Professional Documents
Culture Documents
Article
Abstract
The publication of information about public service performance has expanded dramatically in
recent decades. Multiple experiments have shown that relative performance information influ-
ences citizens’ evaluations of local services. It is unclear whether this information leads citizens
to update their beliefs or temporarily employ different criteria, however, and we do not know if
the influence of performance information will extend to contexts in which citizens have access to
other information sources. We examine these questions using two experiments within nationally
representative surveys. In the first, we find that providing respondents with information about the
performance of their local schools relative to schools in the state, the nation, and other developed
countries depresses average evaluations of local school quality.The second experiment shows
that relative performance information depresses average evaluations in part by priming
respondents to consider how the academic performance of their communities’ schools
compares to others, but also by leading some respondents with erroneous prior beliefs about
school performance to update their beliefs.
© The Author 2016. Published by Oxford University Press on behalf of the Journal of Public Administration Research and Theory, Inc. 1
All rights reserved. For permissions, please e-mail: journals.permissions@oup.com.
2 Journal of Public Administration Research and Theory, 2016, Vol. xx, No. xx
Journal of Public Administration Research and Theory, 2016, Vol. xx, No. xx 3
examined the effects of local public schools, public service Scholars have also long
different types of rather than hypothetical performance. Simon and highlighted the potential
performance services. Residents have Ridley (1938, 466–7) importance of relative
information and have access to numerous note that policymakers comparisons. Simon
consistently found that sources of information play a crucial role in (1939,
infor- mation about about public school deciding what 106) argues that “the
performance relative to performance, includ- ing information to provide only sound basis for
other localities influences their own experiences because “the voter has decisions about
people’s perceptions of and state test results, but neither the time nor the numbers is numerical
and attitudes toward rel- ative performance information to factual information
service quality in their information is difficult interpreted undigested about past experiences
own communities. to acquire. We isolate statistical tables in an or the experiences of
The extent to which learning effects by intelligent manner.” others— nothing more
the finding that relative providing information or less than comparative
per- formance about the rank of local statistics” (see also
information influences schools, rather than March and Simon 1958;
citizens’ perceptions of priming a specific Simon 1937). Yet in the
a school’s quality based profound implications positive or negative light, information about the
for example, describing local
on an inaccurate for the design of public services as being in the
true performance of the
perception of the accountability systems. “best” or “worst” third of school districts in which
school’s academic If information about the services nationally they live, rather than
performance but adjusts rela- tive performance of (Baekgaard 2015; James and hypothetical schools. In
Mosely 2014). There are
his judg- ment when he local services does not order to separate the
strong theoretical and
is provided accurate lead citizens to learn and empirical grounds to expect effect of citizens learn-
information would be an update their beliefs, this such framing to mediate ing from priming effects,
example of learning. In would call into question citizens’ responses to we provide respondents
contrast, priming occurs the value of providing performance information with information about
(Olsen 2015c, 2016). We do not
when an intervention such information. At the the percentile rank of
test framing effects in this
“heightens the salience same time, if priming study. their local schools, rather
of a par- ticular specific comparisons 2 There have been numerous than a comparison to a
consideration and pushes explains the substantial scandals concerning the particular refer- ence
other relevant consid- effects observed, this manipulation of school point, and employ a
performance information, for
erations to the would likely rein- force example, in Indiana in 2012
recently developed
background of concerns about the (Associated Press 2013). methodo- logical
Second, an
inconsistency between 3 Although not as large,
citizens’ evalua- tions of surveys consistently reveal
local schools and similar disparities across a
schools nationally range of topics including
crime rates, environmental
suggests that people do
quality, and quality of
not evaluate schools on healthcare (Daniels et al. 2012;
the basis of accu- rate Newport 2013; Saad 2011).
knowledge about
relative academic
performance. Surveys
routinely find that when
asked to grade public
schools with the five-
point scale traditionally
used to grade students,
more people assign an
“A” or “B” to the schools
in their own community
than to schools across
the United States as a
whole. Indeed, this
evaluation gap has
widened since Gallup
began asking respond-
ents to grade schools
nationally in 1981,
reaching a high of 35
percentage points in
2013 (figure 1).3 This
gap represents a
paradox, as services
nationally are the sum of
local services, and thus
indicates that citizens
cannot be evaluating
local schools and schools
nation- ally on the basis
of accurate information
about relative academic
performance.
A number of factors,
however, could mitigate
the impact of relative
school performance
information. First,
citizens may use
dimensions of quality
other than relative
academic performance
to evaluate local schools.
Americans tend to value
multiple qualities in their
schools, including
academic quality, safety,
demo- graphic mix, and
extracurricular
activities, each of which
may be emphasized to a
greater or lesser extent
8 Journal of Public Administration Research and Theory, 2016, Vol. xx, No. xx
Figure 1. Proportion of respondents in national poll grading local schools and schools nationally A or B.
Source: PDK/Gallup poll on the public’s attitudes toward the public schools, 1974–2013.
Journal of Public Administration Research and Theory, 2016, Vol. xx, No. xx 9
Students in different parts of the country perform respondents’ prior beliefs and the true value of the
differently in math. The average student in your performance measure. In order to do this, we employ
district performs better than what percent of stu- models including the underlying performance measure,
dents across the country? or the difference between respondents’ prior beliefs
and this performance measure, as well as an interac-
A treatment group was then informed of the national
tion between this variable and the treatment dummy.
percentile in math of their local school district
Further details are provided below.
(n = 2,562), using identical wording to that employed
in the 2013 survey, and a control group was given no
information (n = 2,628). We then asked respondents to Results
evaluate the quality of their local schools, again assign- Prior Information About School Quality
ing a letter grade on the A to F scale.
Before reporting the experimental results, we first use
This procedure allows us to isolate learning effects,
responses to the 2014 survey to examine the perfor-
first, by priming all respondents and, second, by explor-
mance-related information that individuals already
ing how treatment effects vary with respondents’ prior
have. Figure 2, which displays the difference between
beliefs. To the extent that there is a learning effect, the
respondents’ perceptions of the national percentile rank
effect of treatment should become increasingly posi-
of their local schools and the actual percentile, provides
tive in the degree to which respondents underestimate
Table 1. Likelihood Ratio Tests to Compare Models for Grade Assigned to Local Schools
(1) st 1 .0004
(2) nt 1 <.0001
(3) it 1 <.0001
(4) gt 1 .575
(5) st + sp + st: sp st + sp .067
(6) nt + np + nt: np nt + np .012
(7) nt + np + ng + nt: np nt + np + ng .012
(8) it + ip + it : ip it + ip .077
(9) it + ip + ig + it : ip it + ip + ig .076
(10) gt + gr + gt: gr gt + gr .447
Note: st, nt, it, and gt are the state, national, international, and graduation rate treatments. sp, np, ip, and gr are the state, national, and interna-
tional percentiles, and the graduation rate. ng and ig are the gap between the national or international percentile, and the state percentile. LRT
p-value reports the p-value for a likelihood ratio test that compares the likelihood of observing the data under the full and reduced models.
A low p-value indicates that the data were not as likely to occur under the reduced model as under the full model.
P(Grade Schools A/B)
1.00
0.75
T rea tment
0.50 No Information
State Percentile
0.25
0.00
0 25 50 75 100
State Percentile
1.00
P(Grade Schools A/B)
0.75
T reatm ent
0.50 No Information
National Percentile
0.25
0.00
0 25 50 75 100
1.00
P(Grade Schools A/B)
0.75
T reatm ent
0.50 No Information
Internat. Percentile
0.25
0.00
0 25 50 75 100
Internat. Percentile
1.00
P(Grade Schools A/B)
0.75
T rea tm ent
0.50 No Information
Graduation Rate
0.25
0.00
0 25 50 75 100
Graduation Rate
Table 2. Likelihood Ratio Tests to Compare Models across respondents. Figure 5 plots predicted probabili-
for Grade Assigned to Local Schools Using 2014 ties of assigning local schools A or B against the actual
Survey Data
national percentile rank, using a model including an
Full Model Reduced Model
interaction between treatment and the national per-
LRT p
centile rank (table 2, Row 2, full model). As with the
(1) nt 1 .149 2013 data, the effect of treatment on the probability
(2) nt + np + nt: np nt + np .047 of assigning A or B becomes increasingly positive as
(3) nt + gp + nt: gp nt + gp <.0001 the national percentile rank of local schools increases.
Note: nt is an indicator for the national percentile treatment,
A likelihood ratio test confirms that the interaction
np is the actual national percentile, and gp is the gap between the between treatment and the national percentile rank is
actual and the estimated national percentile. LRT p-value reports a statistically significant predictor (table 2, Row 2).
the p-value for a likelihood ratio test that compares the likelihood of A test to identify learning effects in the context of
observing the data under the full and reduced models. A low p-value potential priming requires examining how the treat-
indicates that the data were not as likely to occur under the reduced
model as under the full model.
ment effect varies with prior misperception. Larger
treatment effects among respondents who lack relative In figure 6, we observe that the more a respondent
performance information (as measured by their own underestimates the national percentile rank of local
estimates of where their local schools rank nationally) schools, the greater the amount by which providing
indicate a learning effect. Because both groups received the respondent with true percentile rank increases the
the prime, these effects cannot be attributed to prim- probability of grading local schools A or B. Likewise,
ing. We therefore fit a model that includes the differ- the more a respondent overestimates the true percen-
ence between actual national percentile and estimated tile, the more treatment depresses the probability of
percentile along with an interaction between this differ- assigning local schools A or B. A likelihood ratio test
ence and the treatment indicator (table 2, Row 3, full confirms that the interaction between treatment and
model). Figure 6 uses this model to plot how the prob- the respondent estimated percentile is a statistically
ability of assigning local schools A or B varies with the significant predictor (table 2, Model 3). These results
gap between the estimates and true national percentile. indicate that respondents who err in their prior beliefs
Figure 6. Predicted probabilities of grading local schools A or B against the gap between estimated and actual national percentile: 2014
survey data.
about local school Presenting respondents nation’s, rather than on data. Indeed, our results
performance update with information about other considerations. are consistent with a
those beliefs, and in turn the academic Specifically, the average scenario in which the
their evaluations, in performance of local treatment effect of public evaluates local
response to new schools relative to information found in the schools based on high
information. schools in the state, 2013 sur- vey shrinks school graduation rates
This learning effect nation- ally, or and loses statistical and has reasonably
also helps to explain internationally depresses significance in the 2014 accurate information on
why the effect of average evaluations of survey, when both them. We likewise
treatment on local school quality. The treatment and control cannot exclude the
respondents’ effect is especially large groups are primed to possibility that in the
evaluations becomes among those living near consider relative absence of treatment
increasingly positive as the lowest scoring performance. respondents evaluate
the national percentile schools. Absolute The observed learning schools based on criteria
rank of local schools information about the effects occur in a other than academic
increases, as reported in performance of local context in which citizens performance, such as a
figures schools, as measured by already have access to school’s facilities or