u n c t i o n coefficient a quantitative
able. If, however, the three variables in value that multiplies a variable and that
the matrix were verbal intelligence, non- can change depending on other vari-
verbal intelligence, and full scale intelli- ables or COVARIATES. A function coeffi-
gence (i.e., some combination of verbal cient differs from other coefficients (e.g.,
and nonverbal intelligence), the matrix a REGRESSION COEFFICIENT) in that
would not be of full rank. it can vary whereas the others are con-
stant over all entities or participants. A
f u l l y recursive model a statistical function coefficient is often used in sta-
representation in which one variable tistical methods such as REGRESSION
predicts a second variable and the sec- ANALYSIS and TIME-SERIES ANALYSIS,
ond variable predicts the first variable. particularly when the data change over
In other words, all the variables in a fully time or space.
recursive model are reciprocally linked.
f u n d a m e n t a l science see BASIC SCI-
f u n c t i o n n. (symbol: f) a mathematical ENCE.
procedure that relates or transforms one
number, quantity, or entity to another f u n n e l plot a graph of EFFECT SIZES
according to a defined rule. For example, along the horizontal X-AXIS and some
if y = 2;t -I-1, y is said to be a function of x. measure of sample size along the vertical
This is often written y = f(x). Here y is the Y-AXIS. It is used to check for the possi-
dependent variable and x is the inde- bility of bias in META-ANALYSIS studies.
pendent variable. A funnel plot that has a pyramidal, up-
side-down cone shapeas in the follow-
f u n c t i o n a l data analysis (FDA) an ing exampleis thought to indicate low
area of statistics in which mathemati- levels of bias.
cal FUNCTIONS are used to study how The funnel shape reflects the fact that
things change across time. It usually in- most of the studies found approximately
volves the use of derivatives to form
curves that model the pattern of change
in some phenomenon (e.g., health,
achievement).
f u n c t i o n a l p r i n c i p a l components
analysis the reducfion of data to a
smaller set of dimensions or components,
where the data are FUNCTIONS or curves
rather than simple variables. See PRINCI-
PAL COMPONENTS ANALYSIS. -0.4 -0.2 0 0.2 0.4 0.6 0.8
147
funnel sequence
medium-sized effects and used some- and gradually narrows the focus to more
what large sample sizes, with a few small specific items.
or large effects found with smaller sam-
FU plot abbreviation for FOLLOW-UP
ple sizes. Consider, however, a graph of
PLOT.
effect sizes and sample sizes that depicts
a growing, slanted mountain shape in- f u z z y set theory an approach to SET
creasing from lower left to upper right. THEORY that allows gradations of mem-
This positive slanted pattern would indi- bership in a set, instead of only assign-
cate a direct and positive relationship ing a 0 or a 1 that would indicate
(i.e., bias) between effect size and sample nonmembership or absolute member-
sizes. Conversely, if the graph revealed a ship, respectively. Some phenomena are
left-slanted mountain shape, from upper more readily modeled with fuzzy set the-
left to lower right, this would indicate a ory than conventional binary set theory,
negative relationship (and bias) between particularly when there is incomplete
effect size and sample size. This ne- information or varying degrees of the
gative, inverse pattern would reveal entity being studied. For example, in
that smaller effects could only be no- designation of a mental illness fuzzy set
ticed with large sample sizes and larger theory may be employed to allow
effects could be seen with smaller sam- for cases in which individuals could be
ple sizes. classified somewhere between an ab-
sence or a full diagnosis of a particular
f u n n e l sequence a method of stmctur- condition.
ing the order of questions in surveys and
interviews that starts with general items F value see F RATIO
148
Gg
G symbol for LIKELIHOOD-RATIO CHI- then the first player gets the entire prize,
SQUARE. (b) If both players decide to share the
prize, then the prize is split, (c) If both
test see LIKELIHOOD-RATIO TEST. players decide to keep the entire prize,
gain score see DIFFERENCE SCORE. then both leave empty-handed. The
players must therefore base their deci-
G a l b r a i t h plot see RADIAL PLOT. [Rex sions on what they think the other will
F. Galbraith, British statistician] choose to do.
G A M abbreviation for GENERALIZED AD- G A M M abbreviation for GENERALIZED
DITIVE MODEL. ADDITIVE MIXED MODEL.
gambler's f a l l a c y a failure to recog- g a m m a (symbol: y) n. any of several dif-
nize the independence of chance events, ferent statistical indices, most com-
leading to the mistaken belief that one monly GOODMAN-KRUSKAL'S GAMMA.
can predict the outcome of a chance
event on the basis of the outcomes of g a m m a d i s t r i b u t i o n a pattern of data
past chance events. For example, a per- that can be depicted on vertical and hor-
son might think that the more often a izontal axes where information is
tossed coin comes up heads, the more known about two parameters, scale and
likely it is to come up tails in subsequent shape, both of which are values equal to
tosses, although each coin toss is inde- zero or higher. For example, a standard
pendent of the others, and the true prob- gamma distribution is one in which the
ability of the outcome of any toss is still shape and location parameters of the
just .5. continuous probability function are 0
and 1, respectively. Gamma distribu-
game theory a branch of mathematics tions are common in Bayesian statistics
concerned with the analysis of the be- and often used to describe waiting times
havior of decision makers (called players) and reaction times.
whose choices affect one another. Game
garbage i n , garbage out (GIGO) an
theory is often used in both theoretical
expression indicating that if the data
modeling and empirical studies of con-
used in analyses are not reliable or co-
flict, cooperation, and competition, and
herent, the results will not prove useful.
it has helped to stmcture interactive de-
The phrase was first used in computer
cision-making situations in numerous
science, cautioning users against tmst-
disciplines, including economics, politi-
ing computer output derived from unre-
cal science, social psychology, and eth-
liable input. Although the term is not
ics. A simple game theory example could
used much currently, it remains a rele-
be a scenario from a reality television
vant reminder to researchers to pay at-
show in which the final two players are
tention to how studies are designed and
each asked privately to make a decision
how data are collected, and not to de-
to share or keep the entire game prize.
pend on analyses to produce meaningful
The players are told that the outcome
results where the input data are not
will be decided by the following rules:
themselves meaningful.
(a) If one player decides to keep the en-
tire prize and the other decides to share, G A R C H abbreviation for GENERALIZED
149
Gaussian distribution
150
generalized inverse
151
generalized least squares regression
152
Gini coefficient
niques employed in the behavioral sci- MEAN except that the latter involves the
ences can be subsumed under the sum of the quantities divided by the
general linear model. number of quantities, whereas the for-
mer is the product of the quantities with
general n o r m a standard that is widely
the product raised to the power of 1 di-
accepted, such as a regularly occurring vided by the number of quantities. Thus,
level of performance on an intelligence for the numbers 1, 2, and 3 the arithme-
test or an expected pattern of alcohol tic mean is 6/3 = 2, and the geometric
use on college campuses. mean is the cube root of (1)(2)(3) = 6^'^
general systems theory an interdisci- 1.817. See also HARMONIC MEAN.
plinary conceptualframeworkfocusing
geometric progression a series of
on wholeness, pattern, relationship, hi-
numbers in which each successive num-
erarchical order, integration, and orga-
ber is the product of the previous num-
nization. It was designed to move
ber times a constant. For example, the
beyond the reductionistic and mecha-
geometric progression 5, 10, 20, and 40
nistic tradition in science (see REDUC-
would have 80 as the next number and
TIONISM) and integrate the fragmented
the constant, called the common ratio,
approaches and different classes of phe-
is 2 (i.e., 2 X 5 = 10; 2 X 10 = 20; 2 X 20 =
nomena studied by contemporary sci-
40; 2 X 40 = 80; etc.). Also called geo-
ence into an organized whole. In this
metric sequence; geometric series.
framework, an entity or phenomenon
should be viewed holistically as part of a geostatistics n. a set of methods for an-
set of elements interacting with one an- alyzing data related to the physical envi-
other; the goal of general systems theory ronment. In psychology, for example,
is to identify and understand the princi- geostatistics could be used to under-
ples applicable to all entities in the set. stand which factors in the environment
The impact of each element depends on can bring about or exacerbate behavioral
the role played by other elements in- conditions, such as substance abuse, or
volved and order arises from interaction spread diseases, such as AIDS.
among these elements. Also called sys-
tems theory. Gibbs s a m p l i n g a statistical method of
randomly generating sets of data that in-
generating f u n c t i o n a mathematical corporate the patterns of values or PROB-
method of finding the number of occur- ABILITY DISTRIBUTIONS from two or
rences in a large sequence. It is formed more variables. Also called Gibbs sam-
by a power series where the k^^ variable is pler, [losiah W. Gibbs (1839-1903),
raised to the k^^ power. For example, a U.S. mathematician and physicist]
generating function may take the form
GIGO abbreviation for GARBAGE IN, GAR-
UQX^ + AjX' -I- a2X^ + a^X^ where a is
BAGE OUT.
a weight and the generating function is a
weighted sum. See also MOMENT GENER- G i n i coefficient a measure of variance
ATING FUNCTION; PROBABILITY GENER- in a FREQUENCY DISTRIBUTION, provid-
ATING FUNCTION. ing a numerical summary of a LORENZ
CURVE. It ranges in value from 0, indi-
geometric d i s t r i b u t i o n the PROBA- cating no variance such that all entities
BILITY DISTRIBUTION of the number of
are equal, to 1, indicating maximum
failed trials before the first success in a variance such that every entity is differ-
series of BERNOULLI TRIALS. ent. Also called coefficient of con-
geometric mean a measure of CEN- centration; G i n i concentration;
TRAL TENDENCY Calculated for k num- Gini index; G i n i ratio; Gini statis-
bers Xi...Xi. as (x-i xx2'x X2...xi^f'^. A geo- tic. [Corrado Gini (1884-1965), Italian
metric mean is similar to an arithmetic statistician and sociologist]
153
g inverse
g inverse abbreviation for GENERALIZED and ultimately very low sales as the
INVERSE. fad dies out. Gompertz curves often are
Glass's d an EFFECT SIZE measure that used in SURVIVAL ANALYSIS. Also called
represents the standardized difference Gompertz distribution; Gompertz
between means (i.e., the difference in function. [Benjamin Gompertz (1779-
average values for two samples divided 1865), British mathematician]
by the STANDARD DEVIATION of the sec- Goodman-Kruskal's g a m m a a mea-
ond sample). It is often used in META- sure of association between two CATE-
ANALYSIS and other research in which it GORICAL VARIABLES, where 0 represents
is important to determine whether an ef- no relationship and 1 represents perfect
fect persists across studies in order to association. For example, a marketing
consolidate a result. Also called Glass's researcher may want to assess the rela-
delta. See also COHEN'S D; HEDGES'S G. tionship between the location of a store
[Gene V. Glass (1940- ), U.S. statisti- and the type of clothes purchased. A
cian] Goodman-Kruskal's gamma of .30 or
G L M 1. abbreviation for GENERAL LINEAR higher would indicate a medium-sized
MODEL. 2. abbreviation for GENERAL- correlation such that clothes distribu-
IZED LINEAR MODEL. tors desiring to maximize sales should
provide specific kinds of clothes at spe-
G L M M abbreviation for GENERALIZED cific locations. Also called Goodman-
LINEAR MIXED MODEL. Kruskal's tau. [Leo A. Goodman
global m a x i m u m the largest possible (1928- ) and William Henry Kruskal
value in a set of numbers. For example, (1919-2005), U.S. statisticians]
the global maximum for a CORRELATION
goodness of f i t the degree to which
COEFFICIENT is 1, whereas the global values predicted by a model agree with
maximum for a STANDARD DEVIATION empirically observed values. For exam-
or VARIANCE is not finite (i.e., infinity). ple, a researcher may wish to assess
global m i n i m u m the smallest possible whether a pattern of frequencies from
value in a set of numbers. For example, a study is the same as theoretically ex-
the global minimum for a CORRELATION pected, whether two CATEGORICAL
COEFFICIENT is -1 and for a STANDARD VARIABLES are independent, or whether
DEVIATION or VARIANCE it is 0. a REGRESSION EQUATION correctly pre-
dicts obtained data. A small, non-
GLS abbreviation for generalized least significant value from a GOODNESS-OF-
squares. See GENERALIZED LEAST SQUARES
FiT TEST indicates a well-fitting model.
REGRESSION.
G M M abbreviation for generalized mixed goodness-of-fit test any of various sta-
model. See GENERALIZED LINEAR MIXED tistical procedures that assess how
MODEL. closely data obtained from a sample
match an expected population pattern
Gompertz curve a function or curve or model. The CHI-SQUARE GOODNESS-
used to describe a pattern of growth that OF-FIT TEST is a commonly used exam-
is slow at the beginning and end but ple.
rather rapid or steep in the middle, thus
forming a shape resembling a leaning good-subject role the behavior dis-
letter S. For example, the adoption of played by a participant in a research
a fashion fad may follow a Gompertz study who is eager to perform as he
curve: Few people purchase the product or she thinks is desired or wanted.
initially but there is a subsequent period The adoption of such a role could
of highly increased sales that precedes bias study results as the true effect of
a lessening of interest in the product the INDEPENDENT VARIABLE cannot
154
graphic rating scale
155
gray literature
1 2 3 4 5
poor below average above excellent
average average
156
grouped frequency distribution
ing conclusions about its nature and ex- ual level. For example, educational re-
istence. searchers may be interested in providing
a group dimension by depicting results
group n. a collection of participants in a
at the classroom level rather than by in-
research study whose responses are to be
dividual students.
compared to the responses of one or
more other collections of research par- group f l i s t r i b u t i o n the scores for a
ticipants. Participants in a particular group of individuals. For example, re-
group all experience the same experi- searchers may be interested in providing
mental conditions or receive the same the group distribution performance
treatment, which differs from the exper- scores for separate classrooms or
imental conditions or treatments partic- schools.
ipants in other groups experience or
receive. group-divisible design a research de-
sign in which separate groups each con-
group-average clustering see AVER- tain a unique set of BLOCKS whose
AGE-LINKAGE CLUSTERING. members are assigned a unique pair of
group class 1. in statistical methods treatments. The group-divisible design
such as ANALYSIS OF VARIANCE, the des- provides a balanced structure for assign-
ignation of the response options for a ing multiple treatments across blocks
CATEGORICAL VARIABLE as the levels of and across two or more groups.
an INDEPENDENT VARIABLE. For exam- grouped data information that is
ple, if a researcher is studying religious grouped into one or more sets in order to
affiliation, he or she may include gender analyze, describe, or compare outcomes
as a variable in the analysis and use its at a combined level rather than at an in-
two distinct options (i.e., female and dividual level. For example, data from a
male) as levels for comparison. 2. the re- FREQUENCY DISTRIBUTION may be ar-
sults derived from such groups. ranged into CLASS INTERVALS. See also
group-comparison design a type of GROUP-COMPARISON DESIGN; GROUP-
research approach that investigates po- DIVISIBLE DESIGN.
tential differences across sets of individ-
grouped frequency d i s t r i b u t i o n a
uals who are often randomly assigned to
a CONTROL CONDITION or to One or description of how often a set of specific
more specific EXPERIMENTAL CONDI- responses, organized into equal-sized
TIONS. Data from a group-comparison
subsets of possible responses, occur in a
design are often analyzed with such sta- sample. It is a type of FREQUENCY DISTRI-
tistical methods as a T TEST or ANALYSIS BUTION that is particularly useful when
OF VARIANCE. Also called group- there are a large number of response
difference design. choices (e.g., 10-20 or more) and re-
searchers wish to present the informa-
group design any of a variety of proce- tion more concisely. For example,
dures intended to compare two or more suppose that a researcher wants to sum-
collections of individuals or entities. marize the individuals in a sample in
which age ranges from 18 to 85 years. In-
group difference any observed varia- stead of listing all 68 different ages, the
tion between groups of participants in researcher could combine the data into
an experiment when considering each 5-year intervals, which would produce
group as a single entity. See GROUP- 14 subsets of ages. More specifically, the
COMPARISON DESIGN. researcher would create two columns,
group dimension the perspective listing in the left-hand one the 14 age
gained by viewing results at a larger, subsets (e.g., 16-20, 21-25, 26-30, ...
combined level instead of at an individ- 81-85). In the right column, the re-
is?
group effect
searcher would list the frequencies of being used to evaluate the data. It is sim-
each age subset; that is, how often each ilar to a RANDOMIZED CLINICAL TRIAL
age subset was endorsed by the individu- but distinct in assigning intact clusters
als in the sample, thus reducing a large of participants to the specific conditions
number of scores into smaller, more rather than individual participants. Also
manageable groups. called cluster-randomized trial.
group effect a research finding specific group sequential design a research
to the group of individuals to which a design in which a treatment is randomly
participant belongs. A group effect could assigned to different groups at varying
appear in an assigned subset, such as a times, such that every group receives the
treatment or intervention, or in a natu- treatment eventually. A group sequen-
rally occurring subset, such as age level tial design is useful when it is important
or classroom. For example, a researcher for each participant to have an opportu-
might be interested in a group effect of a nity to benefit from the treatment but a
specific reading intervention, or in a comparison group nonetheless is re-
group effect of book reading for students quired: The portion of the sample that
in the current decade who may be read- receives the treatment later acts as a
ing less than previous sets of students CONTROL GROUP initially.
owing to more common use of the com-
puter and television. g r o w t h curve a graphic representation
of progress over time on a specific vari-
group experiment an experiment in able within a specific group (e.g., the de-
which subsets of individuals serve as the velopment of an organism, as in the
unit of analysis. For example, a re- hypothetical example opposite).
searcher could compare the perfor- Data in a growth chart often are evalu-
mance of different classrooms rather ated with TIME-SERIES ANALYSIS, LATENT
than of specific individuals. GROWTH CURVE ANALYSIS, or Other lon-
group i n t e r v i e w an interview in gitudinal methods.
which one or more questioners elicit in- g r o w t h curve analysis a statistical
formation from two or more respon- procedure for assessing the level and rate
dents in an experimental or real-life (SLOPE) of change over time on a specific
situation. The participation and interac- variable in a sample of individuals. See
tion of a number of people, particularly also LATENT GROWTH CURVE ANALYSIS.
if they are acquainted with each other as
members of a club or similar group, is g r o w t h f u n c t i o n the relationship be-
believed to yield more informative re- tween a DEPENDENT VARIABLE and Sev-
sponses than are typically obtained by eral levels of an INDEPENDENT VARIABLE
interviewing individuals separately. defined in units of time (e.g., days,
weeks, months, years).
group m a t c h i n g see MATCHING.
G statistic 1. a value used to assess the
group-randomized t r i a l a research presence of spatial clustering or
design in which groups of individuals AUTOCORRELATION in a data set. A large
are randomly assigned to TREATMENT
G statistic with a small accompanying p
conditions or CONTROL CONDITIONS. In
VALUE (e.g., < .05) suggests there are sub-
educational or organizational settings,
sets of dependence or clustering. 2. see
for example, a group-randomized trial
G^
could be used to study classrooms or de-
partments, with HIERARCHICAL LINEAR G study generalizability study: research
MODELING and similar analyses that about the widespread applicability of a
take the larger group membership (e.g., particular measure or test. A G study as-
classroom or department) into account sesses several facets or factors (e.g., in-
158
Guttman scale
growth curve
159
Hh
h symbol for COMMONALITY COEFFI- portant, whereas those that are further
CIENT. removed from zero (e.g., two or more
standard deviations) are considered sig-
HQ symbol for NULL HYPOTHESIS. nificant. Alternatively, when plotting
RESIDUALS, points far from zero reveal
H j symbol for ALTERNATIVE HYPOTHESIS.
the existence of OUTLIERS that indicate a
Ha symbol for ALTERNATIVE HYPOTHESIS. poorly fitting regression model. The re-
lated full-normal plot (see NORMAL
h a b i t u a t i o n n. the weakening of a re- PROBABILITY PLOT) shows both the origi-
sponse to a stimulus, or the diminished nal positive and negative variable val-
effectiveness of a stimulus, following re- ues.
peated exposure to the stimulus. Com-
pare DISHABITUATION. halo effect a rating bias in which a gen-
h a l f - n o r m a l plot a graph in which the eral evaluation (usually positive) of a
ABSOLUTE VALUES Obtained for a vari- person, or an evaluation of a person on a
able are given along the horizontal x- specific dimension, influences judg-
axis and their corresponding probabili- ments of that person on other specific
ties of occurrence under a STANDARD- dimensions. For example, a person who
IZED DISTRIBUTION are given along the is generally liked might be judged as
vertical y-axis. For example, a half- more intelligent, competent, and honest
normal plot could be used to assess the than he or she actually is.
importance of observed experimental ef- h a n g i n g rootogram a type of HISTO-
fects: Points that fall close to the mean GRAM in which the horizontal X-AXIS is
of zero are considered minor or unim- slightly elevated and the vertical Y-AXIS
^ 140
>
U
>
C 120
3
O- 100
0)
t. eoi
I 6(
ra
V)
-H 1I 1h- -HI 1-
o o o o o o o
o o o o o o o
o o o o o o o
Debt (dollars)
hanging rootogram
160
hazard analysis
161
hazard function
same for all values of x; that is, the vari- estimate having a value that is impossi-
ance in y is a function of the variable x. ble or very rare (e.g., a negative ERROR
Heterogeneity of variance violates one VARIANCE estimate). Heywood cases
of the basic assumptions of REGRESSION may indicate any of the following: a
ANALYSIS and other stafisfical prodedures. sample that is too small to adequately es-
Also called heteroscedasticity. Com- timate the parameters; data that do not
pare HOMOGENEITY OF VARIANCE. have a NORMAL DISTRIBUTION or that
contain OUTLIERS; a misspecified model
heteromethod b l o c k in a MULTI- that is not appropriate for the data; or a
TRAIT-MULTIMETHOD MATRIX, the block parameter whose true value is so close to
of values representing the correlations a boundary (e.g., 1 or 0) in the popula-
between traits (e.g., depression, self- tion that its estimate exceeded this limit
esteem) assessed with different proce- due to sampling fluctuation.
dures (e.g., paper-and-pencil survey,
clinical interview, physiological mea- h i d d e n M a r k o v model a statistical
sure). If the correlations are similar re- representation of a series of time-
gardless of the type of procedure used to ordered events in which (a) each event is
assess the traits, a researcher may con- related to the immediately prior event
clude he or she is assessing the true value but not to earlier events and (b) the pro-
of the traits rather than obtaining scores cess that is driving the series is not di-
that are merely ARTIFACTS of the mea- rectly observable (i.e., is a LATENT
surement procedure used. See HETERO- VARIABLE). For example, consider a
TRAIT-HETEROMETHOD COEFFICIENT. study of the number of cigarettes a day
heteroscedasticity n. see HETEROGENE- that are smoked by individuals in a
ITY OF VARIANCE. hcteroscedastic adj.smoking cessation program. If a hidden
Markov model were used to represent
heterotrait-heteromethod coeffi- the process, this would indicate the pres-
cient in a MULTITRAIT-MULTIMETHOD ence of unobservable states or events
MODEL, a number that indicates the (e.g., peer pressure to smoke; stress or en-
strength of the association between dif- vironmental cues to smoke, as in the so-
ferent characteristics or variables mea- cial context of a bar) that are linked to
sured with different procedures. See also the number of cigarettes smoked, and
HETEROMETHOD BLOCK. that the best prediction of the number of
cigarettes that will be smoked the fol-
heterotrait-monomethod coeffi- lowing day is the number smoked on the
cient in a MULTITRAIT-MULTIMETHOD current day rather than the number
MODEL, a number that indicates the smoked on any prior days. See MARKOV
strength of the association between dif-
CHAIN.
ferent characteristics or variables mea-
sured with the same procedures. See also h i d d e n variable an undiscovered
HETEROMETHOD BLOCK. causative variable. When a relationship
heuristic n. in the social sciences, a con- is found between variables x and y, vari-
ceptual device, such as a model or work- able X may erroneously be thought to be
ing hypothesis, that is intended to the cause of y. However, the cause of y
explore or limit the possibilities of a may be a hidden variable z that is corre-
question rather than to provide an ex- lated with variable x. Also called lurk-
planation of the facts. See also CON- ing variable; t h i r d variable. See
STRUCT. THIRD-VARIABLE PROBLEM.
163
hierarchical clustering
164
hinge
165
histogram
20.000
18,000
16.000
14 000
12,000
10:000 mmi
8.000
6,000
A 000
rl".
2. 000
o o
o o o o o o o o
o ooo_o
oo_o
oo_o
o oo
o oo
o oo o
o o o o o
o_ oo_oo_oo_o o_
o o_
o o_ o o. o
o oo oo_oo_ o_ o_
CN co" Tj-" irT (D CO Ci O" CM" CO "C" in to" f^" co" O) o "
Debt (dollars)
histogram
hinge) and the upper 25% of cases (the searcher could examine the health re-
upper hinge) from the remainder of the cords of, say, smokers and nonsmokers
cases. to follow the path of an illness from be-
fore its manifestation through to its di-
histogram n. a graphical depiction of agnosis and treatment. It is distinct from
continuous data using bars of varying PROSPECTIVE RESEARCH generally, which
height, similar to a BAR GRAPH but with begins with individuals who are appar-
blocks on the A;-axis adjoining one an- ently healthy in the present and moves
other so as to denote their continuous forward to investigate whether a specific
nature. disorder will occur over time.
Consider the example above, showing
the average credit card debt of indivi- h i s t o r i g r a m n. a graph of quantitative
duals. Bars along the A-axis represent TIME-SERIES data, with the value of
amount of debt and are connected to the variable being examined given on
one another, while the heights of the the vertical Y-AXIS and time given on
bars represent the number or frequency the horizontal X-AXIS. For example, a
of individuals with each debt amount. historigram for a weight reduction pro-
gram might depict daily weight along
historical analysis research that ex- the y-axis over a six-month time period
amines past events to understand cur- along the x-axis (see overleaf).
rent or future events. For example,
If the weight reduction program was
researchers could perform a historical
effective then the historigram would be
analysis of an individual's or a family's
expected to slope downward from a high
substance use experiences to understand
point at the upper left.
the present substance use behavior of
that person or group. history effect the influence of events
or circumstances outside an experiment
h i s t o r i c a l c o n t r o l group a CONTROL
on an outcome variable of interest.
GROUP whose participants are selected
QUASI-EXPERIMENTAL RESEARCH Often
to be similar to those in the treatment
attempts to take history effects into ac-
group on the basis of data collected
count in order to rule out potential CON-
some time in the past.
FOUNDS to a posited or apparent link
historical prospective study a re- between two variables. An example
search project that examines LONGITU- would be the occurrence of an actual
DINAL DATA obtained in the past to track earthquake during a field study of the ef-
the incidence of a particular disorder fects of training in earthquake prepared-
over time and its association with vari- ness; this would likely increase news
ous risk factors. For example, a re- coverage of earthquakes, thereby result-
166
honestly significant difference test
250
2 3 4
Time (months)
historigram
167
horizontal axis
168
hypothetico-deductive method
0.8
o
0.6
g 0.4
Q
.d
P 0.2
Logx
hyperbolic distribution
169
Ii
I symbol for WILLIAMS'S AGREEMENT this case, an infinite number of combi-
MEASURE. nations of their values may yield the
same model prediction, such that one
ICC 1. abbreviation for INTRACLASS COR- cannot determine from a given set of ob-
RELATION coefficient. 2. abbreviation for
servations which values of the parame-
ITEM CHARACTERISTIC CURVE. ters actually generated the data.
IDA abbreviation for INITIAL DATA ANAL- identity m a t r i x (symbol: I) a SQUARE
YSIS. MATRIX with values of one along its
idealized experiment see THOUGHT main diagonal and zeros elsewhere. The
EXPERIMENT. following is an example.
i d e n t i f i c a t i o n n. in CONFIRMATORY
1 0 0
FACTOR ANALYSIS and STRUCTURAL EQUA-
0 1 0
TION MODELING, a situafion in which the
model contains a sufficient number of 0 0 1
both fixed and free PARAMETERS to result
in unique estimates from the observed Identity matrices are used in FACTOR
data. A model is said to be identified or ANALYSIS solutions and matrix algebra.
identifiable if a unique set of its parame- Also called unit matrix.
ter values can be determined from observa- i d i o g r a p h i c adj. relating to the descrip-
fions. OVERIDENTIFICATION occurs when tion and understanding of an individual
there are more knowns than free param- case, as opposed to the formulation of
eters, and UNDERIDENTIFICATION OCCUrS general laws describing the average case.
when it is not possible to estimate all An idiographic approach involves the
of the model's parameters. See also IDEN- thorough, intensive study of a single
TIFICATION PROBLEM; OVERIDENTIFIED person or case in order to obtain an in-
MODEL; UNDERIDENTIFIED MODEL. depth understanding of that person or
i d e n t i f i c a t i o n p r o b l e m in CONFIR- case, as contrasted with a study of the
MATORY FACTOR ANALYSIS and STRUC- universal aspects of groups of people or
TURAL EQUATION MODELING, the problem cases. In those areas of psychology in
as to whether the PARAMETER estimates which the individual person is the unit
obtained are stable given the number of of analysis (e.g., in personality, develop-
subjects and variables in the observed mental, or cUnical psychology), the
sample data. The problem arises from idiographic approach has appeal be-
the fact that more than one set of param- cause it seeks to characterize a particular
eter values may be associated with the individual, emphasizing that individ-
same PROBABILITY DISTRIBUTION of a ual's characteristic traits and the unique-
statistical model. When the identifica- ness of the individual's behavior and
tion problem arises, a model is said to be adjustment, rather than to produce a
unidentifiable and its parameter values universal set of psychological constructs
cannot be interpreted. For example, as- that might be applicable to a popula-
sume that the two parameters b and c in tion. Compare NOMOTHETIC
the linear regression model y = ax + b + c IID independent and identically distri-
-I- N(0, o^) are fully interchangeable. In buted: denoting a collection of random
170
incidental stimulus
variables that have the same PROBABIL- reotypes, and emotions in social cogni-
ITY DISTRIBUTION but are statisfically un- tion research. Typically, an implicit
related to one another. For example, a measure is assessed as a response out-
sequence of random tosses of a fair coin come of an experimental procedure in
is IID, as is a sequence of colored balls which the participant is engaged in a
randomly drawn from an urn with re- cognitive task. For example, a word-stem
placement after each draw. IID is a key completion task might be employed to
assumption in the CENTRAL LIMIT THEO- assess emotion implicitly, such that jo_
REM and also a standard assumption in could be completed to form a positive
probability and statistical theory. emotional word (e.g., joy) or a neutral
word (e.g., jog).
ill-conditioned adj. denoting a
SQUARE MATRIX in which the ratio of importance s a m p l i n g a MONTE
the largest EIGENVALUE to the smallest CARLO RESEARCH method for computing
eigenvalue is much larger than 1, such expected values of a RANDOM VARIABLE
that small changes to the entries in the in which samples are generated from a
matrix result in a significantly large distribution similar to the distribution
change in its INVERSE MATRIX. An ill- of interest, which cannot be directly
conditioned matrix thus is very sensitive sampled. The method is particularly use-
to inconsequential changes (e.g., round- ful for high-dimensional models with
ing errors in computer calculations), many variables or parameters and is fre-
thereby making the calculations it yields quently used in BAYESIAN statistics.
difficult to interpret. See also COLLIN-
imposed etics the imposition of one's
EARITY.
own cultural perspective and bias when
illusory correlation 1. the appear- studying another culture. See ETICS.
ance of a relationship that in reality does i m p r o p e r p r i o r in BAYESIAN statistics,
not exist. 2. an overestimation of the de- a PRIOR DISTRIBUTION that does not in-
gree of relationship (i.e., correlation) be- tegrate to one. Although an improper
tween two variables. For example, if an prior in itself is not necessarily problem-
unusual action occurred at the same atic, especially when the corresponding
time that an adolescent was present, the POSTERIOR DISTRIBUTION is proper, it
assumption that the action was carried can sometimes lead to an uninterpret-
out by the adolescent would be an illu- able and misleading estimate of the
sory correlation. BAYES FACTOR.
impact analysis a quantitative ana- i m p u t a t i o n n. a procedure for filling in
lytic procedure used to assess the net missing values in a data set before ana-
success or failure of a program, usually lyzing the resultant completed data set.
through controlled experimentation. It There are several methods of imputa-
is appropriate only if the program's ob- tion, including HOT-DECK IMPUTATION,
jectives are specifiable and measurable, in which missing valiies are replaced
the program is well implemented for its with values drawn randomly from a sim-
intended participants, and the outcome ilar sample in the same data set; and re-
measures are reliable and valid. Also gression imputation, in which missing
called impact assessment. See also values are replaced with predicted values
SUMMATIVE EVALUATION. estimated from a statistical model of the
nonmissing values in the data. See also
i m p l i c i t measure a measurement of a
COLD-DECK IMPUTATION; MULTIPLE IM-
psychological constmct that is obtained
PUTATION.
while the individual being assessed is
unaware that the measurement is taking i n c i d e n t a l stimulus an unintentional
place, often used to assess attitudes, ste- or coincidental stimulus that may occur
171
inclusion-exclusion criteria
172
independent variable
PARTIALING out Ot holding Constant the these groups. Also called independent
effects of all other explanatory variables. groups. Compare DEPENDENT SAMPLES.
See also COLLINEARITY.
independent-samples analysis o f
independent diagonals model see variance see BETWEEN-SUBJECTS ANAL-
DIAGONALS MODEL. YSIS OF VARIANCE.
independent events the situation in independent-samples design see BE-
which observing one event does not pro-
TWEEN-SUBJECTS DESIGN.
vide any additional information about
the occurrence or outcome of another independent-samples t test a T TEST
event. For example, the outcome of a used to analyze data from a BETWEEN-
coin flip and the Dow Jones Industrial SUBJECTS DESIGN, in which the different
Average are independent events, whereas groups of individuals or other entities
the temperatures of two consecutive measured are not associated with one
days are not necessarily independent. another. In the independent-samples t
Formally, two events, A and B, are inde- test, one calculates the value of the TEST
pendent if the probability of event A is STATISTIC t from the means, standard
the same as the conditional probability deviations, and sizes of the two groups
of event A given event B, that is, P(A) = of interest, and compares its value to a
P(A\B), or equivalently, P(B) = P(B\A). T DISTRIBUTION for a given DEGREE OF
Compare DEPENDENT EVENTS. FREEDOM under the assumptions of nor-
independent groups see INDEPEN- mality and INDEPENDENCE of Obser-
DENT SAMPLES. vations, among others. Also called inde-
pendent-measures t test; unrelated
independent-groups analysis o f t test. Compare DEPENDENT-SAMPLES T
variance see BETWEEN-SUBJECTS ANAL- TEST.
YSIS OF VARIANCE.
independent s a m p l i n g a process for
independent-groups design see BE-
selecting a sample of study participants
TWEEN-SUBJECTS DESIGN.
from a larger potential group of individ-
independent-measures analysis o f uals such that the probability of each
variance see BETWEEN-SUBJECTS ANAL- person being selected for inclusion is not
YSIS OF VARIANCE. influenced by which people have been
independent-measures design see chosen already. The resulting samples
BETWEEN-SUBJECTS DESIGN. will be INDEPENDENT SAMPLES.
independent r a n d o m s a m p l i n g see independent variable (IV) the vari-
SIMPLE RANDOM SAMPLING. able in an experiment that is specifically
manipulated or is observed to occur be-
independent random variables fore the occurrence of the dependent, or
RANDOM VARIABLES that exhibit a com- outcome, variable, in order to assess its
plete lack of relationship, such that no effect or influence. Independent vari-
information about one variable, x, con- ables may or may not be causally related
veys any information about another to the DEPENDENT VARIABLE. In Statisti-
variable, y. Any events related to these cal analysessuch as PATH ANALYSIS,
variables are INDEPENDENT EVENTS.
REGRESSION ANALYSIS, and STRUCTURAL
independent samples groups of indi- EQUATION MODELINGan independent
viduals or sets of data that are unrelated variable is likely to be referred to as a
to one other. For example, experimental causal variable, explanatory variable,
groups consisting of different and unre- regressor variable, or PREDICTOR VARI-
lated participants are independent sam- ABLE. See also EXOGENOUS VARIABLE;
ples, as are the data sets obtained from TREATMENT.
173
indeterminacy
174
informative prior
175
informed consent
tics, a type of PRIOR DISTRIBUTION based dents, finances, technology, campus cli-
on large amounts of previously accumu- mate, and other characteristics so as to
lated knowledge about the phenome- facilitate better decision making.
non of interest. An informative prior
i n s t i t u t i o n a l review b o a r d (IRB) a
tends to provide more specific informa-
committee named by an agency or insti-
tion and have a stronger influence upon
tution to review research proposals orig-
calculations of an associated POSTERIOR
inating within that agency for ethical
DISTRIBUTION than a NONINFORMATIVE
acceptability and compliance with the
PRIOR.
organization's codes of conduct. IRBs
i n f o r m e d consent a person's volun- help protect research participants and
tary agreement to participate in a proce- are mandatory at any U.S. institution re-
dure on the basis of his or her ceiving federal funds for research.
understanding of its nature, its potential i n s t i t u t i o n a l survey in INSTITU-
benefits and possible risks, and available TIONAL RESEARCH, a questionnaire or in-
alternatives. Informed consent is a fun- terview used to collect information
damental requirement of research with about a university and its operations.
humans and typically involves having
participants sign documents prior to the i n s t r u c t i o n a l variable 1. any INDE-
start of a study describing specifically PENDENT VARIABLE that is studied in
what their involvement would entail researching the outcomes and effective-
and noting that they are free to decline ness of learning programs. Examples of
participation or to withdraw from the re- instmctional variables include the spe-
search at any time. See also INSTITU- cific activities used, instructor quality,
TIONAL REVIEW BOARD. learner workload, and learner motiva-
tion. 2. more generally, any set of direc-
i n i t i a l data analysis (IDA) an exami- tions manipulated in an experiment. For
nation of the quality of a data set that is example, a researcher investigating goal-
undertaken before beginning more com- setting may wish to assess how the per-
plex statistical tests of hypotheses. It formance of participants who are in-
often involves calculating DESCRIPTIVE structed simply to do their best differs
STATISTICS such as the mean and STAN- from that of participants told to try to
DARD DEVIATION, identifying minimum achieve a specified target level.
and maximum values, and graphing the
data points to determine whether their instrument n. any tool, device, or other
arrangement follows a known or ex- means by which researchers assess or
pected distribution. IDA is important for gather data about study participants. Ex-
ensuring that data conform to assump- amples include tests, INTERVIEWS, QUES-
tions of the statistical models that will TIONNAIRES, SURVEYS, RATING SCALES,
be used for subsequent formal analysis. and reaction-time apparatus. See also AS-
SESSMENT INSTRUMENT.
i n l i e r n. an incorrect value in a data set.
Inliers may arisefromSYSTEMATIC ERROR, i n s t r u m e n t a l i s m n. 1. in the philoso-
respondent error, or processing error. phy ofscience, the position that theories
For example, a researcher may inadver- are not to be considered as either true or
tently report a value in the wrong unit, false but as instmments of explanation
such as including a RAW SCORE in a dis- that allow observations of the world to
tribution of STANDARDIZED SCORES.
be meaningfully ordered. 2. a theory of
knowledge that emphasizes the prag-
i n s t i t u t i o n a l research a study con- matic value, rather than the truth value,
ducted to obtain information about an of ideas. In this view, the value of an
academic setting. For example, a univer- idea, concept, or judgment lies in its
sity may research its faculty, staff, stu- ability to explain, predict, and control
interaction contrast
177
interaction effect
teract with Bj and B2, how Aj and A3 i n t e r a c t i o n sum o f squares the SUM
interact with B, and B2, and so forth. In- OF SQUARES associated with the INTER-
teraction contrasts may be undertaken ACTION EFFECT between two or more
as suggested by ongoing evaluation of variables. In an ANALYSIS OF VARIANCE
study data (see POST HOC COMPARISON) or MULTIPLE REGRESSION, the total sum
or may be decided upon in advance (see of squares can be divided into that for
A PRIORI COMPARISON). Also called in- each variable individually, that for the
teraction comparison. interaction of the variables, and that due
to error.
interaction effect in a FACTORIAL DE-
SIGN, the joint effect of two or more in- i n t e r a c t i o n variance in a FACTORIAL
dependent variables on a dependent DESIGN, the amount of variance in the
variable above and beyond the sum of dependent variable that is explained by
their individual effects: The indepen- an INTERACTION EFFECT between two or
dent variables combine to have a differ- more independent variables.
ent (and multiplicative) effect, such that
intercept n. the point at which either
the value of one is contingent upon the
axis of a graph is intersected by a line
value of another. This indicates that the
plotted on the graph. For example, it is
relationship between the independent
the value of y when ;if = 0 in an equation
variables changes as their values change.
of the form y = a + bx, or the value of x
For example, if a researcher is studying
where a REGRESSION LINE crosses the y-
how gender (female vs. male) and diet-
axis, as shown below. See also X-INTER-
ing (Diet A vs. Diet B) influence weight
CEPT; Y-INTERCEPT.
loss, an interaction effect would occur if
women using Diet A lost more weight interclass correlation the degree of
than men using Diet A. Interaction ef- bivariate relationship between two vari-
fects contrast withand may obscure ables from different measurement classes
MAIN EFFECTS. See also HIGHER ORDER (i.e., the variables have distinct metrics
INTERACTION. Compare ADDITIVE EFFECT. and variances). For example, one may
intercept
178
interquartile range
179
interrater reliability
180
intraclass contingency table
181
intraclass correlation
182
item discriminability
183
item distractor
ranging from -1 to -i-l in value, with the Item response theory models differ in
latter indicating a perfect discriminator. terms of the number of parameters in-
INTERITEM CORRELATIONS also may be cluded in the model. For example, the
used in determining item discrim- RASCH MODEL Is based on the single pa-
inability. See ITEM ANALYSIS. rameter of item difficulty, whereas other
i t e m distractor one of the alternatives models additionally examine ITEM
to the correct answer provided for each DISCRIMINABILITY and the chances of
item on a MULTIPLE-CHOICE TEST. Well- successful guessing. Compare CLASSICAL
chosen distractors can differentiate TEST THEORY.
among the different classes of test takers: i t e m scaling the assignment of a test
Those who know the material should item to a scale position on some dimen-
select the correct alternative despite sion, often that of item difficulty.
the presence of plausible distractors, i t e m selection the selection of test
whereas those with limited knowledge items for inclusion in a test battery based
are likely to guess randomly from upon the final psychometric properties
among the available distractors. of the test battery, the ITEM RESPONSE
i t e m homogeneity see INTERNAL CON- THEORY parameters of the individual
SISTENCY. items, and the clarity and fairness of the
individual items.
item nonresponse see NONRESPONSE.
i t e m - t o t a l correlation the degree to
i t e m r e l i a b i l i t y see INTERNAL CONSIS- which scores on a single item from a test
TENCY.
are related to the total score for the test
i t e m response theory (IRT) a psy- as a whole, used in determining the IN-
chometric theory of measurement based TERNAL CONSISTENCY of the test. Low
on the concept that the probability that item-total correlations (i.e., less than
an item will be correctly answered is a 0.2) indicate poor associations of an
function of an underlying trait or ability item with others on the test, such that it
that is not directly observable, that is, a is not measuring the same construct and
latent trait (see LATENT TRAIT THEORY). should be discarded.
1.0
0.9
10 0 . 8 -
c
o
a. 0.7-
in
0)
te. 0.6-
u
u 0.5-
i-
o 0.4
o
o 0.3-
.>
O
k. 0.2-
0.
0.1 _
"T" 1 \ T
-3.5 -2.5 -1.5 0.5 0.5 2.5
Ability
item characteristic curve
IV
185
Ji
j a c k k n i f e n. a statistical procedure used ERROR between actual and predicted
to estimate the variability of a PARAME- scores. [Willard James, U.S. mathemati-
TER associated with a set of data, such as cian; Charies M . Stein (1920- ), U.S.
the STANDARD ERROR or CONFIDENCE statistician]
INTERVAL. It is particularly appropriate
J curve a pattern on a graph that resem-
when the VARIANCE or underlying distri-
bles the letter J: There is an initial drop
bution is not known. A number of sam-
in value for a short period of time fol-
ples are obtained from the original data
lowed by a continuous increase. For ex-
by eliminating one or more observations
ample, a high achieving person training
at a time, the parameter in question is
for a difficult job might show a J curve
calculated for each sample, and the indi-
for performance (as shown in the dia-
vidual parameters are combined to pro-
gram below). His or her performance ini-
vide an estimate of the overall parameter
tially might drop below that of other
for the entire data set. A jackknife is sim-
employees while he or she learns diffi-
ilar to BOOTSTRAPPING, except that
cult skills but rapidly increase to excel-
bootstrapping involves replacing obser-
lence after fully attaining the required
vations after they have been sampled,
job skills. Also called J-shaped flistri-
such that after each observation is cho-
bution.
sen it is reinserted into the data for possi-
ble selection again. Jeffreys's p r i o r in BAYESIAN statistics,
a method used to estimate the PROBA-
James-Stein estimator a formula used BILITY DISTRIBUTION for a PARAMETER of
to approximate a population PARAMETER interest, such as the population mean.
from sample data. It is an improved ver- Jeffreys's prior is based on the square
sion of the ordinary LEAST SQUARES ESTI- root of the DETERMINANT of FISHER'S IN-
MATOR, providing a smaller and more FORMATION, which provides an approxi-
accurate measure of the MEAN SQUARED mation of the degree of variability that
4 (high)
c 3 -
ra
ce.
0) 2 -
u
c
n
E
4) 1 (iow) -
0.
I r
2 3
Month
/ curve
186
joint probability
og oO O ^<D O
- 08
oocoooooDCO oo ooccoococco
T 1 1 r T 1 1 1 r
Variable x Variable x
jittering
could be expected for the parameter. See rate the points and make it easier to dis-
also PRIOR DISTRIBUTION. [Sir Harold cern a pattern or relationship.
Jeffreys (1891-1989), British mathema-
tician and geophysicist] joint d i s t r i b u t i o n the pattern of val-
ues obtained when estimating the prob-
Jensen's i n e q u a l i t y a mathematical ability of occurrence of two or more
FUNCTION stating that a convex TRANS- RANDOM VARIABLES. For example, the
FORMATION of a mean value will be probability values for drawing a heart
equal to or smaUer than the mean of a and a jack in a 'deck of cards would form
set of convex transformed values. In the following joint distribution: (13/52)
other words, a transformation that cre- X (48/52) to obtain a heart but not a jack;
ates a set of smaller values that follow a (39/52) X (48/52) to obtain a nonheart
convex pattern will yield a smaller over- and a nonjack; (39/52) x (4/52) to obtain
all value than will the average of a set of a nonheart and a jack; and (13/52) x
values that have already undergone a (4/52) to obtain a heart that is also a
convex transformation. [Johan Jensen jack. This is illustrated below.
(1859-1925), Danish mathematician] A joint distribution for two variables is
referred to more specifically as a BI-
j i t t e r i n g n. deliberately adding a small VARIATE DISTRIBUTION, whereas a joint
amount of random noise to each obser- distribution for more than two variables
vation in a SCATTERPLOT where vari- is called a MULTIVARIATE DISTRIBUTION.
ables take on relatively few values,
resulting in identical data points that joint p r o b a b i l i t y the chance that
stack on top of each other and thus are two events will occur simultaneously.
hard to interpret. Consider the generic For example, the joint probability of
example above. Primarily a data visual- drawing a heart that is a jack from a deck
ization technique, jittering helps sepa- of 52 cards is equal to (13/52) x (4/52), or
joint distribution
187
joint probability density function
188
Kk
k 1. symbol for COEFFICIENT OF ALIEN- who are expected to survive a specific ill-
ATION. 2. symbol for the number of in- ness over varying numbers of years; the
dividuals, groups, or other units in a graph of values from a Kaplan-Meier es-
statistical analysis. timator presumably would show a de-
cline over time, with fewer and fewer
Kaiser-Meyer-Olkin test of sam- surviving patients. Also called product-
p l i n g adequacy (KMO test) a measure l i m i t estimator. [Edward L. Kaplan,
of whether the PARTIAL CORRELATIONS
U.S. mathematician; Paul Meier (1924-
among a set of variables are sufficient to 2011), U.S. biostatistician]
conduct a FACTOR ANALYSIS. KMO test
values may range from 0 to 1, with val- k a p p a n. see COHEN'S KAPPA.
ues of .50 or higher indicating there is a
sufficient degree of relationship among K e n d a l l robust line-fit method see
the variables that they reasonably could THEiL-SEN ESTIMATOR. [Sir Mauticc
be analyzed to reveal one or more under- Kendall (1907-1983), British statistician]
lying factors. [Henry F. Kaiser (1927- Kendall's coefficient of concor-
1992), U.S. psychometrician and statisti- dance see COEFFICIENT OF CONCOR-
cian; Michael Meyer, U.S. statistician; DANCE. [Sir Maurice Kendall]
Ingram Olkin (1924- ), U.S. stafistician]
Kendall's tau (symbol: x) a nonpara-
K a l m a n f i l t e r a statistical method used metric measure of the degree of associa-
in TIME-SERIES ANALYSES and other lon- tion between two ordinal variables (i.e.,
gitudinal procedures that combines esti- rank-ordered data). For example, a re-
mates from several sources (e.g., actual searcher could calculate Kendall's tau to
measures, predicted values) to obtain a assess how much relationship there is
more precise indication of a set of values between the rankings of students' per-
over time. It is similar to a HIDDEN formance provided by two observers
MARKOV MODEL but based on CONTINU- (e.g., a teacher and a teaching assistant).
OUS rather than DISCRETE measures. For Also called Kendall's rank correla-
example, a Kalman filter could be ap- tion coefficient. [Sir Maurice Kendall]
plied to assess the amount of drug use
across time for one or more individuals, Kendall's W see COEFFICIENT OF CON-
whereas a hidden Markov model would CORDANCE. [Sir Maurice Kendall]
be used to assess the distinct stages of
kernel density estimator a method
substance use across time. [Rudolf E.
for approximating a smooth, bell-shaped
Kalman (1930- ), Hungarian-born U.S.
distribution from the choppy representa-
electrical engineer and mathematician]
tion in a bar graph or histogram. For ex-
Kaplan-Meier estimator a statistical ample, one might apply a kernel density
method used in SURVIVAL ANALYSIS with estimator to the possible values when
CENSORED DATA for estimating the per- throwing two dice (e.g., one frequency
centage of individuals (or other entities) of 2: a 1 and a 1; two frequencies of 3: a 1
that will remain after a set amount of and a 2 or a 2 and a 1; three frequencies
time. For example, the Kaplan-Meier es- of 4: a 3 and a 1, a 1 and a 3, and a 2 and a
timator could be used in health studies 2; etc.), as they yield a boxy histogram
to estimate the percentage of patients that peaks in the middle at 7 and has low
189
kernel function
frequencies of one to the far left and far happy individuals, and high scores for
right for 2 and 12, respectively. Also individuals with depressed affect who
called kernel estimator; Parzen win- demonstrated little activity or motiva-
dow estimator. tion. It is a form of CONSTRUCT VALID-
ITY.
k e r n e l f u n c t i o n a NONPARAMETRIC
method used to estimate the distribu- Kolmogorov-Smirnov goodness-
tion for a RANDOM VARIABLE. For exam- o f - f i t test a NONPARAMETRIC method
ple, a kernel function may estimate a for comparing the distribution from
rectangular-shaped UNIFORM DISTRIBU- a sample data set to an expected distri-
TION, indicating approximately the bution or the known distribution of a
same number of low, medium, and high given population. If the test yields a dis-
scores. Such a kernel function would be crepancy (D) larger than the CRITICAL
in contrast to the bell-shaped pattern of VALUE, then the sample data are consid-
a NORMAL DISTRIBUTION, in whlch most ered to be significantly different
scores pile up in the middle and gradu- from the reference distribution. Also
ally decline on either side. called Kolmogorov-Smirnov D test;
Kolmogorov-Smirnov one-sample
kernel regression a NONPARAMETRIC test. Compare KOLMOGOROV-SMIRNOV
method of forming a smooth pattern TWO-SAMPLE TEST. [Andrei Nikolaevich
when estimating the relationship be- Kolmogorov (1903-1987) and Nikolai
tween two variables. For example, kernel Vasilevich Smirnov (1900-1966), Soviet
regression could be used to find a rela- mathematicians]
tively smooth (i.e., not choppy) pattern
in a SCATTERPLOT depicting the age and K o l m o g o r o v - S m i r n o v two-sample
height of each chOd in a classroom. Also test a NONPARAMETRIC method for
called kernel regression smoothing; comparing the distributions from two
kernel smoother. samples to see if they are similar. A signif-
icant result from a Kolmogorov-Smirnov
k-means clustering a type of CLUSTER two-sample test indicates that the two
ANALYSIS that separates a large set of en- samples are derived from different distri-
tities into a smaller number of k subsets butions. In contrast, the KOLMOGOROV-
(i.e., clusters) by assigning each entity to
SMIRNOV GOODNESS-OF-FIT TEST IS USed
the subset closest in mean value. For ex-
when only one sample is available
ample,fc-meansclustering might take a
but the larger population distribution is
group of individuals and separate them known. [Andrei Kolmogorov and Nikolai
into three subsets or clusters depending Smirnov]
on the closeness of their IQ scores to the
cluster means of, say, 70, 100, and 130, K-R 20 abbreviation for Kuder-Richard-
for low, medium, and high IQ, respec- son formula 20. See KUDER-RICHARD-
tively. soN FORMULAS.
191
Ll
laboratory analogue see ANALOGUE value at a given point in time is related
STUDY. to its value at an earlier time point.
laboratory research scientific study lagged independent variable in
conducted in a laboratory or other such TIME-SERIES ANALYSIS, an explanatory
workplace, where the investigator has variable exhibiting AUTOCORRELATION
some degree of direct control over the in that its value at a given point in time
environment and can manipulate the is related to its value at an earlier time
INDEPENDENT VARIABLES. Although lab- point.
oratory research generally has greater IN-
TERNAL VALIDITY than FIELD RESEARCH, Lagrange m u l t i p l i e r test in REGRES-
it tends to be less generaUzable to the SION ANALYSIS, a procedure used to esti-
real world (i.e., has less EXTERNAL VALID- mate the improvement in the fit of a
ITY). See EXPERIMENTAL RESEARCH. model to observed data that is gained by
the inclusion of additional variables. For
lack o f f i t the degree to which the val- example, a researcher may hypothesize
ues predicted by a modeltypically, that high-school grade point average
one developed in REGRESSION ANALYSIS, and college admission test scores have
STRUCTURAL EQUATION MODELING, Or more influence on grade point average
CONFIRMATORY FACTOR ANALYSISdi- in college than do gender and ethnicity.
verge from the corresponding empiri- Thus, he or she could develop a model
cal values. A large, significant value from that includes only those two items as
a GOODNESS-OF-FiT TEST indicates a predictor variables and then use a
poorly fitting model. Lagrange multiplier test to determine
whether adding gender and ethnicity re-
l a d d e r i n g n. a knowledge elicitation
sults in a significantly better fitting
technique used in interviews to impose a
model. [Joseph Louis Lagrange (1736-
systematic framework upon questioning
1813), Italian-born French mathemati-
so as to reveal complex themes across
cian]
answers. In laddering, a respondent re-
plies to a series of "why?" probes, thus l a m b d a 1. (symbol: X) see EIGENVALUE.
requiring him or her to expose and ex- 2. (symbol: A) see WILKS'S LAMBDA.
plain choices or preferences and justify
behavior in terms of goals, values, and Laplace d i s t r i b u t i o n see DOUBLE EX-
personal constructs. Laddering is con- PONENTIAL DISTRIBUTION. [Pierre Simon
cerned with establishing links between Laplace (1749-1827), French mathema-
concepts elicited from the participant tician]
(e.g., attitudes and beliefs associated
with a particular consumer product) and large-N design a research approach in
provides greater scope for probing sa- which groups comprising numerous in-
lient issues while optimizing the often dividuals are studied. Generally, the
limited time available with respondents. larger the number of observations or
participants in a research study, the
lagged dependent variable in TIME- greater the statistical POWER in hypothe-
SERIES ANALYSIS, an outcome variable sis testing during data analysis. Compare
exhibiting AUTOCORRELATION in that its SMALL-N DESIGN.
192
latent variable
193
latent vector
194
Levene test for equality of variance
tions, (b) postulate the existence of the l e f t censoring the situation in which a
fewest entities, and (c) invoke the fewest researcher cannot determine the precise
unobservable constructs. Also called time at which a target event occurred for
principle of parsimony. See OC- some individuals within a sample be-
CAM'S RAZOR. cause those participants experienced the
event prior to the observation period.
l a w of t o t a l p r o b a b i l i t y see TOTAL
For example, left censoring might arise
PROBABILITY LAW.
in a study of alcohol initiation among
LCA abbreviation for LATENT CLASS high school students because a subset of
ANALYSIS. the sample is Ukely to have initiated use
prior to high school entry. See also
leading diagonal see MAIN DIAGONAL. RIGHT CENSORING.
least significant difference (LSD) l e p t o k u r t i c adj. describing a frequency
see FISHER LEAST SIGNIFICANT DIFFER- distribution that is more peaked than the
ENCE TEST. NORMAL DISTRIBUTION.
least squares c r i t e r i o n in REGRES-
SION ANALYSIS and similar statistical
procedures, the principle that one
should estimate the values of PARAME-
TERS in a way that wUl minimize the
squared error of predictions from the
model. That is, one should strive to
build models that minimize the squared
differences between actual scores (ob-
served data) and expected scores (those
predicted by the model). Also called
least squares principle. See LEAST
SQUARES REGRESSION.
least squares estimator in LEAST It has more scores in the center and
SQUARES REGRESSION, the model or pro- fewer at the two exfremes. See also MESO-
cedure that yields the smallest differ- KURTIC; PLATYKURTIC.
ences between observed data points and level n. in an experimental design, the
their values as predicted by the model. quantity, magnitude, or category of the
See also LEAST SQUARES CRITERION. INDEPENDENT VARIABLE (or Variables)
least squares regression a type of RE- being studied. For example, if a re-
GRESSION ANALYSIS in which the re- searcher is assessing the effect of alcohol
searcher strives to develop a least- on cognition, each specific amount of
squared-errors solution or least squares alcohol included in the study is a level
solution to describe the relationship be- (e.g., 0.0 oz, 0.5 oz, 1.0 oz, 1.5 oz).
tween an outcome or DEPENDENT VARI- level of analysis see UNIT OF ANALYSIS.
ABLE and one or more predictors or
INDEPENDENT VARIABLES. That is, one level o f measurement see MEASURE-
develops a model that is the best fit for MENT LEVEL.
the data because it yields the smallest
squared differences between the actual level of significance see SIGNIFI-
observations and their values as pre- CANCE LEVEL.
dicted by the model. Also called least Levene test f o r equality o f vari-
squares estimation; ordinary least ance a method for evaluating whether
squares (OLS) regression. See LEAST the VARIANCE in a set of scores is equiva-
SQUARES CRITERION. lent across two or more groups being
195
leverage point
196
linear model
is small. Also called G test; G test. See the requirement that the total of the
also LOG-LIKELIHOOD RATIO. weights used in any comparison equals
zero (e.g., weights of +1 and -1).
l i k e l i h o o d residual see DELETION RE-
SIDUAL. l i n e a r correlation a measure of the
degree of association between two vari-
L i k e r t scale a type of direct attitude
ables that are assumed to have a LINEAR
measure that consists of statements re-
RELATIONSHIP, that is, to be related in
flecting strong positive or negative eval-
such a manner that their values form a
uations of an object. Five-point scales
straight Une when plotted on a graph. It
are common and a neutral middle point
provides an index of the degree of con-
may or may not be included. For exam-
stant change in the value of one variable
ple, an assessment item using a Likert
(y) for each unit change in the value of
scale response format (i.e., a Likert-type
another variable (;*:).
question) might include the following
statement choices: strongly disagree, dis- linear equation an equation that re-
agree, neither disagree nor agree, agree, and lates elements in a summative fashion
strongly agree. The respondent chooses and involves no exponential powers
the option most representative of his or greater than 1. For example, y = a + bx is
her view (e.g., on whether same-sex mar- a linear equation and will form a straight
riages should be permitted) and these line when depicted graphically. Com-
ratings are summed to provide a total at- pare NONLINEAR EQUATION.
titude score for a topic of interest. Also
called Likert summated rating pro- linear function a mathematical proce-
cedure. [Rensis Likert (1903-1981), U.S. dure that transforms one number, quan-
psychologist] tity, or entity to another through
summation, without involving any ex-
linear adj. describing any relationship ponential powers greater than 1. For ex-
between two variables (x and y) that can ample, y = 2>: + 4A: is a linear function. In
be expressed in the form y = a + bx, contrast, a nonlinear function involves
where a and b are numerical constants. interactions (exponents greater than 1)
No COEFFICIENT Can be raised to a power among the quantities, as in y = 2)^ + 4x^.
greater than 1 or be the denominator of
linear i n t e r p o l a t i o n a method of IN-
a fraction. When depicted graphically,
TERPOLATION (i.e., constmcting new
the relationship is a straight line. Com-
data points between two existing data
pare NONLINEAR.
points) that uses a summative formula
linear combination a pattern in without any exponential powers greater
which the value of one variable is de- than 1. Graphically, linear interpolation
rived by multiplying the values of sev- involves connecting two adjacent known
eral other variables in a set by a constant values with a straight line. Compare
quantity and then adding the results. NONLINEAR INTERPOLATION.
For example, the REGRESSION EQUATION
l i n e a r i t y n. a relationship in which one
y = a + b^Xi + biXx + b^x-^ represents a lin-
variable is associated with another via a
ear combination: y is obtained by adding
the constant (a) and the values of the summative formula in which all COEFFI-
CIENTS are to thefirstpower. See LINEAR
three x variables, each of which is first
COMBINATION; LINEAR CORRELATION.
multiplied by a unique weight (b).
l i n e a r i z i n g n. see LINEAR TRANSFORMA-
linear contrast a method of compar-
TION.
ing the mean values on a variable for
two or more groups using CONTRAST linear model any model for empirical
CODING. The mean value for each group data that attempts to relate the values of
is multiplied by a chosen weight with an outcome or dependent variable to the
197
linear regression
1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000
Year
line graph
198
logarithm
from statistical analysis if values are fluence on one another and that any re-
found to be missing for any variable of lationship between them is to be
interest. For example, consider a re- explained by the latent variable underly-
searcher examining the relationships ing them. In ITEM RESPONSE THEORY
among grade point averages in high local independence is the assumption
school, scores on college admissions that an examinee's responses to the vari-
tests, and grade point averages in col- ous items on a test have no influence
lege. If, for whatever reason, a value has upon one another, such that an answer
not been recorded for one of those vari- to one item will not affect the answer to
ables for certain participants, all of the another. It is assumed that the only fac-
information for those individuals is re- tor influencing the respondent's an-
moved from the data set and analyses swers is the latent variable that the test is
are performed only on the remain- designed to measure.
ing (complete) records. This approach,
which assumes data are MISSING C O M - local m a x i m u m the largest value of
PLETELY AT RANDOM, reduces the effec-
a FUNCTION within a given range of
tive sample size and thus is likely to values. When the function is plotted,
overestimate STANDARD ERRORS and the local maximum is the highest point
reduce POWER. Also called casewise on the graph. Compare LOCAL MINI-
deletion; complete-case analysis. M U M .
Compare PAIRWISE DELETION. local m i n i m u m the smallest value of a
l i t e r a l r e p l i c a t i o n see REPLICATION. FUNCTION within a given range of val-
ues. When the function is plotted, the
L j u n g - B o x test a method used in TIME- local minimum is the lowest point on
SERIES ANALYSIS to evaluate the NULL the graph. Compare LOCAL M A X I M U M .
HYPOTHESIS that any AUTOCOR-
RELATIONS among data values observed local regression (loess; lowess) a
across multiple occasions are random. It form of REGRESSION ANALYSIS In which
is a more sophisticated version of the a model of the relationship between out-
BOX-PIERCE TEST. Also called Box- comes and predictors is obtained by fit-
Ljung test; Ljung-Box Q test. [Greta ting different linear or quadratic
M. Ljung, U.S. mathematician; George E. functions to different segments or inter-
P. Box (1919- ), British statistician] vals ofdata. Variables are assigned differ-
ent WEIGHTS in the REGRESSION
L L R abbreviation for LOG-LIKELIHOOD
EQUATIONS to reflect their relative im-
RATIO. portance at each point and no assump-
In abbreviation for natural LOGARITHM. tions are made about the associations
among the variables. The overall curve
l o a d i n g n. see FACTOR LOADING. obtained by combining the individual
l o a d i n g m a t r i x see FACTOR STRUC- fitted curves for the different data seg-
TURE MATRIX. ments shows the general shape of the re-
local effect any outcome from some lationship between the variables. Also
manipulation in a system or experiment called locally weighted regression.
that has a direct but limited influence. LOCF abbreviation for LAST OBSERVA-
For example, a local effect of an analge- TION CARRIED FORWARD.
sic dmg is numbing around the injec-
tion site. Compare REMOTE EFFECT. loess n. see LOCAL REGRESSION.
local independence in FACTOR ANAL- l o g a r i t h m (log) n. the number of times
YSIS and other models involving LATENT that a given value (the base) has to be
VARIABLES, the basic assumption that in- multiplied by itself to produce a specific
dividual observations have no direct in- quantity. For example, the logarithm for
199
logarithmic distribution
the quantity 81 and the base 3 is the more closely matches the assumptions
number4 ( 3 x 3 x 3 x 3 = 81). Logarithms required in particular statistical analy-
are given by the generic formula y = ses, particularly that of NORMALITY.
log^j;, thus 4 = log381. When the base
logical inference see INFERENCE.
is the particular value known as the
Eulerian number (e = 2.71828...), the logical positivism a philosophical
function is a natural logarithm (or perspective that is committed to the
hyperbolic logarithm) and denoted principle of verification, which holds
y = logeX. that the meaning and truth of all non-
tautological statements is dependent on
logarithmic distribution a PROBA- empirical observation. The positivist view
BILITY DISTRIBUTION derived from a log- of science is evident in psychology's com-
arithmic function. It is used to describe mitment to empirical scienfific methods
the probability that a random variable and focus on testing hypotheses de-
(e.g., the number of words understood duced from theory. Logical positivism
by a young child) equates to a given had waned by the middle of the 20th
value. Logarithmic distributions are century. See POSITIVISM. See also POST-
skewed, with long right tails, and usually POSITIVISM; REDUCTIONISM.
have small MEANS and large VARIANCES:
Values are discrete, peak at x = 1, and logistic d i s t r i b u t i o n a THEORETICAL
cannot be negative. Also called log DISTRIBUTION of contlnuous values that
series distribution. describes a set of relationships between
variables as defined by a LOGISTIC FUNC-
logarithmic function the function TION. Frequently used in studying
y = loggX, which gives the relation be- growth rates, it is similar to the NORMAL
tween a number (x) and its LOGARITHM DISTRIBUTION in Shape but has greater
(y) to the base A. It is equivalent tox = a^. KURTOSIS.
l o g a r i t h m i c scale a measurement that logistic f u n c t i o n a basic function of
uses the LOGARITHMS of the values ob- the form y = c/(l + a exp[-bx]), where y
tained rather than the scale values them- and X are variables, a, b, and c are con-
selves. Logarithmic scales are useful stants, and exp is the EXPONENTIAL
when the obtained data cover a very FUNCTION. When graphed, values de-
large range of values and are more easily rived from a logistic function form an S
described or interpreted when reduced shape called the logistic curve. The logis-
to a smaller, more restricted set of num- tic function is used to describe LONGITU-
bers. For example, loudness is assessed DINAL DATA and growth rates in
on a logarithmic scale, as there is a con- particular, in which an initial stage of
siderable difference between the loudest growth is approximately exponential
and softest sounds that may be detected and then, as saturation begins, slows
by the human ear. Compare LINEAR and ultimately ceases altogether.
SCALE.
logistic regression (LR) a form of RE-
l o g a r i t h m i c t r a n s f o r m a t i o n the GRESSION ANALYSIS used when the out-
conversion of raw data values into an- come or DEPENDENT VARIABLE may
other form via the use of LOGARITHMS. assume only one of two categorical val-
In turning multiplicative relationships ues (e.g., pass or fail) and the predictors
into additive ones and eliminating ex- or INDEPENDENT VARIABLES are either
ponential trends, logarithmic transfor- categorical or CONTINUOUS. For exam-
mations allow researchers to analyze ple, a researcher could use logistic regres-
data using LINEAR MODELS. Addi- sion to determine the likelihood of
tionally, such transformations may be graduating from college (yes or no)
used to convert raw data into a form that given such student information as high-
200
log-rank test
school grade point average, college ad- rence of an event. Logits are the inverse
missions test score, number of advanced of LOGISTIC TRANSFORMATIONS.
placement courses taken in high school, logit analysis a statistical procedure
socioeconomic status, and gender. The similar to LOGISTIC REGRESSION that
analysis yields a LOGISTIC REGRESSION presents outcome probabilities in terms
EQUATION with an INTERCEPT and RE- of LOGITS.
GRESSION COEFFICIENTS that indicate
the probability of a specific outcome oc- logit model a mathematical model pos-
curring according to the values of the tulating some relation between the
predictors. Also called logistic model- LOGIT of observed probabilities in a sam-
ing. ple and certain unknown but estimated
population PARAMETERS. A logit model
logistic regression equation the for- often fits the data in cases where a linear
mula produced in a LOGISTIC REGRES- model using probabilities does not.
SION analysis. Given in the form y = e"/(l
log l i k e l i h o o d the LOGARITHM of a
- e"), in which e = the natural LOGA-
probability value or estimate.
RITHM and JT = the probability of an out-
come as determined by a + b^x^ + ^2^2 + log-likelihood r a t i o (LLR) a ratio
... bi,x (where a is the INTERCEPT and b whose numerator and denominator
is the regression weight of a predictor comprise LOG LIKELIHOODS. It is used (in
variable), it is the mathematical expres- the log-likelihood test) to evaluate the
sion of the relationship between a bi- GOODNESS OF FIT Of the NULL HYPOTHE-
nary outcome variable and one or more SIS and the ALTERNATIVE HYPOTHESIS in
predictors. A positive regression coeffi- explaining sample data. The simpler
cient indicates that the predictor in- LIKELIHOOD RATIO indicates how many
creases the probability of the outcome, more times likely the observed values are
whereas a negative regression coefficient to occur under one model than under
indicates that the predictor decreases the other, and the LOGARITHMS of those
the probability of that outcome. A large values yield a log-likelihood ratio.
regression coefficient indicates a predic-
log-linear analysis a method of exam-
tor that strongly influences the probabil-
ining relationships between two or more
ity of the outcome, whereas a near-zero CATEGORICAL VARIABLES that involves
coefficient indicates a predictor that has an analysis of the natural LOGARITHMS
little influence on the probabihty of that of frequency counts within a CONTIN-
outcome. GENCY TABLE. Log-linear analyses do not
distinguish between INDEPENDENT VARI-
logistic t r a n s f o r m a t i o n a TRANS- ABLES and DEPENDENT VARIABLES but
FORMATION in which measurements on
rather attempt to model all significant
a LINEAR SCALE are converted into prob-
associations among all variables, includ-
abilities between 0 and 1. It is given by
ing interactions between any combina-
the formula y = e^/(l + e"), where x is the
tion of the variables, using sets of ODDS
scale value and e is the Eulerian number
and ODDS RATIOS for different category
(see LOGARITHM). The inverse of LOGITS,
outcomes. Also called log-linear mod-
logistic transformations are used in such
eling.
statistical procedures as LOGISTIC RE-
GRESSION. log-normal d i s t r i b u t i o n a THEORET-
ICAL DISTRIBUTION in whlch the LOGA-
log^t n. a nonlinear probability function. RITHMS of values on a variable follow
It is given by the formula logit(p) = log the bell-shaped NORMAL DISTRIBUTION.
(p/[l - p]), where p is the probability of
an event. In other words, a logit is the log odds see LOGIT.
logarithm (log) of the ODDS of the occur- log-rank test a NONPARAMETRIC method
201
log series distribution
202
lurking variable
penalty or "cost" associated with obtain- lower real l i m i t see REAL LIMIT.
ing an incorrect value estimate from a lower-tail p r o b a b i l i t y the probabil-
statistical model. It provides a quantifi- ity that a random variable will take a
cation of error and DECISION RULES that value considerably lower than the mean
serve to minimize the risk of drawing in- for that variable (i.e., one appearing at
accurate conclusions, the extreme left on the ;ic-axis in the plot
loss to f o l l o w - u p the situation in o f a P R O B A B I L I T Y D I S T R I B U T I O N ) . L o w e r -
which researchers lose contact with tail probabilities are important in certain
some participants in a study and thus statistical tests of experimental hypothe-
can no longer obtain data for them. Loss ses. Compare UPPER-TAIL PROBABILITY.
to follow-up may occur, for example, in lower w h i s k e r in a BOX-AND-WHISKER
a clinical study of a new dmg to treat PLOT, the line extending from the value
cancer if some patients unexpectedly at the 25th PERCENTILE to the smallest
move or otherwise become unreachable. value within one INTERQUARTILE RANGE
A common problem in LONGITUDINAL of that percentile score. Compare UPPER
DESIGNS, loss to foUow-up results in WHISKER.
missing data that may bias study find-
lowess n. see LOCAL REGRESSION.
ings.
LR 1. abbreviation for LIKELIHOOD RATIO.
lot acceptance s a m p l i n g see ACCEP- 2. abbreviafion for LOGISTIC REGRESSION.
TANCE SAMPLING.
LSD abbreviation for least significant dif-
lower hinge the point in a distribution ference. See FISHER LEAST SIGNIFICANT
of values below which lie one fourth of DIFFERENCE TEST.
the data. It is equivalent to the first LSD test see FISHER LEAST SIGNIFICANT
QUARTILE and lies midway between the DIFFERENCE TEST.
MEDIAN and the minimum point of the LTA abbreviation for LATENT TRANSITION
distribution. Compare UPPER HINGE.
ANALYSIS.
lower quartile see QUARTILE. l u r k i n g variable see HIDDEN VARIABLE.
203
Mm
M symbol for SAMPLE MEAN. scured by, an INTERACTION EFFECT
between variables.
M A abbreviation for MOVING AVERAGE.
M A D 1. abbreviation for MEAN ABSOLUTE M A model abbreviation for MOVING-
AVERAGE MODEL.
DEVIATION. 2. abbreviation for MEDIAN
ABSOLUTE DEVIATION. M A N C O V A acronym for MULTIVARIATE
ANALYSIS OF COVARIANCE.
m a g n i t u d e o f effect see EFFECT SIZE.
Mahalanobis distance (symbol: D^) M a n h a t t a n distance see CITY-BLOCK
a MULTIVARIATE measure of the degree DISTANCE.
to which two sets of values differ from manifest variable a variable whose
one another: the mean distance between values can be directly observed or mea-
two sets of DEVIATION SCORES, relative sured, as opposed to one whose values
to the variances and covariances among must be inferred. In STRUCTURAL EQUA-
the respective variables. Larger values TION MODELING and FACTOR ANALYSIS
indicate greater distance, and hence less manifest variables are used to study LA-
similarity between sets. The Mahalanobis TENT VARIABLES. Also Called indicator
distance is useful for identifying OUTLI- variable.
ERS in a data set and is often applied in
m a n i p u l a t e d variable see EXPERI-
CLUSTER ANALYSIS, DISCRIMINANT ANAL-
MENTAL VARIABLE.
YSIS, and similar techniques. Also called
generalized distance. [Prasanta C. m a n i p u l a t i o n n. in an EXPERIMENTAL
Mahalanobis (1893-1972), Indian statis- DESIGN, the researcher's adjustment of
tician] an independent variable such that one
or more groups of participants are ex-
m a i n diagonal the line of values run-
posed to specific treatments while one
ning from the upper left to the lower
or more other groups experience a CON-
right of a SQUARE MATRIX.
TROL CONDITION. For example, a health
researcher could introduce a manipula-
tion such that a portion of the partici-
pants in a study randomly receive a new
drug, whereas the remaining partici-
pants receive only a PLACEBO. See EX-
As shown, the first value in the main di- PERIMENTAL VARIABLE.
agonal is the number in the first row and m a n i p u l a t i o n check any means by
thefirstcolumn of the matrix, the sec- which an experimenter evaluates the ef-
ond value is the number in the second ficacy of an EXPERIMENTAL VARIABLE,
row and second column, and so on. Also that is, verifies that a manipulation af-
called leading diagonal. fected the participants as intended.
m a i n effect the consistent total effect M a n n - W h i t n e y U test a NON-
of a single INDEPENDENT VARIABLE on a PARAMETRIC TEST of centrality for ORDI-
DEPENDENT VARIABLE over all Other in- NAL DATA that contrasts scores from two
dependent variables in an experimental INDEPENDENT SAMPLES to assess whether
design. It is distinct from, but may be ob- there are significant differences between
204
marginal homogeneity
205
marginal mean
two sets of data, such as performance any CMV, the researcher introduces a
ratings given by two independent fig- marker variable into the study: This is a
ure-skating judges. new variable that is theoretically unre-
lated to at least one of the main variables
m a r g i n a l mean the average score
of interest. The correlation between the
across two or more groups on one vari-
marker variable and the unrelated vari-
able, particularly as calculated in an
able of interest is taken as an estimate of
ANALYSIS OF VARIANCE to determine the
the common method variance in the
existence of a MAIN EFFECT. See also
study.
MARGINAL DISTRIBUTION.
M a r k o v c h a i n a sequence of steps or
m a r g i n a l p r o b a b i l i t y the chance of events in which the probability of each
occurrence of one variable across each transition depends only on the immedi-
level of another variable. For example, a ately preceding step, and not on any ear-
researcher interested in cancer diagnosis lier step. For example, when a person
and smoking status might list diagnosis considers a behavior change such as los-
(cancer vs. no cancer) in two rows and ing weight he or she experiences certain
smoking status (smoker vs. nonsmoker) stages: The individual could move from
in two columns. The marginal probabili- not thinking about losing weight
ties for cancer diagnosis would be ob- (precontemplation) to considering a
tained by summing the individual weight loss plan (contemplation),
probabihties across both levels of smok- through to maintaining a reasonable
ing status, whereas for smoking status weight (maintenance) and conceivably
the marginal probabilities would be ob- back to not wanting to think about
tained by summing the individual prob- weight loss. This process could be de-
abilities across both levels of cancer scribed as a Markov chain. Also called
diagnosis. In other words, the marginal fUscrete-time Markov chain; Mar-
probabilities are the total probability kov process. Compare RANDOM WALK.
values given at the ends of the rows and [Andrei Markov (1856-1922), Russian
the bottoms of the columns. See also mathematician]
MARGINAL DISTRIBUTION.
Markov chain Monte Carlo
m a r g i n a l sum o f squares see TYPE III
method (MCMC method) a statistical
SUM OF SQUARES.
simulation procedure that tries to find a
m a r g i n of error (MOE) a statistic ex- reasonable pattern of probabilities for
pressing the CONFIDENCE INTERVAL as- moving from one designated state to an-
sociated with a given measurement; it is other in a MARKOV CHAIN. It involves
an allowance for a slight miscalculation sampling from various THEORETICAL
or an acceptable deviation. The larger DISTRIBUTIONS to see which is the best
the margin of error for the sample data, fit for observed data. For example, a
the less confidence one has that the re- Markov chain Monte Carlo method
sults obtained are accurate for the entire could be used to find the plausible pat-
population of interest. tern of probabilities of moving from sub-
stance dependence to recovery. See also
mark-and-recapture s a m p l i n g see MONTE CARLO RESEARCH. [Andrei
CAPTURE-TAG-RECAPTURE SAMPLING. Markov]
m a r k e r variable technique in FAC- M a r k o v m a t r i x see TRANSITION MA-
TOR ANALYSIS, a technique used to con- TRIX.
trol for common method variance
(CMV), that is, the systematic error that M a r k o v r a n d o m f i e l d a graphic rep-
can arise from using a single method of resentation of two or more discrete
measurement. In order to account for states or stages in a MARKOV CHAIN with
206
matching
207
matching coefficient
a disease; one group would be given a is broken down into its simpler compo-
treatment while the other would not. nent elements. See also CHOLESKY
Such an approach would enable the re- FACTORIZATION; SINGULAR VALUE DE-
searcher to rule out sex and family his- COMPOSITION.
tory as potential explanations of the
m a t u r a t i o n n. naturally occurring
study outcome, thereby allowing greater
time-related. changes in a participant
validity to attributing any changes be-
(e.g., growth, aging, fatigue, boredom,
tween the groups to the treatment.
attention shifts) that pose a threat to the
When individuals are paired according
INTERNAL VALIDITY of a smdy, particu-
to their PROPENSITY SCORES, the process
larly a longitudinal one. These pro-
is referred to as caliper matching. See
cessesas opposed to the specific
also MATCHED-PAIRS DESIGN. treatment or interventionmay explain
m a t c h i n g coefficient an index of the any changes in participants during the
correspondence between two sets of experiment. For example, a researcher
items, such as scores or other sample may study substance use in a set of indi-
data. For example, a matching coeffi- viduals from young adolescence to late
cient could be used during an investiga- adulthood. In the study, substance use
tion of possible plagiarism to assess how may naturally decline as a function of
much overlap there is between an essay the development of the participants
submitted by a student and an essay rather than because of the influence of
posted online. an experimental intervention. Thus, the
investigator would want to assess and
mathematical statistics see THEO- possibly control for this maturation
RETICAL STATISTICS.
effect in order to maintain the internal
mathematico-deductive method validity of the study.
see HYPOTHETICO-DEDUCTIVE METHOD.
Mauchly's sphericity test in
m a t r i x n. a rectangular ordered arrange- WITHIN-SUBJECTS ANALYSES OF VARI-
ment (ARRAY) of numbers in rows and ANCE, a technique to determine whether
columns. The following is a simple ex- there is HOMOGENEITY OF VARIANCE
ample: across all possible levels of the inde-
1 6 pendent variable (a condition known as
SPHERICITY). Sphericity is a basic as-
4 1
sumption of such analyses and lack of it
L3 7
can distort the calculations of variance.
Individual items in a matrix are caUed The Mauchly test is based on the LIKELI-
ELEMENTS or entries. Many different HOOD RATIO criterion and involves a
types of matrices are used in statistics, scaled comparison between the DETER-
such as the CORRELATION MATRIX and MINANT and the TRACE of the sample
the COVARIANCE MATRIX. covariance matrix. When the signifi-
cance level of Mauchly's test is < .05
m a t r i x algebra a set of mathematical then equality of variance cannot be as-
rales for analyzing large numbers of sumed. Dohn W. Mauchly (1907-1980),
variables arranged in matrices. Matrix U.S. computer scientist]
algebra procedures may involve basic
calculations, such as adding and sub- m a x i m i n strategy in GAME THEORY or
tracting matrices, and also more com- decision making, a tactic in which an in-
plex processes, such as multiplying and dividual chooses the best of a set of
dividing matrices. worst possible outcomes or payoffs. For
example, participants in a study may
m a t r i x decomposition in linear alge- need to choose whichever outcome
bra, a process in which a complex matrix would best maximize a minimum ad-
208
mean absolute percentage error
209
mean deviation
SERIES values, given as the average dif- SQUARES divided by its DEGREES OF FREE-
ference of a set of such values from their D O M . It is used primarily in the ANALYSIS
expected values, regardless of the direc- OF VARIANCE, in which an F RATIO is ob-
tion of that difference (i.e., whether it is tained by dividing the mean square be-
positive or negative). For example, con- tween groups by the mean square within
sider the following time series of the groups (see BETWEEN-GROUPS MEAN
number of exercise hours per week for SQUARE; WITHIN-GROUPS MEAN SQUARE).
an individual over five weeks: 3, 1, 5, 4, The mean square also is used to deter-
and 7. If a researcher predicts the weekly mine the accuracy of REGRESSION ANAL-
exercise hours to be 2,2, 5,5, and 6, then YSIS models, indicating the amount of
he or she might wish to determine the variance explained by a model (MS^egres-
general degree of inaccuracy of his or her sion) compared to the amount of error or
predictions. Thus, the mean absolute unexplained variance (MSresiduai)-
percentage error would be calculated as
follows: (13 - 21/3 + 11 - 21/1 + 15 - 51/5 +mean
4 square between see BETWEEN-
GROUPS MEAN SQUARE.
- 51/4 + 7 - 6l/7)/5 = (.333 + 1 + 0 + .25 +
.143)/5 = 1.726/5 = .35 x 100 = 35%. mean squared error (symbol: MSE)
the average amount of ERROR VARIANCE
mean deviation see MEAN ABSOLUTE within a data set, given as the typical
DEVIATION. squared distance of a scorefromthe mean
mean difference a measure of variabil- score for the set. Mean squared error
ity in a data set calculated as the average may be calculated in both ANALYSIS OF
of the distances between each score and VARIANCE and REGRESSION ANALYSIS. In
each of the other scores, disregarding the former it is referred to more specifically
whether the deviation is positive or neg- as the WITHIN-GROUPS MEAN SQUARE and
ative. For example, consider the follow- used as the denominator when calculat-
ing three scores: 1, 3, and 9. The mean ing an F RATIO; in the latter it is known
difference would be calculated as: (11-31 as a residual mean square (or mean-
+ 11-91 + 13-11 + 13-91 + 19-11 + 19- square residual) and gives the mean dif-
3l)/6 = (2 + 8 + 2 + 6 + 8 + 6)/6 = 32/6 = ference between actual scores and those
5.33. predicted by a regression model. A large
mean squared error indicates that scores
mean effect size in a META-ANALYSIS, a are not honiogeneous within groups or
measure of the average EFFECT SIZE are not consistent with prediction, such
across multiple studies. For example, an that there is more "noise" than "signal."
investigator analyzing several studies as- For example, a large mean squared error
sessing a new treatment may determine in gender research would show no sig-
a mean effect size by calculating the av- nificant differences between groups of
erage standardized difference between males and groups of females. Also called
treatment and control groups over all of error mean square.
the studies. Calculations of mean effect
size often take variance and RELIABILITY mean-square deviation the typical
into account by assigning different difference between a set of scores and
WEIGHTS to the values derived from dif- the MEAN of those scores, raised to the
ferent studies. A study with a large sam- second power. It is a measure of variabil-
ple, for instance, or one that uses more ity in a data set generally equivalent to
precise measurement techniques may be the VARIANCE.
weighted to have a greater impact in de- mean-square residual (symbol: MSR)
termining the mean effect size. see MEAN S Q U A R E D ERROR.
tion between each score and the pre- TEST THEORY, any difference between an
vious score in a L O N G I T U D I N A L DESIGN. observed score and the T R U E SCORE.
Used to obtain a more accurate estimate Measurement error may arise from flaws
of a trend over time in simations when in the assessment instrument, mistakes
the mean shifts slightly and could cause in using the instrument, or random or
bias, it is calculated in the same manner chance factors. For example, an investi-
as VARIANCE except that scores at succes- gator may obtain biased results from a
sive time points are used i n place of the survey because of problems with ques-
MEAN. tion wording or response options, ques-
tion order, variability in administration,
m e a n square w i t h i n see WITHIN-
and so forth.
GROUPS MEAN SQUARE.
m e a s u r e m e n t i n v a r i a n c e the situa-
m e a n s u b s t i t u t i o n a method of I M P U -
tion in which a scale or constract pro-
T A T I O N i n which the average value for a
vides the same results across several
set of scores is inserted for each missing
different samples or populations. For ex-
score in the data set. Although this prac-
ample, an intelligence test could be said
tice retains the original sample size, it
to have measurement invariance if it
can artificially reduce the variation
yields similar results for individuals of
among the scores. Mean substitution
varying gender, ethnicity, or age. Mea-
may involve inserting either the G R A N D
surement invariance may apply to a sin-
MEAN or the mean for a specific group;
gle relevant characteristic (e.g., gender
the latter approach typically yields a
only), several characteristics (e.g., gen-
more accurate estimate of any given in-
der and ethnicity), or all possible charac-
dividual's score.
teristics and situations. See also
m e a s u r e n. an item or set of items that MEASUREMENT EQUIVALENCE. Compare
provides an indication of the quantity or SELECTION INVARIANCE.
nature of the phenomenon under smdy.
m e a s u r e m e n t l e v e l the degree of
It is sometimes necessary in research to
specificity, accuracy, and precision in a
have more than one measure for each of
particular set of observations or scores,
the main variables of interest.
as reflected in the M E A S U R E M E N T S C A L E
m e a s u r e m e n t n. the act of appraising used.
the extent of some amount, dimension,
m e a s u r e m e n t m o d e l in STRUCTURAL
or criterionor the resultant descriptive
EQUATION MODELING, a model that
or quantified appraisal itself. A measure-
quantifies the association between ob-
ment is often, but not always, expressed
servations obtained during research (in-
as a numerical value.
dicators) and theoretical underlying
m e a s u r e m e n t e q u i v a l e n c e the situ- constracts or factors. When carrying out
ation in which two instruments yield a CONFIRMATORY FACTOR ANALYSIS, for
virraally the same results for the same example, one assesses a hypothesized
group of individuals, such that either measurement model that specifies the
one could replace the other. For exam- relationships between observed indica-
ple, a researcher constructing ALTER- tors and the LATENT VARIABLES that sup-
NATE FORMS of the same test would want port or affect them.
to ensure measurement equivalence for
the FACTOR LOADINGS of each item m e a s u r e m e n t scale any of four com-
within a subscale, as well as for each test mon methods for quantifying attributes
as a whole. See also M E A S U R E M E N T of variables during the course of re-
search, listed in order of increasing
INVARIANCE.
power and complexity: N O M I N A L SCALE,
measurement error in CLASSICAL ORDINAL SCALE, INTERVAL SCALE, and
211
measurement theory
213
method of least squares
microgenetic method a research ap- most an opponent will gain. For exam-
proach that examines developmental ple, a health researcher may propose an
change within a single set of individuals intervention that would be the least
over relatively brief periods of time, usu- aversive treatment for a serious disease,
ally days or weeks, in order to determine thereby minimizing the adverse effects
how it occurs. For example, health re- patients may expect to experience as a
searchers could use a microgenetic result of the disease. Compare MAXIMIN
method to study the weekly improve- STRATEGY.
ment of patients with traumatic brain
injury during a one-year period. m i n i m u m chi-square a method of
identifying acceptable PARAMETER esti-
middle-range theory an approach to mates that involves ttying to find the
the construction of THEORY that aims to smallest possible CHI-SQUARE value
combine existing formulations with ex- when assessing the differences between
perimental findings and in the process observed and expected data points. Min-
to generate new hypotheses that are imum chi-square is an alternative to
open to empirical testing. Representing MAXIMUM LIKELIHOOD for finding pop-
a level of investigation below that of ulation values that are close to observed
a METATHEORY, middle-range theories sample values.
often are used in sociology and psychol-
ogy to identity the constracts and mech- m i n i m u m convex polygon (MCP) a
anisms that best explain relevant data. procedure for estimating the boundaries
For example, a researcher could identity of an area in which a designated POPU-
the main concepts related to organiza- LATION may be found, based on sets of
tional change by investigating previous location points collected over a period of
formulations and findings in this area. time for different individuals. For exam-
ple, a health researcher may draw a min-
m i d p o i n t n. see MEDIAN. imum convex polygon around an area
believed to hold the set of individuals
midrange . the average of the lowest who may have been exposed to a serious
and highest scores in a set of data. The illness, based on reports about individu-
midrange is a MEASURE OF CENTRAL TEN- als with the illness.
DENCY more prone to bias than the
MEAN, MEDIAN, and MODE since it relies minimum description length
solely upon the two most extreme (MDL) in STRUCTURAL EQUATION MOD-
scores, which potentially are OUTLIERS. ELING, the principle that the best model
For example, consider the following for a given set of data is the one that pro-
hours per week spent using a computer vides the most compact description of
for five individuals: 6,10,15, 21, and 35. this data. See also OCCAM'S RAZOR.
The midrange is 20.5, the mean is 17.4,
m i n i m u m variance b o u n d see
and the median is 15. The midrange
CRAMER-RAO LOWER BOUND.
score thus is larger than the mean or me-
dian due to the influence of the outHer m i n i m u m variance unbiased esti-
of 35 hours. mator (MVUE) see EFFICIENT ESTIMA-
TOR.
midspread rz. see INTERQUARTILE RANGE.
MIMIC model acronym for MULTIPLE
m i n i m u m volume ellipsoid a statis-
INDICATORS-MULTIPLE CAUSES MODEL.
tical procedure for finding the smallest
space that would encompass most of the
m i n i m a x strategy in GAME THEORY or points in a MULTIVARIATE data set, as
decision making, a tactic in which indi- in the generic depiction overleaf.
viduals attempt either to minimize their Identitying the minimum volume el-
own maximum losses or to reduce the lipsoid, for example, would help a re-
215
misclassification cost
216
mixed-effects model
217
mixed-methods research
218
modification index
model by adjusting the number or na- tors have a significant interaction be-
ture of its PARAMETERS until the output tween them, and this interaction
from the model is seen to match an ob- changes the narare of the relationships
served sample of data. (e.g., the REGRESSION COEFFICIENTS)
among the other predictors when it is
model f i t assessment the process of
included in the analysis, then moder-
determining how close a proposed
ated multiple regression would be more
model of a relationship between vari-
appropriate.
ables is to the actual sample data ob-
tained. Model fit assessment often moderating effect the effect that oc-
involves a GOODNESS-OF-FIT TEST, an ex- curs when a third variable changes the
amination of the RESIDUALS between nature of the relationship between a pre-
the proposed representation and the ac- dictor and an outcome, particularly in
tual data, and an evaluation of whether analyses such as MULTIPLE REGRESSION.
the proposed representation is consis- For example, STRUCTURAL EQUATION
tent with relevant theory. Statistical pro- MODELING can be used to assess whether
cedures, such as STRUCTURAL EQUATION a predicted association between quanti-
MODELING and HIERARCHICAL LINEAR tative skill and performancefitsequaUy
MODELING, use some or all of these vari- well across different teaching style
ous methods to gauge model fit. groups (e.g., lecture based vs. hands-on
learning). If the prediction is different
model-free test a statistical procedure across the two groups, then teaching
for identifying the underlying structure style is said to have produced a moderat-
or pattern of associations for a phenom- ing effect. Also called moderator ef-
enon that does not require any initial as- fect. See also MODERATED MULTIPLE
sumptions about the nature of this REGRESSION.
pattern. A NONPARAMETRIC TEST is an
example of a model-free test. m o d e r a t o r n. an INDEPENDENT VARI-
that changes the narare of the rela-
ABLE
model i d e n t i f i c a t i o n see IDENTIFICA-
tionship between other variables. For
TION.
example, if a researcher examined the
m o d e l i n g effect a type of EXPERI- relationship between gender and math
MENTER EFFECT in which a participant is performance, a significant difference
unwittingly influenced to give responses might emerge. However, if teaching
similar to the responses the experi- style were taken into account, such that
menter would give if the experimenter those who learned math by applied,
were a participant. hands-on methods performed better
model misspecification see MIS- than those who learned with traditional
lecrare styles, regardless of gender, one
SPECIFICATION.
could say that teaching style was a mod-
model-to-data f i t see FIT. erator of the relationship between gen-
der and math performance. Also called
moderated m u l t i p l e regression a
moderating variable. See also MOD-
statistical procedure that is appropriate
ERATING EFFECT.
when there is a single CONTINUOUS out-
come, two or more PREDICTOR VARI- m o d i f i c a t i o n index (MI) a measure
ABLES, and one or more interactions that indicates the extent to which a
between the predictor variables. For ex- model could be improved if a specific
ample, a researcher might use tradi- PARAMETER were added, or alternatively
tional MULTIPLE REGRESSION to predict whether a parameter could be reason-
mathematics performance from the vari- ably deleted without significantly alter-
ables of quantitative skill, gender, and ing model fit. Although procedures such
teaching style. If the latter two predic- as STRUCTURAL EQUATION MODELING
219
modified replication
often provide a set of modification in- able to draw accurate conclusions. For
dexes as part of the output, researchers example, a researcher could conduct
should only consider amending a model momentaty time sampling to assess the
when using INDEPENDENT SAMPLES and occurrence of smoking at 100 specific
when the modification is consistent time points during a month-long pe-
with relevant theory. riod, using a small portable device that
m o d i f i e d r e p l i c a t i o n see REPLICA-
prompts participants to record their
smoking behavior at each point.
TION.
modulus n. see ABSOLUTE VALUE. moment generating f u n c t i o n a for-
mula for calculating the various MO-
MOE abbreviation for MARGIN OF ERROR. MENTS of a distribution for a random
m o l a r approach any theory or variable. The values obtained from a mo-
method that stresses comprehensive ment generating function describe the
concepts or overall frameworks or struc- range and shape of the set of possi-
tures. For example, METATHEORY uses a ble scores on that variable. See also
molar approach to understanding and CUMULANT G E N E R A T I N G FUNCTION;
developing ways of thinking about PROBABILITY GENERATING FUNCTION.
knowledge and research.
monomethod bias the lack of RELI-
molecular approach any theory or ABILITY or VALIDITY that may occur
method that stresses the components of when measuring a phenomenon with
a phenomenon, process, or system, mak- a single item, scale, or observation. For
ing use of elemental units in its analysis. example, monomethod bias could be
In a molecular approach the specific de- present if a researcher used only one
tails are more important than the overall questionnaire to measure a characteris-
perspective. tic or trait, or if a teacher used scores on
moment n. the power to which the EX- only one exam to evaluate students' per-
PECTED VALUE of a RANDOM VARIABLE is
formance for the entire academic year.
raised. Thus, E(x'') is the *c"^ moment of Also called mono-operation bias.
X. The first moment is usually the MEAN monotonic adj. denoting a variable that
of a variable, the second moment refers either increases or decreases as a second
to VARIANCE, the third moment relates variable either increases or decreases, re-
to SKEWNESS, and the fourth moment spectively: The relationship is not neces-
concerns KURTOSIS. Knowing each of sarily LINEAR but there are no changes
these moments provides a complete pic- in direction. A monotonically increasing
ture of the DISTRIBUTION for a set of variable is one that rises consistently as a
scores: A researcher knows the center second variable increases, for example,
point of the data, how spread out the level of performance in relation to
values are, whether they are lopsided, amount of practice if this were observed
and whether they are peaked (LEPTO- to be the case. In contrast, depression
KURTIC) or flat (PLATYKURTIC). would be a monotonically decreasing
moment about the mean see CEN- variable if its severity were found to fall
TRAL MOMENT.
consistently as a person's level of
perseveration declined.
momentary t i m e s a m p l i n g a proce-
dure in which the researcher indicates monotonic regression a NONPARA-
whether a particular behavior occurred METRIC method used when an outcome
during a designated interval of observa- variable is expected to systematically in-
tion. It is important to collect momen- crease or decrease as a FUNCTION of one
tary time sampling data for a large or more predictor variables. For exam-
number of time periods in order to be ple, monotonic regression could be used
moving average
to assess the relationship between hours shape, and structure are important not
of exercise and body mass index, only in biology but also in neuro-
whereby the latter consistently decreases psychology and medical psychology.
as the former consistently increases. For example, researchers interested in
morphometries could study the form
monotonic relationship any associa- and stracture of the brain in individuals
tion between two variables in which in- who have various physical or psycholog-
crease or decrease in one produces a ical disorders. By contrast, PSYCHO-
corresponding increase or decrease in METRICS specificaUy focuses on the
the other. measurement of psychological phenom-
m o n o t r a i t - m u l t i m e t h o d model see ena.
MULTITRAIT-MULTIMETHOD MODEL.
m o r t a l i t y effect the degree to which
Monte Carlo m a x i m u m l i k e l i h o o d circumstances or behavior increase or
estimation (MCML estimation) in decrease the incidence of death. For ex-
MONTE CARLO RESEARCH, a Simulation ample, a behavioral health researcher
procedure for obtaining a good approxi- could srady the mortality effects of a
mation of a PARAMETER value (i.e., one fatty diet and lack of exercise, which
that closely matches the sample values could lead to early death from heart dis-
generated in the simulation). MCML es- ease.
timation often is used in BAYESIAN IN- m o r t a l i t y odds ratio a measure of the
FERENCE. chance of dying from a specific illness as
Monte Carlo research a SIMULATION the result of a specific behavior or cir-
technique in which a large number of cumstance. For example, a medical psy-
samples with specific selected properties chology researcher could examine the
(e.g., NORMALITY, slze, modeltype)are ODDS RATIO for dying from cancer for
generated by computer in order to assess individuals who smoke a pack or more
the behavior of a statistical procedure or of cigarettes per day. The mortality odds
PARAMETER Under vatying conditions. ratio wiU be greater than 1.0 (as in this
For example, an investigator might con- example) when the chance of the out-
duct Monte Carlo research with a large come (i.e., death) occurring is high and
number of normally distributed samples less than 1.0 when the characteristic
of various sizes (e.g., N = 50, 100, 200, would lead to a smaller chance of death
400, 800) in which a structural model is (e.g., as for the characteristic of a healthy
applied to characterize the data. Results diet).
would help the researcher determine the m o r t a l i t y rate a measure of how often
conditions under which the model be- death occurs, usually with respect to a
haves correctly (i.e.,fitsthe data) as well specific illness, characteristic, behavior,
as shows its limits (e.g., not fitting well or population. For example, a researcher
with sample sizes less than 200). Also could estimate the mortality rate for in-
called Monte Carlo method. dividuals who have been diagnosed with
m o r b i d i t y rate the incidence of dis- cancer.
ease, expressed as a ratio denoting the m o r t a l i t y table see LIFE TABLE.
number of people in a population who
are iU or have a specific disease com- m o v i n g average (MA) a form of aver-
pared with the number who are well. age comprising the means of successive
subsets of data within a longer set of ob-
morphometries n. the theory and servations. For example, a three-term
technique associated with the physical moving average of the sequence 1, 3, 2,
measurement of living organisms and 4, 3, and 5 would be (1 + 3 + 2)/3 = 2, (3 +
their component parts. Studies of size. 2 + 4)/3 = 3, (2 + 4 + 3)/3 = 3, and (4 + 3 +
221
moving-average model
222
multimodal distribution
for mapping, usually along two axes, amount of caffeine intake (e.g., 0 mg, 50
how a set of individuals rate their prefer- mg, 150 mg) affect rats' performance on
ences for or similarity to various charac- a particular maze task. When only one
teristics, siraations, or entities. For independent variable with multiple con-
example, a researcher could use multidi- ditions is involved, it is termed a SINGLE-
mensional unfolding to graph the pref- FACTOR MULTILEVEL DESIGN.
erences of a set of journal editors for
m u l t i l e v e l model see HIERARCHICAL
research article characteristics such as
LINEAR MODEL.
experimental srady, theoretical review,
qualitative inquity, and methodological m u l t i m e t h o d approach a design
application. Editors closer to one axis that uses more than one procedure for
might prefer theoretical or qualitative measuring the main characteristic or
articles, whereas those along the other construct of interest. For example, a re-
axis would favor publishing articles that searcher could use a multimethod ap-
are experimental or methodological. proach to understanding relationship
m u l t i d i m e n s i o n a l variable a con- satisfaction by simultaneously collect-
struct that cannot be fully described or ing data from a survey, asking one or
measured using a single underlying FAC- both of the partners to give their own
TOR. For example, intelligence could be self-report, and systematically observing
described as a multidimensional variable the degree of relatidnship satisfaction.
that involves verbal, quantitative, and See also MULTITRAIT-MULTIMETHOD
other aspects. MODEL.
m u l t i f a c t o r design see FACTORIAL DE- multimethod-multitrait model
SIGN. see MULTITRAIT-MULTIMETHOD MODEL.
m u l t i l e v e l design any research study m u l t i m o d a l d i s t r i b u t i o n a set of
involving two or more INDEPENDENT data in which there is more than one
VARIABLES with multiple conditions MODE or score that occurs most fre-
(levels) of investigation. An example quently, as shown in the generic graphic
is a study examining how time of day overleaf.
(e.g., morning, afternoon, evening) and For example, the ages of a sample of
Stable
Instinct
Knowledge Family upbringing
Financial
Necessity Age of mother
R o l e of ="PP'*
Personality partner
Internal External
Breastfeeding
Emotional support
Complications
Amount of time
spent with child
Desire to Amount of
have child time spent at
hospital
Temporary
multidimensional scaling
223
multinomial
m u l t i n o m i a l logistic regression a
statistical procedure to characterize the
relationship between a set of predictors
college students would form a multi- and a single outcome with several cate-
modal distiibution since the largest num- gories, usually expressed as the odds of
ber of people are either 18, 19, or 20 years falling into one of the outcome catego-
old, with the remaining individuals aged ries (usually the most extreme) given
17 or 21 through 70. See also BIMODAL each individual predictor. For example,
DISTRIBUTION; UNIMODAL DISTRIBUTION. a multinomial logistic regression could
be conducted to assess the likelihood of
m u l t i n o m i a l adj. describing a measure- being included in a high-risk category
ment that can have more than two cate- on a four-category risk outcome (where
gories or outcomes. For example, a 1 = low and 4 = high) depending on pre-
professor assigning grades of A, B, C, D, dictors of smoking history, cholesterol
or F to students in his or her course is level, and weight.
making a multinomial decision, whereas
assigning grades of pass or fail would be m u l t i n o r m a l d i s t r i b u t i o n see MUL-
a binomial decision. TIVARIATE NORMAL DISTRIBUTION.
224
multiple indicators-multiple causes model
225
multiple linear regression
health, which in rarn has several indica- sessment Test (SAT) reading score, and
tors, such as systolic and diastolic blood SAT mathematics score. The multiple re-
pressure readings, pulse, and cholesterol gression equation would be
level. This multiple indicators-multiple
College GPA = intercept +
causes model would be depicted as four
B,(High School GPA) +
(predictor) measures having arrows
B2(SAT Reading) +
pointing toward the factor, which in
B3(SAT Mathematics) +
turn would have four lines emanating
prediction error
outward to represent its four indicators.
Also called multiple regression
m u l t i p l e linear regression see MUL-
model.
TIPLE REGRESSION.
multiple time series a set of measures
m u l t i p l e R see MULTIPLE CORRELATION
on two or more variables or individuals
COEFFICIENT.
taken over numerous occasions. For ex-
m u l t i p l e range test see DUNCAN'S ample, a researcher could examine the
MULTIPLE RANGE TEST. daily recordings of speech and interac-
tion for children with autism across a
multiple regression a statistical tech-
one-year period. The use of several paral-
nique for examining the linear relation-
lel-running TIME SERIES to gather data in
ship between a continuous DEPENDENT
a LONGITUDINAL DESIGN Is Called a mul-
VARIABLE and a set of two or more INDE-
tiple time-series design, whereas the
PENDENT VARIABLES. It Is Often used to
MULTIVARIATE ANALYSIS used to identity
predict a single outcome variableftoma
relationships among the variables in
set of predictor variables. For example,
such data is called multiple time-series
an educational psychology researcher
analysis. Alternatively, the researcher
could use multiple regression to predict
could use an INTERRUPTED TIME-SERIES
college achievement (e.g., grade point
DESIGN to see if the pattern of data for
average) from the variables of high
the two variables changed noticeably
school grade point average. Scholastic
after the introduction of an intervention
Assessment Test (SAT) reading score, SAT
during the recording period. Also called
mathematics score, and SAT writing
multivariate time series.
score. When a single predictor and a sin-
gle outcome are involved the process is m u l t i p l e t test a T TEST used to make
known as LINEAR REGRESSION. Also more than one comparison between
called multiple linear regression. pairs of group MEANS. The increased
number of calculations results in an in-
m u l t i p l e regression equation a RE-
creased chance of making a TYPE I ERROR
GRESSION EQUATION generated from the
(i.e., of concluding there is a significant
process of MULTIPLE REGRESSION, which
difference between groups when there is
represents the relationship between a
not).
single outcome variable and the best
LINEAR COMBINATION of a Set of predic- multiplication rule a rule stating that
tors. In the equation an outcome score the JOINT PROBABILITY of two indepen-
(y) is presented as a function of a con- dent events occurring together or in suc-
stant called an INTERCEPT, plus the cession is equal to the probability of the
products of a REGRESSION COEFFICIENT first event times the probability of the
(B) times the values of various predictor second event. For example, the multipli-
variables (x), plus some prediction error. cation rule would indicate that the prob-
For example, assume a researcher is ability of drawing a heart followed by
studying the relationship of college the probability of drawing a spade is
grade point average (GPA) to the predic- equal to 13/52 x 13/52 = .25 x .25 =
tors of high school GPA, Scholastic As- .0625. Also called and rule; multipli-
multitrait-multimethod matrix
cation law; multiplicative law. sure a sample of the subunits in any cho-
Compare ADDITION RULE. sen unit to avoid uneconomically mea-
suring all of them. When two grouping
m u l t i p l i c a t i v e model a description sets are involved, the process is also
of the effect of two or more predictor known as two-stage sampling, when
variables on an outcome variable that al- three sets are involved it is also called
lows for INTERACTION EFFECTS among three-stage sampling, and so on.
the predictors. This is in contrast to an
ADDITIVE MODEL, which sums the indi- multistate model a statistical repre-
vidual effects of several predictors on sentation of several possible stages for an
an outcome. For example, a health re- event, either over the course of time or
searcher could use a multiplicative across a set of individuals; such models
model to examine the interaction effect are often used to understand disease pro-
of number of cigarettes smoked per day gression or health promotion. For exam-
and length of smoking habit on the ple, a researcher may hypothesize a
onset of cancer; the results might be multistate model of behavior change,
compared to those obtained from an ad- involving various stages from pre-
ditive model that examines the separate contemplation, when an individual is
effects of amount of cigarettes and not thinking of making a change,
length of a smoking habit on cancer through to maintenance, when an indi-
onset. vidual has successfully made a behavior
multisite study research and data col- change (e.g., quitting smoking) for more
lection conducted over several locations than six months. A MARKOV C H A I N is
or geographical areas; this allows greater similar to a multistate model but in-
generalization offindingsthan research volves stages with some degree of de-
conducted in a single place. Also called pendence among them.
cooperative study.
multitrait-multimethod matrix
multistage s a m p l i n g a technique in (MTMM) a matrix showing correlations
which samples are drawn first from among two or more measurement tech-
higher order groupings (e.g., states) and niques used to assess two or more con-
then from successively lower level structs or traits, as obtained from a
groupings (e.g., counties within states, MULTITRAIT-MULTIMETHOD MODEL. It
towns within counties) in order to avoid includes correlations among the same
the necessity of having a SAMPLING traits with different methods (i.e.,
FRAME for the entire population. That is, monotrait-heteromethod) and among
if subunits within a selected unit give different traits with the same method
similar results, one can select and mea- (i.e., heterotrait-monomethod). The for-
Association test
Courtesty Al .13 .14 .10 .14 (.28)
Honesty B, .06 .12 .16 .08 .27 (.38)
Poise Cl .01 .08 -10 .02 .19 .37 (.42)
School drive D, .12 .15 .14 .16 .27 .32 .18 (.36)
multitrait-multimethod matrix
227
multitrait-multimethod model
mer are expected to be the largest, thus ing assumptions about their underlying
demonstrating CONVERGENT VALIDITY, functional relationships. For example, a
whereas the latter are expected to be researcher studying the relationship be-
smallest, demonstrating DISCRIMINANT tween achievement motivation and aca-
VALIDITY. Consider the example below, demic performance could use MARS to
in which VALIDITY DIAGONALS are more accurately model the inverted U-
shown in italics and RELIABILITY DIAGO- shaped pattern expected to emerge. That
NALS in parentheses. is, individuals who have very low
HETEROTRAIT-MONOMETHOD COEFFI- achievement motivation would be ex-
CIENTS as large or larger than monotrait- pected to have relatively low perfor-
heteromethod coefficients indicate that mance scores, which would increase as
some method variance is present, sug- achievement motivation increases. After
gesting that participants are responding a certain point, however, performance
similarly across different traits simply scores could be expected to decline for
because they are being assessed with the individuals who have too much achieve-
same method. ment motivation, thus forming a non-
linear pattern. See SPLINE FUNCTION.
multitrait-multimethod model
(MTMM) a procedure for examining m u l t i v a r i a t e analysis 1. a set of sta-
CONSTRUCT VALIDITY that assesses the tistical procedures for studying the rela-
correlations among two or more charac- tionships between one or more
teristics where these are each measured predictors and several outcome or DE-
in two or more ways. For example, a re- PENDENT VARIABLES. Examples include
searcher studying self-concept and CANONICAL CORRELATION ANALYSIS for
achievement as measured by both self- assessing the relationships among two
reports and teacher .evaluations could sets of variables; FACTOR ANALYSIS for as-
use a multitrait-multimethod model to sessing the relationships among a large
evaluate the associations between the set of measures and a small set of under-
self-reports and teacher evaluations on lying factors; MULTIVARIATE ANALYSIS
each one of the characteristics. The set of OF VARIANCE and MULTIVARIATE ANALY-
correlations resulting from such an anal- SIS OF COVARIANCE for assessing poten-
ysis is displayed in a MULTITRAIT- tial group differences on several
MULTIMETHOD MATRIX. In the related dependent variables; and STRUCTURAL
monotrait-multimethod model, a re- EQUATION MODELING, which examines
searcher examines a single characteristic a theoretically based pattern of relation-
using several different methods. Also ships among multiple independent, de-
called multitrait-multimethod an- pendent, and even mediating variables.
alysis. Also called multivariate statistics.
Compare UNIVARIATE ANALYSIS. 2. more
m u l t i v a r i a t e adj. consisting of or oth-
erwise involving a number of distinct generally, any procedure for under-
variables. For example, a multivariate standing any large set of variables,
srady of ability could involve multiple whether dependent or independent.
measures of intelligence and achieve-
m u l t i v a r i a t e analysis of co-
ment. Compare UNIVARIATE. See also
variance (MANCOVA) a statistical
BIVARIATE. procedure for assessing possible group
m u l t i v a r i a t e adaptive regression differences on a set of outcome or DE-
spline (MARS) a type of nonpara- PENDENT VARIABLES, after taking into ac-
metric REGRESSION ANALYSIS that exam- count the scores on one or more
ines NONLINEAR aS Well aS LINEAR COVARIATES. For example, a researcher
relationships among dependent and in- could conduct a multivariate analysis of
dependent variables but without mak- covariance to assess whether two groups
multivariate normal distribution
229
multivariate normality
the plot shows a series of points along a when comparing the mean values be-
diagonal line, indicating a lack of any tween two groups on two or more DE-
discrepant scores (i.e., any MULTI- PENDENT VARIABLES. It Is an extension of
VARIATE OUTLIERS), then the data have a the T DISTRIBUTION used when testing
multivariate normal distribution. Also for potential differences in situations
called multinormal distribution. that involve more than one outcome.
See also MULTIVARIATE NORMALITY. For example, a researcher might consult
a multivariate t distribution in order to
m u l t i v a r i a t e n o r m a l i t y the situa-find the CRITICAL VALUE that would in-
tion in which the values for a set of vari- dicate whether treatment and control
ables have an even distribution, with groups differ in their physical health
most scores falling in the middle of the and psychological well-being outcomes.
range and a smaller number of high and
low scores. In other words, a set of scores m u l t i v a r i a t e test any of various statis-
demonstrating multivariate NORMALITY tical procedures involving two or more
follows a MULTIVARIATE NORMAL DIS- outcome or DEPENDENT VARIABLES. An
TRIBUTION and lacks OUTLIERS or ex- example is the MULTIVARIATE ANALYSIS
treme scores. OF VARIANCE, whlch extends the ANALY-
SIS OF VARIANCE examining one de-
m u l t i v a r i a t e outlier a data point pendent variable to include additional
whose values on several variables are dependent variables, FACTOR ANALYSIS
very different from the others in a set, and MULTIPLE REGRESSION sometlmes
such that it may BIAS the results of a sta- are considered multivariate tests as well,
tistical analysis, particularly one based even though the former involves a sin-
on an ASSUMPTION of NORMALITY. For gle set of variables that are not necessar-
example, an individual participating in ily independent or dependent and the
a study who obtained an IQ score of 150 latter allows for multiple INDEPENDENT
(when the mean IQ is 100 and the stan- VARIABLES but only one outcome.
dard deviation is 15), a performance
m u l t i v a r i a t e time series see MULTI-
score of 100 (when the mean is 50 and
PLE TIME SERIES.
the standard deviation is 20), and a so-
cial skills score of 1 (when the mean is 10 m u l t i v a r i a t e t test see HOTELLING'S T^
and the standard deviation is 5) would TEST.
be a multivariate outlier, MAHALANOBIS m u n d a n e realism the extent to which
DISTANCE, COOK'S DISTANCE, and Other an experimental siraation resembles a
statistical procedures may be used to de- real-life siraation or event. See also EX-
termine whether multivariate outliers PERIMENTAL REALISM.
are present in a data set. Compare
UNIVARIATE OUTLIER. m u t u a l l y exclusive events 1. two or
more events that have no common ele-
m u l t i v a r i a t e research a study con- ments, that is, they are disjoint (see DIS-
ducted to simultaneously assess the rela- JOINT SETS). 2. in probability theory, two
tionships among multiple DEPENDENT or more events that cannot co-occur:
VARIABLES and INDEPENDENT VARI- The occurrence of one precludes the si-
ABLES. Compare UNIVARIATE RESEARCH. multaneous or subsequent occurrence of
the other(s). For example, the alterna-
m u l t i v a r i a t e statistics see MULTI- tives "heads" and "tails" in a single toss
VARIATE ANALYSIS. of a coin are mutually exclusive events.
m u l t i v a r i a t e t d i s t r i b u t i o n the dis- M V U E abbreviation for minimum vari-
tribution of possible values for the statis- ance unbiased estimator. See EFFICIENT
tic obtained from HOTELLING'S T^ TEST ESTIMATOR.
Nn
n symbol for the number of scores or ob- experiments. Also called naturalistic
servations obtained from a particular ex- design; naturalistic research.
perimental condition or subgroup.
naturalistic observation data collec-
N symbol for the total number of cases tion in a field setting, without labora-
(participants) in an experiment or study. toty controls or manipulation of
variables. These procedures are usually
naive p a r t i c i p a n t a participant who carried out by a trained observer, who
has not previously participated in a par- watches and records the everyday be-
ticular research study and has not been havior of participants in their natural
made aware of the experimenter's hy- environments. Examples of naturalistic
pothesis. observation include an ethologist's
study of the behavior of chimpanzees
narrative analysis atypeof QUALITA- and a developmental psychologist's ob-
TIVE ANALYSIS in which a researcher col- servation of playing children. See also
lects and examines stories from SYSTEMATIC NATURALISTIC OBSERVA-
individuals about a variety of concrete TION. Compare ANALOGUE OBSERVA-
life siraationsranging from first ro- TION; SELF-MONITORING OBSERVATION;
mantic involvements to larger issues, STRUCTURED OBSERVATION.
such as divorce, aging, and life satisfac-
tion. The goal is to understand how indi- n a t u r a l l o g a r i t h m (In) see LOGA-
viduals experience certain events, RITHM.
structure them into coherent sequences, nay-saying n. answering questions nega-
and give them subjective meaning. Also tively regardless of their content, which
called narrative inquiry; narrative can distort the results of surveys, ques-
research. tionnaires, and similar instmments. Also
called response deviation. Compare
n a t u r a l experiment the study of a Y E A - S A Y I N G .
naturally occurring siraation as it un-
folds in the real world. The researcher NCE abbreviation for NORMAL CURVE
does not exert any influence over the sit- E Q U I V A L E N T .
uation but rather simply observes indi-
viduals and circumstances, comparing N=l design see SINGLE-CASE DESIGN.
the current condition to some other nearest neighbor see SINGLE-LINKAGE
condition. For example, an investigator C L U S T E R I N G .
might evaluate the influence of a new
community policing program by observ- negative b i n o m i a l d i s t r i b u t i o n a
ing neighborhood activities after it has T H E O R E T I C A L D I S T R I B U T I O N Of d i s c r e t e
been implemented and comparing the values that describes the number of tri-
outcome to that for neighborhoods in als that will occur before a success in a
which the policy has not yet been imple- sequence of n BERNOULLI TRIALS, with a
mented. Since such real-life events can- given likeliness of failure and success
not be manipulated or prearranged, across the sequence. Also called Pascal
natural experiments are QUASI-EX- distribution. Compare POSITIVE BINO-
PERIMENTAL DESIGNS rather than true MIAL DISTRIBUTION.
231
negative case analysis
232
nominal data
233
nominal scale
two terms are often used interchange- that characterize the average person or
ably. case. Compare IDIOGRAPHIC
n o m i n a l scale a sequence of numbers nonadditive adj. describing values or
that do not indicate order, magnitude, measurements that cannot be meaning-
or a true zero point but rather identity fully summarized through addition be-
items as belonging to mutually exclusive cause the resulting total does not
categories. For example, a nominal scale correctly reflect the underlying proper-
for the performance of a specific group ties of and associations between the
of people on a particular test might arbi- component values. For example, if two
trarily use the number 1 to denote pass variables a and b interact to influence
and the number 2 to denote fail. Since another variable y, the addition of the
the numbers represent categoty labels, separate effects of a and b will not equal
they cannot be manipulated mathemat- the total effect since the contribution of
ically or otherwise quantitatively com- the interaction needs to be included.
pared. A nominal scale is one of four Compare ADDITIVE.
types of measurement scale, the others
being an ORDINAL SCALE, an INTERVAL noncentral d i s t r i b u t i o n a DISTRIBU-
SCALE, and a RATIO SCALE. See also CATE- TION in which the NONCENTRALITY PA-
GORICAL SCALE. RAMETER is not equal to zero: In
statistical SIGNIFICANCE TESTING, such a
n o m i n a l variable a variable whose . distribution is obtained when the NULL
possible values are unordered categories HYPOTHESIS under test is false. The
or labels. For example, choice of college noncentral version of a distribution has
major is a nominal variable. a different mean, SKEWNESS, and VARI-
nomological n e t w o r k a conceptual ANCE (among other properties) from its
network: a broadly integrative theoreti- corresponding central distribution, as
cal framework' that identifies the key well as a larger proportion of numbers
constructs associated with a phenome- beyond the CRITICAL VALUES. For exam-
non of interest and the associations ple, the CHI-SQUARE DISTRIBUTION, F
DISTRIBUTION, a n d T D I S T R I B U T I O N a l l
among these constructs. For example,
have corresponding noncentral versions
psychopathy is a complex notion in-
signified by a noncentrality parameter
volving a significant nomological net-
that is not equal to zero.
work of knowledge and speculations
about components, causes, correlates, noncentrality parameter in many
and consequences as well as their inter- PROBABILITY DISTRIBUTIONS USed i n SIG-
relationships and means of measure- NIFICANCE TESTING, a PARAMETER that
ment or evaluation. has a value different from zero when the
nomological v a l i d i t y the degree to NULL HYPOTHESIS under test is false.
which a measure assesses the specific This parameter is important in deter-
construct it is designed to assess, as for- mining the POWER of a statistical proce-
mulated hom the NOMOLOGICAL NET- dure. See NONCENTRAL DISTRIBUTION.
WORK for the construct being measured.
n o n d i r e c t i o n a l hypothesis a hy-
See CONSTRUCT VALIDITY.
pothesis that one experimental group
nomothetic adj. relating to the formu- will differ from another without specifi-
lation of general laws as opposed to the cation of the expected direction of this
study of the individual case. A effect. For example, a researcher might
nomothetic approach involves the hypothesize that college students will
study of groups of people or cases for the .perform differently from elementary
purpose of discovering those general school students on a memory task with-
and universally valid laws or principles out predicting which group of students
234
nonlinear interpolation
235
nonlinear model
known range of such points that in- n o n m e t r i c adj. describing data that are
volves use of a NONLINEAR EQUATION. N O M I N A L or ORDINAL, as opposed to IN-
Compare LINEAR I N T E R P O L A T I O N . See TERVAL DATA or RATIO DATA. Such data
INTERPOLATION. cannot be precisely quantified. Exam-
ples include yes/no answers or a list
n o n l i n e a r m o d e l any model that at-
ranking individuals on some attribute.
tempts to relate the values of an out-
come or dependent variable to the n o n m e t r i c s c a l i n g an analytic tech-
explanatory or independent variables nique i n which similarities among data
using a NONLINEAR E Q U A T I O N . Compare points are determined based on a rela-
LINEAR MODEL. tive ordering of their indexed difference
values. A variation of standard (metric)
n o n l i n e a r r e g r e s s i o n a procedure MULTIDIMENSIONAL SCALING, which
for analyzing the relationship between uses the actual quantified spatial dis-
an INDEPENDENT VARIABLE (x) and a DE- tances among points (i.e., DISSIMILAR-
PENDENT VARIABLE (y) where the RE- ITY COEFFICIENTS and SIMILARITY
GRESSION EQUATION involves exponen- COEFFICIENTS), nonmetric scaUng exam-
tial forms of X. That is, the changes in y ines only the rankings of those distance
are not consistent for unit changes in values to identify similarities.
the X variable(s) but are a function of the
particular values of x. For example, a n o n n o r m a t i v e adj. not conforming to
nonlinear regression model is given by y or reflecting an established N O R M . For
= a + b^Xi + ^2^1^ + e, where a indicates example, an individual test score well
the place where the line of BEST FIT above or well below the mean for a class-
crosses the y-axis, b^ indicates the num- room of students is a nonnormative
ber of units that y changes when x is value. Compare N O R M A T I V E .
changed by 1 point, and e is error. Also
called c u r v i l i n e a r regression. Com- n o n p a r a m e t r i c adj. describing any an-
pare LINEAR REGRESSION. See REGRES- alytic method that does not involve
SION ANALYSIS. making ASSUMPTIONS about the data of
interest. Compare P A R A M E T R I C
n o n l i n e a r r e l a t i o n s h i p an associa-
n o n p a r a m e t r i c h y p o t h e s i s test see
tion between two variables in which the
N O N P A R A M E T R I C TEST.
direction and rate of change flucraate.
That is, the amount of change i n a D E - n o n p a r a m e t r i c r e g r e s s i o n a form of
PENDENT VARIABLE (y) varies as a func- REGRESSION ANALYSIS in which the rela-
tion of the particular value or level of the tionship between an outcome or D E -
INDEPENDENT VARIABLE (x). Compare PENDENT VARIABLE and one or more
LINEAR RELATIONSHIP. predictors or I N D E P E N D E N T V A R I A B L E S is
analyzed without the assistance of a pre-
n o n l i n e a r t r a n s f o r m a t i o n a TRANS- existing model. KERNEL REGRESSION is
F O R M A T I O N of a data set that uses a func- an example. In contrast to traditional re-
tion to change the linear relationship gression, i n which the structure of the
between variables. For example, a nor- REGRESSION EQUATION is known and
malizing transformation creates a new only the REGRESSION COEFFICIENTS are
set of scores that approximate a N O R M A L estimated from the data set, non-
DISTRIBUTION from the original non- parametric regression requires that both
normal data set, and a L O G A R I T H M I C the equation and coefficients be deter-
TRANSFORMATION may be used to con- mined from the data.
vert raw data into a form that more
closely matches the ASSUMPTIONS re- n o n p a r a m e t r i c s t a t i s t i c s statistical
quired in particular statistical analyses. procedures in which the narare of the
Compare LINEAR T R A N S F O R M A T I O N . data being analyzed is such that certain
236
norm
237
normal approximation
238
nuisance parameter
239
nuisance variable
240
Oo
objective adj. 1. having verifiable exis- ments, techniques of control, and obser-
tence in the external world, indepen- vations are as free from bias as possible.
dently of any opinion or judgment. 2. Compare SUBJECTIVITY.
impartial or uninfluenced by personal
o b l i m i n n. a class of OBLIQUE ROTATION
feelings, interpretations, or prejudices.
methods used in FACTOR ANALYSIS.
Compare SUBJECTIVE.
oblique adj. describing a set of axes that
objective e x a m i n a t i o n see OBJEC-
do not form right angles; in graphical
TIVE TEST. representations of mathematical com-
objective p r i o r see PRIOR DISTRIBU- putations (such as FACTOR ANALYSIS),
TION. this indicates correlated (not independ-
ent) variables. Compare ORTHOGONAL.
objective scoring scoring a test by
means of a key or formula, so that differ- oblique r o t a t i o n a transformational
ent scorers will arrive at the same score system used in FACTOR ANALYSIS when
for the same set of responses. It is con- two or more factors (i.e., LATENT VARI-
trasted with subjective scoring, in which ABLES) are correlated. Oblique rotation
the score depends on the scorer's opin- reorients the factors so that they fall
ion or interpretation of participant re- closer to clusters of vectors representing
sponses to items. MANIFEST VARIABLES, thereby simplify-
ing the mathematical description of the
objective test a type of assessment in-
manifest variables. It is one of two types
stmment consisting of a set of items or
questions that have specific correct an- of FACTOR ROTATION used to identify
swers (e.g., How much is 2 + 27), such a simpler stracrare pattern or solution,
that no interpretation, judgment, or per- the other being ORTHOGONAL ROTATION.
sonal impressions are involved in scor- oblique s o l u t i o n in FACTOR ANALYSIS,
ing. Examples include MULTIPLE-CHOICE any of various FACTOR STRUCTURE pat-
TESTS and TRUE-FALSE TESTS. In contrast, terns identified using OBLIQUE ROTA-
short-answer and essay examinations are TION methods.
SUBJECTIVE TESTS. Also called objective
examination. observation n. 1. the careful, close ex-
amination of an object, process, or other
objectivism n. the position that judg- phenomenon for the purpose of collect-
ments about the external world can be ing data about it or drawing conclusions.
established as true or false independent See CONTROLLED OBSERVATION; NATU-
of personal feelings, beliefs, and experi- RALISTIC OBSERVATION; PARTICIPANT OB-
ences. Compare SUBJECTIVISM. ob- SERVATION. 2. a piece of informafion (see
jectivist n., adj. DATA). observational adj.
objectivity n. 1. the tendency to base observational study research in
judgments and interpretations on exter- which the experimenter passively ob-
nal data rather than on subjective fac- serves the behavior of the participants
tors, such as personal feelings, beliefs, without any attempt at intervention or
and experiences. 2. a quality of a re- manipulation of the behaviors being ob-
search study such that its hypotheses, served. Such studies fypically involve
choices of variables studied, measure- observation of cases under naturalistic
241
observation coding system
conditions rather than the random as- EXPECTED FREQUENCY for the major. In-
signment of cases to experimental condi- deed, in the CHI-SQUARE TEST and other
tions: Specially trained individuals record statistical procedures observed frequen-
activities, events, or processes as precisely cies are compared with expected fre-
and completely as possible without quencies, with large differences suggest-
personal interpretation. Also called ob- ing a poorfitfor models or explanations
servational design; observational proposed to describe the data. Also
method; observational research. See called obtained frequency.
also NATURALISTIC OBSERVATION. observed score a data value recorded
observation coding system a for a variable via actual measurement or
scheme or list of mutually exclusive la- observation, as opposed to an estimated
bels, categories, and so fortheach of or predicted score (or to a score on a LA-
which characterizes a coherent dimen- TENT VARIABLE, which is derived or ap-
sion of interestused for classifying in- proximated).
formation obtained by observing others. observed-score e q u a t i n g see SCORE
It is an essential component of any OB- EQUATING.
SERVATIONAL STUDY as it clarifies what
data should be collected and how, pro- observed variable see MANIFEST VARI-
viding definitions of each code along ABLE.
with examples. For example, a re- observer bias any expectations, beliefs,
searcher investigating infant behavior or personal preferences of a researcher
might develop the following observa- that unintentionally influence his or her
tion coding system: (1) quietly alert, (2) recordings during an OBSERVATIONAL
crying, (3) fussing, and (4) sleeping. STUDY. See EXPERIMENTER EFFECT.
Thus, for every time during a specific ob-
servation period that an observer sees observer d r i f t gradual, systematic
the baby acting alert, he or she would re- changes over a period of time by a partic-
cord a 1 on his or her data form; for ular observer in his or her application of
every time the observer sees the baby criteria for recording or scoring observa-
crying, he or she would record a 2, and tions. See EXPERIMENTER DRIFT.
so forth. o b t a i n e d f r e q u e n c y see OBSERVED
observed d i s t r i b u t i o n the DISTRIBU- FREQUENCY.
TION of values on a variable as actually Occam's razor the maxim that given a
obtained from a SAMPLE, as opposed to a choice between two hypotheses, the one
THEORETICAL DISTRIBUTION based on involving the fewer assumptions should
the laws of probability. For example, one be preferred. In other words, one should
might draw a card from a standard deck apply the LAW OF PARSIMONY and
of playing cards, note the suit, replace choose simpler explanations over more
the card in the deck, and repeat the pro- complicated ones. [WUUam of Occam or
cess 100 times. An observed distribution Ockham (c. 1285-1347), English Francis-
would show how many times a heart, a can monk and Scholastic philosopher]
diamond, a club, and a spade were cho-
sen in the trial of 100 draws. odd-even r e l i a b i l i t y a method of as-
sessing the reliability of a test by corre-
observed frequency the counts of val- lating scores on the odd-numbered
ues or categories on a variable as ob- items with scores on the even-numbered
tained from a sample. For example, the items. The CORRELATION COEFFICIENT
number of people majoring in psychol- between the two halves is adjusted using
ogy at a particular college would be the the SPEARMAN-BROWN PROPHECY FOR-
observed frequency of that major, which MULA, which accounts for the decreased
may or may not match a theoreticaUy size of the odd and even item sets com-
one-bend transformation
243
one-factor analysis o f variance
245
operationism
246
orthogonal
247
orthogonal coding
dicates uncorrelated (unrelated) vari- particular set of payments, one set being
ables. Compare OBLIQUE. 2. denoting a paid to each participant.
research design in which there are an
equal or proportional number of partici- outcome research a systematic inves-
pants across study conditions. See OR- tigation of the effectiveness of a type or
THOGONAL DESIGN. technique of intervention (e.g., a new
form of psychotherapy for treating de-
orthogonal coding see CONTRAST pression) or of the comparative effec-
CODING. tiveness of different intervention types
or techniques (e.g., cognitive behavior
orthogonal contrast in a FACTORIAL therapy vs. drug therapy for depression).
ANALYSIS OF VARIANCE, a set of non- In other words, outcome research fo-
redundant comparisons among the cuses on determining whether partici-
mean values for different samples. In pants benefit from receiving the
other words, the set contains no overlap- intervention. Compare PROCESS RE-
ping information, such that the results SEARCH.
from one comparison provide no infor-
mation about the results of the second outcome variable see DEPENDENT
comparison. Also called orthogonal VARIABLE.
comparison.
outlier n. an extreme observation or
orthogonal design an experimental measurement, that is, a score that signif-
design involving multiple INDEPENDENT icantly differs from all others obtained.
VARIABLES in which each level of one For instance, assume a researcher ad-
variable is combined with each level of ministered an intelligence test to a
every other and all of the resulting con- group of people. If most individuals ob-
ditions contain an equal or proportional tained scores near the average IQ of 100
number of participants or observations. yet one person had an IQof 150, the lat-
ter score would be an outlier. OutUers
JM orthogonal rotation a trans- can have a high degree of influence on
formational system used in FACTOR summaty statistics (e.g., the MEAN and
ANALYSIS in which the different under- STANDARD DEVIATION can be pulled se-
lying or LATENT VARIABLES are required verely toward outliers) and on estimates
to remain separated from or un- of PARAMETER values, and they may dis-
correlated with one another. It is one of tort research findings if they are the re-
t w o types o f F A C T O R R O T A T I O N USCd tO sult of error.
identify a simpler structure pattern or
solution, the other being OBLIQUE ROTA- overdispersion n. in CATEGORICAL
TION. Also called rigid rotation. DATA ANALYSIS, 3 Situation in which ob-
tained observations display more varia-
orthogonal solution in FACTOR tion (DISPERSION) than is predicted by a
ANALYSIS, any of various FACTOR STRUC- model. Overdispersion distorts STAN-
TURE patterns identified using OR- DARD ERROR and CONFIDENCE INTERVAL
THOGONAL ROTATION methods. estimates and often is associated with
OUTLIERS or model MISSPECIFICATION.
outcome n. 1. the result of an experi-
ment, treatment, intervention, or other overfitted model a model of a data set
event. For example, the test scores of stu- in which there are too many predictor
dents in a classroom who have been variables or PARAMETERS relative to the
taught with a new lecture method com- number of sample observations. An
prise an outcome, as does any individual overfitted model is unnecessarily com-
element from a SAMPLE SPACE. 2. in plex for the amount of avaUable data
GAME THEORY, the factor determining a upon which it is based and does not rep-
248
o-x-o
licate well nor accurately predict re- to the treatment rather than to sex or
sponses. age differences. If, however, the re-
searcher were to pair the groups on such
o v e r i d e n t i f i c a t i o n n. in STRUCTURAL additional factors as area of residence
EQUATION MODELING and Similar statis- and household income, overmatching
tical techniques, the presence of more would be present and likely to mask the
PARAMETERS in a model than are re- true narare of the relationship under in-
quired to correctly specity the relation- vestigation and lead to statistical BIAS,
ships among the variables of interest. In such as by reducing the POWER and EFFI-
other words, the number of known pa- CIENCY of analyses.
rameters exceeds the number of free pa-
rameters, allowing one or more values to oversampling n. a sampling strategy in
be estimated in more than one way. which certain subsets of participants are
Compare UNDERIDENTIFICATION. overrepresented in a study group com-
pared to the larger population from
overidentified model a model of a which they are drawn. Oversampling in-
data set in which the number of PARAM- volves deliberately selecting greater
ETERS estimated exceeds the number of numbers of such participants than
data points or pieces of unique informa- would be obtained via RANDOM SAM-
tion in the data. In other words, an PLING so as to enhance the accuracy of
overidentified model contains elements PARAMETER values estimated through
that are redundant, with more known statistical procedures. For example, a re-
than free parameter values. Compare searcher selecting firefighters from a
JUST-IDENTIFIED MODEL; UNDERIDENTI- candidate pool might include equal
FIED MODEL. numbers of males and females to ensure
a sufficient sample of females for statisti-
overmatching n. unnecessary MATCH- cal analysis, even though males com-
ING: the pairing of research participants prise the majority of firefighters in the
on an excessive number of characteris- pool.
tics or on characteristics having little
or no potential influence upon the out- oversaturated model see SATURATED
come of interest. For example, a re- MODEL.
searcher investigating a new drug O-X-O shorthand for a ONE-GROUP PRE-
treatment for cancer might create two TEST-POSTTEST DESIGN, in which the re-
groups whose members are of the same searcher observes and measures a single
age and sex, administering the drag to set of participants (O), introduces an in-
one group while the other receives a pla- tervention (X), and then measures the
cebo. Such group comparability would participants (Q) again to determine
allow the researcher greater validity in whether the intervention resulted in
attributing any changes between them any change.
249
Pp
p symbol for PROBABILITY. different levels of these attributes. Also
called paired comparison design.
paired comparison method a sys- See PAIRED COMPARISON METHOD.
tematic procedure for scaling and com-
paring a set of stimuli or other items. A paired sample any sample in which
pair of stimuli is presented to the partici- each participant is matched on a partic-
pant, who is asked to compare them on a ular variable to a participant in a second
particular dimension: This is often a sample. This ensures that any differ-
physical characteristic, such as size, ences on an outcome variable cannot be
loudness, or brightness, but may also in- due to differences between participants
clude personal traits, abilities, or perfor- on the matching variables. A simple ex-
mance on some task. The process is ample of a paired sample is a PRETEST-
continued until evety item in the set has POSTTEST DESIGN in which a variable is
been compared with evety other item; in measured before and after an interven-
a work setting, for example, the process tion; in this case the participant is
would be complete when evety em- matched to him- or herself. More fre-
ployee has been evaluated relative to quently, paired samples are achieved by
every other employee supervised by the matching individuals on personal char-
rater. The number of pairs is found by acteristics such as age and gender.
the formula n x ( n ~ 1)12, where n is the Typically, each member of the pair is
number of objects or people to be rated. randomly assigned to the treatment or
Thus, if there are five objects or people control group. See MATCHED-PAIRS DE-
to be compared, the number of pairs is SIGN.
10; if there are 20 objects or people to be
rated, the number of pairs is 190. This paired-samples t test see MATCHED-
latter example illustrates a limitation of PAIRS T TEST.
the procedure; when the number of ob-
jects to be rated becomes large, the num- pairwise comparison in a FACTORIAL
ber of paired comparisons becomes DESIGN where the variable being investi-
excessive and burdensome for the rater. gated has more than two levels, a proce-
See PAIRED COMPARISON EXPERIMENT. dure in which the data obtained from
each level of the variable are compared
paired comparison experiment a separately to the data from evety other
study design in which sets of paired al- level. For example, where the research
ternatives are offered to respondents, interest is the differences in student
who have to indicate which of the alter- achievement resulting from three differ-
natives they prefer or rank more highly ent types of content presentation, statis-
on a specified dimension. The alterna- tical tests would be used to evaluate the
tives offered may be real or hypothetical different outcomes of each pair of con-
and judgments can be qualitative or tent methods (i.e.. Method 1 and
quantitative. The purpose of the paired Method 2, 1 and 3, and 2 and 3).
comparison experiment is to assess the DUNNETT'S MULTIPLE COMPARISON
importance of various attributes to the TEST, T U K E Y ' S H O N E S T L Y SIGNIFICANT
respondents and to determine the DIFFERENCE TEST, a n d the N E W M A N -
weights the respondents attach to the KEULS MULTIPLE COMPARISON TEST aU
250
parameter estimation
251
parameter space
parameter space the set of experimen- ables. This can help a researcher to get
tal variables that are used as input into a a clearer understanding of the relation-
model and serve as the basis for estima- ship between x and y. Also called semi-
tion of PARAMETERS in a population. partial correlation. Compare PARTIAL
CORRELATION; ZERO-ORDER CORRELA-
parametric adj. describing any analytic
TION.
method that makes ASSUMPTIONS about
the data of interest. Compare NON- p a r t i a l autocorrelation in TIME-
PARAMETRIC SERIES DATA, the relation between a
parametric hypothesis test see measure at one time point and the same
PARAMETRIC TEST. measure at a subsequent time point,
controlling for study variables that are
parametric statistics statistical pro- thought to affect the magnirade of the
cedures that are based on assumptions over-time relation. For example, mood
about the distribution of the attribute ratings from one day to the next may be
(or attributes) in the population being affected by whether or not a positive so-
tested, for example, that there is a NOR- cial interaction occurred. See AUTO-
MAL DISTRIBUTION of values. Compare CORRELATION; PARTIAL CORRELATION.
NONPARAMETRIC STATISTICS.
parametric test a HYPOTHESIS TEST p a r t i a l correlation the association
that involves one or more assumptions between two variables, x and y, with the
about the underlying distribution of the influence of one or more other variables
population (typicaUy assumed to be a (zj, Z2) statistically removed, controlled,
NORMAL DISTRIBUTION) from which the or held constant; the effect of the z vari-
sample is drawn. Common parametric able is removed from both x and y. For
hypothesis tests include ANALYSIS OF example, partial correlation between sal-
VARIANCE, REGRESSION ANALYSIS, CHI- ary and education level can be examined
SQUARE TESTS, TTESTS, and Z TESTS. Also
after the effects of age on each are re-
called distribution-dependent test. moved. It is often of interest to learn
Compare NONPARAMETRIC TEST. whether a correlation is significantly re-
duced in magnitude once a third vari-
Pareto d i s t r i b u t i o n a distribution able is removed. Also called higher
whose values follow a POWER FUNCTION, order correlation; higher order
such that there is a narrow peak at the partial correlation; partial associ-
extreme left of the x-axis and a long tail ation; partial relationship. Com-
to the right. Pareto distributions occur pare PART CORRELATION; ZERO-ORDER
in a diverse range of phenomena, such CORRELATION.
as the distribution of city populations or
income, that are not adequately de- p a r t i a l eta squared (symbol: y\p) in
scribed by a NORMAL DISTRIBUTION. The ANALYSIS OF VARIANCE, an EFFECT SIZE
marked SKEWNESS of the Pareto distribu- measure given as the ratio of the BE-
tion is often summarized in the so-called TWEEN-GROUPS S U M OF SQUARES (SSB)
80:20 rule, which states that, for exam- relative to the sum of the SSg and the
ple, 20% of cities will have 80% of the ERROR SUM OF SQUARES. It expresses the
total population, 20% of people will proportion of variance in the dependent
earn 80% of total income, and so on. or outcome variable that is accounted
for by the factors or independent vari-
p a r s i m o n y n. see LAW OF PARSIMONY. ables in the model. A higher value
part correlation the association be- means that a higher proportion of the
tween two variables, x and y, with the in- variance was accounted for by the fac-
fluence of a third variable, z, removed tors included in the model. See also ETA
from one (but only one) of the two vari- SQUARED; GENERALIZED ETA SQUARED.
partial replication
253
partial residual plot
an empirical study in which only a sub- researcher can become enmeshed with
set of the srady's design and methodol- the group to the extent that he or she is
ogy are repeated. Often, a researcher will no longer able to document it in an un-
choose to conduct a partial replication biased way. Also called participative
to show that the general findings of research; participatory research.
a study remain the same, despite the
methodological changes. p a r t i c i p a n t report a SELF-REPORT
provided by an individual who takes
partial residual plot a graphical rep- part in a study. These data are useful be-
resentation of the relationship between cause an individual has a unique access
an INDEPENDENT VARIABLE (x^) and a to his or her own thoughts, emotions,
DEPENDENT VARIABLE (y) with the other and behaviors. On the other hand, an
independent variables in the model (X2, individual may show biases due to mem-
x^, ... x^) statistically removed or con- oty recall and the desire to be seen in a
trolled for. It allows the researcher to see particular light.
the unique effects of an independent
variable on the outcome. These plots are participants' rights in a study ap-
less informative when the independent proved by an INSTITUTIONAL REVIEW
variables are highly correlated. Also BOARD, a set of conditions relating to
called added-variable plot; compo- participants in the srady and their role
nent-plus-residual plot. in the research. Participants normally
should be informed about the purpose
p a r t i c i p a n t n. a person who takes part of the study (but see DECEPTION RE-
in an investigation, study, or experi- SEARCH), its procedures (i.e., what spe-
ment, such as by performing tasks set by cifically is expected to occur) and the
the experimenter or by answering ques- associated costs and benefits; that their
tions set by a researcher. In an experi- data from the study will be kept confi-
mental design, the person may be dential; whom they can contact if they
further identified as an experimental have any concerns; and that they can
participant (see EXPERIMENTAL GROUP) leave the study at any time without pen-
or a control participant (see CONTROL alty. Also called subjects' rights. See
GROUP). Participants may also be re- also INFORMED CONSENT; RESEARCH
ferred to as SUBJECTS, although the for- ETHICS.
mer term is now often preferred when
referring to humans rather than ani- p a r t i c i p a n t variable see SUBJECT
mals. See also RESPONDENT. VARIABLE.
participative research see PARTICI-
p a r t i c i p a n t m a t c h i n g see MATCH- PANT OBSERVATION.
ING.
participatory a c t i o n research a
p a r t i c i p a n t observation a QUASI- form of ACTION RESEARCH that empha-
EXPERIMENTAL RESEARCH method in
sizes collaboration between researchers
which a trained investigator studies a and members of the disadvantaged com-
preexisting group by entering it as a
munity of interest.
member, while avoiding a conspicuous
role that would alter the group processes p a r t i c i p a t o r y evaluation a type of
and bias the data. The researcher's role PROGRAM EVALUATION in which the in-
may be known or unknown to the other dividuals who provide a service (e.g.,
members of the group. Cultural anthro- professional staff) take a direct role in
pologists become participant observers evaluating the service or program; cli-
when they enter the life of a given cul- ents of the service provided may also be
rare to study its stracture and processes. involved. The official evaluator facili-
A downside to the approach is that the tates proceedings rather than actually
254
path diagram
255
path modeling
path diagram
period effect
257
periodicity
ies. Period effects may be difficult to can provide basic information about the
distinguish from AGE EFFECTS and CO- underlying process that may otherwise
HORT EFFECTS in research. be obscured.
periodicity n. 1. the state of recurring period prevalence see PREVALENCE.
more or less regularly, that is, at inter-
p e r m u t a t i o n n. an ordered arrange-
vals. Phenomena that exhibit periodic-
ment of elements from a set. A permuta-
ity include the cycle of the seasons,
tion is similar to a COMBINATION but
human circadian rhythms, and con-
distinguished by its emphasis on order.
sumer spending. 2. in SYSTEMATIC SAM-
For example, if there are three colored
PLING, a problem that can arise when
objectsred (R), white (W), and blue
the process of selection (e.g., choosing
(B)there are six possible permutations
every fifth unit from a list) interacts with
of these objects: RWB, RBW, WBR, WRB,
a hidden periodic trait (e.g., gender)
BRW, and BWR.
within the population. If the sampling
technique coincides with the periodicity permutation test see RANDOMIZA-
of the trait (e.g., evety fifth person on TION TEST.
the list is a male), the sampling tech-
nique will no longer be random and the perpendicular n. a line that is at a right
representativeness of the sample is com- angle (90) to another line or plane.
promised. Typically, in plotting a graph, the y-axis
is perpendicular to the x-axis.
periodic survey an assessment, often a
personality p r o f i l e 1. a presentation
questionnaire, that is administered with
of results from psychological testing in
some regularity but not necessarily at
graphic form so as to provide a summary
equal time intervals and that generally
of a person's traits or other unique at-
includes new respondents each time. tributes and tendencies. Because various
The national census is an example of a scores appear in one display, a researcher
periodic survey with equal time inter- can see the pattern of high and low scale
vals. scores for a given person. For example, a
periodogram n. in SPECTRAL ANALYSIS profile of a client's test results from the
or TIME-SERIES ANALYSIS, a plot that Minnesota Multiphasic Personality In-
breaks down the complex sources of ventory would display his or her scores
variance and noise in the function into on measures of depression, hysteria,
component elements. A periodogram masculinity/femininity, paranoia, social
8 50 -
(A
1 1 1 1 1~
Depression Hysteria Masculinity/ Paranoia Introversion
femininity
Attribute
personality profile
Phillips-Perron test
259
phone survey
TEST. [Peter C. B. Phillips (1948- ), Brit- and meet at nodal points to form a con-
ish econometrician; Pierre Perron tinuous line. This is an approach that is
(1959- ), Canadian econometrician] vety useful for identifying abrapt changes
phone survey see TELEPHONE INTER-
or discontinuities in a process. For exam-
VIEW.
ple, piecewise regression might be ap-
propriate where a researcher is studying
physical determinism see STATISTI- a group of participants before treatment,
CAL DETERMINISM. during treatment, and after treatment.
physiological measure any of a set of Also called segmented regression.
instraments that convey precise infor- pie chart a graphic display in which a
mation about an individual's bodily circle is cut into wedges with the area of
functions, such as heart rate, skin con- each wedge being proportional to the
ductance, skin temperature, Cortisol percentage of cases in the category repre-
level, palmar sweat, and eye tracking. sented by that wedge. For example, a re-
Studies using these measures typically searcher might present the results of a
obtain measurements before and after survey on sources of psychology-related
the introduction of a stimulus condition literature used by the public in the man-
as a way to document an individual's re- ner shown below.
sponse to that stimulus. It generally works best when there are
p i n. (symbol: TC) 1. a ratio expressing the not many categories (with thin wedges)
circumference of a circle to its diameter, being shown. A downside of the graphic
given as 3.141592. 2. a symbol denoting is that it is not very efficient because it
the probability of success on a trial, used uses significant space to show the fre-
often in ODDS RATIO calculations as well quencies of a single variable.
as in PROBIT ANALYSIS and LOGIT M O D -
P i l l a i - B a r t l e t t trace a statistic used in
ELS.
MULTIVARIATE ANALYSES OF VARIANCE
piecewise regression a variant on or- and other procedures to determine the
dinary LEAST SQUARES REGRESSION In STATISTICAL SIGNIFICANCE of differ-
which a REGRESSION LINE consisting of ences among levels of the INDEPENDENT
several different lines isfittedto the data. VARIABLES while simultaneously taking
The several pieces have different slopes into account multiple DEPENDENT VARI-
Academic sources 6%
Internet 30%
pie chart
PMP
261
PMLE
262
polynomial regression
a couple interrapts the other during a egories. For example, the quality of a
problem-solving task. See also ZERO- training program could be described as
INFLATED POISSON REGRESSION. [Simeon polychotomous: poor, fair, good, or ex-
D. Poisson] cellent. Also called polytomous. See
also DICHOTOMOUS VARIABLE.
Poisson variable a study variable that
is described by a count distribution polychotomous variable a variable
within a particular time period. See POIS- having more than two possible catego-
SON DISTRIBUTION. [Simeon D. Poisson] ries, either ordered or unordered. For ex-
ample, college matriculation could be
policy analysis a collection of tech-
described as a polychotomous variable:
niques for synthesizing information (a)
freshman, sophomore, junior, or senior.
to specify alternative policy and pro-
Also called polytomous variable. See
gram choices in cost-benefit terms, (b)
to assess organizational goals in terms of also DICHOTOMOUS VARIABLE.
input and outcome, and (c) to provide p o l y n o m i a l n. a mathematical expres-
a guide for future decisions concerning sion consisting of multiple terms, each
research activities. Policy analysis is of which is the product of a constant (a)
a useful approach for examining the im- and a variable (x) raised to a whole num-
plications of different potential scenar- ber exponent:
ios. For example, an organization n-l
considering aflexibletime option for its + a_ix + a
employees might choose to engage in See also BINOMIAL.
this process before implementing the
change. p o l y n o m i a l contrast a comparison of
mean values for more than two different
policy research empirical studies con- levels or time points of an INDEPENDENT
ducted to guide the formulation of cor- VARIABLE to determine whether they
porate or public policies. For example, a follow a particular mathematical pat-
nonprofit organization might investi- tern, such as linear, quadratic, cubic, or
gate how to help students from lower in- quartic. For example, a researcher might
come urban areas gain access to higher have a specific prediction about how cli-
education. ents' average degree of motivation to at-
tend therapy will differ over the course
polychoric correlation coefficient
of treatment (e.g., will decrease, then in-
an index showing the degree of associa-
crease, and then decrease again), and he
tion between two variables that are
or she could use a polynomial contrast
scored as ordered categories but assumed
to be manifestations of underlying CON- to determine whether this is indeed the
TINUOUS RANDOM VARIABLES. In Other
case. See also REPEATED CONTRAST.
words, it is an alternative to the PROD- p o l y n o m i a l f u n c t i o n a mathematical
UCT-MOMENT CORRELATION COEFFI- function that relates one number, quan-
CIENT that is applied to ordered tity, or entity to a POLYNOMIAL. For ex-
POLYCHOTOMOUS VARIABLES. For exam- ample, y = 2x' + 4^" + x i s a polynomial
ple, a researcher might use a polychoric function.
correlation coefficient to assess the rela-
tionship between teacher ratings of stu- p o l y n o m i a l regression atypeof LIN-
EAR REGRESSION analysis in which the
dent interest in a topic (none, moderate,
substantial) and student self-reports of relationship between an explanatory or
INDEPENDENT VARIABLE and an outcome
interest.
or DEPENDENT VARIABLE Is modeled as a
polychotomous adj. describing a single POLYNOMIAL. Polynomial regression is
measurement or item scored with more appropriate for expressing higher order
than two unordered or ordered cat- NONLINEAR effects among variables (via
263
polynomial trend
265
postdictive validity
tial in the early development of psy- comparison (or contrast); post hoc
chology and helped to form its commit- contrast; unplanned comparison
ment to empirical methods. Despite the (or contrast).
advent of POSTPOSITIVISM, it continues
to be a major force in contemporaty psy- post hoc power see RETROSPECTIVE
POWER.
chology. positivist adj.
postdictive v a l i d i t y see RETROSPEC- post hoc test a statistical procedure
TIVE VALIDITY. conducted on the basis of the findings
obtained from previous analyses. Most
posterior d i s t r i b u t i o n in BAYESIAN commonly, the phrase refers to compari-
approaches, an estimated distribution of sons of the mean values obtained on a
values for a population characteristic of variable by different study groups that
interest that is obtained by combining are made only after an overall ANALYSIS
empirical data with prior expectations OF VARIANCE has revealed a significant F
based on existing knowledge or opinion RATIO, indicating that there is some ef-
(the PRIOR DISTRIBUTION). fect or difference across the groups that
should be examined further. Also called
posterior p r o b a b i l i t y the probability follow-up test. See POST HOC COM-
that a certain event will occur given that PARISON.
other related events have been observed.
For example, assume that Basket A con- postpositivism n. 1. the general posi-
tains 10 white cards and 20 red cards, tion of U.S. psychology since the mid-
while Basket B contains 20 white cards 20th centuty, when it ceased to be domi-
and 10 red cards. If one selects a card nated by LOGICAL POSITIVISM, HYP
from a basket at random, without noting THETICO-DEDUCTIVE METHODS,
the color, the probability that it will OPERATIONALISM. Postpositivistic ps
be from Basket A is .5, since the only chology is a broader and more human
possibilities are A or B. If, however, endeavor, influenced by such develop-
the color of the card is noted then ments as social constructionism and
one must take this additional informa- PHENOMENOLOGY. 2. more generally,
tion into account in determining the approach to science and the philosophy
probability that the card came from Bas- of science that has moved away from a
ket A. For example, the probability that position of strict POSITIVISM. post-
a white card will be pulled from Basket A positivist adj., n. postpositivistic
would be revised downward fiom .5 adj.
according to BAYES THEOREM: Since
there are fewer white cards in A, a white posttest 1. n. an assessment carried out
card is more likely to be drawn from B after the application of some interven-
instead. tion, treatment, or other condition to
measure any changes that have oc-
post hoc comparison any examina- curred. Posttests often are used in re-
tion in which two or more quantities are search contexts, in conjunction with
compared after data have been collected PRETESTS, to isolate the effects of a vari-
and without prior plans to carty out able of interest. For example, in a study
the particular comparison. For example, examining whether a new therapy helps
after obtaining a significant F RATIO for a to alleviate depression, participants
data set, a researcher may perform post might receive the therapy and then
hoc comparisons to follow up on and complete a short symptom inventory,
help explain the initial findings. Differ- the results of which would be compared
ent statistical tests are required for post to those from an inventory taken prior
hoc comparisons than for A PRIORI COM- to the treatment. 2. vb. to administer a
PARISONS. Also called a posteriori posttest.
practical significance
267
practice effect
268
pretest
269
pretest-posttest design
270
principle of beneficence
srady that was gathered directly by the nical features. Also called principal-
researcher, from his or her own experi- components factor analysis.
ments or from first-hand observation.
principal components regression
Compare SECONDARY DATA. 2. original
a prediction model that uses a set of
experimental or observational data, that
uncorrelated variables obtained from a
is, RAW DATA.
PRINCIPAL COMPONENTS ANALYSIS aS
p r i m a r y f a c t o r see FIRST-ORDER FAC- predictor or INDEPENDENT VARIABLES.
TOR. The benefit of this approach is that the
original set of predictors in the model
p r i m a r y s a m p l i n g u n i t an item or may have been so highly interrelated as
case chosen in the initial stage of a to result in COLLINEARITY. The drawback
multistep process for selecting elements is that if the uncorrelated variables are
for study from a larger group. For exam- not interpretable then the problem of
ple, in a national srady of medical collinearity is not solved.
school students, the first major step
might be choosing a random pool of p r i n c i p a l f a c t o r analysis an ap-
medical schools from among those proach to identifying the dimensions
available throughout the countty; each underlying associations among a set of
medical school so identified is a primaty variables using a COVARIANCE MATRIX of
sampling unit. Compare SECONDARY estimated COMMUNALITIES as input.
Principal factor analysis assumes that all
SAMPLING UNIT; TERTIARY SAMPLING
variables have been measured with some
UNIT.
degree of error and requires that dimen-
principal-axis f a c t o r analysis in sions be extracted in a particular way.
EXPLORATORY FACTOR ANALYSIS, an ex- Specifically, the first dimension ex-
traction method in which the COEFFI- tracted must account for the maximum
CIENT OF MULTIPLE DETERMINATION of possible variance, having the highest
one variable with all other variables in squared correlation with the variables it
the system is used as the initial underlies; the second dimension must
COMMUNALITY estimate for that vari- account for the next maximal amount of
able. variance and be uncorrelated with the
previously extracted dimension; and so
p r i n c i p a l component a single linear forth. The researcher retains a certain
combination extracted as part of a PRIN- number of dimensions based on various
CIPAL COMPONENTS ANALYSIS. criteria, including interpretations of
FACTOR LOADINGS.
principal components analysis
(PCA) a data reduction approacfi in p r i n c i p l e o f beneficence in RE-
which a number of independent linear SEARCH ETHICS, the requirement of IN-
combinations of underlying explana- STITUTIONAL REVIEW BOARDS that
tory variables are identified for a larger studies "do good" with respect to the
set of original observed variables. PCA work being conducted, the benefits to
reproduces all of the information in the society at large, and the treatment of
original CORRELATION MATRIX and does participants. Thus, the researcher should
not assume that variables are measured maximize the possible benefits of each
with any degree of error. Thus, the result srady and consider its potential impact
is a new set of variables that are in the broadest sense. For example, in a
uncorrelated with each other and or- study of implicit attitudes a researcher
dered in terms of the percentage of the might note that understanding more
total variance for which they account. about the measurement of attitudes re-
The technique is similar in its aims to garding sensitive topics may lead to re-
FACTOR ANALYSIS but has different tech- duced societal prejudice.
271
principle of nonmaleficence
272
probit transformation
273
procedure
erage individual scores. Also called pro- well as sentence completion, word asso-
file proximity measure. ciation, and drawing tests. The use of
projective techniques has generated
prognostic variable in medical re-
considerable discussion among re-
search, a BASELINE CHARACTERISTIC Of
searchers, with opinions ranging from
an individual that may lead to a clinical
the belief that personality assessment is
diagnosis in the future. It is important
incomplete without data from at least
for researchers to define such character-
one or more of these procedures to the
istics early in a study to help customize
view that such techniques lack impor-
interventions to be as effective as possi-
tant psychometric features such as RELI-
ble for each person. For example, child-
ABILITY and VALIDITY. Also called
hood obesity is a prognostic variable for
projective method.
adult-onset diabetes, such that many in-
terventions focus on promoting healthy promax r o t a t i o n in EXPLORATORY
eating and exercise in youth. FACTOR ANALYSIS, atypeOf PROCRUSTES
ROTATION used to fit a solution to a sim-
p r o g r a m evaluation an appraisal pro-
pler structure or target. More specifi-
cess that contributes to decisions on
cally, it is a form of OBLIQUE ROTATION
installing, continuing, expanding, certi-
in which a researcher starts with a solu-
fying, or modifying social programs, de-
tion from a VARIMAX ROTATION and
pending on their effectiveness. Program
raises the FACTOR LOADINGS to a certain
evaluation also is used to obtain evi-
power to make them more extreme; he
dence to rally support or opposition for
or she then finds the rotation to the tar-
the organization providing services and
get using the LEAST SQUARES CRITERION.
to contribute to basic knowledge in the
A researcher subsequently interprets the
social and behavioral sciences about so-
loadings of the variables on the factors
cial interventions and social experimen-
and also a correlation matrix of the fac-
tation. See EVALUATION RESEARCH.
tors.
projection n. in linear algebra and
proof n. 1. the establishment of a propo-
EXPLORATORY FACTOR ANALYSIS, the
sition or theory as true, or the method
mapping of a set of points in multi-
by which it is so established. There is
dimensional space. project vb.
much debate as to whether propositions
projection pursuit in EXPLORATORY or theories can ever be truly proven. In
DATA ANALYSIS, a method for identify- logic and philosophy, even a valid argu-
ing the most interesting clusters in LIN- ment can be untrue if its first premise is
EAR COMBINATIONS of variables in a false. For example, it is a valid argument
multivariate data set. If a nonnormal to say that All trees are pines: I have a tree
cluster is identified, its corresponding in my garden: Therefore my tree is a pine. In
data are removed and a subsequent clus- empirical sciences such as psychology,
ter is obtained. The process continues both logical and methodological prob-
until no further clusters can be found. lems make it impossible to prove a the-
ory or hypothesis true. Disciplines that
projective technique any assessment
rely on empirical science must settle for
procedure that consists of a fixed series
some type of probabilism based on em-
of relatively ambiguous stimuli designed
pirical support of its theories and hy-
to elicit unique, sometimes highly idio-
potheses. See also FALSIFIABILITY. 2. in
syncratic, responses that reflect the per-
mathematics and logic, a sequence of
sonality, cognitive style, and other
steps formally establishing the truth of a
psychological characteristics of the re-
theorem or the validity of a proposition.
spondent. Examples of this type of pro-
cedure are the Rorschach Inkblot Test propensity score in an experimental
and the Thematic Apperception Test, as design lacking RANDOM ASSIGNMENT, a
275
prophecy formula
measure of the probability that a partici- searcher calculates an ODDS RATIO for
pant will be part of the treatment group. each category, which indicates the likeli-
It is a type of BALANCING SCORE used to hood of that specific outcome occurring
adjust for SAMPLING BIAS by matching according to the values of the predictors
differences among participants across and COVARIATES of interest. For exam-
conditions. Several different algorithms ple, one might use the proportional
exist for determining propensity scores. odds model to determine employees'
likely degree of job satisfaction (poor,
prophecy f o r m u l a see SPEARMAN-
fair, or good) according to such variables
B R O W N PROPHECY FORMULA.
as salary, work hours, length of com-
mute, work responsibilities, decision-
p r o p o r t i o n n. a decimal value that ex-
presses the size of a subset of cases of in-making authority, age, and sex.
terest relative to the set as a whole. For proportional reduction of error
example, if 10 students out of 40 stu- (PRE) an index of the extent to which
dents in a class complete their assign- the addition of one or more INDEPEN-
ments on time then the proportion of DENT VARIABLES In a REGRESSION EQUA-
students who are not late is 10/40 = .25. TION reduces inaccuracy in predicting
p r o p o r t i o n a l adj. having a constant scores on the DEPENDENT VARIABLE. It is
ratio between quantities, such that the based on a LOSS FUNCTION and generally
overall relationship does not change. ranges from 0 to 1 in value. The concept
For example, consider a researcher who may be appUed to GENERALIZED LINEAR
is examining the differences between MODELS and other statistical approaches
two treatments among males and fe- as well. Also called proportional re-
males as follows: 10 males in Treatment duction of prediction error.
A and 20 males in Treatment B, and 20 proportional sampling an approach
females in Treatment A and 40 females in which one draws cases for srady from
in Treatment B. Although there are un- certain groups (e.g., gender, race/ethnic-
equal numbers of people in each treat- ity) in the amounts that are observed in
ment condition, the ratio or proportion the larger population. For example, if a
between them remains the sametwice university has 60% female students and
as many females as males. In UNBAL- 40% male students, a researcher would
ANCED DESIGNS, the presence or absence obtain a sample comprising the same
of such proportional cell frequency is percentages or proportions, such as 120
critical to a researcher's choice of ana- females and 80 males in a 200-student
lytic strategy. subset. A significant drawback of this
proportional hazards assumption form of STRATIFIED SAMPLING is that
in cox REGRESSION ANALYSIS, the situa- small sample sizes may limit statistical
tion in which two individuals with simi- modeling and inference. Also called
lar COVARIATES have the same ratio of proportionate sampling.
estimated hazards over time. A variety of proportion of variance accounted
tests of proportionality exist for re- f o r the extent to which certain factors
searchers to evaluate whether this as- or INDEPENDENT VARIABLES in a pro-
sumption holds. posed model are associated with the out-
come or DEPENDENT VARIABLE of
p r o p o r t i o n a l hazards model see
interest. The total amount of possible
cox REGRESSION ANALYSIS. variance in an outcome is 1.00 or 100%,
p r o p o r t i o n a l odds model a type of hence the proportion of variance ac-
LOGISTIC REGRESSION used when the counted for indicates what fraction of
outcome variable has more than two or- that 100% is explained by the model.
dered response categories. The re- For example, in a REGRESSION EQUATION
pseudorandom
277
pseudoscience
278
P-technique factor analysis
279
publication bias
281
quadratic discriminant function
Quadrant C Quadrant A
(-,+) 2 {+.+)
-4 -3 -2 -1 1 2 3
H 1 1 1 -H 1 h
Quadrant B Quadrant D
(-, -) (+. -)
quadrant
ables into two or more groups. A set of used process in statistical procedures in-
PREDICTOR VARIABLES Is Combined opti- volving the LEAST SQUARES CRITERION
mally to maximize the association with or other principles for minimizing error;
a particular grouping outcome (e.g., dif- it sequentially determines the difference
ferent clinical diagnostic groups) with- between a provisional estimate and a
out the strong assumptions required of final estimate until some stopping point.
other, more simple linear versions of the See LOSS F U N C T I O N .
model (see DISCRIMINANT ANALYSIS).
These assumptions include MULTI- quadrature n. the numerical computa-
VARIATE NORMALITY and HOMOGENEITY tion of an integral of a function, either
OF V A R I A N C E .
by traditional rules or by an adaptive
process involving approximation. It is
quadratic discriminant function used extensively during estimation in
the set of optimal PREDICTOR VARIABLE modeling approaches, such as ITEM RE-
weights that emerge from QUADRATIC SPONSE THEORY and GENERALIZED LIN-
DISCRIMINANT ANALYSIS. T h e weights
EAR MODELS.
help the researcher understand and in-
terpret which of the predictor variables qualitative adj. referring to a variable,
in the analysis are particularly strong in study, or analysis that involves a method
accounting for membership in a particu- of inquity based on descriptive data
lar outcome group and which variables without the use of numbers. Qualitative
have little or no influence in this predic- methods and approaches focus on un-
tion. derstanding open-ended responses, such
as those found in written narratives,
quadratic form 1. a specific mathe-
interviews, focus groups, observation,
matical form that is central to MULTI-
and case studies (see QUALITATIVE RE-
VARIATE ANALYSIS. If X is a vector and A
SEARCH). Compare QUANTITATIVE.
a square matrix, then the quadratic form
is given by x'Ax. 2. a POLYNOMIAL of de-
qualitative analysis the investigation
gree two; for example, 4x^ + 2;i;y - 3y^ is a
of open-ended material and narratives
quadratic form of the variables x and y.
by researchers or raters who describe
quadratic loss f u n c t i o n a commonly dominant themes that emerge in the
282
quality control angle chart
data. In many cases specialized com- years to death, time to relapse) but also
puter programs are used to identity these recognizes aspects of the treatment that
themes with researcher-provided search affect a patient's quality of life. For ex-
terms. A major component of describing ample, the analysis would identify a
the data is ttying to understand the rea- treatment condition (e.g., introduction
sons behind the observed themes. Com- of a new drag) that might lead to longer
pare QUANTITATIVE ANALYSIS. but much lower quality of life for the pa-
tient.
qualitative data information that is
not expressed numerically, such as de- q u a l i t y adjusted s u r v i v a l time in
scriptions of behavior, thoughts, atti- QUALITY ADJUSTED SURVIVAL ANALYSIS,
tudes, and experiences. If desired, a patient's days of survival in good
qualitative data can often be expressed health, that is, discounting days with
quantitatively through a CODING pro- side effects, sickness, or low productivity
cess. See QUANTITATIVE RESEARCH. Com- due to the treatment.
pare QUANTITATIVE DATA.
q u a l i t y assurance in health adminis-
qualitative observation a formal de- tration or other areas of service delivety,
scription of a phenomenon that takes a systematic process that is used to mon-
into account the context in which that itor and provide continuous improve-
phenomenon occurs but does not rely ment in the quality of health care
on numbers in the description. See services. It involves not only evaluating
QUALITATIVE RESEARCH. the services in terms of effectiveness, ap-
qualitative research a method of re- propriateness, and acceptability but also
search that produces descriptive (non- offering feedback and implementing so-
numerical) data, such as observations of lutions to correct any identified defi-
behavior or personal accounts of experi- ciencies and assessing the results.
ences. The goal of gathering this QUALI- quality control processes associated
TATIVE DATA is to examine how in-
with research, production, or services
dividuals can perceive the world from that are designed to reduce the number
different vantage points. A variety of of defective measurements and products.
techniques are subsumed under qualita-
tive research, including CONTENT ANAL- q u a l i t y c o n t r o l angle chart a picto-
YSES of narratives, in-depth INTERVIEWS, rial representation of a TIME SERIES in
FOCUS GROUPS, PARTICIPANT OBSERVA- which a change in the mean or variabil-
TION, and CASE STUDIES, often con- ity is indicated by a change in angular
ducted in naturalistic settings. Also direction. It shows the behavior of a sta-
called qualitative design; qualita- tistic (on the y-axis) across many differ-
tive inquiry; qualitative method; ent samples over time (on the ;t-axis). A
qualitative study. Compare QUANTI- researcher can see whether the statistic
TATIVE RESEARCH. settles to a particular point over time
and where extreme values of that statis-
qualitative variable a descriptive
tic are over samples. The figure usually
characteristic or attribute, that is, one includes the average of the statistic over
that cannot be numerically ordered. Ex- the samples and upper (UCL) and lower
amples are gender, eye color, and pre- (LCL) control limits shown as lines that
ferred sport. Compare QUANTITATIVE are two or three standard deviations
VARIABLE. from the mean.
quality adjusted survival analysis Consider the hypothetical example
in controlled clinical trials, a type of overleaf, which depicts mood change for
SURVIVAL ANALYSIS that not only pre- a group of 25 individuals undergoing 12
dicts time to an event (e.g., number of weeks of psychotherapy for depression.
283
quantification
UCL
baseline
LCL
10 15 20
Participant
quality control angle chart
measurements of reaction time. The goal quartiles: (Q3 + Qi - 2Q2)/(Q3 - Q,). The
of gathering this QUANTITATIVE DATA is value is negative (or positive) if few val-
to understand, describe, and predict the ues are at the negative (or positive) side
nature of a phenomenon, particularly of the distribution and is zero when the
through the development of models and data distribution is symmetric.
theories. Quantitative research tech-
niques include experiments and surveys. quartile deviation a measure of DIS-
Also called quantitative design; PERSION that is defined as the value half-
quantitative inquiry; quantitative way between the first and third
method; quantitative study. Com- QUARTILES (i.e., half the INTERQUARTILE
RANGE). Also called semi-interquartile
pare QUALITATIVE RESEARCH.
range.
quantitative variable a characteristic
or attribute that can be measured nu- q u a r t i m a x r o t a t i o n in FACTOR ANAL-
YSIS, an ORTHOGONAL ROTATION that
merically using a score obtained from
any of a variety of data sources. Exam- maximizes the variance across the rows
ples are age, height, and weight. Com- of the factor matrix by raising the load-
pare QUALITATIVE VARIABLE. ings to the fourth power; the effect is to
make large loadings especially large and
quartile n. one of the three values in a small loadings especially small. The ob-
series of values that divide it into equal- jective is to increase the interpretability
sized fourths. For example, the first (or of a factor solution by satisfying the SIM-
lower) quartile of a distribution is the PLE STRUCTURE ideal.
data value below which are the lowest
25% of scores, the second quartile is the quasi-experimental control group
data value below which are 50% of in a QUASI-EXPERIMENTAL DESIGN, any
scores, and the third (or upper) quartile group of participants who are assigned
is the data value below which are 75% of to a condition in which the treatment is
scores (or, conversely, above which are not introduced. These individuals will
25% of scores). These values provide in- receive the standard of care, receive a
formation to researchers about the rela- placebo, or participate in an activity that
tive spread of the distribution. See Q; involves them in the srady but does not
include the treatment in any way.
QUANTILE.
quartile coefficient of dispersion quasi-experimental design an ex-
a descriptive index that allows research- perimental design in which assignment
ers to compare the spread of two data of participants to an EXPERIMENTAL
distributions. It is computed by finding GROUP or to a CONTROL GROUP cannot
the difference between the first and be made at random for either practical or
third QUARTILES for each of the distribu- ethical reasons; this is usually the case in
tions and then comparing the values in FIELD RESEARCH. Assignment of partici-
a ratio: (Q, - Q3)/(Q] - Q3). For exam- pants to conditions is usually based on
ple, if the quartile coefficient of disper- self-selection (e.g., employees who have
sion is 10 for the first set and 4 for the chosen to work at a particular plant) or
second set, then the quartUe coefficient selection by an administrator (e.g., chil-
of dispersion is 2.5 times as great for the dren are assigned to particular class-
first set as for the second set (10/4 = 2.5). rooms by a superintendent of schools).
Such designs introduce a set of assump-
quartile coefficient o f skewness a tions or threats to INTERNAL VALIDITY
descriptive index of SKEWNESS in a data that must be acknowledged by the re-
set that is a ratio between a function of searcher when interpreting srady find-
the first, second, and third quartiles over ings. A study using this design is called
the difference between thefirstand third a quasi-experiment. Examples include
285
quasi-experimental research
interaction between the respondent and dividuals than are needed for a srady. It
a computer or website. also allows a researcher to obtain a de-
quick-and-dirty adj. describing a re- sired balance of sample sizes across
search design or data analysis that is ad- groups for statistical testing. Most com-
mitted to be informal and imperfect. puterized surveys include an automatic
The researcher who adopts such an ap- quota control function. See also QUOTA
proach recognizes that there is a more SAMPLING.
refined way to set up the experiment or quota s a m p l i n g a method of forming
to analyze the findings. In research, a a sample in which a prespecified number
quick-and-dirty approach is more infor- of individuals with specific background
mal than a PILOT STUDY; in data analy- characteristics, such as a particular age,
sis, it is regarded as a first peek at data race, ethnicity, sex, or education, are se-
from a specific analysis. lected for inclusion. Often, participants
q u i n t i l e n. one of four values in a score are recruited as they arrive; once the
distribution that divides it into five quota for a given demographic group is
equal parts. For example, the first filled, the researcher stops recraiting
quintile of a distribution is the value subjects from that group. A researcher
below which are the lowest 20% of who uses this approach can obtain a
scores. See QUANTILE. final srady sample that has the same
proportional characteristics as the target
quota c o n t r o l in survey methodology, population, enabling statistical testing
an approach that imposes a limit on the to be performed on a subset of cases that
number of respondents that are ob- is appropriately representative of the
tained either in the total sample or in larger group of interest. See DELIBERATE
substantively meaningful subgroups, SAMPLING. See also QUOTA CONTROL.
such as gender or ethnicity. The ap-
proach is efficient because it does not quotient n. the value that is the result of
overase resources, such as respondent dividing one number or quantity by an-
time or good will, by obtaining more in- other.
287
Rr
r symbol for SAMPLE CORRELATION COEF- for each srady involved in the meta-
FICIENT, which is typically in the form of analysis. If an effect is large and the stan-
a PRODUCT-MOMENT CORRELATION CO- dard error is small, the point for that
EFFICIENT. study will appear in the upper right cor-
ner of the radial plot. A wide range of
r^ symbol for COEFFICIENT OF DETERMI-
points indicates that effect sizes and
NATION.
standard errors vaty across studies (i.e.,
R 1 . abbreviation for RESPONSE or respon- there is heterogeneity), which could
dent. 2. symbol for MULTIPLE CORRELA- limit conclusions regarding the average
TION COEFFICIENT. effect size in the meta-analysis and sug-
gests that variables potentially causing
symbol for COEFFICIENT OF MULTIPLE
this heterogeneity should be explored.
DETERMINATION.
Also called Galbraith plot.
r a d i a l plot a SCATTERPLOT used in
META-ANALYSIS to assess Whether there r a d i a n (rad) n. a unit of angular mea-
is HETEROGENEITY In the findings from sure. It is calculated as the length of an
different studies, which could limit in- arc divided by its radius. 1 rad = 180/jt.
terpretation of results. It depicts an EF- R^adj symbol for ADJUSTED R^.
FECT SIZE divided by its STANDARD
ERROR on the vertical y-axis and the r a n d o m adj. 1. without order or predict-
ratio of 1 divided by the standard error ability. 2. determined by chance alone,
on the horizontal x-axis, as in the hypo- as in RANDOM SAMPLING or a RANDOM
ERROR.
thetical example below.
Each point represents the strength of r a n d o m assignment in experimental
the effect, relative to the standard error design, the assignment of participants or
95% confidence
limit
N
95% confidence
w 0 limit
U
T" T"
2 4 6
Precision (1/S)
radial plot
288
random error
289
random error variance
ment. Such errors are generally assumed tions components of VARIANCE at sev-
to form a NORMAL DISTRIBUTION around a eral levels. The similar random coeffi-
TRUE SCORE. AIso Called unbiased error; cient model (or random slope model)
unreUabiUty; unsystematic error; allows each data cluster to have its own
variable error. See also ABSOLUTE SLOPE, indicating the explanatory vari-
ERROR; CONSTANT ERROR. Compare SYS- able has a different effect upon each.
TEMATIC ERROR.
r a n d o m i z a t i o n n. see RANDOM AS-
r a n d o m error variance that part of SIGNMENT.
the variability of a test or other score randomization test an inferential ap-
that is not accounted for by other mod- proach that combines observed data
eled factors. It reflects RANDOM ERRORS across all participants and experimental
in measurement arising from the re- conditions and then randomly sorts the
spondent, the situation, evaluator idio- data into new samples. A test of STATIS-
syncrasies, the measure itself, or TICAL SIGNIFICANCE is performed, and
interactions among these factors. Differ- the value obtained is compared with
ent statistical frameworks handle this the value that was obtained when the
type of variation differently; some com- data were in their original form. This
bine it with SYSTEMATIC ERROR VARI- process is repeated many times, theoreti-
ANCE whereas others treat it separately. cally for all possible rearrangements
See ERROR VARIANCE. (permutations) of the data, although the
r a n d o m event an EVENT that is gener- sheer number of possible permutations
ated by a chance process and cannot be generally precludes this and a subset of
predicted from any other event. permutations (10,000 often is recom-
mended) is used instead. For example,
r a n d o m event generator (REG) see suppose there were 1,000 recalculated
RANDOM NUMBER GENERATOR. mean differences between the experi-
r a n d o m f a c t o r see RANDOM EFFECT. mental (E) and control (C) groups, de-
rived from 1,000 permutations. Com-
r a n d o m group design see RANDOM- parison of the 1,000 recalculated means
IZED GROUP DESIGN. to the original means provides an exact
probability of getting the original means.
r a n d o m intercept model a type of If 25% of the mean differences between
HIERARCHICAL LINEAR MODEL USed tO
the E and C group permutations differ
describe the correlations among obser- by as much or more than the difference
vations or scores within a cluster. For ex- between the original E and C means,
ample, a researcher examining the then the probability value for the ori-
average number of work hours and ginal mean difference is p = .25. A ran-
burnout among nurses at several differ- domization test is a NONPARAMETRIC
ent hospitals might use this approach to approach; that is, it does not make as-
account for dependence among certain sumptions about the distribution of the
variables, determining both the overall data. Also called permutation test.
score variation across all nurses and the
variation among the nurses in each of r a n d o m i z e d b l o c k design (RBD) an
the individual hospitals. When plotted, approach to assigning participants to
the data for each nurse grouping have treatment conditions in which mean-
a different INTERCEPT, thus providing ingful discrete strata within the sample
a more accurate and comprehensive de- (e.g., gender, experience) are used to
scription of relationships that corrects identity homogeneous subsamples; in-
for the underestimated STANDARD ER- dividuals from each subsample or
RORS obtained when other models are "block" are then assigned randomly to
applied to nested data and that parti- the different conditions. In this way.
randomized-response technique
291
random model
than the entire set; the subset might be crepency between the low and high
chosen for each participant using a pro- scores, the range is generally perceived
cess determined by chance. as less informative than other measures
of dispersion, such as the STANDARD DE-
r a n d o m slope model see RANDOM IN-
VIATION.
TERCEPT MODEL.
r a n d o m variable a variable that takes range restriction see RESTRICTION OF
on different values according to a chance RANGE.
process. These values cannot be predicted r a n k 1. n. a relative position along an or-
with certainty and are assumed to vaty dered continuum. See RANK ORDER. 2.
across sradies; however, their frequency vb. to arrange items in a graded order,
can be described in terms of probability. such as from highest to lowest value. In a
Also called stochastic variable. Com- peer nomination study, for example, a
pare FIXED VARIABLE. child might be asked to order individu-
r a n d o m v a r i a t i o n differences in a DE- als in a class from most disruptive to least
PENDENT VARIABLE that are due to disruptive. See also ORDINAL DATA. 3. n.
chance, rather than to the factors being the maximum number of linearly inde-
studied. Causes of random variation in pendent row vectors or column vectors
test results may include respondent in a CORRELATION MATRIX: These values
factors, such as health, motivation, at- are always equal. For example, a 10 x 10
tention, concentration, and fatigue; sit- correlation matrix of personality scores
uational factors, such as room tempera- might have a rank of 6, indicating that
ture, noise, and working environment; there are not 10 independent pieces of
or respondent-by-siraation factors, such information present in the 10 scores.
as a respondent not being prepared for r a n k biserial correlation coeffi-
the specific rating task. Researchers try cient an index of association between a
to estimate the extent to which these DICHOTOMOUS VARIABLE and an ORDI-
factors may be involved in the srady (see NAL VARIABLE. Its interpretation is the
RANDOM ERROR VARIANCE) to under-
same as for other standardized measures
stand the true impact of the factors of association. For example, a researcher
being assessed. See also CHANCE DIFFER- might relate experimental condition
ENCE.
(experimental vs. control group) to an
r a n d o m w a l k a series of values plotted ordinal measure of task performance.
over time that reflects the workings of a Compare POINT BISERIAL CORRELATION
chance process, such that each value has COEFFICIENT.
an unpredictable relation to the preced-
r a n k correlation coefficient a nu-
ing value and the series has no definable
merical index reflecting the degree of re-
pattern. The concept is used in many re-
lationship between two variables that
search settings, including simulation
have each been arranged in ascending
sradies and models of price movements
or descending order of magnitude (i.e.,
on the stock market. Compare MARKOV
ranked). It does not reflect the associa-
CHAIN. See also DRUNKARD'S WALK; STO-
tion between the actual values of the
CHASTIC
variables but rather that between their
range n. a measure of DISPERSION ob- relative position in the distribution. For
tained by subtracting the lowest score example, placement in a marathon race
from the highest score in a distribution. could be correlated with the runners'
For example, if the highest score on heights but in this case the two vari-
a test is 100 and the lowest score is ablesrace outcome and heightwould
10, then the range is (100 - 10) = 90 take the form first place, second place,
points. Because it describes a raw dis- and so on; and tallest, next tallest, and
293
rank-flifference correlation
295
rational number
the mean vector, and a CONFIDENCE IN- reactivity n. the condition in which a
TERVAL around the mean angle. [Lord participant being observed is changed
Rayleigh] in some way by the act of observation.
Within an experimental setting reactiv-
r^ symbol for BISERIAL CORRELATION CO- ity is viewed as a threat to INTERNAL VA-
EFFICIENT.
LIDITY because the change in behavior is
RBD abbreviation for RANDOMIZED not due to the experimental manipula-
BLOCK DESIGN. tion. See also REACTIVE MEASURE.
'bis symbol for BISERIAL CORRELATION reactivity effect any of several specific
COEFFICIENT. ways in which a target's performance
may change as a result of being observed
Rc symbol for CANONICAL CORRELATION
or receiving increased attention within
COEFFICIENT, A further subscript may be
a research setting. Examples include EX-
added to show which canonical variates
PERIMENTER EFFECTS, the HAWTHORNE
are being correlated. EFFECT, the PYGMALION EFFECT, and
R correlation see MULTIPLE CORRELA- SELF-FULFILLING PROPHECY effects.
TION COEFFICIENT.
real l i m i t the lower or upper bound of a
RCT abbreviation for RANDOMIZED CLIN- given value for a continuous variable
ICAL TRIAL. measured on a RATIO SCALE. For exam-
ple, a test score of 95 has the lower real
R D D 1. abbreviation for RANDOM-DIGIT
limit of 94.5 and the upper real limit of
DIALING. 2. abbreviation for REGRES-
95.5 since any value within that range
SION-DISCONTINUITY DESIGN.
will equal 95 when rounded to a whole
reaction time (RT) the time that number.
elapses between the onset or presenta-
real-world setting conditions for data
tion of a stimulus and occurrence of a
collection that closely resemble condi-
specific response to that stimulus. There
tions in the natural flow of life (e.g., ob-
are severaltypes,including simple reac-
serving workers at their workplace rather
tion time (single stimulus and single re-
than having them report about what oc-
sponse) and choice reaction time (two or
curs at work). A srady conducted in a
more stimuU, each with, a different re-
real-world setting is often viewed as
sponse). Reaction time can be used to as-
more ecologically valid and gen-
sess various psychological constructs. To
eralizable than one conducted in a labo-
assess negative affect, for example, a re-
ratoty setting.
searcher might measure the time be-
tween presentation of various words recall bias the type of BIAS that often
with emotional connotations and a par- occurs when an individual reports about
ticipant's indication that the word was a past behavior or event. Although such
either "positive" or "negative." Also retrospective reporting may have accu-
called response latency; response rate fearares, it also tends to include in-
time. accurately remembered aspects, such as
a systematic undercount or overcount of
reactive measure a measure that alters
the ftequency with which a certain be-
the response under investigation. For ex-
havior occurred. This type of distortion
ample, if participants are aware of being
is discussed in the literatures associated
observed, their reactions may be influ-
with survey methodology and eyewit-
enced more by the observer and the fact
ness testimony.
of being observed than by the stimulus
object or situation to which they are os- recall data responses to questions that
tensibly responding. Compare UNOB- ask what a participant can remember
TRUSIVE MEASURE. about something. For example, a partici-
297
receiver-operating characteristic curve
pant may be asked to read a passage and l/x and that of 1/x is x. So, for example,
later list specific features of it from mem- 1/4 is the reciprocal of 4.
ory. Another type of memory task might
involve asking the participant whether reciprocal relationship 1. a correla-
or not specific items have been pre- tion between two variables such that the
sented before (the recognition method). value on one variable is the RECIPROCAL
of the value on the other. For example, if
receiver-operating characteristic a researcher is studying the average time
curve (ROC curve) in a detection, dis- taken to complete a task, then tasks
crimination, or recognition task, the re- completed per unit time (e.g., two per
lationship between the proportion of hour) has a reciprocal relationship with
correct "yes" responses (hit rate) and the unit time taken per task (0.5 hours). 2.
proportion of incorrect "yes" responses the situation in which two variables can
(false alarm rate). This is plotted on a mutually influence one another, that is,
graph to show an individual's sensitivity each can be both a cause and an effect.
on the particular task: The axes are hit
and false alarm rates, points are marked reciprocal transformation a TRANS-
to denote the different rates obtained FORMATION of raw data that involves (a)
under different condifions, and the points replacing the original data units with
are connected to form a smooth area (see their RECIPROCALS and (b) analyzing
iUustration below). the modified data. It can be used with
For example, an ROC curve may be used nonzero data and is commonly used
to indicate how well a person detects a when distributions have SKEWNESS or
specific tone in the presence of noise. A clear OUTLIERS. Unlike other transfor-
single quantitative INDEX of performance mations, a reciprocal transformation
may be calculatedfromthe curve as well. changes the order of the original data.
Also called isosensitivity function. For example, if the original unit of a
study variable is time, a researcher might
reciprocal n. the number that when transform the raw data to produce an
multiplied by another number gives a re- analysis of rate. Also called inverse
sult of 1. The reciprocal of x is therefore transformation.
1 1 1 r
0.0 0.2 0.4 0.6 0.8 1.0
False Alarm Rate
298
redundancy coefficient
299
reference axis
300
regression-discontinuity design
score: This could then be used as the out- with a unit change in a specific
come variable in a regression analysis. independent (predictor) variable on the
dependent (outcome) variable, given
regression analysis any of several sta- the relationship of that predictor to
tistical techniques that are used to de- other independent variables already in
scribe, explain, or predict (or all three) the model. This value may be standard-
the variance of an outcome or DEPEND- ized (see STANDARDIZATION) with a vari-
ENT VARIABLE using scores on one or ance equal to 1 (in which case it is called
more predictor or INDEPENDENT VARI- a BETA COEFFICIENT), or it may be un-
ABLES. Regression analysis is a subset of standardized and expressed in the units
the GENERAL LINEAR MODEL. It yields a of the outcome variable being measured
REGRESSION EQUATION as well as an (in which case it is called a B coefficient).
index of the relationship (R or r) be- Also called regression weight. See
tween the dependent and independent also PARTIAL REGRESSION COEFFICIENT.
variables. In addition, the regression
weights obtained for the various inde- regression constant the value of a re-
pendent variables provide information sponse or DEPENDENT VARIABLE in a RE-
about their relative predictive contribu- GRESSION EQUATION when its associated
tion to the outcome. For example, a re- predictor or INDEPENDENT VARIABLES
gression analysis could show the extents equal zero (i.e., are at baseline levels).
to which first-year grades in college (out- Graphically, this is equivalent to the Y-
come) are predicted by such factors as INTERCEPT, or the point at which the RE-
standardized test scores, courses taken in GRESSION LINE crosses the y-axis.
high school, letters of recommendation,
and particular extracurricular activities. regression curve see REGRESSION LINE.
Also called slope analysis. regression diagnostics a set of graph-
regression a r t i f a c t an experimental ical and numerical techniques routinely
finding that has been distorted by ex- used by researchers to check for VIOLA-
treme measurements and the associated TIONS OF ASSUMPTIONS in the applica-
influence of REGRESSION TOWARD THE
tion of REGRESSION ANALYSIS tO
MEAN. particular data sets. For example, one as-
sumes that the relationship between the
regression c a l i b r a t i o n 1. an ap- INDEPENDENT VARIABLES and the DE-
proach in which a researcher uses spe- PENDENT VARIABLE is linear, that the
cific values of a DEPENDENT VARIABLE to variables have been measured accu-
determine the associated values of an IN- rately, and that any prediction errors re-
DEPENDENT VARIABLE when the rela- sulting from the REGRESSION EQUATION
tionship between the two variables is are independent and normaUy distrib-
already known. For example, a re- uted with equal variance and a mean of
searcher sradying university faculty sala- zero. If the data do not possess such
ries might choose a very high salaty characteristics, the analysis may not be
(dependent variable) to see how many appropriate and thus its results may not
years it takes for a new faculty member be valid. See DIAGNOSTICS; RESIDUAL
at that instiration to achieve such a rate ANALYSIS.
of pay (independent variable). 2. a
method of adjusting POINT ESTIMATES regression-discontinuity design
and INTERVAL ESTIMATES for COEFFI- (RDD) a type of QUASI-EXPERIMENTAL
CIENTS in a prediction model to account DESIGN in which a specific threshold
for known MEASUREMENT ERROR. value or CUTOFF SCORE is used to assign
participants to treatment conditions.
regression coefficient in a REGRES- Theoretically, individuals near the
SION ANALYSIS, the WEIGHT associated threshold value are comparable and
301
regression effect
only differ on the basis of their treat- called regression formula; regres-
ment assignment, which enables a re- sion model.
searcher to estimate treatment effects.
For example, a researcher might use a re- regression estimate an EXPECTED
gression-discontinuity design to investi- VALUE for an outcome or DEPENDENT
gate worker performance, assigning VARIABLE that is calculated from a RE-
employees who work more than a cer- GRESSION EQUATION. For a typical uni-
tain number of hours to receive a reward variate linear model, in which multiple
while those below that threshold num- INDEPENDENT VARIABLES predict a single
ber do not. The analysis of such a design outcome, this value is obtained using
involves examining the REGRESSION LEAST SQUARES REGRESSION.
LINES for those receiving the treatment
(e.g., receiving a reward) versus those regression f o r m u l a see REGRESSION
not receiving the treatment (e.g., no re- EQUATION.
ward). A continuous straight line for the
two groups indicates no effect of reward regression i m p u t a t i o n see IMPUTA-
on performance, whereas any break or TION.
jump (discontinuity) in the line across
the groups indicates a treatment effect. regression line a straight or curved line
fitting a set of data points and usually
regression effect see REGRESSION TO- obtained by a LEAST SQUARES REGRESSION.
WARD THE MEAN. A graphic representation of the REGRES-
SION EQUATION expressing the hypothe-
regression equation the mathematical sized relation between an outcome or
expression of the relationship between DEPENDENT VARIABLE and One or more
a dependent (outcome or response) vari- predictors or INDEPENDENT VARIABLES, a
able and one or more independent (pre- regression line summarizes how well the
dictor) variables that results from proposed model acraally fits the sample
conducting a REGRESSION ANALYSIS. It data obtained.
often takes the form y = a + bx + e, in Data points that do not fall exactly on
which y is the dependent variable, x the line indicate deviations in model fit,
is the independent variable, a is the IN- as in the hypothetical example below
TERCEPT, b is the REGRESSION COEFFI- showing parent ratings of aggressive be-
CIENT, and e is the ERROR TERM. Also havior in their children as a function of
15-
o
regression line
O)
O)
< 5-
n\II\\ I I I I
0 1 2 3 4 5 6 7 8 9 10
Amount of TV Violence
regression line
regression tree
303
regression weight
eluded in or excluded from the srady to the measured value. For example, if
sample depending on whether they fall the value of a particular characteristic in
within a particular range. Also called ac- the population is known to be .7 but a
ceptance-rejection method; rejec- recently developed statistical procedure
tion sampling. estimated a value of .5, the relative error
rejection region see CRITICAL REGION. would be (.7- .5)/.7 = .2/.7 = .286. A rela-
tive error may be converted to a percent
rejection value see CRITICAL VALUE. error by multiplying by 100thus, for
related-measures design see WITHIN- the example given, the percent error is
SUBJECTS DESIGN. .286 X 100, or 28.6%.
related samples see DEPENDENT SAM- relative frequency the frequency of a
PLES. type or categoty of event expressed as a
proportion of the total frequency of all
related-samples design see WITHIN-
types or categories. For example, if 47 out
SUBJECTS DESIGN.
of 100 participants answered "yes" to a
related-samples t test see DEPEN- particular question on a survey, the rela-
DENT-SAMPLES T TEST. tive frequency of "yes" responses would
equal .47 (i.e., the number of "yes" re-
r e l a t i o n a l research see CORRE-
sponses divided by the total number
LATIONAL RESEARCH.
of "yes," "no," and "I don't know" re-
relationship n. an association or con- sponses).
nection between objects, events, vari-
ables, or other phenomena. Research relative frequency d i s t r i b u t i o n a
often involves the srady of associations tabular display of the number of obser-
between and among variables. See also vations at each level of a variable com-
CORRELATION. pared to the total number of observations
obtained. For example, a researcher asks
relative efficiency 1. for two tests (A 200 participants to describe their level
and B) of the same hypothesis operating of extraversion on a 5-point scale and
at the same SIGNIFICANCE LEVEL, the obtains the following results: 32 individ-
ratio of the number of cases needed by uals indicated they are extremely intro-
Test A compared to the number needed verted; 24 individuals indicated they are
by Test B for each to have the same sta- somewhat introverted; SO individuals
tistical POWER. The relative efficiency indicated they are neither introverted
value enables a researcher to determine nor extraverted; 38 individuals indi-
whether there is a preferred statistical cated they are somewhat extraverted;
approach for evaluating a particular and 56 individuals indicated they are ex-
phenomenon. For example, when hy- tremely extraverted. These values can be
pothesis testing involves NORMAL DIS- expressed as both proportionsthe
TRIBUTIONS, parametric inferential sta-
number of observations per level di-
tistics have a more favorable ratio than vided by all observations (i.e., .16, .12,
nonparametric inferential statistics. 2. .25, .19, and .28)and as percentages
for two parameter ESTIMATES (A and B), the number of observations per level di-
a value reflecting the ratio of the STAN- vided by all observations multiplied by
DARD ERROR of Estimate A compared to 100 (i.e., 16%, 12%, 25%, 19%, and
the standard error of Estimate B. See REL- 28%). Thus, in the corresponding rela-
ATIVELY EFFICIENT ESTIMATOR.
tive frequency distribution overleaf, the
relative error in measurement, the first column lists the different extra-
ratio of the ABSOLUTE ERROR (i.e., the version levels, the second column lists
positive difference between an exact the number of people at each of the dif-
measured value and the estimated value) ferent extraversion levels, the third col-
305
relative frequency polygon
umn lists each value in the second the horizontal ;t-axis with the interval
column as a proportion of the total, and midpoints from lowest to highest; (c) la-
the fourth column lists each value in the beling the vertical y-axis with the range
second column as a percentage of the of proportional score values; (d) plotting
total. the specific proportion associated with
This type of table is useful in identify- each midpoint; and (e) connecting aU of
ing which scores or values are most the plotted points with lines. A hypo-
likely to occur at which variable level, thetical example is given below.
and it provides an organized display of Also called relative frequency
data that could also be graphed in a REL- curve; relative frequency diagram;
ATIVE FREQUENCY POLYGON. AISO Called relative frequency function; rela-
relative distribution; relative fre- tive frequency graph. See FRE-
quency table. See FREQUENCY DISTRI- QUENCY POLYGON.
BUTION.
relatively efficient estimator given
two ESTIMATORS (approximated popula-
relative frequency p o l y g o n a graph-
tion values) for a given model, the esti-
ical representation of a RELATIVE FRE-
mator that has the smallest variance in
QUENCY DISTRIBUTION, conveying the
its sampling distribution and thus pro-
number of individuals responding in a
vides a more precise estimate of the pa-
particular manner on a given variable as
rameter. See RELATIVE EFFICIENCY.
compared to the total number of people
responding. For a CONTINUOUS VARI- relative measurement a context-
ABLE, the polygon is constructed by (a) specific assessment approach in which
creating intervals of scores and identify- the value of an individual score depends
ing each interval midpoint; (b) labeling on its comparison to other scores within
Extraversion Level
relative frequency polygon
306
reliability diagonal
a group or subgroup. For example, one diction. In ethics, relativism is the claim
might evaluate beauty by assigning a that there are no moral absolutes. Thus,
person a number that reflects his or her judgments of right and wrong are based
attractiveness compared with others of on local culrare and tradition, on per-
the same age, the same occupation, and sonal preferences, or on artificial princi-
so forth. Comparison of scores within ples. Standards of conduct vaty
subgroups can sometimes be useful but enormously across individuals, cultures,
also may produce bias. Compare ABSO- and historical periods, and it is impossi-
LUTE MEASUREMENT. ble to arbitrate among them or to pro-
duce universal ethical principles because
relative risk see RISK RATIO.
there can be no means of knowing that
relative standing the position of a these are trae. In this way, relativism in
particular score in the context of a distri- epistemology and relativism in ethics
bution of scores. There are several ways are related. relativist adj.
to describe relative standing, including
PERCENTILES (e.g., deciles, quartiles), relevance-sensitivity tradeoff in
STANDARDIZED SCORES (e.g., T SCORES, Z research, the balance strack between
SCORES), and staras as an OUTLIER. measuring dependent variables accu-
Graphical approaches such as BOX-AND- rately and specifically and obtaining re-
WHISKER PLOTS and FREQUENCY POLY- sults with applicability to contexts
GONS alSO can provide valuable informa- beyond the original study. The rele-
vance-sensitivity tradeoff should be
tion about relative standing.
considered by a researcher when design-
relative survival a value indicating ing a study.
the number of people with a specific dis-
ease who are alive compared to the num- r e l i a b i l i t y n. the trustworthiness or
ber of surviving people who are disease consistency of a measure, that is, the de-
free but otherwise have the same charac- gree to which a test or other measure-
teristics (age, sex, etc.). Relative survival ment instrument is free of RANDOM
often is calculated for several different ERROR, yielding the same results across
time points, such as one year since a sig- multiple applications to the same sam-
nificant milestone (e.g., diagnosis, treat- ple. See ALTERNATE-FORMS RELIABILITY;
ment, remission), five years since the INTERNAL CONSISTENCY; RELIABILITY
milestone, and 10 years since the mile- COEFFICIENT; RETEST RELIABILITY.
stone. reliability coefficient (symbol: r^x)
relative v a r i a b i l i t y an assessment of an index describing the consistency of
the DISPERSION within a sample, typi- scores across contexts (e.g., different
cally expressed as a COEFFICIENT OF times, items, or raters). Its value, ranging
VARIATION or a DIVERSITY INDEX. from 0 to 1, provides an estimate of the
amount of obtained score variance that
relativism n. any position that chal- is due to TRUE VARIANCE rather than to
lenges the reaUty of absolute standards error. The larger this coefficient, the
of truth or value. In EPISTEMOLOGY, rel- more confident a researcher may be that
ativism is the assertion that there are no scores obtained at different times under
absolute grounds for truth or knowledge similar conditions with the same partici-
claims. Thus, what is considered trae de- pants wUl be alike. Typically, reUabUity
pends solely on individual judgments coefficients are considered to be accept-
and local conditions of culture, reflect- able if they are above .80. Also called co-
ing individual and coUective experience. efficient of reliability; index of
Such relativism challenges the validity reliability; reliability index.
of science except as a catalog of experi-
ence and a basis for ad hoc empirical pre- r e l i a b i l i t y diagonal the MAIN DIAGO-
307
reliability index
308
representational validity
309
representative conclusion
311
research protocol
312
response acquiescence
Variable 1 2 3 4 5 6 7 8 9 10 II 12 13 14 15 16
1. JobSall 0.20
2. JobSal2 1.04 1.89
3. JobSat3 -0.49 0.87 0.33
4. HlConl -0.09 0.56 0.09 0.02
5. HlCon2 -1.81 -0.04 -0.35 6.39 0.01
6. HlConS 0.41 0.54 0.81 -3.00 -2.98 0.02
7. Psych] -1.25 -0.69 -0.51 -8.55 -4,73 -SJ< 0.70
8. Psycii2 -2.30 -1.41 -1.21 -8.46 -4.57 -9.05 0.67 0.35
9. Psych3 -1.38 -0.60 -0.56 -7.93 -^.77 -8.79 0.86 0.34 0.43
10. HelSall 1.50 1.49 1.92 -1.45 -1.79 1.88 -8.19 -8.00 -8.29 0.04
11. HclSat2 0.05 0.22 0.26 -2.82 -0.55 2.10 -6.47 -6.53 -5.93 1.79 0.02
12. HCISBL; 1.72 0.29 0.54 0.51 -2.16 3.77 - 9 J 6 -9.37 -8.62 -0.86 -1.02 0.02
13. JwithI -3.04 -1-31 -3.58 1.44 1.25 1.40 0.45 2.12 0.75 1.01 0.15 -0 27 0.00
14. Jwilh2 -5.54 -5.67 -5.49 -0.16 0.09 -0.20 4.43 4.78 4.54 -1.05 -0 65 -1.00 -0.02 0.00
15. Jwilh3 -5.74 -6.01 -5.66 1.69 0.50 1.09 299 3.96 2.61 0.63 -1.16 -0.44 1.14 -0.88 0.00
16. Wwilhl -O30 1.34 0 02 -3.36 -2.81 -3.93 5.99 6.19 6.15 -5.63 -4.26 -4.80 2.30 1.96 2.33 0.00
17. Wwith2 -2.60 -1 68 -2.20 -3.37 -2.10 -4.71 7.52 7.00 6.70 -4.75 -3.99 -1.51 2.98 4.41 3.51 0.00 0 00
18. Wwilh3 -0.30 1.81 1.23 -4 15 -3.21 -4.37 6.12 6.72 5.99 -5.50 -4.29 -5.15 2.03 2.26 1.76 0.00 -0.40
residual matrix
313
response bias
r e s p o n s e - f e a t u r e a n a l y s i s any of
several methods of evaluating L O N G I T U -
DINAL DATA that involve calculating the
same S U M M A R Y STATISTICS, such as the
MEAN, STANDARD DEVIATION, SLOPE,
and AREA UNDER THE CURVE, for a data
set at different time points.
r e s p o n s e l a t e n c y see R E A C T I O N T I M E .
314
retrospective sampling
315
retrospective validity
comparison to others who were not ex- LEAST SQUARES REGRESSION designed to
posed to the risk or did not complete the remedy problems that arise from
particular process. Compare PROSPEC- MULTICOLLINEARITY. It involves modi-
TIVE SAMPLING. fying the MAIN DIAGONAL Of the CORRE-
LATION MATRIX before calculating
retrospective v a l i d i t y the extent to
the coefficients, thus eliminating the as-
which an instrament that purports to
sociations among the INDEPENDENT
measure a particular behavior or phe-
VARIABLES. A researcher can use the in-
nomenon of interest can be shown to
formation obtained from this analysis to
correlate with past behaviors or occur-
determine whether certain independent
rences that demonstrate this behavior.
variables should be removed from the
For example, a researcher evaluating
final model. Also called damped re-
a new measure of accident proneness
gression.
might administer it to a sample of
respondents and then check the individ-
r i g h t censoring inability of a re-
uals' archived medical records to deter-
searcher to document when all partici-
mine if higher test scores correlate with
pants have reached a target event (e.g.,
the number of actual treated incidents.
achieving a specific milestone, experi-
It is one of several types of CRITERION
encing a relapse) at the conclusion of the
VALIDITY. Also called postdictive va-
study period. This may occur for one of
lidity. See also CONCURRENT VALIDITY;
three reasons: A participant may never
PREDICTIVE VALIDITY. experience the target event or experi-
reversal design an experimental de- ence it after the observation period has
sign, generally used when only a single ended; a participant may experience a
group is being sradied, that attempts competing event that prevents him or
to counteract the confounding effects her from experiencing the target event;
(see CONFOUND) of sequence, order, and or a participant may be lost to ATTRI-
treatment by alternating baseline condi- TION. Compare LEFT CENSORING.
tions with treatment conditions. Exam-
ples include the A-B-A DESIGN, A-B-A-B r i g i d r o t a t i o n see ORTHOGONAL ROTA-
DESIGN, and other similar combina- TION.
tions. See also ALTERNATING TREAT-
MENTS DESIGN. risk level the probability of making a
TYPE I ERROR that One is willing to accept
reverse counterbalancing a specific in null hypothesis SIGNIFICANCE TEST-
procedure for ordering stimulus materi- ING.
als in a research study that involves ad-
ministering' one order (A-B-C) for one risk ratio, the comparison of one
half of the participants and the opposite group's probability of experiencing an
order (C-B-A) for the other half of the event (e.g., being diagnosed with lung
participants. Reverse counterbalancing cancer) to a second group's probability
is used to minimize any potential influ- of experiencing that event. It is often
ence of presentation upon results, so as used to describe health status following
to ensure it is the stimuli themselves exposure to some stimulus (e.g., lead in
that are producing any effect that may water) or clinical intervention. A value
be seen. greater than 1 indicates the group under
study has a higher probability than the
r h o c o r r e l a t i o n (symbol: p) 1. see POP- control group of experiencing the event;
ULATION CORRELATION COEFFICIENT. 2. a value less than 1 indicates the group
see SPEARMAN CORRELATION COEFFI- under srady has a lower probability of
CIENT. experiencing the event; and a value of
ridge regression a variant on ordinaty exactly 1 indicates the two groups are
316
rootogram
317
Rosenbaum's test of unidimensionality
160
o o o o o o o o o o o o o o o o o o o o o
o o o o o o o o o o o o o o o o o o o o
o o o o o o o o o o o o o o o o o o o o
T-" CM'm in CD r^' cd a i o T-" CN in co r^' co'
Debt (dollars)
rootogram
318
run-in period
319
r u n n i n g average
100-1
80-
^ 60-
.a
.2
^ 40-
20-
rugplot
320
Ss
s symbol for SAMPLE VARIANCE. observations in this subset, such as the
mean, median, or STANDARD DEVIATION,
S abbreviation for a SUBJECT (i.e., individ- are called STATISTICS.
ual or case) in a srady. People partaking
in research are now more commonly re- sample correlation coefficient
ferred to as PARTICIPANTS. (symbol: r) an index of the degree of as-
sociation between two variables based
St. Petersburg paradox a paradox on the data in a studied subset (sample)
often cited in PROBABILITY THEORY and of cases from a larger group of interest. It
DECISION THEORY, which shows how a is a variant of the PRODUCT-MOMENT
decision criterion that only takes an EX- CORRELATION COEFFICIENT, such that
PECTED VALUE into account may result
the same symbol is used for both statis-
in a recommended course of action that tics. See also POPULATION CORRELATION
no rational person would follow. The
COEFFICIENT.
paradox involves a theoretical game
called the St. Petersburg Lottety, which sample d i s t r i b u t i o n the DISTRIBU-
is played by flipping a coin until it comes TION of scores in a particular subset
up tails, and the total number offlips,n, (sample) drawn from the wider popula-
determines the prize, which equals $2 x n. tion. Of interest is the general shape of
Thus, if the coin comes up tails the first the distribution, reflecting the fre-
time, the prize is $2 and the game ends. quency of particular scores (its KUR-
If the coin comes up heads the first time, TOSIS, SKEWNESS, etc.). Compared to the
it isflippedagain. If it comes up tails the theoretical POPULATION DISTRIBUTION,
second time, the prize is $4 and the the distribution of scores in a sample is
game ends. If it comes up heads the sec- often jagged and not smooth.
ond time, it isflippedagain, and so forth.
The expected value of the game is the sample mean (symbol: X, M) the arith-
sum of the potential payoffs; since the metic average (MEAN) of a set of scores
payoff of each possible consequence is from cases or observations in a subset
$2/2 = $ 1, and since there are potentially from a wider population. Because each
score contributes equally to this index of
an infinite number of payoffs, the sum is
CENTRAL TENDENCY, it can be affected
an infinite number of dollars. In theoty,
greatly by OUTLIERS.'Many widely used
because the amount one could win is in-
STATISTICAL TESTS are based on the com-
finite, a rational gambler should be will- parison of sample means.
ing to pay any finite amount to play (i.e.,
any price of entry is smaller than the ex- sample of convenience see CONVE-
pected value of the game). In practice, NIENCE SAMPLING.
however, this would clearly not be the
case. sample overlap the situation in which
two or more subsets drawn from a popu-
sample n. a subset of a POPULATION of lation feature items in common (i.e., the
interest that is selected for study with same individual or observation features
the aim of making inferences to the pop- in more than one sample). A researcher
ulation. It is important to ensure that a can design a selection strategy in which
sample is representative of the larger the probability of finding the same indi-
population. Characteristics that describe viduals across two sradies is maximized.
321
sample reliability
thereby reducing costs associated with QUOTA SAMPLING. Each approach has a
data collection. Alternatively, this prob- different potential of obtaining a sample
ability could be minimized, thereby en- appropriately representative of the POP-
suring independence across studies and ULATION under study.
reducing respondent burden.
s a m p l i n g bias a systematic and direc-
sample r e l i a b i l i t y the degree to which tional error involved in the choice of
a SAMPLE is representative of the POPU- units, cases, or participants from a larger
LATION from which it is drawn. It is typi- group into a research study. This is a po-
cally indexed by the STANDARD ERROR tential problem whenever the researcher
OF THE MEAN. has latitude in selecting individual units
for the sample. Selection bias can pose
sample size the number of observations a threat to the INTERNAL VALIDITY of a
(cases, individuals, units) included in study if there is a possibility that pre-
the sample to be sradied. This is usually existing differences arising from the
denoted N (for the study as a whole) or n sampling process may interact with the
(for subgroups from the study). variable of interest. Similarly, if the pro-
sample space the collection of all possi- cedure used to choose participants tends
ble outcomes of an experiment of chance. to favor specially motivated individuals
For example, for a toss of a single coin or people from a certain segment of soci-
the sample space is heads and tails, ety, there would be a threat to the srady's
whereas for a toss of two coins the sam- EXTERNAL VALIDITY (I.e., Inferences to a
ple space is heads-heads, heads-tails, larger population would not be viable).
taUs-tails, and tails-heads. Selection bias is associated with a lack
of RANDOM SAMPLING and with non-
sample standard deviation (sym- random assignment to conditions. Also
bol: 5) see STANDARD DEVIATION. called selection bias; recruitment
bias. See also SELF-SELECTION BIAS.
sample variance (symbol: 5^) the dis-
persion of scores within a group selected sampling design the specific ap-
for study, as opposed to the POPULATION proach, method, or strategy that a re-
VARIANCE. It is calculated by determin- searcher decides to use to select a sample
ing each score's difference from the aver- from the larger population. Formulating
age for the set, squaring and summing a design involves determining the na-
these differences, and then dividing by ture of the target population, a suitable
the total number of scores minus one. SAMPLING FRAME for drawing the cases,
the desired sample size, whether ran-
s a m p l i n g n. the process of selecting a
dom or nonrandom selection will be
limited number of units from a larger set
used, and whether there are any impor-
for a study. The term most often refers
tant variables on which to stratity selec-
to the selection of respondents, obser- tion. Also called sampling plan.
vations, or cases for inclusion in ex-
periments, surveys, interviews, or other s a m p l i n g d i s t r i b u t i o n the distribu-
research. However, sampling can also in- tion of a statistic, such as the mean,
volve selecting theoretical constructs to over repeated samples drawn from a
study, selecting measurement instru- population, SIMULATION studies allow
ments from a broad set of potential op- researchers to specify known population
tions, or selecting time points at which information, conduct a very large num-
to observe individuals or cases. There are ber of repeated draws on the population,
various different methods of selecting and buUd an empirical distribution of
participants for a study, including SIM- the statistic based on these draws (e.g., t,
PLE RANDOM SAMPLING, STRATIFIED F, or distributions). For example, the
SAMPLING, CONVENIENCE SAMPLING, and means calculated from samples of 100
322
sampling variability
observations, repeatedly and randomly cords in light of the inferences that will
drawn from the population, yield a sam- be made: Some relevant questions would
pling distribution for the mean. Knowl- be whether part-time as well as full-time
edge about the distribution of a statistic students are included in the institution's
allows researchers to say when a finding files and whether students currently on
from a sample is unusual (e.g., statisti- leave are included. The key point is that
cally significant) and when it would any sampling frame should be represen-
be expected from the statistic's known tafive of the target population as a whole.
behavior, thus enabUng the sampling
s a m p l i n g i n t e r v a l in SAMPLING DE-
distribution of a statistic to be used in
SIGNS that involve the researcher select-
testing hypotheses about variables and
ing every th case from a list or table,
their relationships. See also INFERENTIAL
the value of n. The chosen interval will
TEST. usually depend on the nature of the
s a m p l i n g error the predictable MAR- SAMPLING FRAME. For example, if the re-
GIN OF ERROR that occurs in sradies that searcher has access to an alphabetized
draw samples of cases or observations list, he or she might want to ensure that
from a larger POPULATION: It indicates all parts of the alphabet are represented
the possible variance between the trae and select a value to maximize that goal.
value of a parameter in the population Alternatively, for a study conducted in
and the estimate of that value made from the field that involves the researcher ap-
the sample data. The larger the sample, proaching individuals in an office build-
the smaller the sampling error (if the ing, the value could be selected based on
entire population was sampled, there the average number of people who are
would be no error in the sample esti- likely to be available and eligible for se-
mate). Large national surveys, such as lection.
those reporting political attitudes, state s a m p l i n g p l a n see SAMPLING DESIGN.
the sampling error along with their find-
ings. For example; a survey finding that s a m p l i n g theory the body of princi-
65% prefer a particular policy with a ples underlying the drawing of SAMPLES
margin of error of 3% means that the that accurately represent the population
true figure could be anywhere between from which they are taken and to which
62% and 68%. inferences wUl be made. This conceptu-
alization of the sampling process pro-
s a m p l i n g f r a m e the specific source vides guidance to researchers about which
used in drawing a subset of cases or indi- SAMPLING DESIGN to choose in the par-
viduals from the larger POPULATION. In ticular circumstances and how best to
many cases the samplingframewill be a account for subsets of cases that are not
complete list of all the elements in a well represented (or are overrepresented)
population (e.g., the electoral register). in the population (e.g., by using WEIGHTS).
In other cases this will be impossible or
s a m p l i n g u n i t any of the elements se-
the issue may be less straightforward
than it appears. In a study of college sra- lected from a population to make up a
dents' aspirations for life after gradua- sample. For instance, if classrooms are
tion, for example, the researchers might selected at random from the population,
decide to use a listing of students then the classroom, not the individual
obtained from the institution's office student, is the sampling unit. See UNIT
of evaluation and assessment and ran- OFANALYSIS.
domly select from this list evety fifth s a m p l i n g v a r i a b i l i t y the extent to
student. In this example, however, the which the value of a statistic differs across
researcher would need to be assured of a series of samples, such that there is
the completeness of the institution's re- some degree of uncertainty involved in
323
sampling without replacement
making inferences to the larger POPULA- nitude and not direction. Compare VEC-
TION. See SAMPLING ERROR. TOR. 2. n. in MATRIX ALGEBRA, a quan-
tity that can multiply a vector in a vector
s a m p l i n g w i t h o u t replacement a
space to produce another vector. 3. adj.
SAMPLING technique that involves se-
describing a variable that can be repre-
lecting an item from the larger set and
sented by positions on a scale. 4. adj. de-
removing it from the general pool; thus,
scribing a matrix in which the entries in
this particular case cannot be redrawn.
the MAIN DIAGONAL are equal and all
Compare SAMPLING WITH REPLACEMENT.
other entries are zero.
s a m p l i n g w i t h replacement a SAM-
PLING technique in which each item se- scalar analysis the evaluation of data
lected from the larger set is returned to obtained from a test or measure using an
the general pool so that it may subse- ITEM RESPONSE THEORY model, often a
quently be redrawn. This means that a RASCH MODEL.
particular case may be drawn more than scale n. 1. a system for ordering test re-
once for a given sample. Compare SAM- sponses in a progressive series, so as to
PLING WITHOUT REPLACEMENT. measure a trait, ability, attirade, or the
saturated model a model that fits the like. For example, an agreement scale
data perfectly because it has as many es- used on an attitude survey might have
timated parameters as there are values to seven response options ranging from
be fitted. In STRUCTURAL EQUATION strongly disagree (1) to strongly agree (7),
MODELING this type of model is called a with neither disagree nor agree (4) as the
JUST-IDENTIFIED MODEL: It poses a prob- middle point. See LIKERT SCALE; RATING
lem as a researcher cannot evaluate fit, SCALE. 2. a sequence of ordered values
as would be the case with the preferred used as a reference in measuring a physi-
OVERIDENTIFIED MODEL (or oversatu- cal property (e.g., weight, temperature).
rated model), in which there are fewer See I N T E R V A L S C A L E ; R A T I O S C A L E . See
estimated parameters than DEGREES OF also MEASUREMENT LEVEL. 3. more gen-
FREEDOM. Relatedly, the underidentified erally, any test or other assessment in-
model (or unsaturated model) has more strament as a whole.
parameters than can be estimated from scale attenuation the situation in
the available data. In FACTOR ANALYSIS, which the response format on a measure
the problem of saturation occurs when includes too few options to reflect a re-
the number of factors that will be esti- spondent's actual behavior, opinion, or
mated is the same as the number of vari- belief. For example, if an item asks about
ables in the data set. average time spent watching television
saturated test in EXPLORATORY FAC- each day and a scale ranging from none
TOR ANALYSIS, a scale that is highly cor- (1) to 7 hour or more (5) is used, the upper
related with one of the factors being end of the scale is likely to be selected by
measured. Such a test helps define the the great majority of respondents, some
meaning of the factor. of whom will watch many hours per
day; this results in little ability to differ-
saturation n. see SATURATED MODEL. entiate the sample in terms of how
scalability n. the ability of an item on a many hours are actually watched. See
test or scale to elicit responses that repre- CEILING EFFECT; FLOOR EFFECT.
sent identifiable positions on an ordered
scale development the process of cre-
progression of scores or values describ-
ating a new INSTRUMENT for measuring
ing an underlying constract, such as an
an unobserved or latent constract, such
ability or attitude. See GUTTMAN SCALE.
as depression, sociability, or fourth-
scalar 1. n. a quantity having only mag- grade mathematics ability. The process
324
scatterplot matrix
includes defining the construct and test tendency for a set of scores for an indi-
specifications, generating items and RE- vidual to vaty from the average of scores.
SPONSE SCALES, piloting the items in a
scatter analysis 1. a study of the rela-
large sample, conducting analyses to
tionship between two variables carried
fine-tune the measure, and then re-
out using a SCATTERPLOT. 2. any study
administering the refined measure to de-
of SCATTER, especially in intelligence
velop NORMS (if applicable) and to assess
tesfing and PROFILE ANALYSIS.
aspects of RELIABILITY and VALIDITY.
scatterplot n. a graphical representa-
scale homogeneity see INTERNAL CON- tion of the relationship between two
SISTENCY. continuously measured variables in
which one variable is arrayed on each
scale parameter see DISPERSION PA- axis and a dot or other symbol is placed
RAMETER.
at each point where the values of the
scale value a number that represents an variables intersect. For instance, the hy-
individual's position on an underlying pothetical example below depicts the
construct (trait, ability, attitude, etc.), as weight and average daily calorie con-
measured by an item on a SCALE or the sumption for a sample of individuals.
scale as a whole. The overall pattern of dots provides an
indication of the extent to which there
scaling n. the process of constructing an is a LINEAR RELATIONSHIP between vari-
instrament to measure, assess, and order ables. A round mass of points shows no
some quantity or characteristic (e.g., linear relation, an elliptical cloud of
height, weight, happiness, empathy). A points with a positive slope shows a pos-
researcher must evaluate how a given itive relation, and an eUiptical cloud
CONSTRUCT of interest should be mea- with a negative slope shows a negative
sured and how optimally to obtain scores relation, OUTLIER points are also clearly
for individuals on these measures. See visible. Also called dot plot; scatter-
SCALE DEVELOPMENT. gram. See also REGRESSION LINE.
scalogram n. see GUTTMAN SCALE. scatterplot m a t r i x a matrix in which
all variables being studied are arrayed as
scatter n. 1. the tendency of data points both rows and columns, so that each
to diverge from each other. An example cell reflects the relationship between
is the variation in scores across a set of two variables and the values along the
tests on the same individual; another is MAIN DIAGONAL reflect the distribution
the variation among test takers on a par- of each variable. For example, the
ticular test. See also DISPERSION; VARI- scatterplot matrix overleaf shows re-
ANCE. 2. in PROFILE ANALYSIS, the spondent age, respondent gender, and
3,500 1
3,000
a re 2,500
2,000
I " 1,500
1,000
500
0 I 1 1 1 I
325
scedasticity
326
screening sample
327
screening test
business owners over 60 years of age. See ently sized rocks (screes) that form at the
also SCREENING TEST. foot of a mountain. Also caUed Cat-
tell's scree test.
screening test any testing procedure
designed to separate out people or items SD symbol for STANDARD DEVIATION.
with a given characteristic or property.
Screening tests aretypicallyused to dis- SDT abbreviation for SIGNAL DETECTION
tinguish people who have a disease, dis- THEORY.
order, or predisease conditionfromthose SE symbol for STANDARD ERROR.
who do not; they may be used, for exam-
ple, in primary health care settings at seasonal adjustment in TIME-SERIES
intake to identity people who are de- ANALYSIS, the removal of that compo-
pressed and need further clinical atten- nent of the variance that is associated
tion. Screening tests are designed to be with a systematic pattern occurring at
broadly sensitive, and subsequent highly regular intervals over time. For example,
specific or focused testing is often re- if highway accidents occur about 5%
quired to confirm the results. They are more in the winter than in the summer
often designed to be brief to facilitate months, winter figures would have to be
broad classifications. Also called screener; seasonally adjusted to give a true in-
screening study. crease that may be caused by other fac-
tors. See SEASONAL COMPONENT.
scree plot in EXPLORATORY FACTOR
ANALYSIS, a graphic that shows (in de- seasonal component in TIME-SERIES
scending order) the EIGENVALUES of DATA, that element of the variance that
the coMMUNALiTY-adjusted correlation can be accounted for by patterns reoc-
matrix. Consider the generic example curring regularly over time (e.g., certain
below. months of the year, days of the week,
Researchers examine the plot to iden- times of day). Usually the variance asso-
tify a "break" in the curve between ciated with this component can be de-
strong, dominant factors at the top and scribed and modeled. For example,
other smaller factors at the bottom: In researchers studying substance use in
this way they can determine which fac- college sradents might need to account
tors to retain in the analysis. The plot for the days in the week (Thursday, Fri-
was named by Raymond B. Cattell, in day, and Sararday) that are associated
analogy with the sloping heaps of differ- with most use.
scree plot
selection
329
selection bias
ing, classifying, or working (hiring em- ple of data that could be appropriate for
ployees). this method is women's salaries: Some
aspects of the data could be modeled
selection bias see SAMPLING BIAS.
using standard regression methods, but
selection invariance in choosing the preponderance of zero salaries (re-
among applicants for, say, employment flecting a decision by some women not
or admission to college, the property of a to work) would also need to be incorpo-
selection procedure such that it is rated into the analysis. In such cases the
equally efficient (i.e., makes a similar two-stage regression method uses a model
number of errors) for all subgroups of that relies on BIVARIATE NORMALITY to
applicants (e.g., ethnic or income estimate the selection equation and an
groups). Selection invariance is the em- ordinaty LEAST SQUARES REGRESSION to
pirically testable assumption that there estimate the remainder. See also TRUN-
is equal SENSITIVITY and equal SPECIFIC- CATED DISTRIBUTION; ZERO-INFLATED
ITY across all groups. Compare MEA- POISSON REGRESSION.
SUREMENT INVARIANCE.
selection ratio the proportion of all
selection methods i n regression in those eligible to be selected for a purpose
t h e G E N E R A L LINEAR M O D E L , v a r i O U S SCtS who actually are selected. In personnel
of decision rules that help researchers selection, for example, it is the number
determine which predictors or INDE- of applicants hired to perform a job di-
PENDENT VARIABLES are related to the vided by the total number of applicants.
outcome or DEPENDENT VARIABLE. Some The lower the selection ratio, the more
methods are theoty driven and involve competitive the hiring situation will be
entering variables into the model in a and the more useful, all other factors
preconceived order (i.e., HIERARCHICAL being held constant, any given predictor
REGRESSION). Other approaches apply a will be in making selection decisions.
rule (usually based on STATISTICAL SIG- selection threat the THREAT TO VALID-
NIFICANCE) for the inclusion of a vari-
ITY arising from any form of SAMPLING
able into the model or its exclusion from BIAS (i.e., nonrandomly selecting units
the model (i.e., STEPWISE REGRESSION). for a study).
For example, FORWARD-SELECTION ap-
proaches begin with no variables in the selective dropout the nonrandom loss
model and proceed by including the of participants from a srady that occurs
variable (or variables) with the strongest when an identified fearare of the study
association with the outcome until there design (e.g., topic sradied, number of
is no increase in EXPLAINED VARIANCE tasks) interacts with respondent char-
by the addition of remaining variables. acteristics (e.g., depression, education
Conversely, BACKWARD-ELIMINATION ap- level). See ATTRITION.
proaches begin with all possible vari-
ables in the model and proceed by suc- selective observation a process in
cessive deletion of the variable (or which an individual attends to behav-
variables) that contribute least to the iors, attitudes, and interpersonal set-
prediction or explanation of the out- tings that correspond to his or her
come. current beliefs or self-interests. When
listening to a speech by a presidential
selection model a two-stage REGRES- candidate, for example, a person with
SION ANALYSIS used to estimate PARAME- strong views on national spending for
TERS and STANDARD ERRORS in an social programs may listen for aspects of
. unbiased way when scores for one aspect the speech that address those specific be-
of the sample may be affected by a liefs. Compare UNSELECTIVE OBSERVA-
nonrandom selection process. An exam- TION. See also CONFIRMATION BIAS.
330
semantic differential
331
semi-interquartile range
is one of the most widely used methods probability that a test gives a positive
of assessing attitudes and may be used in diagnosis given that the individual actu-
psychometric testing or in a wide array ally has the condition for which he or
of settings, such as marketing and poli- she is being tested. Compare SPECIFIC-
tics, to gauge public reactions to a prod- ITY.
uct, issue, or personality.
sensitivity analysis an analysis that
semi-interquartile range see QUAR- measures the extent to which the overall
TILE DEVIATION. outcome of a model or system will be af-
fected by potential changes to the input.
semi-Markov process a STOCHASTIC This type of analysis is often used where
PROCESS with a finite set of states in the values of key variables are uncertain
which the "jump" (transition) from each or subject to change: In financial plan-
state to the next is memoryless but the ning, for example, it might be used to
process as a whole is not. determine whether a potential change
semiparametric model a model that in mortgage rates, cash flow, local house
combines a PARAMETRIC component that prices, or all of these might alter the
has strict assumptions relating to the decision to purchase a new home. In re-
distribution of the variables and a NON- search studies, sensitivity analysis en-
PARAMETRIC component that does not ables researchers to understand the
rely on distribution parameters. An ex- boundaries of their statistical models
ample of this flexible hybrid model is and design updated models that can ac-
the Cox proportional hazards model (see count for the data at hand.
cox REGRESSION ANALYSIS). Sen's slope estimator see THEIL-SEN
semipartial correlation see PART ESTIMATOR.
CORRELATION. separate-variances t test a variant
semistructured interview an inter- o n t h e I N D E P E N D E N T - S A M P L E S T TEST
view format that involves the inter- used where the samples being compared
viewer asking a specified set of questions have differing VARIANCES. It involves
making a special adjustment to the DE-
in a particular order, while also allowing
GREES OF FREEDOM.
for open-ended responses by the inter-
viewee and a more natural conversa- sequence effect in WITHIN-SUBJECTS DE-
fional style. In a study of health SIGNS, a difference in scores that emerges
professionals, for example, interviewers because of a particular arrangement of
might follow a guide that includes sev- treatments; that is, the presentation of
eral open-ended questions in which the one level of the independent variable
professionals are asked to reflect on their has an effect on responses to another
patients' use of their medications. By level of that variable. A researcher can
using prompts such as "Tell me more" test for a sequence effect by administer-
and "Why is that?", the interviewers ing the treatments in various different
would encourage a full and rich re- arrangements (e.g., the arrangement ABC
sponse. This approach recognizes that a VS. ACB, vs. BCA, etc.). The sequence ef-
researcher may still learn valuable infor- fect is not to be confused with the
mation even when the interviewee ORDER EFFECT.
moves away a little from the topic at
hand. See also STRUCTURED INTERVIEW. sequential analysis a class of statisti-
cal procedures in which decisions about
sensitivity n. 1. the capacity to detect sample size and the type of data to be
and discriminate differences. In SIGNAL collected are made or modified as the
DETECTION THEORY, sensitivity is mea- srady proceeds, based on the cumulative
sured by the index D PRIME (d'). 2. the findings to date. This approach contrasts
332
Shannon index
333
Shapiro-Wilk test
334
Simes modified Bonferroni procedure
335
SIMEX
336
simultaneous comparison method
337
simultaneous confidence intervals
338
skewness
339
skew-symmetrical matrix
skewness
the MODE. Approximately 68% of the slope analysis see REGRESSION ANALY-
scores lie within 1 STANDARD DEVIATION SIS.
of the mean and approximately 95% of
the scores lie within 2 standard devia- smallest space analysis (SSA) a sta-
tions of the mean. When a distribution tistical technique for creating a visual
has a few extreme scores toward the high representation of-data, in which more
end relative to the low end (e.g., when a closely correlated variables are grouped
test is difficult and few test takers do together. Smallest space refers to the
well), it has a positive skew (or is posi- fewest number of geometric dimensions
tively skewed), such that the mean is (e.g., one dimension may be denoted by
greater than the mode. When a distribu- a line, two by a square, three by a cube)
tion has a few extreme scores toward the by which a body of data may be ade-
low end relative to the high end (e.g., quately represented. Similar in purpose
when a test is easy and most test takers to FACTOR ANALYSIS and PRINCIPAL
do well), it has a negative skew (or is COMPONENTS ANALYSIS, Smallest space
negatively skewed). analysis may be appUed to the assess-
ment of objects, persons, attitudes, test
skew-symmetrical m a t r i x a SQUARE constracts, and other similar variables.
MATRIX whose TRANSPOSE is equal to its
negation. small expected frequencies an issue
that arises when the CHI-SQUARE TEST is
original skew-symmetrical used to study GOODNESS OF FIT for vari-
ables in a CONTINGENCY TABLE. The chi-
1 2 0 -1 -2 square test is based on the assumption
0 1 1 0 - 1 that the obtainedfrequencieswdthin any
-1 0 2 1 0 cell of the table will be distributed nor-
maUy around the theoretically EXPECTED
It is symmetrical about its MAIN DIAGO- FREQUENCY. If the expectedfrequencyis
NAL, with each element negated, as small, however, the observed frequen-
shown in the example above. See also cies cannot be normally distributed and
SYMMETRICAL MATRIX. thus the chi-square test is inappropriate.
Generally, the expected frequency for a
slope n. the steepness or slant of a line on cell should be five or more.
a graph, measured as the change of value
on the Y-AXis associated with a change small-N design a research approach in
of one unit of value on the X-AXIS. In a which only a few participants are stud-
REGRESSION EQUATION, slope is repre- ied in order to focus attention upon the
sented by the variable b, with +b indicat- individual rather than the group as a
ing an upward slope to the line and -b whole. Typically, such designs are con-
indicating a downward slope. See also ducted over a longer period of time, to
ACCELERATION. allow for the behavior of interest to sta-
340
socially sensitive research
bilize, and employ QUALITATIVE ANALY- asked to nominate others, and so on until
SES as opposed to INFERENTIAL STAT- a sufficient number of participants is ob-
ISTICS iri analyzing results. Although tained. Researchers generally use snow-
small-N designs have several advan- ball sampUng if the population of inter-
tagesincluding reduced cost since est is hard to locate, rare (e.g., people
fewer participants are involved, better who have an infrequent condition or
control of EXTRANEOUS VARIABLES, and disease), or otherwise limited. Although
more comprehensive records of partici- this nonprobability strategy of chain re-
pant performancea significant disad- ferral is simple and cost efficient, there is
vantage is that EXTERNAL VALIDITY may the potential for SAMPLING BIAS in that
. be lower if parficipants are not representa- initial participants may tend to nomi-
tive of the larger population from which nate people they know well and thus are
they are drawn. Compare LARGE-N DESIGN. likely to share the same traits and char-
acteristics.
small-sample theory the understand-
ing and application of appropriate social desirability the bias or ten-
methods for analyzing data from groups dency of individuals to present them-
(samples) comprising 30 or fewer indi- selves in a manner that will be viewed
viduals or cases. Although there is some favorably by others. In an experiment,
SAMPLING ERROR in any group under for example, it manifests as the social
srady, the theory accounts for the fact desirability response set, which is the
that this error tends to be greatest in tendency of participants to give answers
smaller samples. that are in accordance with social norms
or the perceived desires of the researcher
Smirnov test see KOLMOGOROV-SMIRNOV rather than genuinely representative of
GOODNESS-OF-FIT TEST; KOLMOGOROV- their views. This is a CONFOUND to be
SMIRNOV TWO-SAMPLE TEST. controlled for in certain research, as it
smoothed curve a graphical represen- often reduces the validity of interviews,
tation of the relationship between two questionnaires, and other self-reports.
variables that has been adjusted to elimi- social i n d i c a t o r any fearare of a soci-
nate erratic or sudden changes in SLOPE, ety that can be measured over time and
so that its fundamental shape and direc- is presumed to reveal some underlying
tion will be evident. Also called smooth aspect of social reality and quality of life.
curve. For example, the retail price index is
smoothing n. a collection of techniques used as a measure of inflation, which in
used to reduce the irregularities (RAN- turn is taken as a key indicator of eco-
DOM VARIATION) in a data set or in a plot nomic performance. Other commonly
(curve) of that data, particularly in TIME- used indicators are derived from unem-
SERIES analyses, so as to more clearly see ployment figures, per capita income,
the underfying trends. The use of a MOV- poverty levels, labor conditions, hous-
ING AVERAGE is one example of smooth- ing costs, mental health, general health
ing. See SMOOTHED CURVE. and mortality data, nutrition informa-
tion, pollution levels, crime rates, educa-
Snedecor's F d i s t r i b u t i o n see F DIS- tion levels, opportunities for leisure and
TRIBUTION. [George W. Snedecor (1881- recreation, and the status of the elderly.
1974), U.S. statistician] Researchers and policymakers fre-
quently use social indicators to assess
s n o w b a l l s a m p l i n g a technique to the extent to which a society is "pro-
identify and recruit candidates for a gressing" as well as make predictions
study in which existing participants rec- about its furare.
ommend additional potential partici-
pants, who themselves are observed and socially sensitive research research
341
social science
METRIC TEST of whether there is an asso- SERIES ANALYSIS, where the purpose is to
ciation between two ORDINAL VARI- identity the nature of the phenomenon
ABLES when the data are presented in a represented by a sequence of measure-
C O N T I N G E N C Y T A B L E . T h e foCUS o f t h e ments, typically taken at successive
analysis is on the number of agreements points in time under the assumption
and disagreements in the set of data. For that the sequence is not random, and to
example, suppose a researcher interested forecast or predict future values. In a
in the relationship between age and atti- STARIMA model, the value of each ob-
tudes conducted a srady in which the servation in the series depends, at least
former were measured as ranges (e.g., in part, on the value of one or more of
20-25 years, 26-30 years, etc.) and the the immediately preceding observa-
latter were measured as good, neutral, or tions, and each observation in the series
poor. Somers's d, which ranges from -1 is predicted by a linear combination of
to +1, could then be used to obtain a prior observations as well as RANDOM
measure of association between the two. ERROR. When a variable, such as the
Also called Somers's delta. [Robert H. number of unemployed, is plotted over
Somers] tirhe there are likely to be considerable
seasonal or cyclical components to the
sorting test a format for assessing the variation that make it difficult to see the
ability to conceptualize, often used in underlying pattern. These components
adult neuropsychological assessments or can be eliminated by taking a suitable
in determining a child's level of cogni- moving average, which reduces random
tive development. The participant is fluctuations and makes long-term trends
asked to arrange an assortment of com- clearer.
mon objects by categoty. A common ex-
ample is the Wisconsin Card Sorting space-time autoregressive moving-
Test, in which the participant is asked to average model (STARMA model)
match a set of cards depicting shapes of an extension of the AUTOREGRESSIVE
differing color, quantity, and design; he MOVING-AVERAGE MODEL to include ad-
or she is given no matching rales but is ditional spatial and temporal dimen-
told whether a particular match is cor- sions in the exploration and forecasting
rect or incorrect. During the course of of relationships in a TIME SERIES. For
the test the matching rules are changed example, a researcher could apply a
and the time taken for the participant to STARMA model to predict traffic flow at
learn the new rules and the mistakes several adjacent intersections in a neigh-
made during this learning process are borhood using data gathered through-
analyzed to arrive at a score. out a particular year. There are likely to
soundproof r o o m a room designed to be multiple associations among multiple
balance the absorption of sound with its variables at multiple points in time that
controlled reflection. Mostly used for must be accounted for in such a situa-
the study of acoustic issues, soundproof tion, which can best be understood
rooms eliminate unwanted outside noise using a STARMA model.
and ensure sounds are sufficiently con-
spatial autocorrelation a measure-
fined to and evenly distributed through-
out the space within. , ment of a variable's association with
itself throughout space. Although statis-
space-time autoregressive inte- tical approaches often assume that mea-
grated moving-average model sured outcomes are independent of each
, (STARIMA model) an extension of the other, this may not be true for observa-
AUTOREGRESSIVE INTEGRATED MOVING- tions made at different locations. For ex-
AVERAGE MODEL to STATIONARY PRO- ample, measurements made at nearby
CESSES. The model is used in TIME- locations may be closer in value than
343
spatial data
measurements made at locations farther changes in test size. Also called Spear-
apart. Spatial autocorrelation thus pro- man-Brown prediction formula.
vides an index of the similarity of ob- [Charies Edward Spearman (1863-1945),
jects within an area, the level of inter- British psychologist and psychometri-
dependence between the variables, and cian; W. Brown, 20th-cenraty British
the nature and strength of that interde- psychologist]
pendence. Suppose one wanted to srady
Spearman correlation coefficient
the'relationships among cultures. In this
research the distance between societies (symbol: r^; p) a nonparametric measure
could be either spatial (based on phy- of statistical dependence between two
sical distance) or cultural (based on lan- variables that were measured on an OR-
DINAL SCALE; that is, the individual ob-
guage). Thus, spatial autocorrelation
models would examine the relationship servations (cases) can be ranked into two
among values of a single variable that ordered series. The Spearman correla-
arises from the geographic arrangement tion coefficient assesses how well the re-
of the areas in which these values occur. lationship between the variables can be
Spatial autocorrelation may be positive described using a MONOTONIC function.
or negative. Positive spatial autocorre- It ranges in value from +1 to -1. Also
lation occurs when similar values occur called Spearman's rank correlation
near one another, whereas negative spa- coefficient; Spearman's rho. [Charles
tial autocorrelation occurs when dissim- Spearman]
ilar values occur near one another. Spearman footrule a simpler version
When no statistically significant spatial of the SPEARMAN CORRELATION COEFFI-
autocorrelation exists, the pattem of spa- CIENT: a procedure for measuring rank
tial distribution is considered random. association that assesses the absolute dif-
See also AUTOCORRELATION. ference (D) between pairs of values. For
example, one supervisor ranks a set of
spatial data any observations or mea-
employees as 1, 2, 3, and 4, respectively,
surements with a direct or indirect refer-
on overall performance while another
ence to a specific location or geographi-
supervisor rates the same set of employ-
cal area.
ees as 2, 3,1, and 4, respectively. The ab-
spatial experiment a research design solute sum of the differences for the sets
that studies variables with respect to their of ranks is obtained by subtracting one
physical distance or separation from one rank in a pair from the other, disregard-
another. For example, a spatial experi- ing the sign of the difference, and add-
ment might examine the rates of mal- ing the values together. Thus, [1 - 2 = 1;
treatment referrals for children of a 2-3 = l;3-l=2;4-4 = 0], yielding a D
specific age across different residential = 4. The lower the absolute value of D,
areas. the closer the relationship between the
two sets of data. [Charles Spearman]
Spearman~Brown prophecy for-
m u l a the mathematical formulation of Spearman's r h o see SPEARMAN CORRE-
a basic tenet of CLASSICAL TEST THEORY LATION COEFFICIENT.
concerning the length (number of items)
speciesism n. discriminatoty, prejudi-
of a test and its influence on reliabiUty,
cial, or exploitative practices against
whereby increasing the number of items
nonhuman animals, often on the basis
with similar content results in increased of an assumption of human superiority.
reliabiUty for the test; similarly, decreas- See also ANIMAL RIGHTS. speciesist
ing the number of items leads to de- n., adj.
creased reliability. The formula allows a
researcher to estimate the gains or losses specification error in REGRESSION
in reliability that would occur with ANALYSIS, atypeof error that may occur
spherical data
345
sphericity
346
SSA
347
SSCP
SSCP symbol for SUM OF SQUARES OF sign, all samples have a common ori-
CROSS-PRODUCTS. ginthey start at the same timeand
there is a single, terminal event (typi-
SSE symbol for sum of squared errors. See
cally, death). By contrast, in a staggered
ERROR SUM OF SQUARES. entry design each participant begins the
S-shaped curve see OGIVE. study at a particular time zero and has an
ending point corresponding to the
SSregression Symbol for REGRESSION SUM
length of time that he or she was part of
OF SQUARES.
the study (i.e., until he or she experi-
Stability n. see STATISTICAL STABILITY. enced the event of interest or otherwise
ceased participation). For example, con-
stability coefficient an index of RELI- sider a study of a new treatment method
ABILITY determined via a test-retest
implemented at a specific hospital: There
method, in which the same test is ad- will be patients who survived over the
ministered to the same respondents at entire study period, others who survived
two different points in time. For exam- but entered the hospital after the study
ple, the stability coefficient of a psycho- had already begun, and still others who
logical test may be estimated by moved away and lost contact with the
determining the degree of similarity be- researcher. Although the latter two groups
tween participants' scores across time: of patients contributed only partial, in-
The more the two scores for each partici- complete information to the research
pant are alike, the higher the correlation (see CENSORED OBSERVATION), their data
between the two administrations and nonetheless should not be excluded from
the greater the stability coefficient of the the srady since those individuals are
test. A critical issue in using this strategy "survivors" and therefore reflect on the
is the interval of time between the two success of the new treatment method.
administrations. It should not be so long Staggered entry allows for increased pre-
that participants are likely to change on cision in estimating the SURVIVAL FUNC-
the construct being assessed nor so short TION.
that memory of responses on the initial
administration infiuences responses on Stahel-Donoho robust m u l t i v a r i -
the second administration (see PRAC- ate estimator in FACTOR ANALYSIS,
TICE EFFECT). Also called coefficient of DISCRIMINANT ANALYSIS, and Other sta-
stability. tistical techniques for examining multi-
stable rate 1. a constant change in the variate data, a strategy for handling
value of a DEPENDENT VARIABLE as. the OUTLIERS that weights each observation
value of an INDEPENDENT VARIABLE according to its overall "outlyingness."
changes. 2. in a REGRESSION ANALYSIS, The Stahel-Donoho estimator is neces-
the relative consistency of a BETA COEF- sary since the most common methods
FICIENT over time and across popula-
for identitying averages and variances,
tions, conditions, and so forth. the sample MEAN arid the sample CO-
VARIANCE MATRIX, are sensitive to out-
stacked bar g r a p h see COMPONENT liers and thus tenable only when the
BAR GRAPH. data foUow a NORMAL DISTRIBUTION.
[Werner Stahel, Swiss statistician; David
stage s a m p l i n g see MULTISTAGE SAM-
Donoho (1957- ), U.S. statistician]
PLING.
Observation Number
1 2 3 4
1234567890123456789012345678901234567890
6 XX X X X XX X X X XX X X
8 XX X X X XX X X X XX X X
10 XXX X X X XX X X X XX X X
12 X X X X X X X XX X X X X XX
14 XXX X X X X X XX X X X XX X X
16 XXX X X X X X X XX X X XX X X
18 XXX X X X X X X XX X X XX X X
20 XXX X X X X X X XX X X XX X
22 XXX X X X XX X X XX X
0) 24 XXX X X X X X X XX X X XX X
N 26 XXX X X X X X X XX X X XX
28 XXX X X X X X XX X X XX
vi
4-1 30 XXX X X X X X X XX XXX
0) 32 XX X X X X X
sqn
34 X X X X X X X X XXX
36 X X X X X XX
(0 38 X X X X X X X X XX
40 X X X X X XX
42 XX X XX
44 X X X X X X
46 X X
48 X X
50 X
stalactite plot
OUTLIERS in MULTIVARIATE data. In the mean value X, the sample standard de-
hypothetical example above, xs are used viation is taken to be
to denote those observations that can be
V[i(x,- x)2/(-i)],
considered outliers within each subset
size, with observation 23 emerging as that is, the divisor is (n -1) rather than n.
the most extreme outlier. See also ROOT MEAN SQUARE.
349
standard error of estimate
means over all of the samples taken. any given measurement opporranity
Standard error is expressed in units represents his or her TRUE SCORE. It is
given in the same scale of measurement one of several types of STANDARD ERROR
that was used for the sample data (e.g., that may be calculated.
for a set of means that are given in
standard error o f p r e d i c t i o n see
weight the standard error unit is also a
S T A N D A R D E R R O R OF E S T I M A T E .
weight). The more samples involved in
determining the standard error, the standard error of the difference
smaller its value; the smaller the stan- an index of the degree to which statisti-
dard error, the more reliable the calcu- cal distinctions between two SAMPLES
lated population value. from a larger POPULATION reflect real
disparities between those samples. For
standard error of estimate (symbol: example, suppose a researcher analyzes
S) for a relationship between two vari- two samples, A and B, randomly drawn
ables (x and y) given by a REGRESSION from the same normally distributed
EQUATION, an index of how closely the source population and obtains a differ-
predicted value of y for a specific value of ent MEAN for Sample A than for Sample
X matches its actual value. If y' is an esti- B. The standard error of the difference
mated value from a regression line and y between the two means is used to deter-
is the actual value, then the standard mine whether those sample values truly
error of estimate is are different or instead reflect random,
V[Z(y-y')%], irrelevant dataflucraations.It is equal to
350
standardized mean difference
of college graduates. See also STANDARD intended. The arrangement of scores ob-
ERROR OF THE DIFFERENCE. tained by such a standardization group
subsequently provides a point of com-
Standard error of the mean (sym- parison for the scores of other older
bol: SM; OM) a statistic that indicates adults who take the test. It is important
how much the average value (MEAN) for to note, however, that the standardiza-
a particular SAMPLE is likely to differ tion group must be representative of the
from the average value for the larger intended population of test takers in
POPULATION from which it is drawn. It is order to yield valid information. Also
equal to aHn, where a is the standard called norm group; standardization
deviation of the original distribution sample.
and n is the sample size. Less commonly
called standard error of the popula- standardized coefficient any index
tion mean. derived from an analysis of two or more
variables that have been transformed via
standard g a m m a d i s t r i b u t i o n see STANDARDIZATION, whlch ensures their
GAMMA DISTRIBUTION. value RANGES and VARIANCES are equiv-
Standardization n. 1. the process of alent and thus appropriate for compari-
establishing NORMS for a test. 2. the use son. The term is most commonly used to
of uniform procedures in test adminis- denote the standardized regression coef-
tration to ensure that all participants ficient or BETA COEFFICIENT.
take the same test under the same condi- standardized distribution a NOR-
tions and are scored by the same criteria, MAL DISTRIBUTION whose values have
which in turn ensures that results can be undergone TRANSFORMATION SO as to
compared to each other. 3. the transfor- have a MEAN of 0 and a STANDARD DEVI-
mation of data into a distribution of ATION of 1. Also called standard nor-
STANDARDIZED SCORES, often One hav- mal distribution; unit normal
ing a mean of 0 and a STANDARD DEVIA- distribution.
TION of 1, which produces derived'
measures of relative standing and allows standardized instructions direc-
comparison of raw scores from different tions for a measuring instrument that
distributions. The Z-SCORE TRANSFOR- are to be presented to all participants ex-
MATION is an example of standardiza- actly as prepared as part of the standard
tion. experimental or assessment procedure.
See STANDARDIZATION.
standardization group a sample
Standardized i n t e r v i e w see STRUC-
used to establish reliable norms for the
TURED INTERVIEW.
population that it represents. This is
done by analyzing the results of the test standardized mean difference a
administered to the sample and ascer- summaty statistic used in a META-
taining the average performance level ANALYSIS of studies that assess the same
and the relative frequency of each devia- outcome but measure it in a variety of
tion from the mean. The NORMAL DIS- ways. For example, many studies mea-
TRIBUTION thus created is then used for sure the constract of anxiety and its rela-
comparison with any specific furare test tionship to school performance, but they
score. For example, the standardization do so using different psychometric
group for a new test of computer literacy scales. In order to examine the anxiety-
in older adults might comprise a large school performance effects found across
set of test takers above age 60 whose these multiple studies, it is necessaty to
characteristics (e.g., sex, demographics, first standardize the results of each of the
ethnicity, race) reproduce those of the studies to a uniform scale and then com-
larger population for whom the test is bine the results. The standardized mean
351
standardized measure
352
statistical conclusion validity
change the intensity of another sound TION strucrare remain constant over
to match. time. Most statistical forecasting meth-
ods are based on the assumption that
standard treatment c o n t r o l in a the time series of interest can be mathe-
CLINICAL TRIAL, atypeof CONTROL CON- matically transformed into approximate
DITION in which participants receive the stationarity (i.e., can be "stationarized").
treatment typically administered to a
group of participants with a given medi- Stationary d i s t r i b u t i o n a distribu-
cal problem. Another group of partici- tion describing ah invariant equilibrium
pants receives the experimental treat- state in a MARKOV CHAIN. The process
ment. After the trial, the outcomes of being modeled is associated with a sin-
the two groups are compared to see if the gle, time-independent MATRIX.
experimental treatment is better than, stationary process a STOCHASTIC
as good as, or worse than the standard PROCESS whose joint probability distri-
treatment. bution does not change when shifted in
stanine n. a method of scaling scores on time or space. In other words, the pa-
a nine-point scale that ranges from a low rameters of a stationary process, such as
of 1 to a high of 9, with a mean of 5 and a the mean and the variance, do not
STANDARD DEVIATION of 2. A Stanine is a change over time or position. See also
standard ninth, referring to the interval STATIONARITY.
used in dividing the results into nine statistic n. 1. a number measuring some
more or less equal parts. A stanine score characteristic, construct, variable, or
of 1, 2, or 3 is below average; 4, 5, or 6 is other item of interest. 2. any function of
average; and 7,8, or 9 is above average, 'A the observations in a SAMPLE that may
stanine is a type of STANDARDIZED be used to estimate the unknown but
SCORE and is mainly used with school corresponding value in the POPULA-
achievement tests. | TION. Examples include measures of
STARIMA model abbreviation for CENTRAL TENDENCY (e.g., the MEAN, ME-
SPACE-TIME AUTOREGRESSIVE INTEGRATED DIAN, MODE), measures of DISPERSION
MOVING-AVERAGE MODEL. (e.g., STANDARD DEVIATION, VARIANCE),
and distributional attributes (e.g., SKEW-
S T A R M A model abbreviation for NESS, KURTOSIS). Statistics often are as-
SPACE-TIME AUTOREGRESSIVE MOVING- signed Roman letters (e.g., M, s), whereas
AVERAGE MODEL. ] the equivalent values in the population
(called PARAMETERS) are assigned Greek
State space 1. multidimensional space,
letters (e.g., p, a).
particularly as related to the depiction of
the results of classification methods that statistical analysis any of a wide
group objects with similar characteris- range of techniques used to describe, ex-
tics and patterns of behavior. 2. in sys- plore, understand, explain, predict, and
tem behavior, an abstract representation test HYPOTHESES about data. It involves
of the potential states of the system that the examination of data collected from
emerges from the complex interactions SAMPLES within POPULATIONS as well as
of the variables that make up the system. the use of probabilistic models to make
The interactions are based on a set of inferences and draw conclusions.
nonlinear relations.
statistical conclusion v a l i d i t y the
stationarity n. in TIME-SERIES ANALY- degree to which the conclusions drawn
SIS, the property of being unchanging or from statistical analyses of data are accu-
"flat," such that the data are without rate and appropriate. In other words, sta-
trends or periodic fluctuations and the tistical conclusion validity addresses
MEAN, VARIANCE, and AUTOCORRELA- whether inferences about relationships
353
statistical control
(i.e., whether the independent variable given decision. Specific equations are
and dependent variable covary and, if used to calculate the degree of loss asso-
so, how strongly) are reasonable or not, ciated with each course of action in
given the observed data. It is related to order to determine the most advanta-
but distinct from INTERNAL VALIDITY, geous choice.
which is concerned with the causality of
statistical dependence the situation
the relationship. Statistical conclusion
in which the conditional probabUity of
validity is enhanced when there is good
one event given another event is greater
statistical POWER, RELIABILITY of mea-
than the probability of that flrst event.
sures, and use of good experimental
In other words, knowing information
methods and procedures. Conversely, it
about one variable provides information
is threatened by such factors as (a) viola-
about the other. Compare STATISTICAL
tions of the ASSUMPTIONS embedded in
INDEPENDENCE. See also DEPENDENT
a statistical test; (b) problems associated
EVENTS.
with the EXPERIMENT-WISE ERROR RATE;
(c) RESTRICTION OF RANGE; (d) use of in- statistical determinism the position
appropriate EFFECT-SIZE measures; and that the laws of probability can predict
(e) extraneous variation in the experi- the likely number of events of a given
mental setting. kind that will occur in a given popula-
tion under certain defined conditions
statistical c o n t r o l the use of statistical (e.g., the number of coin tosses per 1,000
procedures to remove the influence of a that will be heads; the number of male
particular factor that could not be elimi- Americans age 70 who will die in a 12-
nated or controlled by the experimental month period). This is analogous to, but
design in order to better analyze the rela- logically distinct from, physical deter-
tionship between two variables. For ex- minism, which is the proposition that all
ample, the relationship between age (x) eventsincluding human behaviors
and income earned (y) could be influ- are caused by prior events, conditions,
enced by a third variable, years of educa- and natural laws.
tion (z). Thus, if a researcher did not first
remove the effects of education he or she statistical difference see SIGNIFI-
might derive erroneous conclusions CANT DIFFERENCE.
about the influence of age on income statistical error see ERROR.
from his or her analysis. One type of sta-
tistical control is PARTIAL CORRELATION, statistical evaluation see STATISTI-
which shows the association between CAL TEST.
two quantitative variables after statisti- statistical hypothesis a research
cally controlling for one or more extra- question posed in a statistically testable
neous variables. A second type of form. For example, if a researcher is in-
statistical control is ANALYSIS OF CO- terested in whether one treatment leads
VARIANCE, a technique that shows the to a more positive outcome than an-
relationship between a categorical inde- other treatment, he or she could reframe
pendent variable and a quantitative the question in terms of mean differ-
dependent variable after statistically ences, such that the NULL HYPOTHESIS is
controlling for one or more extraneous 0 (no difference between treatments)
variables. and the ALTERNATIVE HYPOTHESIS is not
0 (the difference between treatments is
statistical decision theory a branch
greater or lesser than zero). See also HY-
of statistical science concerned with the
POTHESIS TESTING.
use of data to arrive at decisions. It fo-
cuses upon identifying the values, un- statistical independence the condi-
certainties, and other issues relevant to a tion in which the occurrence of one
354
statistical table
event makes it neither more nor less participants might be required than in
probable that another event will occur. alternative strategies.
In other words, knowing information
statistical significance the degree to
about one variable provides no informa-
which a research outcome cannot rea-
tion about the other variable. Compare
sonably be attributed to the operation of
STATISTICAL DEPENDENCE. See also IN-
chance or random factors. It is deter-
DEPENDENT EVENTS.
mined during SIGNIFICANCE TESTING
statistical inference see INFERENCE. and given by a critical p value, which is
the probability of obtaining the ob-
statistical model a formal description served data if the NULL HYPOTHESIS (i.e.,
of the relationships between two or of no significant relationship between
more variables in the form of a mathe- variables) were true. Significance gener-
matical equation. It is statistical in that ally is a function of sample sizethe
the variables are related in a STOCHASTIC larger the sample, the less likely it is that
rather than a deterministic manner, one's findings will have occurred by
with each set of possible observations on chance. See also CLINICAL SIGNIFICANCE;
a variable linked to a set of probability PRACTICAL S I G N I F I C A N C E ; SIGNIFICANCE
distributions. Many STATISTICAL TESTS LEVEL.
involve comparing a particular model
with the observed data. Statistical significance testing see
SIGNIFICANCE TESTING.
statistical power see POWER.
statistical s t a b i l i t y consistency of re-
statistical p r e d i c t i o n the process of sults across samples, srady designs, and
using correlations between variables to analyses. A META-ANALYSIS may be used
hypothesize about future events and to examine the stability of means, COR-
outcomes. For example, a university ad- RELATIONS, and other PARAMETER esti-
ministrator may use a REGRESSION mates obtained from different studies of
EQUATION to predict a student's coUege the same population. See also STABILITY
grade point average with reasonable ac- COEFFICIENT.
curacy from measures of performance in
statistical surveillance continuous
h i g h school, such as scores o n tests a n d
observation and analysis of a set of data
f i n a l grades i n classes.
with the goal of detecting any important
Statistical procedure see STATISTICAL changes (e.g., in the direction of a pro-
TEST. cess) as soon as they occur. For example,
a statistical surveillance system to deter-
statistical psychology the applica- mine the end of the influenza season
tion of statistical methods and models to might rely upon NONPARAMETRIC meth-
organize, summarize, and interpret data ods to identify the decline in incidence
so as to derive descriptions and explana- that signals the season's end.
tions of cognitive, behavioral, social,
and other psychological phenomena. statistical table any table of values
used by a researcher to plan experiments
statistical sequential test any statis- or interpret results. For example, there
tical procedure in which the sample size are tables of CRITICAL VALUES in HY-
is not fixed in advance but rather the POTHESIS TESTING, tables to estimate the
process of coUecting more data (more sample size needed to obtain a desired
observations or more participants) is level of POWER, and tables to convert
continued until there is enough infor- one distribution to another (e.g., from a
mation available either to accept or to correlation to a standardized value). A
reject the NULL HYPOTHESIS. A parficular CONTINGENCY TABLE Is often used in the
benefit of this approach is that fewer CROSS-CLASSiFiCATiON of items: For ex-
355
statistical test
ample, a 2 x 2 table might show the aver- on the dependent variable. Repeated
age scores on a test according to runs are therefore undertaken and the
individuals' gender (male and female) path of steepest ascent is identified via the
and education level (high school and use of a REGRESSION EQUATION. Where
college graduates). improvement is defined as the minimiz-
ing of a particular variable, then the
statistical test any mathematical tech- steepest descent will be sought in the
nique or procedure used to evaluate the same way.
correctness of an empirical hypothesis
by determining the likelihood of the stem n. the introductory part of a FIXED-
sample results occurring by chance. Sta- ALTERNATIVE QUESTION, often an in-
tistical testing will reveal the probability complete statement that the respondent
of committing a TYPE i ERROR if the NULL is asked to complete. It is followed by a
HYPOTHESIS is rejected. See HYPOTHESIS set of response options consisting of the
TESTING; STATISTICAL SIGNIFICANCE. correct answer and several plausible but
incorrect choices (DISTRACTORS).
statistical uncertainty the degree of
inaccuracy inherent to the process of es- stem-and-leaf plot a graphical me-
timating PARAMETERS from sampled thod of presenting data that have been
data. When researchers make such esti- measured on an INTERVAL SCALE. A basic
mates they generally assume that some stem-and-leaf plot comprises two col-
exact or TRUE SCORE exists and often umns separated by a vertical line; the
specify a CONFIDENCE INTERVAL or range right column lists the last digit of each
of values that they expect this'true value data point (the "leaves") and the left col-
to fall within. For example, an estimate umn lists all of the other digits from
of 8 2 means that the researcher is con- each data point (the "stems"). Each stem
fident that the actual parameter value is listed only once and no numbers are
lies between 6 and 10. The uncertainty is skipped, even if that means some stems
the experimenter's best estimate of how have no leaves. The leaves are listed in
far an estimated quantity might be from increasing order of magnitude in a row
the true value. to the right of each stem. For example,
consider the following hypothetical val-
statistics n. the branch of mathematics ues for participants measured on a par-
that uses data descriptively or inferen- ticular variable:
tially to find or support answers for sci-
entific and other quantifiable questions. 55 57 58 59 74 75 77 79 79 83 83 86 87
It involves various techniques and proce- 92 95 99 107
dures for recording, organizing, analyz-
The stem-and-leaf plot for these values is
ing, and reporting quantitative infor-
mation. See also DESCRIPTIVE STATISTICS; 5 15 7 8 9
INFERENTIAL STATISTICS; SUMMARY STA-
TISTICS. statistical adj. statisti- 6I
cian n. 714 5 79 9
steepest ascent when conducting a 8 13 3 6 7
study with repeated runs, that region of 91259
the INDEPENDENT VARIABLE in which
there is maximum response or improve- 101 7
ment on the DEPENDENT VARIABLE of Also called stem-and-leaf diagram;
interest. If research is being carried out stem-and-leaf display; stemplot.
in a new area, it is likely that initial levels
of the independent variable will not stepdown selection see BACKWARD
provide a satisfactory level of response ELIMINATION.
356
stochastic process
stepup selection see FORWARD SELEC- and females to elicit donations from par-
TION. ticipants instead of using a single male
and a single female.
stepwise correlation see STEPWISE
REGRESSION. stimulus value 1. the strength of a
given stimulus, measured in standard
Stepwise d i s c r i m i n a n t analysis a units (e.g., a shock of 40 volts). 2. a theo-
type of DISCRIMINANT ANALYSIS in retical characteristic of a stimulus said to
which a model that can differentiate index its effectiveness as a reinforcer.
maximally between categories is built
step by step. In forward stepwise dis- stochastic adj. 1. random or undeter-
criminant analysis, all possible predictor mined; arising from chance. 2. describ-
variables are reviewed and evaluated at ing a system or process that follows
each step to determine which one con- a probability pattern, such that events
tributes most to the discrimination be- may be analyzed according to their sta-
tween groups. That variable is then tistical probability but not accurately
included in the model, and the process predicted.
starts again. In backward stepwise stochastic matrix a SQUARE MATRIX
discriminant analysis, all variables are in which all of the rows or all of the col-
initially included in the model and then, umns sum to 1. Arightstochastic matrix
at each step, the variable that contrib- is one in which the rows contain
utes least to the prediction of group nonnegative real numbers summing
membership is eliminated. The stepwise to 1, whereas a left stochastic matrix is
procedure is structured so that the re- one in which the columns contain
spective F-TO-ENTER and F-TO-REMOVE nonnegative real numbers summing to
values of each variable can be isolated, 1. See also DOUBLY STOCHASTIC
thereby determining its unique contri-
bution to the prediction of group mem- stochastic model a model in which
bership. one or more of the inputs allow for ran-
dom variation, thus generating a range
stepwise regression a group of RE- of potential outcome values. The ran-
GRESSION ANALYSIS techniques that dom variation is usually based on fluc-
enter predictor (independent) variables tuations observed in historical data. Sto-
into (or delete them from) the REGRES- chastic models are used to estimate the
SION EQUATION one variable (or block of probabilities of various outcomes occur-
variables) at a time according to some ring under varying conditions. They are
predefined criterion. It is contrasted widely used in the social and behavioral
with SIMULTANEOUS REGRESSION, whlch sciences and also in theflnancialworld.
enters all variables at the same time. Also Compare DETERMINISTIC MODEL.
called stepwise correlation; step-
wise multiple regression. stochastic process a random process: a
sequence of events with a random prob-
stimulus s a m p l i n g a procedure for in- ability pattern such that the occurrence
creasing the generalizability of research of any event in the sequence is inde-
results by using multiple stimuli within pendent of past events. For example, the
a categoty as representative of an experi- number of people in a doctor's office
mental condition, as opposed to select- who have colds during a one-month pe-
ing a single stimulus whose unique riod could be said to follow a stochastic
characteristics may distort results. For process. In contrast to deterministic pro-
example, a study investigating the ef- cesses, stochastic processes involve
fects of gender on monetaty generosity some indeterminacy, such that their de-
would demonstrate stimulus sampling if velopment over time may only be de-
it employed a variety of different males scribed by probability distributions. A
357
stochastic variable
358
structured observation
of the amount of change expected in an ships, one might ask each of 15 people
outcome or DEPENDENT VARIABLE given to indicate who among the others is a
a one-unit change in the causal or INDE- friend. This would yield a 15 x 15 con-
PENDENT VARIABLE and no change in tingency table or matrix, with 1 repre-
any other variable.- senting "is a friend" and 0 representing
"is not a friend." Although there would
structural equation modeling be empirical zeros in some cells to repre-
(SEM) any of a broad range of multi- sent not-a-friend responses, all entries
variate analysis methods, including FAC- along the diagonal of the matrix are
TOR ANALYSIS and PATH ANALYSIS, that
structural zeros since they represent one
examine VARIANCES and COVARIANCES
person's response with regard to himself
in order to find interrelationships or herself rather than another person.
among LATENT VARIABLES. For example,
assume a researcher states that job sat- stmcture coefficient 1. in REGRESSION
isfaction leads to happiness. Both are ANALYSIS, a CORRELATION COEFFICIENT
latent variables that are not directly indicating the extent of relationship be-
observable but are defined in terms tween scores on a particular MANIFEST
of other measurable variables, such as VARIABLE and scores on a latent or pre-
judgments of job performance from su- dicted variable derived from the entire
pervisors and peers; self-reports about set of variables. It is used to quantity the
attitudes toward pay, supervision, type importance of an observed variable. 2.
ofwork, and other job characteristics; re- in DISCRIMINANT ANALYSIS, a value rep-
sults from PSYCHOLOGICAL TESTS; and resenting the association between a par-
so forth. The researcher could use the ticular variable being considered as a
measurable data to generate an equation differentiator of group membership and
representing the strength and nature of the DISCRIMINANT FUNCTION scores de-
the links among the latent variables. rived from the full set of predictor vari-
Also called structural modeling. ables.
359
structured Q sort
361
subjective prior
362
summated rating scale
363
summative evaluation
364
survival probability
the influence of a third variable. Com- methods used to model a variety of time-
pare REINFORCER EFFECT. related outcomes. The simplest applica-
tion of survival analysis involves esti-
suppressor variable 1, a condition or mating the amount of time until the
characteristic that is associated with an occurrence of an event (e.g., death, ill-
INDEPENDENT VARIABLE, SUCh that the ness, graduation, marriage) for a group
correlation between the independent of individuals, but the technique also
variable and the DEPENDENT VARIABLE is may be applied to compare durations for
less than what it would be otherwise. 2. two or more groups and to build
in REGRESSION ANALYSIS, a piedlctor multivariate models that explain varia-
variable that is highly correlated with tion in duration. Survival analysis can be
Other predictors but that has a small cor- more informative than other techniques
relation with the outcome variable. The because it gives insight into the relation-
suppressor variable thus serves to re- ship between time and the outcome of
move the other predictors' variance that interest. That is, it enables the researcher
is irrelevant to the outcome, resulting in to determine not just whether an out-
stronger relationships with the depend- come is likely to occur but whether it
ent variable of interest. See also PAR- will occur early or late and whether the
TIALING.
chances of event occurrence increase
surrogate end p o i n t in CLINICAL TRI- gradually or sharply over time. Also
ALS, a measure that is intended to substi- called duration analysis; event his-
tute for or predict a particular outcome. tory analysis.
Surrogate end points often are used survival curve in SURVIVAL ANALYSIS,
when the primary end point is undesir- a graph showing the probabUity of an
able or rare, thus making it unethical or event's occurrence at specific points in
impractical to conduct a clinical trial to time, as in the illustration overleaf using
gather a statistically significant number hypothetical data.
of end points. For example, a researcher In this example, the survival curve de-
testing a drag to decrease the likelihood picts the number of individuals who
of death from heart disease may use died from a disease of interest at various
blood pressure as a surrogate end point. points in time relevant to the study of
surrogate v a r i a b l e see PROXY VARI- the specific disease. Also called survival
ABLE. plot.
survey n. a study in which a group of s u r v i v a l f u n c t i o n in SURVIVAL ANALY-
participants is selected from a popula- SIS, a mathematical formula that de-
tion and some selected data about or scribes the relationship between .the
opinions of those participants are col- likelihood of a particular event occur-
lected, measured, and analyzed. Infor- ring and a given time frame.
mation typically is gathered by inter- s u r v i v a l model in SURVIVAL ANALYSIS,
view or self-report questionnaire, and a model that shows the probability of
the results thus obtained may then be the occurrence of a well-defined event
extrapolated to the whole population. at different time points while taking
survey error the degree to which the into account the influence of certain
results of a SURVEY are inaccurate, due to predictors or explanatoty variables. See
such factors as SAMPLING BIAS, content CONTINUOUS-TIME SURVIVAL MODEL;
or measurement flaws, or the RESPONSE DISCRETE-TIME SURVIVAL MODEL.
BIASES of participants. See MEASURE- survival plot see SURVIVAL CURVE.
MENT ERROR; SAMPLING ERROR.
survival p r o b a b i l i t y in SURVIVAL
survival analysis a family of statistical ANALYSIS, the likelihood of experiencing
365
survival rate
survival curve
366
systematically biased sampling
rors the other, such that half of the synchronic adj. describing research
values are in the left tail of the distribu- that focuses on events or other phenom-
tion and the other half are in the right ena at a given point in time. For exam-
tail. Compare ASYMMETRICAL CONFI- ple, a synchronic approach to linguistics
DENCE INTERVAL. would seek to characterize linguistic fea-
symmetrical d i s t r i b u t i o n a distribu- tures at a particular time, without refer-
tion in which the frequency of values ence to their historical development.
above the MEAN are a mirror image of Synchronic research contrasts with
those below the mean. Compare ASYM- diachronic research, which focuses on
METRICAL DISTRIBUTION.
processes of change over time.
symmetrical m a t r i x a SQUARE MA- synergism n. the joint action of differ-
TRIX in which the values above the MAIN ent elements such that their combined
DIAGONAL are a mirror image of the val- effect is greater than the sum of their in-
ues below the diagonal. It has the prop- dividual effects. For example, synergism
erty A = (fly), orfly= ajj, with / denoting occurs in organizational behavior when
rows and / denoting columns. a work group's overaU performance ex-
ceeds the aggregate performance of its
individual members. See also INTERAC-
TION EFFECT. synergistic adj.
synthetic approach the combining
Thus, in the example matrix above the (synthesizing) of various processes, sys-
third value in column 2 is the same as tems, skills, or other components into a
the second value in column 3 (i.e., 1). more complex whole as a means of
learning or better understanding the
symmetry n. equality relative to some whole. For example, a synthetic ap-
axis. More specifically, it is a condition proach to learning to read is one in
in which values are arranged identically which the child first learns to recognize
above and below the middle of a data set written letters and understand their as-
(see NORMAL DISTRIBUTION) or abovesociated sounds before learning to com-
and below the diagonal of a matrix (see bine letters into syllables and words.
SYMMETRICAL MATRIX). Many Standard Compare ANALYTIC APPROACH.
statistical techniques are appropriate
only for symmetrical data, such that synthetic risk m a p a graphical dis-
nonsymmetrical data often are trans- play of the results of a HAZARD ANALYSIS,
formed into a roughly symmetrical form which uses multivariate statistical proce-
prior to analysis. Compare ASYMMETRY. dures to identity and understand haz-
symmetrical adj. ardous or risky situations. Synthetic risk
maps provide an estimate of what the
symmetry test a method of determin- prevalence rate of the hazard (e.g., can-
ing whether a graphical representation cer mortality from airborne toxins) in
of a data set demonstrates SYMMETRY a given area is expected to be given the
about its X-axis, y-axis, or origin, such demographics of the area.
that knowing the arrangement of values Consider the example overleaf from
for one portion of the graph (e.g., values the 2002 National-Scale Air Toxics As-
above the mean) enables one to deter- sessment of the U.S. Environmental Pro-
mine the values for the opposite, mirror- tection Agency.
image portion of the graph (i.e., values
below the mean). If the data do possess synthetic variable see LATENT VARI-
ABLE.
such symmetty, they can be analyzed
using statistical techniques that assume systematically biased sampling
a NORMAL DISTRIBUTION. any sampUng method that consistently
367
systematic error
<1 in a Million
1 ^ 1 - 25 in a Million
^ 25 - 50 in a Million
I 50 - 75 in a Million .
I 75 -100 in a Million
H >100 in a Million
368
systems theory
recorded. The intent is to ensure that, strategy, the coding and analysis of the
under the same or similar circumstances, included studies, and other procedures
all observers will obtain the same results. are explicitly defined in advance in
order to ensure that the process is trans-
systematic r a n d o m s a m p l i n g see parent and can be replicated, with the
SYSTEMATIC SAMPLING. aim of minimizing bias in terms of what
is reviewed and the conclusions that are
systematic r e p l i c a t i o n the process of drawn.
conducting a study again but with cer-
tain consistent differences, often in an systematic sampling a type of sam-
attempt to extend the original research pling process in which all the members
to different settings or participants. For of a population are listed and then some
example, a systematic replication could objective, orderly procedure is used to
refine the design (e.g., by using more randomly choose specific cases. For ex-
participants) or the methodology (e.g., ample, the population might be listed al-
by using more standardized procedures phabetically and evety seventh case
or objective measures). Compare DIRECT selected. Also called quasi-random
REPLICATION. sampling; systematic random sam-
pling.
systematic review an organized
method of locating, assembling, and systematic variance see UNSYSTEM-
evaluating a body of literature on a par- ATIC VARIANCE.
ticular topic using a specific set of crite- systems theory see GENERAL SYSTEMS
ria. Standards for inclusion, the search THEORY.
369
Tt
T symbol for the statistic obtained from simple printed form, used for checking,
the WILCOXON SIGNED-RANKS TEST. counting, or scoring a variable. For ex-
ample, a researcher might use a tally
see HOTELLING'S T^ TEST.
sheet to record the frequency of occur-
table n. a presentation of data in the rence of various behaviors or other
form of an ordered arrangement of over- events. Also called tally chart.
laid vertical columns and horizontal T A R abbreviation for THRESHOLD AUTO-
rows. As with a GRAPH, the purpose of a REGRESSION.
table is to communicate information (ei-
ther in words or numerical values) in a target n. 1. an area or object that is the
concise, space-efficient manner that can focus of a process, inquity, or activity. 2.
be assessed at a glance and interpreted the goal object in a task. For example,
easily. The columns have headings (the the target in a visual search might
leftmost column, which usually lists the be to find a letter S in a randomly ar-
independent variable, is referred to as ranged array of letters. In some concept-
the stub column). The intersection of a discovety tasks, the target is the rule that
column and row is called a CELL. Tables classifies objects as belonging or not be-
are often accompanied by explanatory longing to a category. Where a search
notes. tabular adj. has more than one item as its goal, these
are known as the target set.
tachistoscope n. a device that displays
target group the specific group of peo-
(usually by projecting) visual material
ple within a TARGET POPULATION that is
on a screen for a specific amount of time,
the focus of research. The target group
usually at very brief intervals. Words, may be defined by age, gender, marital
numbers, pictures, and symbols can be status, or other similar background vari-
rapidly presented in the right or left vi- ables. Often, a particular combination of
sual field. The device is used in experi- variables, such as men ages 20 to 50,
ments that are concerned with visual may define the target group.
perception, recognition speed, and
memory. It is also widely used in market target p o p u l a t i o n the population
research concerned with advertising, that a study is intended to research and
logos, branding, and so forth. Also called to which generalizations from samples
T-scope. are to be made. Also called reference
population. See also TARGET GROUP.
tactile test any test designed to mea-
sure how people perceive something target stimulus a specific stimulus to
through the sense of touch. which participants in a test or experi-
mental procedure must attend or re-
t a i l of a d i s t r i b u t i o n in a PROBABIL- spond. For example, in tests of hearing
ITY DISTRIBUTION, especially one that is the target stimulus may be a specific
graphically displayed, the region (or re- tone that must be identified.
gions) of least frequently occurring val-
ues. This is often the CRITICAL REGION in Tarone-Ware test in SURVIVAL ANAL-
YSIS, a test that can be used to determine
tests of STATISTICAL SIGNIFICANCE.
whether two survival curves are equiva-
tally sheet an instrument, usually a lent when there are reasons to think that
temporal validity
the traditional survival test may not give is the sampling distribution of the statis-
useful results. This may be because test tic (M-[IQ)/S, where [LQ is the mean of the
assumptions have been violated, such as population from which the sample is
the assumption of independence of CEN- drawn, M is the estimate of the mean of
SORING, or because there are factors un- the population as obtained from sample
accounted for in the analysis that affect data, and s is the standard deviation of
survival or censoring times or both. The the data set. Also called Student's t
Tarone-Ware test is designed to be used distribution.
with a large percentage of censored data
and with nonnormal distributions. telephone i n t e r v i e w an interview
[Robert E. Tarone, U.S. epidemiologist; that is conducted by telephone rather
James H. Ware, U.S. statistician] than face to face, QUESTIONNAIRES and
surveys involving large numbers of geo-
task analysis 1. the breakdown of a graphically dispersed participants are
complex task into component tasks to often carried out in this way. Some-
identify the different skills needed to times, telephone interviews may be used
correctly complete the task. In educa- to screen participants in order to narrow
tion, for example, it entails the break- the pool of those who will be invited for
down of a subject or field of study to in-person interviews.
identify the specific skills the student
must possess in order to master it. 2. a t e m p o r a l consistency the CORRELA-
method of evaluating a product or sys- TION between measurements obtained
tem in which researchers interview ac- when the same test or instrument is ad-
tual or target users in order to find out ministered to the same sample on two
information such as (a) what tasks are different occasions. Temporal consis-
performed, (b) which of these are most tency is an index of the RETEST RELIABIL-
frequently performed and which are ITY of an instrament. This approach
most important, (c) how and in what se- assumes that there is no substantial
quence the tasks are performed, (d) what change in the CONSTRUCT being mea-
standards of performance apply, and (e) sured between the two occasions. The
how different categories of user vaty in longer the time gap, the greater the like-
their answers to the above. Although lihood of a lower correlation. Also called
some scripted questions are asked, the temporal stability. See RELIABILITY.
interviews are otherwise unstrucrared, temporal frequency the number of
the better to reflect users' actual experi- occurrences of a repeating event per unit
ence. time. For example, if 80 repeating events
t a u test see KENDALL'S TAU. occur within 20 seconds, the frequency,
f, is 80/20 = 4.
Taylor series an infinite sum of terms
that are calculated from the values of t e m p o r a l precedence in establishing
a function's DERIVATIVES at a single cause-effect relationships between two
point a. The process provides for a Tay- variables, the principle that the cause
lor expansion of the function f(x) about must be shown to have occurred before
x = a. [Brook Taylor (1685-1731), British the effect. Two other requirements are
mathematician] those of COVARIATION and nonspu-
riousness (i.e., there are no plausible al-
Tchebechev's inequality see CHE- ternative explanations for the observed
BYSHEV'S INEQUALITY.
relationship).
t d i s t r i b u t i o n a theoretical PROBABIL-
temporal stability see TEMPORAL
ITY DISTRIBUTION that plays a central
CONSISTENCY.
role in testing hypotheses about popula-
tion means, among other parameters. It temporal v a l i d i t y a type of EXTERNAL
371
terminal event
372
test f o r two independent proportions
373
testing effect
the difference between two proportions, which some treated participants are
measured in two samples, is statistically likely to experience extreme reactions in
significant. An example is the z TEST for either direction is one that generally is
a population proportion. subject to intentional distortion of re-
sponses. Tests of extreme reactions are
testing effect the research finding that typically tests of ranks and are used
the long-term retention of information when tfiere is an indication beforehand
is significantly improved by testing learn- that the experimental condition may
ers on the information. Exams or tests cause such reactions.
seem to activate retrieval processes that
facilitate the learning of study material test of homogeneity see TEST FOR
and cause knowledge to be stored more EQUALITY OF VARIANCE.
effectively in long-term memory.
test o f significance any statistical test
test interpretation the clinical, edu- or procedure, such as a T TEST, Z TEST, F
cational, vocational, or other practical TEST, or CHI-SQUARE TEST, used in SIG-
implications and inferences given to a NIFICANCE TESTING.
particular test result. Such conclusions test o f simple effects in an ANALYSIS
are typically drawn by an expert in test- OF VARIANCE, a test to determine the ef-
ing or by suitable computer software. fect of one INDEPENDENT VARIABLE On
test i t e m a constituent part, or the the DEPENDENT VARIABLE at a Single level
smallest scoreable unit, of a test. It is the of a second independent variable; the
stimulus (question or task) to which a test examines the effects of one of the in-
test taker responds. dependent variables with the other in-
dependent variable held constant. When
test n o r m the standard of performance a statistically significant interaction is
typically attained in a test, as established found, in which, for example, there
by testing a large group of people (the are two independent variables, a and
STANDARDIZATION GROUP) and analyz- b, with each independent variable hav-
ing their scores. In NORM-REFERENCED ing two levels, a^ and 02 ^nd b^ and
TESTING, subsequent test takers' scores ^2, the question of interest turns to a
on the test are compared with the test
systematic examination of the narare of
norm to provide an estimate of the posi-
the interaction. In this case, a test of sim-
tion of the tested individual in a prede-
fined population, with respect to the ple effects involves two statistical tests:
trait being measured. one of the difference between and fl2
at i>i, and a second of the difference be-
test of association any of a category of tween and 3t ^2- The comparisons
STATISTICAL TESTS that examines the de- can also be undertaken by examining
gree of relationship or dependence be- the difference between and ^2 3t
tween variables. An example is the CHI- and the difference between bj and b2 at
SQUARE TEST. (li-
test power see POWER.-
test o f extreme reactions a proce-
dure used to test for differences in range test p r o f i l e an overall description that
between the responses (scores) of a treat- summarizes an individual's relative
ment group and those of the control standing or characteristics by collating
group. In some experiments, the treat- the findings from a series of tests or
ment is likely to increase the scores of subtests. For example, a personality pro-
some participants and at the same time file may present the data gathered on
decrease the scores of other participants; personality and other tests of interest
in contrast, the control group does not and be used to evaluate the individual in
experience such reactions. A situation in areas related to his or her personal, edu-
tetrachoric correlation coefficient
375
tetrad difference criterion
376
therapy outcome research
TRiBUTiON, and the POISSON DIS- tion. In order to gain a deeper un-
TRIBUTION. In general, the procedures of derstanding of the constructs involved,
INFERENTIAL STATISTICS involve taking the researcher samples new research
one or more EMPIRICAL DISTRIBUTIONS sites, cases, incidents, time periods, or
and referring these to an appropriate data sources to compare with those that
theoretical distribution. When there is have already been studied. In this way
correspondence between an empirical he or she seeks to build a theory from the
and a theoretical distribution, the latter emerging data while continuing to se-
may be used to make inferences (predic- lect new samples to examine and elabo-
tions) about the probability of future rate on the theory. See GROUNDED
empirical events. Also called reference THEORY.
distribution. theoretical statistics the study of sta-
theoretical frequency distribu- tistics from a mathematical and theoret-
t i o n the FREQUENCY DISTRIBUTION that ical perspective involving PROBABILITY
would result if data conformed to the THEORY, DESCRIPTIVE STATISTICS, IN-
values predicted by a theoty or law as op- FERENCES, and MODEL BUILDING. For ex-
posed to their acraal observed values. ample, a researcher could use theoretical
statistics to describe a set of achievement
theoretical p r o b a b i l i t y the mathe- data, conduct HYPOTHESIS TESTING, and
matical PROBABILITY of a particular create models assessing possible predic-
event occurring, as determined by divid- tors of achievement. Also called mathe-
ing the number of positive outcomes by matical statistics. Compare APPLIED
the total number of possible outcomes. STATISTICS.
For example, when flipping a coin, the
theoretical probabihty that a tail will theory n. 1. a principle or body of inter-
occur is 1 divided by 2 (the number of related principles that purports to ex-
possible outcomes), giving a probability plain or predict a number of interrelated
of .5 or 50%. Compare EMPIRICAL PROB- phenomena. See CONSTRUCT; MODEL. 2.
ABILITY. in the philosophy of science, a set of log-
ically related explanatoty hypotheses
theoretical relative frequency dis- that are consistent with a body of empir-
t r i b u t i o n for a discrete variable x, a ical facts and that may suggest more em-
hypothesized or expected distribution of pirical relationships. See SCIENTIFIC
observations or scores that will be ob- EXPLANATION, theoretical adj.
tained at each of the possible values of x
in relation to the total number of obser- theory-led thematic analysis a
vations to be made (i.e., their theoretical form of THEMATIC ANALYSIS in which
relative frequencies). In contrast to a a specific theoretical or epistemological
RELATIVE FREQUENCY DISTRIBUTION, approach is used to provide a direction
which indicates the relative frequencies for the analysis.
with which the values of x actually occur theory t r i m m i n g in PATH ANALYSIS or
in a chosen sample, a theoretical relative STRUCTURAL EQUATION MODELING, de-
frequency distribution specifies the leting nonsignificant paths from the
likely number of occurrences of each of model to improve model fit.
the possible values of x with reference to
the anticipated total. theory v e r i f i c a t i o n the process of de-
veloping and citing empirical evidence
theoretical s a m p l i n g a sampling to increase or bolster the tenability of
strategy, often adopted in QUALITATIVE theories.
RESEARCH, that Involves the PURPOSIVE
therapeutic t r i a l see CLINICAL TRIAL.
SAMPLING of further data while a theo-
retical framework is still under construc- therapy outcome research research
377
thick description
that investigates the end results of treat- incidental relationship between the two
ment or other interventions to which variables, or changes independently in
patients are exposed. The focus is pri- variable a and variable b. For example, as
marily on the "cure" (or not) of patients, the sales of air conditioners increase, the
but the research also evaluates their ex- number of drownings also increases: The
periences, preferences, and values, as unintentional third variable in this case
well as the wider impact on society. The would be the increase in heat. See HID-
aim is to identify shortfalls in practice DEN VARIABLE.
and to develop strategies to prevent or
mitigate problems and improve care. See thought experiment a mental exer-
also TREATMENT OUTCOME RESEARCH; cise in which a hypothesis, theoty, or
TREATMENT PROCESS RESEARCH. idea is put to the test without actually
conducting an experiment or research
t h i c k description in QUALITATIVE RE- project. The purpose is to explore the
SEARCH, a strategy that aims to describe logical consequences of a hypothesis
and interpret observed behavior within or principle. Thought experiments often
its particular context so that the behav- involve arguments about events or
ior becomes meaningful to an outsider. states of affairs of a hypothetical or
The context may be a small unit (a counterfactual nature, which neverthe-
family or work environment) or a larger less have implications for the actual
unit (a community or general culture). world. They can be used to challenge the
The researcher not only accurately de- intellectual status quo, correct misinfor-
scribes observed behavior or social mation, identityflawsin an argument,
actions but also assigns purpose, motiva- or generate ideas as part of a problem-
tions, and intentionaUty to these actions solving exercise. Thought experiments
by explaining the context within which are most famUiar in philosophy but are
they took place; thick description con- also used in the physical sciences, gener-
veys the thoughts and feelings of partici- ally as a step toward designing a physical
pants as well as the complex web of experiment. Also called idealized ex-
relationships among them. In contrast, periment.
thin description is a superficial account
that does not explore underlying mean- threats to v a l i d i t y factors that may
ings. threaten the VALIDITY of inferences
drawn from the results of an experiment
t h i r d moment see MOMENT. or research program. For example, com-
t h i r d quartile see QUARTILE. mon threats to the INTERNAL VALIDITY
of an experiment include (a) lack of clear
t h i r d variable see HIDDEN VARIABLE. TEMPORAL PRECEDENCE among vari-
third-variable p r o b l e m the fact that ables, leading to confusion of cause
an observed correlation between two and effect; (b) SAMPLING BIAS or other
variables may be due to the common nonrandom factors in the assignment of
correlation between each of the vari- participants to the different conditions
ables and a third variable rather than be- of the experiment; (c) the possibility
cause the two variables have any that events extraneous to the experi-
underlying relationship (in a causal ment, including REGRESSION TOWARD
sense) with each other. In other words, THE MEAN and naturally occurring
when two variables, a and b, are found changes over time, could cause the ob-
to be positively or negatively correlated, served effect; and (d) ATTRITION of par-
it does not necessarily mean that one ticipants, especially where differential
causes the other: It may be that changes attrition rates between groups produce
in an unmeasured or unintentional third artificial effects.
variable, c, are causing a random and co- Common threats to the converse EX-
378
three-way analysis of variance
379
three-way classification
Television violence
Stress level None Moderate High
Sugar
High 20 30 51
Low 17 15 16
No sugar
High 10 20 32
Low 10 12 8
three-way table
tors; that is, all participants receive both the behavior or action of a participant)
types of training, for both contents, and changes from one state (e.g., unwilling
are tested three times. Data from such to buy a product) to another (e.g., will-
designs often are evaluated with a three- ing to buy the product). Also called
way within-subjects analysis of variance. threshold-crossing model.
Also called three-way repeated mea-
sures design. Thurstone scaling a method for de-
veloping a scale to assess attitudes to-
threshold autoregression (TAR) an ward a single subject of interest. There
AUTOREGRESSIVE MODEL used in the are six basic steps in the general ap-
analysis of nonlinear TIME SERIES. Non- proach: (a) identify the CONSTRUCT of
linear time-series models have the ad- interest (e.g., attitudes toward providing
vantage of being able to capture public funds for charter schools); (b)
asymmetries, jumps, and time irrevers- generate statements that describe spe-
ibility in complex empirical phenom- cific attirades that people might have to-
ena, such as financial and economic ward such a construct; (c) have judges or
data (e.g., changes in interest rates). The subject matter experts rate each state-
TAR model allows for changes in the ment on a numerical scale in terms of
model parameters when the value of an how far each statement indicates a fa-
exogenous variable (i.e., one from out- vorable attitude toward public funding
side the time series of interest) crosses a for charter schools (e.g., the rating 1 =
particular threshold. extremely unfavorable attitude toward pub-
lic funding for charter schools and the rat-
threshold-crossing data in a THRESH-
ing 11 = extremely favorable attitude
OLD MODEL, data used to determine the
toward public funding for charter schools);
point in time at which a THRESHOLD EF-
(d) compute scale score values for each
FECT occurs in the dependent variable.
item, together with the mean or median
This may have to be imputed (see IM-
and some measure of VARIANCE; (e) se-
PUTATION) if observation has not been
lect the final scale items, choosing state-
continuous.
ments that are at equal intervals across
threshold-crossing model see THRESH- the range of means or mediansfor ex-
OLD MODEL. ample, select one statement for each of
the 11 mean values, with the constraint
threshold effect an effect in a DEPEND-
that there be little variance in ratings for
ENT VARIABLE that does not occur until
the statements; and (f) administer the
a certain level, or threshold, is reached
scale by asking participants to agree or
in an INDEPENDENT VARIABLE. For exam-
disagree with each statement. To get an
ple, a drug may have no effect at all until
individual's total scale score, the scale
a certain dosage level (the threshold
scores of all the items that the person
value) is reached. The change in the de-
agreed with are averaged. [Louis Leon
pendent variable can be either an irre-
Thurstone (1887-1955), U.S. psycholo-
versible dichotomous change, for
gist]
example, from "functional" to "not
functional," or it can be continuous,
such that the dependent variable t i e d ranks in a NONPARAMETRIC TEST
changes each time the independent vari- that involves ranking data, the ranks as-
able crosses the threshold in either direc- signed to two or more data points that
tion. have the same values. Such data are usu-
ally given a rank that is the mean of their
threshold model any model specify- positions in the ascending order of
ing that there is a value in a range of val- ranks. For example, suppose the data re-
ues on an INDEPENDENT VARIABLE at veal that Case 1 has the lowest value, fol-
which the DEPENDENT VARIABLE (e.g., lowed by Cases 2, 3, and 4, each of
381
time-constant covariate
which yields the same value; Cases 5 and ple, a time-lag study of intelligence
6 have the next-to-highest and highest might compare a group of people who
values, respectively. Here Cases 2, 3, and were 20 years old in 2005 with groups
4 can be awarded the tied rank of 3, as who were 20 years old in 2006, 2007,
this is the mean value of 2 + 3 + 4. and 2008. Time-lag designs have the
benefit of controlling for time-of-testing
time-constant covariate see TIME-
effects. Thus, although no differences
INDEPENDENT COVARIATE.
between ages are examined, the re-
time-dependent covariate a PREDIC- searcher can determine differences due
TOR VARIABLE whose values change over to changes in the environment over
time or the course of the study for an in- time. However, there are drawbacks of
dividual (e.g., blood pressure, dose of low INTERNAL VALIDITY and the diffi-
drug). Time-dependent variables are rel- culty in separating COHORT EFFECTS
evant in SURVIVAL ANALYSIS. Also Called f r o m AGE EFFECTS.
time-varying covariate. Compare
time-lagged correlation the correla-
TIME-INDEPENDENT COVARIATE. tion of a measure at one point in time
time-dependent Cox regression with the value of that same measure at a
model see cox REGRESSION ANALYSIS. different point in time. An example is
the correlation of IQ scores of individu-
time-homogeneous Markov chain als at 5 years of age with their IQ scores
a TIME-SERIES model in which an event's when they are 10 years of age. See STA-
probability is dependent only upon the BILITY COEFFICIENT.
immediately preceding event in the se-
ries and the transition probabilities do time-lag study a LONGITUDINAL DE-
not change over time (i.e., the probabil- SIGN in which participants are measured
ity of going from State A to State B today on two variables at two or more different
is the same as it will be at any time in points of time, with a view to determin-
the future). Also called time-homoge- ing whether one is more likely to
neous stationary chain. See MARKOV "cause" the other. Ideally, the measure-
CHAIN.
ments of the variables are separated by
the length of time it takes for the cause
time-independent covariate a PRE- to influence the effect. There are limi-
DICTOR VARIABLE that is measured at the tations to the procedure for inferring
outset of a study and whose values do causality, particularly the OMITTED-
not change over time or the course of VARIABLE BIAS or the THIRD-VARIABLE
the srady for an individual (e.g., gen- PROBLEM.
der, year of birth). Time-independent
covariates are relevant in SURVIVAL time-location s a m p l i n g a method of
ANALYSIS. Also Called time-constant finding research participants in which
covariate. Compare TIME-DEPENDENT members of a hard-to-reach target popu-
lation (e.g., homeless persons, migrant
COVARIATE.
workers) are recraited from specific loca-
time-lag design a type of QUASI- tions at which they may be found dur-
EXPERIMENTAL RESEARCH in which par- ing specific time periods when they
ticipants of the same age are compared are likely to be present. The sample is se-
at different time periods. It is typically lected in stages: (a) The SAMPLING
used in developmental, educational, FRAME is determined, comprising all of
and social psychological research to the locations at which there is sufficient
srady whether there are differences in a attendance by persons in the population
given characteristic for samples of equal of interest to make sampling worth-
age but drawn from different cohorts while, (b) A random sample of locations
measured at different times. For exam- is chosen from this frame. If attendance
382
Tobit analysis
depends on day of the week and time of obtained are age-related only). In a time-
day, a sampling period is then chosen sequential design a second age group is
for each location in the sample, (c) A added to a TIME-LAG STUDY, and two or
sample of participants is chosen, usually more cross-sectional comparisons are
randomly, during each sampling event. made at different times of testing.
Also called time-location cluster
sampling; time-space sampling; time series a set of measures on a single
venue-based sampling; venue sam- attribute, variable, or constract obtained
pling; venue time-based sampling. repeatedly over time.
time-series analysis a branch of statis-
time-reversible M a r k o v c h a i n a tics that involves the analysis of changes
MARKOV CHAIN in whlch it is not possi- in a single variable recorded repeatedly
ble to determine, given the states at a over time. The data may have an inter-
number of points in time after ranning nal stracture (such as AUTOCORRELA-
the STOCHASTIC PROCESS, which state TION, trend, or seasonal variation) that
came first and which state arrived later. should be accounted for and that pro-
time s a m p l i n g in DIRECT OBSERVA- vides input allowing for the prediction
TION, a data collection strategy that in- of furare values of the variable. Compare
volves noting and recording the CROSS-SECTIONAL ANALYSIS.
occurrence of a target behavior when- time-series data a sequence of mea-
ever it is seen during a stated time inter- surements taken at successive times sep-
val. The process may involve fixed time arated by specified tim^ intervals. The
periods (e.g., evety 5 minutes) or ran- data have a natural temporal ordering
dom time intervals. For example, a re- (e.g., chronological order).
searcher may observe a group of
children for 10 seconds evety 5 minutes time-series design an experimental
for a specific 30-minute period each day, design that involves the observation of
noting the occurrence or nonoccurrence units (e.g., people, countries) over a de-
of particular behaviors (overt actions). fined time period. Data collected from
Observations taken during these periods such designs may be evaluated with
are known as time samples. An individ- TIME-SERIES ANALYSIS.
ual score is made on the basis of one or time-series plot a graph on which
more of the following: (a) number of TIME-SERIES DATA are connected by
time units in which the defined behav- straight lines, used to show the changes
ior occurs, (b) total frequency of occur- of the measured variable (y-axis) relative
rence of the defined behavior in the to the different measurement times (x-
total observational time, and (c) average axis).
frequency of the defined behavior per
unit of time. See PARTIAL-INTERVAL RE- time-varying covariate see TIME-
CORDING; WHOLE-INTERVAL RECORD- DEPENDENT COVARIATE.
ING. Tobit analysis a type of REGRESSION
time score a score based on the amount ANALYSIS used when a DEPENDENT VARI-
ABLE with values above or below a cer-
of time used to , complete a particular
task. An example is the number of min- tain threshold takes on the value of that
utes a 3-year-old child requires to solve a threshold; that is, the exact value of the
simple puzzle. variable is unknown or ignored (see CEN-
SORED DATA). For exarnple, in an analy-
time-sequential design an experi- sis of aptitude in coUege students, the
mental design to separate AGE EFFECTS dependent variable could be a score on a
from time of measurement and COHORT standardized aptitude test that has an
EFFECTS (i.e., to determine if the results upper limit of 500 (the threshold value).
383
tolerance
Sradents who answer all questions on will be computed to describe the sample,
the test correctly receive a score of 500, such as the MEAN. Relatedly, each indi-
even though these sradents may not vidual condition has its own DEGREES OF
have equal aptitude. In such a case the FREEDOM as well: 20 - 1 = 19.
influence of INDEPENDENT VARIABLES,
t o t a l effect in the study of causal ef-
such as reading and math scores, on the
fects, the total extent to which the de-
dependent variable of academic apti-
pendent (or outcome) variable is
rade is more appropriately sradied with
changed by the independent (or predic-
Tobit regression analysis than with ordi-
tor) variable, including any indirect ef-
nary LEAST SQUARES REGRESSION be-
fect through a MEDIATOR. In a simple
cause the threshold values would bias
example, if the independent variable, x,
the slope of the obtained REGRESSION
is presumed to cause the outcome vari-
LINE. Also called censored regression;
able, y, the PATH COEFFICIENT of this di-
Tobit model. [James Tobin (1918-
rect effect. A, is the total effect. If there is
2002), U.S. economist]
an intervening variable, linked by two
tolerance n. permissible or allowable path coefficients, B and C, this indirect
deviation from a specified value or stan- effect is BC, and the total effect is A + BC.
dard. tolerant adj. See also PATH ANALYSIS.
tolerance i n t e r v a l a range of values total p r o b a b i l i t y l a w a law relating
within which, with some probability, a to BAYES THEOREM and conditional
specified proportion of a population probabilities, stating that, for two
falls. For instance, one may be 95% con- events, A and B, the probability of A oc-
fident that 90% of the population will curring is equal to the probability that
fall within the range specified by the tol- both A and B occur plus the probability
erance interval. It differs from a CONFI- that A occurs and B does not.
DENCE INTERVAL, which bounds a single total sum of squares (symbol: T55) in
population parameter (e.g., the mean or ANALYSIS OF VARIANCE and REGRESSION
variance). ANALYSIS, the SUM OF SQUARES due to
tolerance l i m i t the upper or lower the effects of treatment plus the sum of
value of a TOLERANCE INTERVAL. squares due to error.
topastic error o f measurement the t o t a l variance in ANALYSIS OF VARI-
partly random, partly systematic error or ANCE and REGRESSION ANALYSIS, the
variance that results when the taker of a variability that is due to the effects of
multiple-choice test has the opportunity treatment (TRUE VARIANCE) plus the
to get some of the answers correct by variability that is due to error (ERROR
VARIANCE).
guessing.
top-down clustering (top-down hi- trace n. in multivariate statistics and lin-
ear algebra, the sum of the elements on
erarchical clustering) see DIVISIVE
the MAIN DIAGONAL of a SQUARE MA-
CLUSTERING.
TRIX.
t o t a l degrees of freedom the total
t r a i n i n g sample a representative sam-
number of observations in an analysis
ple of a larger population that is typi-
minus one. For example, in an experi-
cally used to develop a model or to
ment in which there are four conditions,
obtain conditional probabilities in
with 20 participants randomly assigned
to each condition, there are 80 inde- BAYESIAN analysis.
pendent observations: The total degrees t r a i n i n g set a portion of a TIME SERIES
of freedom is 79 (80 -1) observations, as used to predict furare values, that is,
there is one constraint on any value that to train (fit) a model for prediction.
384
transformed score
Training sets are used in artificial intel- different traits. Also called aptitude-
ligence (e.g., for neural networks), sta- treatment interaction (ATI).
tistical modeling, and other areas of in-
formation science. t r a i t v a l i d i t y the degree to which a test
is capable of measuring a trait. A test has
t r a i t n. 1. an enduring personality char- trait validity if it demonstrates an associ-
acteristic that describes or determines an ation between the test scores and the
individual's behavior across a range of prediction of an underlying trait. See
situations. 2. in ITEM RESPONSE THEORY, also MULTITRAIT-MULTIMETHOD MATRIX.
an individual's level of competence on a
certain task or aptitude measurement. transfer f u n c t i o n model in TIME-
SERIES ANALYSIS, atypeof model used to
t r a i t p r o f i l e a graphic display of test forecast a time series that is influenced
scores in which each score represents an by present and past values of other time
individual TRAIT. These scores or ratings series.
are often arranged on a common scale to
enable them to be interpreted quickly. t r a n s f o r m a t i o n n. the conversion of
The hypothetical example below shows data to a different form through a rule-
a respondent's results (given in T SCORES) based, usually mathematical process, for
on a personality quesfionnaire assess- example, changing Fahrenheit to Cel-
ing neuroticism, extraversion, openness, sius. In statistics, a RAW SCORE is of-
agreeableness, and conscientiousness. ten transformed into a STANDARDIZED
Also called psychogram. See TEST PRO- SCORE for purposes of comparison. See
FILE. also LINEAR TRANSFORMATION; NONLIN-
EAR TRANSFORMATION. transform
t r a i t r a t i n g a technique in which a vb. transformational adj.
given behavioral feature or trait (e.g., a
character trait or attribute) is observed, t r a n s f o r m e d score a score that has
rated, and recorded. been changed into another scale to
allow direct comparison and meaning-
trait-treatment interaction (TTI) ful interpretation with other scores. For
the interaction between traits (e.g., gen- example, a RAW SCORE of 44 on a first
der, aptitude) and conditions of treat- test might not mean the same thing as a
ments (e.g., methods of instruction) as it raw score of 44 on the second test. The
affects a dependent variable (e.g., scores transformation could be to percentages:
on an academic test). Using ANALYSIS 44 on the first test could be out of 100
OF VARIANCE Or MULTIPLE REGRESSION items, and thus represent 44%, while on
analysis, it is possible to determine the the second test it could be out of 50
best type of treatment for people with items, and thus have a transformed
llll
E O
Trait
A C
trait profile
385
transient state
Subsequent behavior
Initial behavior 1 2 3 4 5 6 7 8 9 10
I. Change talk .24 .06 .16 .11 .49 .36 .04 .04 .01 .30
2. Question positive .15 .15 .17 .30 .32 .25 .02 .04 .02 .25
3. Question negative .06 .02 .08 .10 .22 .09 .02 .01 .01 .32
4. Question neutral .12 .02 .36 .32 .15 .05 .01 .01 .01 .02
5. Simple reflection .09 .05 .28 .02 .10 .28 .00 .02 .00 .02
6. Complex reflection .26 .36 .32 .02 .12 .31 .00 .00 .00 .02
7. Raise concem .04 .06 .40 .10 .04 .02 .00 .03 .02 .10
8. Give affirmation .10 .06 .26 .09 .03 .02 .00 .01 .03 .09
9. Give advice .07 .03 .01 .01 ,02 .01 .01 .01 .00 .11
10. Other .06 .01 .03 .02 .02 .04 .01 .01 .00 .50
transition matrix
386
treatment variability
387
treatment variable
Successful Correct -
retrieval recall
Sufficient
storage
1-9,
Unsuccessful. Incorrect
retrieval recall
1-e.
Insufficient Incorrect
storage recall
tree diagram
389
true zero
390
two-by-two factorial design
391
two-by-two table
tied out using the predicted values from DIRECTIONAL HYPOTHESIS). For a signifi-
the first-stage regression analysis in cance value of .05, .025 marks the CRITI-
place of the original dependent variable. CAL REGION at each end of the
The new dependent variable will be un- distribution.
correlated with the error term of the two-tailed test see NONDIRECTIONAL
original dependent variable because it is TEST.
a function of the independent variables.
two-way analysis o f covariance an
two-stage s a m p l i n g see MULTISTAGE ANALYSIS OF COVARIANCE in which
SAMPLING. there are two INDEPENDENT VARIABLES
two-Stage stopping rule in a CLINI- and a COVARIATE whose effects the re-
CAL TRIAL comparing two treatments, a searcher wishes to bring under statistical
strategy in which results are examined control.
after only a fraction of the planned two-way analysis o f variance an
number of participants in each group ANALYSIS OF VARIANCE that Isolates the
have completed the trial (usually either MAIN EFFECTS of two Independent vari-
half or two thirds of the patients). At this ables, a and b, and their INTERACTION
point, the test statistic is computed and EFFECT, a X b, on a dependent variable.
the trial stopped if the difference be- Also called two-factor analysis of
tween treatment means is significant at variance.
the SIGNIFICANCE LEVEL Set for this first
stage. If not, the remaining participants two-way chi-square test a CHI-
in each group are studied, the test statis- SQUARE TEST used to determine whether
tic is recomputed, and the means com- there is a significant relationship be-
pared at a significance level set for the tween the variables summarized in a
second stage. The significance levels for TWO-WAY TABLE. Also called two-by-
the two stages should be such that they two chi-square test.
equal the overall significance level for two-way design atypeof FACTORIAL
the trial, usually .05 or .01. DESIGN in which two INDEPENDENT
VARIABLES are manipulated. Also called
two-tailed alternative hypothesis two-factor design.
see N O N D I R E C T I O N A L H Y P O T H E S I S .
two-way f a c t o r i a l design see TWO-
two-tailed confidence i n t e r v a l in BY-TWO FACTORIAL DESIGN.
statistical testing, a CONFIDENCE INTER-
VAL that specifies both upper and lower two-way interaction in a TWO-WAY
limits to the population parameter (see ANALYSIS OF VARIANCE, the jolnt effect
CONFIDENCE LIMIT). of both independent variables, a and b,
on a dependent variable. A statistically
two-tailed hypothesis see NON- significant two-way interaction indi-
DIRECTIONAL HYPOTHESIS. cates that there are differences in the in-
two-tailed p value in a NONDIREC- fluence of each independent variable at
TIONAL TEST of significance, a probability
their different levels (e.g., the effect of]
value for making a TYPE TERROR that and at b^ is different from the effect of
falls below the SIGNIFICANCE LEVEL for fll and fl23t b2). See also HIGHER ORDER
the test, leading to the rejection of the INTERACTION.
NULL HYPOTHESIS. In Other words, as- two-way m i x e d design an experi-
suming the null hypothesis to be true, mental design in which there are two in-
the p value is the probability that ran- dependent variables of interest, where
domly selected samples would have one variable is a BETWEEN-SUBJECTS FAC-
means that are different, with either sam- TOR (e.g., gender) and the other variable
ple having the larger mean (see NON- is a WITHIN-SUBJECTS FACTOR (e.g., pre-
393
two-way repeated measures design
test and posttest scores for the same par- called the POWER of the test, where
ticipants). Data from such designs may power =1-3. Also called beta error.
be evaluated with a two-way mixed-
design analysis of variance. Type III error 1. the error that occurs
when there is a discrepancy between the
two-way repeated measures design research focus and the hypothesis acra-
see TWO-WAY WITHIN-SUBJECTS DESIGN. ally tested. For example, a Type 111 error
two-way table a table in which the would be committed if a researcher col-
joint FREQUENCY DISTRIBUTION Of tWO lected data on INDIVIDUAL DIFFERENCES
within a sample and determined the
INDEPENDENT VARIABLES is arrayed. See
causes of variation when the question of
also TWO-BY-TWO TABLE.
interest concerned differences between
two-way within-subjects design a populations. In other words, a Type III
type of WITHIN-SUBJECTS DESIGN in error involves providing the right an-
which there are two variables or factors swer for the wrong question. Also called
of interest, a and b, and the same partici- Type 0 error. 2. the error that occurs
pants are studied in each of the levels of when a researcher correctly rejects the
each of the factors. For example, sup- NULL HYPOTHESIS of no difference be-
pose a and b representtypeof training tween samples but then makes an incor-
(lecture vs. reading) and type of content rect inference about the direction of the
(physics vs. psychology), respectively. difference. Researchers investigating the
All participants receive all treatment direction of a relationship (e.g., "Which
conditions of both factors; that is, all is more?" or "Which is better?") will
participants receive bothtypesof train- make a Type III error if they use a NON-
ing and are exposed to both contents. DIRECTIONAL TEST to make a directional
Data from such designs may be evalu- decision. 3. in clinical tests, attributing a
ated with a two-way within-subjects lack of results to the weakness of a treat-
analysis of variance. Also called two- ment when, in fact, the problem was
way repeated measures design. that the treatment was not administered
properly.
Type 0 error see TYPE III ERROR.
Type I error the error of rejecting the Type I sum of squares the reduction
in the ERROR SUM OF SQUARES obtained
NULL HYPOTHESIS when it is in fact true.
Investigators make this error when they by adding a particular factor (e.g., Factor
believe they have detected an effect or a A) to a flt that already includes the other
relationship that does not actually exist. factors in the model (e.g., Factors B, C,
The projected probability of committing and D): In other words, it is the amount
a Type I error is called the SIGNIFICANCE of the sum of squares attributable only
LEVEL or alpha (a) level. Also called to Factor A. A Type I sum of squares de-
alpha error; rejection error. pends on the order in which factors are
listed in the model statement. For some
Type II error the error of failing to re- data sets, there would be different results
ject the NULL HYPOTHESIS when it is in for a model that states B then A than for
fact not true. Investigators make this a model that states A then B; this is be-
error if they conclude that a particular cause each factor is adjusted only for the
effect or relationship does not exist factors that were entered before it. If one
when in fact it does. The probability of flts two models, one with A then B, the
committing a Type II error is called the other with B then A, not only can the
beta (P) level of a test. Conversely, the Type 1 sum of squares for Factor A be dif-
probability of not committing a Type II ferent under the two models, but there
error (i.e., of detecting a genuinely sig- is no way to predict whether the sum
nificant difference between samples) is of squares will go up or down when A
Type III sum of squares
395
ultimate cause the underlying reason RANDOM ERROR and therefore self-
for some observed result or event, as canceling in the long run.
compared to the PROXIMATE CAUSE di-
unbiased error see RANDOM ERROR.
rectly preceding it. For example, in a
study of the link between a social factor, unbiased estimator a quantity calcu-
such as socioeconomic status, and an lated from sample data whose value is
outcome behavior, such as health be- representative of the true quantity in the
havior, the data may show that lower so- larger population. In other words, when
cioeconomic groups practice poorer data from samples are used to make in-
health behaviors. However, the ultimate ferences about unknown quantities (pa-
cause may be that lower socioeconomic rameters) in populations, an unbiased
groups are subject to greater hazard or estimator is one that over repeated sam-
environmental harm, which inramdis- pling has an average equal to the true
courages healthy behavior. Also called value of the parameter in the popula-
distal cause; remote cause. tion. For example, an unbiased estima-
tor of variance provides an accurate
U M P test abbreviation for UNIFORMLY index of the variability of measurements
MOST POWERFUL TEST. for a given phenomenon in a given pop-
unbalanced design an experimental ulation of interest based on calculations
design having multiple INDEPENDENT made from the SAMPLE VARIANCE. Also
VARIABLES in which the number of mea- called unbiased statistic. Compare BI-
surements or observations obtained ASED ESTIMATOR.
is different for each condition under
unbiased s a m p l i n g selecting individ-
study. Although BALANCED DESIGNS
uals for a study using a process that
generally are preferred for their greater
yields a group exemplitying the larger
POWER in statistical analyses, unbal-
population from which it derives. In
anced designs nonetheless may arise due
practice, no strategy produces a com-
to participant ATTRITION or other un-
pletely unbiased sample but RANDOM
avoidable factors. For example, if a re-
SAMPLING yields a good approximation,
searcher is investigating how sleep and
as it introduces the minimum possible
diet influence academic performance
amount of error in representing the pop-
and only 35 of the 40 college undergrad-
ulation. Compare BIASED SAMPLING.
uates recraited to participate were able
to do so through the full term of the pro- unbiased statistic see UNBIASED ESTI-
ject, certain data will be missing for the MATOR.
five people who left the study prema-
rarely and the design will become unbal- uncertainty analysis in an experi-
anced. ment or study, an analysis used to assess
the accuracy of measurements and
unbalanced l o n g i t u f l i n a l data see model predictions, taking into account
BALANCED LONGITUDINAL DATA. such possible sources of error as instm-
mentation, methodology, and the pres-
unbiased adj. impartial or without net ence of CONFOUNDS.
error. In unbiased procedures, studies,
and the like any errors that do occur are u n c o n d i t i o n a l model a type of HIER-
396
underidentified model
ARCHICAL LINEAR MODEL that amalgam- degrees then education would be identi-
ates the effect of explanatory or PRE- cal for all individuals and thus serve as a
DICTOR VARIABLES across the different CONTROL VARIABLE.
levels of analysis. For example, imagine
uncorrected moment a MOMENT cal-
a study having three levels of analysis:
culated on the basis of the raw data in a
students (Level 1), nested within class-
distribution, rather than from the same
rooms (Level 2), further nested within
data after CORRECTION for error or some
schools (Level 3). An unconditional
unwanted effect. The term is often ap-
model would represent variation in the
plied to moments calculated from a
outcorne measure or DEPENDENT VARI-
GROUPED FREQUENCY DISTRIBUTION (as
ABLE across the three levels (sradent,
in a corrected moment).
classroom, and school), whereas a CON-
DITIONAL MODEL would explain the in- underdetermination n. a situation in
dividual influence of each predictor which the available evidence is insuffi-
variable on the dependent variable. cient to know what beliefs to hold in re-
sponse. For example, if one knows that a
u n c o n d i t i o n a l variance the long- person spent $10 on apples and oranges
term trend in a sequence of observations and that apples cost $1 while oranges
made over time. Unconditional variance cost $2, then one knows that the person
implies that knowledge of individual did not buy six oranges but not whether
fluctuations between previous observa- the person bought one orange and eight
tions is irrelevant to identifying and un- apples, two oranges and six apples, and
derstanding the overaU variability of the so forth. underdetermined adj.
sequence as a whole! Compare CONDI-
TIONAL VARIANCE. underestimation n. a situation in
which systematic discrepancies between
u n c o n f o u n d e d comparison a com-
measurements and model outputs lead
parison of two or more groups of re-
researchers to calculate a PARAMETER of
search participants that yields an
interest as being significantly lower than
unbiased estimate of the effect of the
its true value. For example, applying a
treatment or other condition under in-
REGRESSION ANALYSIS In whlch there are
vestigation. For comparisons to be un-
BIASED ESTIMATORS could lead a medical
confounded, studies must be designed
researcher to obtain a lower value for the
to ensure identical handUng of all partic-
presence of a particular illness in the
ipants, in addition to their RANDOM AS-
general population than is acraally the
SIGNMENT to different experimental
case.
groups, such that any differences between
the groups may be attributed solely to u n d e r i d e n t i f i c a t i o n n. a situation,
the experimental manipulation rather such as may occur during STRUCTURAL
than the influence of other related factors. EQUATION MODELING, in whlch it Is not
Compare CONFOUNDED COMPARISON. possible to estimate all of the model's
parameters based on the sample data
uncontrolled variable a characteris-
being analyzed. Compare OVERIDENTI-
tic factor that is not regulated or mea-
FICATION.
sured by the investigator during an
experiment or study, such that it is not underidentified model a theoreti-
the same for all participants in the re- cally identified model in which one or
search. For example, if the investigator more PARAMETER esfimates cannot be
collects data on participants having made because of an insufficient number
varying levels of education, then educa- of available data points. This may occur,
tion is an uncontrolled variable. If the for example, when there is high
investigator,' however, were to collect MULTICOLLINEARITY among the vari-
data only on participants with college ables studied. Empirically underidenfi-
397
underlying dimension
fied parameters are vety unstable. from a larger population in which differ-
Compare JUST-IDENTIFIED MODEL; OVER- ent units have different likelihoods of
IDENTIFIED MODEL. being chosen for inclusion. In SAMPLING
u n d e r l y i n g dimension an explana- WITHOUT REPLACEMENT, for example,
tory or descriptive concept, inferred each time a unit is chosen to be part of
from empirical evidence or theory, that the sample it is removed from further
is used to interpret observed outcomes. consideration, thus altering the likeli-
For example, researchers sradying infant hood of inclusion for the remaining
attachment may conclude that the dif- units. This contrasts with equal proba-
ferent attachment styles demonstrated bility sampling, in which all population
can be explained by two underlying di- units have the same likelihood of being
mensions of avoidance and anxiety. In selected for the sample.
FACTOR ANALYSIS and Other contexts, unexplained variance see ERROR
such underlying dimensions are called V A R I A N C E .
LATENT VARIABLES.
u n f o l d i n g n. a unidimensional SCALING
u n d e r l y i n g d i s t r i b u t i o n the THEO- procedure in which respondents evalu-
RETICAL DISTRIBUTION fot a given popu- ate a set of items and their choices are
lation of interest. used to construct a continuum along
underspecified model in CAUSAL which their relative preferences are
MODELING, a predictive formula from placed. For example, consider a scale
which relevant explanatory variables measuring attitudes toward marijuana
have been excluded, such that the effect sales. A person who wholeheartedly fa-
estimates it provides are both biased vors the item "The city should legalize
and inefficient. See OMITTED VARIABLE the sale of marijuana" would be located
BIAS. at a different point along the continuum
from a person who completely opposes
undirected graph a display in which the sale of marijuana, and a person who
various points (nodes) of information endorses the item to some extent would
are connected by lines (edges) having be located at yet another point along
no direction. The value between two the continuum. See also MULTIDIMEN-
nodes often is called a WEIGHT. Thus, SIONAL UNFOLDING.
the following example has several nodes
(e.g.. A, B, F, E) and several weights (e.g., unidimensionality n. the quality of
5, 9, 12, 18). measuring a single construct, trait, or
other attribute. For example, a uni-
dimensional personality scale, attitude
scale, or other scale would contain items
related only to the respective concept of
interest. Compare MULTIDIMENSIONAL-
ITY. unidimensional adj.
u n i f o r m association model a model
for a CONTINGENCY TABLE in which a
constant ODDS RATIO is assumed across
all of the component cells.
In contrast, a DIRECTED GRAPH uses ar- u n i f o r m d i s t r i b u t i o n a theoretical
rows or numbered nodes to show CONTINUOUS DISTRIBUTION in which
directionality, such as A leading to B, the probability of occurrence is the same
which in turn leads to E, and so forth. for all values of x, represented by f(x) =
unequal p r o b a b i l i t y s a m p l i n g a l/(b -fl),where a is the lower limit of the
strategy for selecting a sample of units distribution and b is its upper limit. For
398
unit-specific model
example, if a fair die is thrown, the prob- the system. It is given as 1 - hj^, where hj^
ability of obtaining any one of the six is the COMMONALITY of the jth variable.
possible outcomes is 1/6. Since all out- Each of the observed variables in the
comes are equally probable, the distribu- data set being analyzed can be expressed
tion is uniform. If a uniform distribution as a combination of a common factor
is divided into equally spaced intervals, shared among all variables plus a unique
there will be an equal number of mem- factor associated with a measurement
bers of the population in each interval. enor or another specific, individual source
Also called rectangular distribu- of variation.
tion.
u n i t m a t r i x see IDENTITY MATRIX.
uniformly most powerful test
u n i t n o r m a l d i s t r i b u t i o n see STAN-
(UMP test) a statistical test of one hy-
DARDIZED DISTRIBUTION.
pothesis against another that has the
greatest POWER among all tests available u n i t n o r m a l variable see STANDARD
at a given ALPHA value. Suppose the two NORMAL VARIABLE.
hypotheses are the NULL HYPOTHESIS of
no difference between two groups (Sam- unit of analysis in research, the group
ple 1 mean = Sample 2 mean) and the of people, things, or entities that are
ALTERNATIVE HYPOTHESIS of some dif- being investigated or studied. For exam-
ference between the groups (Sample 1 ple, in organizational contexts, data can
mean * Sample 2 mean). Since the latter be collected from employees, who in
actually is a composite of several possi- turn are part of departments, which in
bilities, a UMP test is one defined by a turn are part of the larger organization,
CRITICAL REGION that is best suited to which may have multiple sites in several
test the null hypothesis against each of countries. The unit of analysis chosen
the simple hypotheses comprising the influences the methodological and ana-
larger alternative hypothesis. lytical procedures used (e.g., studying
groups within organizations may re-
u n i m o d a l d i s t r i b u t i o n a set of scores quire a HIERARCHICALLY NESTED DE-
with a single peak, or MODE, around SIGN). Also called analysis unit; level
which values tend tofluctuate,such that of analysis.
the frequencies at first increase and then
decrease. See also BIMODAL DISTRIBU- unit root test a statistical procedure
TION; MULTIMODAL DISTRIBUTION. that uses an AUTOREGRESSIVE MODEL to
determine whether TIME-SERIES DATA
u n i p o l a r r a t i n g scale a type of instru- exhibit systematic trends, which must
ment that prompts a respondent to eval- be mathematically removed before any
uate the degree to which a single quality analyses may be conducted. Two com-
or attribute is present. For example, con- monly used unit root tests are the
sider a five-point scale with the foUow- Dickey-Fuller test and the Phillips-
ing anchors or benchmarks: (1) not at all Perron test.
satisfied, (2) slightly satisfied, (3) moder-
ately satisfied, (4) very satisfied, and (5) unit-specific model a HIERARCHICAL
completely satisfied. Since there is no an- LINEAR MODEL that aims to identify those
chor that represents the opposing qual- outcomes specifically associated with a
ity of dissatisfaction, the scale has one parficular level, magnitude, or categoty
pole. Also called unipolar scale. Com- of the independent variable being ana-
pare BIPOLAR RATING SCALE. lyzed. For example, a researcher may be
investigating sradents, who are nested
uniqueness n. in FACTOR ANALYSIS, that within classrooms, who are further
part of the variance of a variable that it nested within schools. A unit-specific
does not share with any other variable in model would address questions about
399
unity
cause relevant variables have not been color. Unselective observation is a more
included in the causal systems investi- objective strategy than SELECTIVE OB-
gated, a problem generally referred to SERVATION.
as OMITTED VARIABLE BIAS. See THIRD-
unstandardized score see RAW
VARIABLE PROBLEM.
SCORE.
unnumbered graphic rating scale
see VISUAL ANALOGUE SCALE. unstructured i n t e r v i e w an inter-
view that is highly flexible in terms of
unobtrusive measure a measure ob- the questions asked, the kinds of re-
tained without disturbing the partici- sponses sought, and the ways in which
pant or alerting him or her that a the answers are evaluated across inter-
measurement is being made. For exam- viewers or across interviewees. For ex-
ple, a researcher may observe passersby ample, a human resource staff member
in a public park from a nearby cafe and conducting an unstructured interview
document their activities. The behavior with a candidate for employment may
or responses of such participants are ask open-ended questions so as to allow
thus assumed to be unaffected by the in- the spontaneity of the discussion to re-
vestigative process or the surrounding veal more of the applicant's traits, inter-
environment. Also called concealed ests, priorities, and interpersonal and
measure; nonreactive measure. verbal skills than a standard predeter-
Compare REACTIVE MEASURE. mined question set would. Also called
unobtrusive observation the process nondirective interview. Compare
STRUCTURED INTERVIEW.
of coUecting UNOBTRUSIVE MEASURES,
such as by the use of hidden cameras.
unstructured stimulus a vague,
u n p l a n n e d comparison see POSTpoorly organized, and not clearly identi-
HOC COMPARISON. fiable stimulus, such as an inkblot in the
Rorschach Inkblot Test. The perception
unpredicted variance see ERROR VAR-
of unstmctured stimuli is often thought
IANCE.
to be influenced more by the character-
unrelated t test see INDEPENDENT- istics of the perceiver than by those
SAMPLES T TEST. of the stimulus. Compare STRUCTURED
unrestricted model see FULL MODEL. STIMULUS.
unsaturated model see SATURATED unsystematic error see RANDOM
MODEL. ERROR.
401
up-and-down method
a research study, there are no longer the extreme right on the x-axis in the
equal numbers of cases in each experi- plot of a FREQUENCY DISTRIBUTION, de-
mental condition. An unweighted means noting the probabUity of obtaining the
analysis takes this inequality into ac- highest value in the distribution. Upper-
count by applying an adjusted sample tail probabilities are useful when con-
size value across all conditions, thus en- ducting statistical tests of experimental
suring that data from all experimental hypotheses. Compare LOWER-TAIL PROB-
groups contribute equally to the deter- ABILITY.
mination of treatment effects. Also
called unweighted means proce- upper w h i s k e r in a BOX-AND-WHISKER
dure. PLOT, the line extending from the value
at the 75th PERCENTILE to the largest
up-and-down method a strategy in value within one INTERQUARTILE RANGE
which the administration of a stimulus of that percentile score. Compare
event or item depends on a participant's LOWER WHISKER.
response to the previous item. The up-
and-down method employs fixed, dis- u p w a r d P y g m a l i o n effect an effect
crete levels and sequential transition in which the expectations of followers
rales of moving one level up, moving or subordinates lead to behavior on the
one level down, or remaining at the cur- part of the leader or superior that is con-
rent level. For example, in PSYCHO- sistent with these expectations. The be-
PHYSICAL RESEARCH, a participant who havior of the leader does not reflect his
correctly detects a stimulus during one or her true abilities or personality traits,
trial would be presented with a reduced but rather the perception of the leader
value of the same stimulus in the next by subordinates. Compare PYGMALION
trial; conversely, if he or she did not de- EFFECT. See SELF-FULFILLING PROPHECY.
tect the original stimulus its value would
u r n model a THOUGHT EXPERIMENT in
be increased in the next trial. Thus, the
which objects of interest (e.g., people,
stimulus value that is presented varies
events) are represented as colored balls
sequentially until a stable estimate is ob-
placed in an urn or other container. In
tained of the individual's sensitivity to
imagination, the experimenter ran-
the stimulus. Similarly, in computer
domly removes one "baU" from the urn,
adaptive testing there is a finite set of
notes its color, and places it back before
possible items with known difficulty lev-
repeating the process; the goal is to de-
els, and depending on the individual's
termine the probability of drawing one
response to the initial item the next item
color or another. The urn model is a con-
administered wiU have the same diffi-
venient way to calculate certain basic
culty or be slightly more or less difficult.
probabilities using CONDITIONAL PROB-
The process continues until an estimate
ABILITIES.
of the individual's ability on the trait
being measured can be obtained. V-shaped d i s t r i b u t i o n a graphical
representation of a FREQUENCY DISTRI-
upper hinge the point in a distribution BUTION that is shaped more or less like
of values above which lie one fourth of the letter U, with the maximum fre-
the data and below which lie the other quencies at both ends of the range of the
three fourths of the data. It is equivalent variable. For example, the number of
to the third QUARTILE. Compare LOWER people infected by the flu each year may
HINGE.
have a U-shaped distribution by age,
upper quartile see QUARTILE. with those who are very young or very
old having the highest frequency of oc-
upper real l i m i t see REAL LIMIT.
currence, as shown in the hypothetical
upper-tail p r o b a b i l i t y the value at illustration opposite.
402
U-shaped distribution
10,000
7,500
w
ra
O
5,000
o
2,500
"T" I
20 40 60 80 100
Age (years)
U-shaped distribution
In an inverted U-shaped distribution, die of the distribution and the least fre-
the most frequent values are in the mid- . quent values are at the extremes.
403
Vv
vague p r i o r in BAYESIAN statistics, an and the legitimacy of included values
imprecise probability specification ap- (e.g., impossible values are not present),
plied in the estimation of a population confirming that information is consis-
parameter when more conclusive infor- tent across records, and confirming that
mation about its true value is lacking. no records are missing.
For example, a researcher may use a
mean obtained previously for another v a l i d i t y coefficient an index, typi-
population to constract a CONFIDENCE cally a CORRELATION COEFFICIENT, that
INTERVAL within which the mean for reflects how well an assessment instru-
the current population of interest may ment predicts a well-accepted indicator
be considered to lie. of a given concept or criterion. For ex-
ample, if a measure of criminal behavior
validation n. the process of establishing is valid, then it should be possible to use
the trath or logical cogency of some- it to predict whether a person (a) will be
thing. An example is determining the arrested in the future for a criminal vio-
accuracy of a research instrument in lation, (b) is currently breaking the law,
measuring what it is designed to mea- and (c) has a previous criminal record. A
sure. validate vb. validity coefficient could be used to re-
late scores on the measure to each of
validity n. the degree to which empiri-
these criteria and thus determine how
cal evidence and theoretical rationales
useful the measure actually is for behav-
support the adequacy and appropriate-
ioral forecasting.
ness of conclusions drawn from some
form of assessment. Validity has multi- valiflity criterion an external concept
ple forms, depending on the research or standard of comparison that is used to
question and on the particular type of define the attribute an instrament is pur-
inference being made. For example, the ported to measure and that is applied in
three major types of test validity are CRI- estimating how well the measurement
TERION VALIDITY, based on correlation instrament acraally fulfills its intended
with an accepted standard; CONSTRUCT purpose. See CRITERION VALIDITY.
VALIDITY, based on the concepraal vari-
able underlying a test; and CONTENT VA- v a l i d i t y d i a g o n a l the MAIN DIAGONAL
LIDITY, based on the subject matter of a in a MULTITRAIT-MULTIMETHOD MA-
test. Other forms of validity prominent TRIX, which represents correlations be-
in the social sciences include ECOLOGI- tween the same construct or trait
CAL VALIDITY, EXTERNAL VALIDITY, IN- measured using different methods. See
TERNAL VALIDITY, and STATISTICAL CON- CONVERGENT VALIDITY.
CLUSION VALIDITY. See also THREATS TO
VALIDITY, valid adj. v a l i d i t y generalization the use of
META-ANALYSIS and other statistical pro-
v a l i d i t y check the process of verifying cedures to assess the evidence of a test's
that a data set is free of errors and ad- adequacy and appropriateness in multi-
heres to standard or intended rales. Per- ple siraations and settings. Validity gen-
formed manually or using software, a eralization typically involves correcting
validity check may involve such things all of the correlations being examined
as verifying the accuracy of calculations for methodological and statistical limi-
variance explained
tations and flaws and providing esti- PARISONS can be made with equal preci-
mates of correlations or results that sion. In other words, the amount of vari-
would have been obtained in the ab- ation or inaccuracy involved when
sence of such limitations. estimating the average effect for one
treatment will be the same as that in-
value analysis a type of CONTENT
volved when estimating the average ef-
ANALYSIS of written material consisting
fect for all other treatments.
of a table, or other systematic notation,
documenting the frequency of appear- variance components analysis any
ance in the material of all expressions re- statistical procedure for examining
ferring to specified values. MIXED-EFFECTS MODELS and RANDOM-
EFFECTS MODELS that decomposes the
v a r i a b i l i t y n. the degree to which
members of a group or population differ total variance on a DEPENDENT VARIABLE
from each other, as measured by such into that stemming from the effects of
statistics as the RANGE, STANDARD DEVI- the INDEPENDENT VARIABLE (the TRUE
ATION, and VARIANCE. VARIANCE) and that produced by extra-
neous factors (the ERROR VARIANCE). An
variable n. a condition in an experi- ANALYSIS OF VARIANCE provides an ex-
ment or characteristic of an entity, per- ample. Assume a researcher is studying
son, or object that can take on different whether caffeine and gender influence
categories, levels, or values and that can students' test performance. The former
be quantified (measured). For example, is a random variable, as it has numerous
test scores and ratings assigned by judges possible levels from which to choose a
are variables. Numerous types of vari- subset to study (e.g., 0 mg, 50 mg, 100
ables exist, including CATEGORICAL mg, 150 mg), and the latter is a flxed
VARIABLES, DEPENDENT VARIABLES, IN- variable, as it may assume only one of
DEPENDENT VARIABLES, MEDIATORS, two values (male or female). The re-
MODERATORS, and RANDOM VARIABLES. searcher might use an analysis of vari-
Compare CONSTANT. ance to apportion the variance in the
variable error see RANDOM ERROR. response measuretest performance
among the two different factorslevel
variable stimulus in PSYCHO- of caffeine and being male or female.
PHYSICAL RESEARCH, any one of a set of The central output is a table that shows
experimental stimuli that are to be sys- the proportion of variance attributable
tematically compared to a constant to the main effects of the factors, the
stimulus. proportion attributable to interactions
variance (symbol: a^) n. a measure of between the factors, and the proportion
the spread, or DISPERSION, of scores attributable to error and other external
within a sample or population, whereby factors.
a small variance indicates highly similar variance components model see
scores, aU close to the sample MEAN, and RANDOM-EFFECTS MODEL.
a large variance indicates more scores at
a greater distance from the mean and varianee-covariance m a t r i x see
possibly spread over a larger range. See COVARIANCE MATRIX.
also STANDARD DEVIATION. variance estimate an index of varia-
variance analysis see ANALYSIS OF VAR- tion in a population that has been calcu-
IANCE. lated using a sample of that population.
For example, a sample STANDARD DEVIA-
variance-balanced design a type of
TION is an estimate of the deviation in
CROSSOVER DESiGN-in which the possi-
the larger population.
ble sequences of treatments are arranged
in such a manner that all PAIRWISE C O M - variance explained an indication of
405
variance inflation factor
how well variation in one variable (or set ability within each of the smaller, more
of variables) can be accounted for by the homogeneous groups is less than that
variation in another variable. For exam- for the undivided larger group.
ple, if the CORRELATION COEFFICIENT
variance ratio a comparison of two in-
between a PREDICTOR VARIABLE (X) and
dices of variance, expressed as a quo-
a DEPENDENT VARIABLE (y) is 0.25, then
tient. The F RATIO, which compares
25% of the variability in y is explained
effect variances to error variances, is an
by the variability in x.
example.
variance i n f l a t i o n f a c t o r (VIF) an variance-ratio test see F TEST.
index of the degree to which the vari-
abiUty of an estimated REGRESSION CO- variance-reduction model see VARI-
EFFICIENT is increased because of in- ANCE-PRESERVATION MODEL.
terrelationships among the variables in variance-stabilizing transforma-
an ordinaty LEAST SQUARES REGRESSION t i o n in REGRESSION ANALYSIS Or ANALY-
model. A variance inflation factor exists SIS OF VARIANCE, any of a class of
for each of the k PREDICTOR VARIABLES mathematical processes that convert data
in the model. A VIF of 1 means that s h o w i n g H E T E R O G E N E I T Y OF V A R I A N C E
there is no correlation among the k^^ to a different form, so as to enable the
predictor and the remaining predictor application of specific analytical tech-
variables (and hence the variance is not niques or to simplify considerations.
inflated at all), a VIF exceeding 4 war-
rants further investigation, and a VIF ex- variate n. 1. a specific value of a particu-
ceeding 10 indicates serious MULTI- lar VARIABLE. 2. a RANDOM VARIABLE it-
COLLINEARITY requiring correction. seff.
v a r i a t i o n n. fluctuation: the degree of
variance-preservation model a
VARIANCE or DISPERSION of values that
method of conducting a META-ANALYSIS is obtained for a specific variable.
to determine EFFECT SIZES that accounts
for variation among study designs. For v a r i a t i o n coefficient see COEFFI-
example, some studies included in the CIENT OF VARIATION.
analysis may involve a single INDEPEN- v a r i a t i o n r a t i o a measure of VARIABIL-
DENT VARIABLE (fl) with a single level
ITY for numerical values that represent
while others may involve two indepen- membership in specific categories (i.e.,
dent variables (b, c) with two levels each. for CATEGORICAL DATA). It indexes the
To ensure that single-factor effect-size proportion of cases that deviate from the
measures are comparable to multiple- MODE.
factor effect-size measures, a researcher
may use a variance-preservation model v a r i m a x r o t a t i o n a statistical proce-
to treat the multiple-variable design as dure applied within FACTOR ANALYSIS
two single-variable designs (i.e., the and PRINCIPAL COMPONENTS ANALYSIS
treatment effect for a is measured at to simplify solutions and enhance inter-
Level 1 of b and again at Level 2 of b). pretation of the results. It is a type of OR-
Variances within each level are then THOGONAL ROTATION intended to make
compared to ensure that they do not dif- each factor have a small number of large
fer significantly between the two types FACTOR LOADINGS and a large number of
of,designs. Alternatively, a researcher zero (or small) factor loadings. Thus, fol-
may use a variance-reduction model, in lowing a varimax rotation, each original
which the participants in the single- variable tends to be associated with a
variable design are stratified into two small number of factors, and each factor
homogeneous subpopulations (e.g., represents only a small number of vari-
males and females), such that the vari- ables. See FACTOR ROTATION.
406
venue sampling
h (distance in km)
variogram
407
verbal protocol
45 1
40
35 -
3
30
a. 25
o 0
a. 2
15
10
5
0
Democrat Republican Independent No Affiliation
Political Party
408
volunteer bias
looks longer at that red box, it is as- function in newborn infants might have
sumed that he or she has some under- items requiring the infants tofixateon a
standing of object permanence and was target of black and white concentric cir-
not expecting the toy to be there. cles or track a colorful object as it moves
visual analogue scale a psychometric past their eyes.
instrument used to evaluate subjective voice key an electronic device that in-
characteristics that extend over a range terfaces between a microphone and a
of continuous values: Respondents spec- computer, used for recording vocal re-
ity their level of the characteristic of in- sponse times in language-production
terest by indicating a position along a tasks, such as word or picture naming.
continuous line anchored at its end When a stimulus is presented the voice
points by word descriptors. For example, key is activated and begins to monitor
a visual analogue scale for performance the sound level from the microphone; if
might have poor on the left end with a the sound level then exceeds a specific
blank line across to excellent on the right threshold level, indicating speech, the
end, as shown above. voice key records the amount of time
The respondent would mark a place that has elapsed since it was activated.
along that line to indicate his or her per- volunteer bias any systematic differ-
ceived performance level. Also called un- ence between individuals who volunteer
numbered graphic rating scale. to be in a study versus those who do not,
visual test any test intended to measure which may potentially render the result-
or srady vision, involving stimuli that ing group or sample of participants un-
are viewed. For example, a test of visual representative of the larger population.
409
Ww
W 1 . symbol for the COEFFICIENT OF CON- (M) and females (F) from 1 to 27 accord-
CORDANCE. 2. symbol for the statistic ing to their performance on a task:
obtained from the WILCOXON RANK-
MMMFFFMMMMFFMMMFFFFFFFMMFMM
SUM TEST.
In this data set there are nine runs. If
symbol for the statistic obtained from the two samples are from the same pop-
the CRAMER-VON MISES GOODNESS-OF- ulation, then the males and females will
FIT-TEST. be well mixed and the number of runs
thus will be large (e.g., close to 25); if the
wait-list c o n t r o l group a group of re- number of runs is small, as it is in this ex-
search participants who receive the ample, the ordering cannot be caused by
same intervention or treatment as those chance fluctuation and the NULL HY-
in the EXPERIMENTAL GROUP but at a
POTHESIS thus is rejected. Indeed, visual
later time. Wait-list CONTROL GROUPS inspection of the data shows that the
commonly are used in therapy outcome males tend to cluster toward the left-
and similar studies to account for the hand side of the scale and the females
potential influence of elapsed time upon toward the right-hand side. Also called
treatment effectiveness; they may also runs test; two-sample runs test.
be used to address the ethical ramifica- [Abraham Wald; Jacob Wolfowitz (1910-
tions of withholding treatment from in- 1981), U.S. psychologist]
dividuals.
Ward's method an approach to HIER-
W a l d d i s t r i b u t i o n see INVERSE GAUS- ARCHICAL CLUSTERING that attempts to
SIAN DISTRIBUTION. [Abraham Wald minimize the SUM OF SQUARES of any
(1902-1950), Hungarian-born mathe- two (hypothetical) clusters that may be
matician] formed at each step of the data analysis.
In other words, Ward's method reveals
Wald's test a PARAMETRIC statistical the cost in increased sum of squares that
procedure used to evaluate the signifi- arises from merging clusters, such that
cance of individual coefficients ((i) in a only those clusters that yield the small-
LOGISTIC REGRESSION model. It yields est increases should be merged to
estimated values of population PARAME- achieve the optimum grouping of cases,
TERS that are compared to a CHI-SQUARE [foe H. Ward (1927-2011), U.S. statisti-
DISTRIBUTION having one DEGREE OF cian]
FREEDOM in order to determine their
theoretical occurrence under the NULL wash-out p e r i o d the time frame allot-
HYPOTHESIS. [Abraham Wald] ted for an administered drug to be elimi-
nated from the body or for a previously
W a l d - W o l f o w i t z test a nonpara- administered intervention to become
metric test of the NULL HYPOTHESIS that ineffective. Wash-out periods are partic-
two samples have been taken from iden- ularly important in medical and other
tical populations, based on whether or clinical research since the CARRYOVER
not the number of rans or sequences in EFFECT between treatments might other-
an ordering is random. For instance, wise confound the estimates of treat-
consider the following ordering of males ment effects.
Wherry's formula
411
whisker
412
within-groups degrees of freedom
413
within-groups design
values: (20 - 1) + (20 - 1) + (20 - 1) = 19 + SIGN). Because such designs involve re-
19 + 19 = 57. The within-groups degrees cording multiple responses from the
of freedom is used to calculate the same person, it is necessary to examine
W I T H I N - G R O U P S M E A N S Q U A R E . Also how each individual varies in his or her
called w i t h i n - c o n d i t i o n s degrees o f responses, so as to separate such unique
f r e e d o m ; w i t h i n - s u b j e c t s degrees fluctuation from variation that is due to
o f f r e e d o m ; w i t h i n - t r e a t m e n t s de- the influence of the treatment under in-
grees o f f r e e d o m . Compare BETWEEN- vestigation. For example, a researcher
G R O U P S DEGREES O F F R E E D O M . studying how amount of daily walking
(e.g., none, 30 minutes, 60 minutes, 90
w i t h i n - g r o u p s d e s i g n see WITHIN-
minutes) affects quality of sleep might
SUBJECTS D E S I G N .
have participants walk each length of
w i t h i n - g r o u p s m e a n s q u a r e an time across consecutive weeks and then
index of random variability or error in evaluate the results using a within-
an ANALYSIS OF VARIANCE. It Is Calcu- subjects analysis of variance. Also called
lated as the WITHIN-GROUPS SUM OF dependent-groups a n a l y s i s o f v a r i -
SQUARES divided by the WITHIN-GROUPS ance; dependent-samples a n a l y s i s
DEGREES OF F R E E D O M , and it forms the o f variance; w i t h i n - g r o u p s a n a l y -
denominator of the F RATIO. Also called sis o f variance; repeated measures
mean square w i t h i n ; within- analysis o f variance; treatments-
c o n d i t i o n s m e a n square; w i t h i n - by-subjects a n a l y s i s o f v a r i a n c e .
groups variance; w i t h i n - s u b j e c t s Compare BETWEEN-SUBJECTS A N A L Y S I S
m e a n square; w i t h i n - s u b j e c t s v a r i - OF VARIANCE.
ance; w i t h i n - t r e a t m e n t s error;
w i t h i n - s u b j e c t s degrees o f f r e e -
w i t h i n - t r e a t m e n t s m e a n square;
d o m see W I T H I N - G R O U P S DEGREES O F
w i t h i n - t r e a t m e n t s variance. Com-
pare B E T W E E N - G R O U P S M E A N S Q U A R E .
FREEDOM.
w i t h i n - s u b j e c t s d e s i g n an experi-
w i t h i n - g r o u p s s u m o f s q u a r e s an
mental design in which the effects of
index of variability i n an ANALYSIS O F
treatments are seen through the com-
V A R I A N C E that is used to determine the
parison of scores of the same participant
WITHIN-GROUPS MEAN SQUARE. It Is Cal-
observed under all the treatment condi-
culated by adding together the squared
tions. For example, teachers may want
deviations of the individual observa-
to give a pre- and postcourse survey of
tions (scores) on the D E P E N D E N T VARI-
skills and attirades to gauge how much
ABLE from the relevant group mean. Also
both changed as a result of the course.
called w i t h i n - c o n d i t i o n s s u m o f
Such a design could be analyzed with a
squares; w i t h i n - s u b j e c t s s u m o f
DEPENDENT-SAMPLES T TEST, a WITHIN-
squares; w i t h i n - t r e a t m e n t s s u m o f
SUBJECTS ANALYSIS OF VARIANCE, OF an
squares. Compare BETWEEN-GROUPS
ANALYSIS OF COVARIANCE. Also Called
SUM OF SQUARES.
correlated-groups design; corre-
w i t h i n - g r o u p s v a r i a n c e see W I T H I N - lated-samples design; dependent-
GROUPS MEAN SQUARE. groups design; dependent-samples
design; related-measures design;
within-subjects analysis o f vari-
related-samples design; repeated
a n c e a variation of the standard A N A L Y -
measures design; treatment-by-
SIS OF V A R I A N C E that is applied to data
subjects design; w i t h i n - g r o u p s de-
from a study in which the independent
sign. Compare BETWEEN-SUBJECTS D E -
variable has multiple levels and each
SIGN.
participant experiences each treatment
level or is otherwise measured more w i t h i n - s u b j e c t s f a c t o r the indepen-
than once (see WITHIN-SUBJECTS D E - dent variable under study i n a W I T H I N -
worm plot
415
X a letter used to symbolize a variable of ample, in the ordered pair (4, 8), the x-
interest. For example, in a REGRESSION coordinate is 4 and the corresponding
EQUATION, X may represent a particular vertical Y-COORDINATE value is 8. See
predictor or INDEPENDENT VARIABLE. CARTESIAN COORDINATE SYSTEM.
416
Yy
y symbol for a variable of interest. For ex- called acquiescence bias; acquies-
ample, in a REGRESSION EQUATION, ycent response set; response acqui-
may represent a particular outcome or escence. Compare NAY-SAYING.
DEPENDENT VARIABLE.
y-intercept n. in an equation represent-
Y symbol for Y PRIME. ing a straight-line relationship between
Yates's correction for continuity two variables, the value of variable y
an adjustment made to a CHI-SQUARE when the value of variable x equals zero.
TEST of data from a CONTINGENCY TABLE For example, in the general LINEAR EQUA-
having only two columns and two rows TION format y = a + bx,a represents the y-
of information. The Yates's correction intercept.
yields a more conservative chi-square
statistic and improves the test's accuracy yoked-control group a CONTROL
by accounting for the fact that it uses a GROUP in which each participant is
CONTINUOUS DISTRIBUTION to approxi- paired with a participant in another
mate a DISCRETE DISTRIBUTION. To group (see MATCHED-PAIRS DESIGN); the
apply the correction, one takes each paired individuals are then exposed to
value within the table (i.e., the observed precisely the same experiences except
frequency) and subtracts the value pre- for the specific treatment or other condi-
dicted to occur by a theoretical model tion under study. This procedure is in-
(i.e., the expected frequency). One then tended to make the control group as
takes the difference so obtained and sub- similar as possible to the experimental
tracts 0.5, squares the total, and divides group. For example, in a study of the ef-
the resulting number by the expected fectiveness of rewards upon children's
frequency. When such a correction is ap- learning, matched individuals in the
plied, the term Yates-corrected chi- control and experimental groups would
square test or Yates chi-square test is complete the same tasks under the same
used to refer to the test itself. [Frank conditions but only those in the experi-
Yates (1902-1994), British statistician] mental group would be praised for good
performance.
y-axis n. the vertical axis on a graph. See
ORDINATE. Youden's index a measure of the ca-
y-coordinate n. the vertical value in a pacity of a diagnostic test to correctly
pair of graph coordinates (x, y), which identify individuals with a certain illness
indicates how far to move up or down and to correctly identify those who do
from the origin along the Y-AXIS. For ex- not have the illness. It ranges from 0 to
1, with values closer to 1 indicating both
ample, in the ordered pair (4, 8), the y-
greater SENSITIVITY and SPECIFICITY of
coordinate is 8 and the corresponding
the test. Also called Youden's J . [Wil-
horizontal X-COORDINATE value is 4. See
liam John Youden (1900-1971), Austra-
CARTESIAN COORDINATE SYSTEM.
lian-born U.S. chemist and statistician]
yea-saying n. answering questions posi-
tively regardless of their content, which y prime (symbol: y') the predicted or EX-
can distort the results of surveys, ques- PECTED VALUE of a given variable of in-
tionnaires, and similar instraments. Also terest, y.
417
Yule-Simpson effect
418
z-axis n. the third dimension in a CARTE- counts of zero (i.e., no boxes sold) be-
SIAN COORDINATE SYSTEM or graph. It is cause several teams did not go out to seU
perpendicular to both the horizontal x- due to inclement weather in their neigh-
AXis and the vertical Y-AXIS. borhoods; however, zero counts also
could occur because of nonsuccess in
Zelen's design a type of RANDOMIZED selling. Consequently, the, number of
CLINICAL TRIAL in which patients are as- zeros may be inflated and the number of
signed to experimental conditions be- teams not selling boxes cannot be ex-
fore they have consented to participate plained in the same manner as the num-
in the research. In a typical trial, patients ber of teams selling more than one box.
are first informed of the nature of the
study and then decide whether to partic- zero m a t r i x see NULL MATRIX.
ipate, which may result in certain biases
that Zelen's design attempts to over- zero-order correlation a simple asso-
come. There are two versions of the de- ciation between two variables that does
sign, both of which remain ethically not control for the possible influence of
controversial: In the single-consent other variables. For example, consider
strategy, patients assigned to the experi- the relationship between success selling
mental treatment are told there is an al- computers and knowledge of how the
ternative (the control) available and Internet works. A zero-order correlation
offered the option to switch or remain in would examine the direct relationship
the experimental group; those in the between these two variables without
control group are not informed of their taking into account other explanatoty
srady participation and their agreement information, such as education level,
to participate thus is not obtained. In sales experience, and so forth. The actual
the double-consent strategy, both the index of the magnirade or degree of
experimental and control groups are of- such a relationship is called the zero-
fered the option to switch from the order correlation coefficient; it ranges in
treatment to which they initially were value from -1 to +1, with the former in-
assigned. Also called randomized con- dicating a perfect negative relationship
sent design. [Marvin Zelen (1927- ), and the latter a perfect positive relation-
U.S. biostatistician] ship. Compare PART CORRELATION; PAR-
TIAL CORRELATION.
zero-inflated Poisson regression
(ZIP) a model for analyzing a data set zero-sum game in GAME THEORY, a
with an excessive number of zero out- type of game in which the players' gains
comes. An alternative to regular POIS- and losses add up to zero. The total
SON REGRESSION for explaining outcome amount of resources available to the par-
variability, it assumes that with proba- ticipants is fixed, and therefore one
bility p the only possible observation is player's gain necessarily entails the oth-
zero (0), and with probability 1 - p, a ers' loss. The term is used particularly in
POISSON VARIABLE Is obsetved. For ex- analyses of bargaining and economic be-
ample, consider a study of 100 teams havior but is sometimes also used in
selling boxes of cookies on a weekend. In other sociocultural contexts (e.g., poli-
this data set, there could be excessive tics).
419
zero-truncated Poisson distribution
423
Abbreviations and Acronyms
IV independent variable
KMO test Kaiser-Meyer-Olkin test of sampling adequacy
K-R 20 Kuder-Richardson formula 20
K-R 21 Kuder-Richardson formula 21
LCA latent class analysis
LLR log-likelihood ratio
In natural logarithm
LOCF last observation carried forward
loess local regression
log logarithm
lowess local regression
LR likelihood ratio; logistic regression
LSD least significant difference
LSD test Fisher least significant difference test
LTA latent transition analysis
MA moving average
MAD mean absolute deviation; median absolute deviation
MA model moving-average model
MANCOVA multivariate analysis of covariance
MANOVA multivariate analysis of variance
MAPE mean absolute percentage error
MAR missing at random
MARS multivariate adaptive regression spline
MCA multiple classification analysis
MCAR missing completely at random
M C M C method Markov chain Monte Carlo method
M C M L estimation Monte Carlo maximum likelihood estimation
MCP minimum convex polygon
MDA multiple discriminant analysis
MDL minimum description length
MDS multidimensional scaling
MI modification index
MIMIC model multiple indicators-multiple causes model
MLE maximum likelihood estimation
ML-EM estimation maximum likelihood-expectation maximization estimation
MOE margin of error
MTMM multitrait-multimethod matrix; multitrait-multimethod
model
MVUE minimum variance unbiased estimator
NCE normal curve equivalent
NFl normed fit index
1
NH null hypothesis
NHST null hypothesis significance testing
425
Abbreviations and Acronyms
426
Abbreviations and Acronyms
Acronyms
427
Entry Illustrations
Andrews plot 11
ANOVA summary table 11
array 14
back-to-back stem-and-leaf plot 18
bagplot 19
balanced Latin square 19
bar graph 20
bimodal distribution 27
biplot 29
box-and-whisker plot 32
bubble plot 34
Cartesian coordinate system 36
centile reference chart 40
chain graph 41
Chernoff faces 42
circumplex 44
classification table 46
classification tree 46
complete block design 53
component bar graph 54
compound bar graph 55
confidence band 57
contingency table 61
contour plot 63
correlation matrix 66
covariance matrix 67
cumulative frequency distribution 75
cumulative frequency polygon 76
cumulative relative frequency distribution 77
cumulative relative frequency graph 77
cumulative sum chart 78
cycle plot 79
decision tree 84
dendrogram 87
differences vs. totals plot 93
428
Entry Illustrations
429
Entry Illustrations
430
Overview of Research Design Considerations
Ol S -S
QJ tl.
S Q >
Q -g EEs
is S E
<j i z
0) J - 0,
5 5 S .2 S -
fig
*S (0 OJ ro 0) ~
C3 Z cc
0)
Sa.
C < op O OJ
S< S
Q o mau 3
i CD s S s I
OS . .
S .5
ol
; UJ
5 '
c
>-
o <^
srtStu
i\ stud
idinal
c
>
'5) uo n
rve
c <b
o a
s .a - "2 E S .2
- c nj u
ra .y o
2 S g- E S^f
if c o
"1 2 w t 3 . c o o
431
Symbols
Latin
adj R^ adjusted R^
C coefficient of contingency
d Cohen's d; Glass's d
d' d prime
D Cook's distance; difference score
D2 Mahalanobis distance
df degrees of freedom
E expected value
f an effect size index; frequency
F F ratio
likelihood-ratio chi-square
communality coefficient
Ho null hypothesis
H, alternative hypothesis
Wa alternative hypothesis
H hat matrix
In Williams's agreement measure
I identity matrix
k coefficient of alienation; number of units in a statistical analysis
K? D'Agostino test statistic
M sample mean
MS mean square
MSE mean squared error
MSR mean-square residual
MSSD mean-square successive difference
n number of scores or observations from a particular
experimental condition
N total number of cases (participants) in an experiment
P probability
q Studentized range statistic; probability of failure in a binary trial
Q Cochran Q test statistic; Yule's Q; quartile
r product-moment correlation coefficient; sample correlation
coefficient
r2 coefficient of determination
R multiple correlation coefficient
coefficient of multiple determination
432
Symbols
adjusted 7?^
biserial correlation coefficient
''bis biserial correlation coefficient
Rc canonical correlation coefficient
^effect size effect-size correlation coefficient
''pb point biserial correlation coefficient
point biserial correlation coefficient
Spearman correlation coefficient
tetrachoric correlation coefficient
reliability coefficient
RK cross-validated multiple correlation
sample standard deviation
sample variance
5D standard deviation
S standard error
SEE standard error of estimate
SEM standard error of measurement; standard error of the mean
SS sum of squares
SSCP sum of squares of cross-products
SSE sum of squared errors
55, regression sum of squares
regression
T Wilcoxon signed-ranks test statistic
J2 Hotelling's test
TSS total sum of squares
U Mann-Whitney U test statistic
V Cramer's V
W coefficient of concordance; Wilcoxon rank-sum test statistic
W2 Cramer-von Mises goodness-of-fit test statistic
X a variable of interest
x' X prime
X sample mean
X data matrix
y a variable of interest
y' y prime
A
predicted value of an outcome or dependent variable
y
Greek
a alpha
P beta
t chi-square
433
Symbols
e a small value
z?- proportion of shared variance
r| correlation ratio; a latent variable
T|2 amount of explained variance
TIG^ generalized eta squared
T|p2 partial eta squared
Y Goodman-Kmskal's gamma
K Cohen's kappa
X eigenvalue
A Wilks's lambda
|i population mean
co^ omega squared
<S) phi coefficient
Is)^ Cramer's V
K ratio expressing the circumference of a circle to its diameter;
probability of success in a trial
p population correlation coefficient; Spearman correlation
coefficient
o population standard deviation
population variance
OM standard error of the mean
S sum
T Kendall's tau
e ability parameter
434