Randomized Response Techniques for Multiple Sensitive Attributes Author(s): Ajit C.

Tamhane Source: Journal of the American Statistical Association, Vol. 76, No. 376 (Dec., 1981), pp. 916923 Published by: American Statistical Association Stable URL: http://www.jstor.org/stable/2287588 Accessed: 21/10/2010 18:01
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=astata. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of the American Statistical Association.

http://www.jstor.org

Randomized Response Techniques forMultiple butes Sensitive Artr

C. TAMHANE* AJIT

forinvestigating Some randomized responsetechniques Suitablestatistical techniques forcollecting and analyzA newtechnique ingdata forsurveys attributes are reviewed. dealing withsuchmultiple t _ 2 sensitive sensitive of requiring do notappearto be available. onlyr attributes is proposedthathas the advantage ofuptor-variate The purpose trials perrespondent (r ' t)ifestimation ofthispaperis two-fold. First, we briefly is desired. The case ofr = 2 is analyzed reviewsome recentliterature jointproportions thathas a bearing on the the restricted maxi- multiple in detail.A procedure forderiving sensitive attributes problem. Second, we promumlikelihood estimators (MLE's) of the proportionspose and developforthegivenproblem a newRR techany set of pairsof niquethat and a testof independence between has somedesirable Thistechnique properties. ofmeasure ofrespondent is an extension attributes aregiven. The notion ofa technique earlier proposed byBarksthismeasure dale (1971),buttheestimation jeopardyis extended to oursetup.Keeping here procedure proposed for thet = 2 case is new. We also givea testofpairwise fixed, we makenumerical comparisons for independence between in terms of the traceof any set of pairsof attributes. competing techniques We extendthe notion of matrix of the esti- respondent the asymptotic variance-covariance jeopardy proposed byLeysieffer and Warner a practical application ofthenew (1976)to themultiple mator vector.Finally, sensitive attributes setup. Keeping technique is described. thismeasure ofrespondent we carry out jeopardy fixed, a numerical comparison of efficiencies of some competing maxiresponse;Restricted KEY WORDS: Randomized Finally, we givetheresults of an actualapsensitive attributes;procedures. mum likelihood estimators; Multiple plication of the technique to demonstrate its feasibility jeopardy function. Samplesurvey techniques; Respondent in practice. In thenumerical itturns outthat comparisons thepro1. INTRODUCTION posed technique does notfareas well as an "optimal' on sensitive foreliciting information or stig- version Surveys of a technique a repeated involving (foreach atare plaguedby the problem of un- tribute) attributes matizing application of the Simmons unrelated question truthful by respondents,technique. responsesor noncooperation it was felt Nevertheless, desirable to publish bothof whichlead to biased estimates. To avoid this theresults because thetechnique does have someprac''evasive answer theprivacy ofthe ticaladvantages bias" andtopreserve at least reasonably and performs well. an innovative tech- In any case, the comparisons respondent, Warner (1965)introduced betweencompeting techresponse niquecommonly referred toas randomized (RR), niquesshouldproveuseful to thepractitioner. Furthertechnique.Since Warner'sarticle,manyauthorshave more,many oftheresults are newand interesting and it madecontributions to this area; a review ofthese is hopedthattheywillattract general other to work researchers in Horvitz, contributions Greenberg, and on theproblem. maybe found Abernathy (1975). is restricted to the Mostofthework on RR techniques 2. SOMEPREVIOUS WORK-ANOVERVIEW ofa single attribute. sensitive Very however, study often, In hisdissertation, Barksdale (1971)proposed andanaare interested in studying severalsensocial researchers lyzedsome'RR -techniques for twosensitive investigating are not Thus the researchers sitiveattributes together. dichotomous In particular, attributes. he considered a andtesting coninestimating hypotheses onlyinterested repeated(for each attribute) application of Warner's the of thepopulation theproportions possessing cerning original and Iglewicz1976), technique (see also Clickner butalso the individual attributes understudy, sensitive a repeated of Simmons's application unrelated question attributes. degreeof associationbetweenthe different et al. 1969),and a third technique (Greenberg technique thatwe describein detailin thenextsection.In thereof Warner's peated application technique (W technique) * AjitC. Tamhane ofIndustrial is AssociateProfessor, Department are performed On theithtrial perrespondent. Ev- twotrials andManagement University, Engineering Sciences,Northwestern wishesto thank the previous editors, (i = 1, 2) the interviewer anston,IL 60201.The author therespondent with presents
H. DeGroot andthe Morris andGeorge T. Duncan, an associate editor, for The referees andsuggesting improvements. many helpful comments author is grateful to Cynthia FredHubbard, andKateRobinson Grant, forcarrying outtheinterviews. Thisresearch was supported byGrant ofEducation. The author NIE-C-74-0115 from theNational Institute is grateful to Robert Boruch forproviding thissupport. ? Journalof the AmericanStatisticalAssociation December 1981,Volume76, Number 376 Theoryand MethodsSection

916

Tomhane: Randomized Response Techniques

917

a pair of statements: "I possess the attribute to compute theestimates ofthe Ai" and yetallowstheresearcher "I do notpossesstheattribute Ai," where and bivariate Aiis a sensitive marginal proportions oftheattributes from attribute. Therespondent picksoneofthetwostatementsthe observed frequencies of "Yes-Yes," "Yes-No," at random according to known probabilities Pi and 1 - "No-Yes," and "No-No" responses. revealing his choiceto theinIn a survey Pi (Pi * ) and, without dealing with t ?-2 sensitive attributes, the terviewer, responds to it. Then from the observedfre- W and S techniques involvet trials perrespondent. If t quencies of "Yes-Yes," "Yes-No," "No-Yes," and is large,thenthesetechniques becometedious,costly, "No-No" responses, and theknowledge ofthePF's,the and lead to degradation incooperation on thepartofthe desiredproportions can be estimated. The repeated ap- respondents. Also, the estimating equationsinvolveall plication of Simmons's technique (S technique) is quite thejointproportions, whichoften the researcher is not similar, exceptthat on theithtrial (i = 1, 2) therespond- interested in. On theother hand,thetechnique described entis presented with a pairofstatements, "I possessthe in theprevious paragraph can be easilyextended to the attribute and "I possess theattribute Ai" Yi," whereY, case of t > 2, withthenumber of trials per respondent is some unrelated and innocuousattribute. From the restricted to r < t iftheresearcher's interest onlylies in knowledge of Pi = the probability of picking the first up to r-variate jointproportions. Quiteoften, r = 2 will statement, ofpopulation possessing suffice forthepurposes oftheresearch. Pi = theproportion the attribute of reYi, and the observedfrequencies it is clearthatfort > 2, sincethe W and Intuitively, sponses,thedesired proportions can be estimated. S techniques involvet trialswhilethe technique to be Some othercontributions to the problem of multiple proposedinvolves onlyr < t trials, thelatter technique sensitive attributes are as follows.Drane (1975, 1976) mustbe less informative. This is indeedso. Partof the studied theproblem oftesting independence between two extrainformation obtained by theformer techniques is sensitive dichotomous attributes, usingrepeatedappli- intheform ofestimates ofhigher order jointproportions cationsof variousRR techniques forsingleattributes. thatare notobtainable withthe latter technique, while Warner (1971) proposeda generallinearRR modelfor therestoftheextrainformation manifests itself interms many attributes butdid notexplicitly consider theprob- of lowervariancesof the estimates. The former techlemofjointdistributions oftheattributes. Another tech- niques,however, would suffer from degradation in coniquefor estimating marginal distributions ofseveral sen- operation fort as low as threeor fourwhilethe latter sitiveattributes thatmakesuse of weighing was technique, designs forfixedr (whichis based on investigator's proposed by Raghavarao and Federer (1979). interests andgoals)wouldsuffer from somewhat inflated Relatedworkon theRR techniques formultiple sen- variances. The exacttrade-off is notclear,noris itclear sitiveattributes has been done in Europe by Eriksson howmuch larger samplesizeswouldbe required with the (1973) and Bourke(1975). Erikssonpresented a theory latter technique to compensate for theinflated variances. for the generalcase of a two-way table. These issues needfurther contingency research. Bourke considered various designs for estimating thecorNow we describe thelatter we refer technique, which responding cell probabilities in a two-way tableformed toas themultiple RR trials technique ortheM technique. bytsensitive attributes, eachhaving c categories ofwhich at most(c - 1) are sensitive. Bourke'sworkdoes not, 3.2 Notationand Descriptionof the Technique however, addressthe problem of estimating joint proportions of different attributes. Considert - 2 dichotomous attributes The detailsof some of A1, A2, . . .. thesetechniques are foundin Horvitz, are sensitive, and A,; we shallassumethatall theattributes Greenberg, butobviously thatneed notbe so. Let Oi, Abernathy (1976). ...i denotethe in thetarget unknown of individuals proportion popula3. MULTIPLE tionthatpossess the attributes RRTRIALS TECHNIQUE Ai,, ... , Aiu(1 < i1 < < iu ? t, 1 - u ? t). The researcher's interestlies 3.1 Barksdale's Third Technique in making and hypothstatistical inferences (estimation The technique we are aboutto propose is an extension esis testing) the 0's. concerning of the thirdtechniqueproposedby Barksdale(1971), For employing the multiple RR trialstechnique, the whichis as follows. The two statements the statements must a "Yes" response to be phrased so that concerning two sensitive attributes are phrased so thata "Yes" re- some statements would be nonstigmatizing, while a wouldbe nonstig- "No" response sponseto one of thetwo statements to theothers wouldbe so. Without loss be "I have shall the first s matizing of we assume that < t state(e.g.,thetwostatements never generality, might smokedmarijuana"and "I am an alcoholic"). The in- mentsare phrased"I possess the attribute Ai" (1 ? i terviewer on ? s), a "No" responseto each one of whichwouldbe bothstatements to therespondent presents two occasions. On each occasion the respondent the remaining t - s statements are picks nonstigmatizing; ? one of the two statements at random, + unknown to the phrased"I do notpossess theattribute 1 i Ai" (s to each one of whichwouldbe interviewer, but according to some knownprobability ? t), a "Yes" response (different foreach occasion), and respondsto it. This so. An appropriate choice of s would be -t/2. Let procedure maintains theprivacy of the respondent in the same manner as Oi..i butwith and .iru4be defined

918

Joumal ofthe American Statistical Assoclotion,December 1981

respect to themodified attributes are either the one (whichcorresponds to the "directresponse"case). Bi, which i the of the original or By = ' choosing _ s) Ph.(l 1 and complements P*J(2) 1 fordifferent Ai(I pairs Ai(s + 1 _ i ' t). It is clearthat the0's can be obtained from (i,j) fordifferent subsamples h(l c h ' b), it is easy to the 7T's we shallconsider see thatall the parameters and vice versa,and therefore can be estimated by using theequivalent of estimation of the u's. problem and no smaller number of subsamples (2) subsamples, in theprevious As remarked we shallassume willdo. An extension section, of thisargument showsthateven that theresearcher is interested P values,at least(2) subsamples and forgeneral onlyin themarginal are required bivariate thatis, i.(1_<i ' t) and rrir(1 proportions; all theparameters. In other ' toestimate words, bysuitably i <j Thus thereare t(t + 1)/2 unknown choosing t), respectively. theP's, thematrix R defined in (3.2) and (3.3) parameters to be estimated and onlytwo trialsmaybe can be madeto have a fullcolumn rankonlyifb _ (). performed We nowdescribe thetechnique. Let us then perrespondent. assumethat b _ (2) andthat R is a full column A totalsampleof n individuals is dividedintob _ 1 rankmatrix. in thefol- We proposeto obtainthe maximum the value of b will be specified subsamples; likelihood esti(MLE) of n from theobserved data{nhv} where lowing section. Let n1, n2, . .. , nb be the subsample mator nhv = thenumber = n. sizes with h ofindividuals from thehth havsubsample all the t statements Each individual is presented and inga scoreof v(Oc v c 3), 2J3=O nhv = nh(i _ h _ b). offirst askedto respond to one statement ac- Theusualmethod theunrestricted obtaining pickedat random MLE to some randomizing cording device,butnotrevealhis (UMLE) of A (i.e., the UMLE of Xhv = nahvnh for0 c choiceofthestatement totheinterviewer. Thisprocedure v c 3, 1 c h c b) and then theUMLE of a by obtaining is repeatedwithanother fortwo reasonsin the randomizing device,and both "solving" (3.1) is notapplicable theresponses context. are recorded. Let Ph/'( denote the(known) present probability thatan individual drawnfrom the hthsub1. Matrix R can be chosento be a square fullrank samplepicks,on the lthtrial,the ithstatement (1 i matrixonly for t = 2. For t > 2, in general,thereis no we have I= Phi"0 = I forI - h ' b t); obviously unique solution in a to (3.1). and 1 = 1, 2. 2. Even in thecase in which theUMLE of a can be obtainedby the above method, the resulting estimator 3.3 Estimation of the w's may not satisfythe naturalrestrictions on the r's, that Suppose thattheresponses are coded so thata score namely, of 2` ' is assignedto a "Yes" response on theIthtrial 0O wi 1 Viand (3.4) and a score of 0 is assignedto a "No" response.Then thetotalscore,say v, completely identifies theindivid- max(O,wi + nj - 1) _rr _min(wi, wj) (i,j). ual's response.For example,v = 3 corresponds to a Froma theoretical viewpoint, theUMLE of T mayeven "Yes-Yes" response, v = 2 corresponds to a "No-Yes" be inadmissible, as shownin thecase of Warner's techresponse, and so on. Let XhVdenotethe probability of nique for a singleattribute by Fligner, Policello,and a score of v foran individual obtaining drawn from the Singh(1977)and Devore (1977); it appearsthatWarner X = (XI,, X12, X13, * * * , 119 Xb2, Ab3) s- (1965)was also awareofthisproblem, hthsubsample, as is evident from and iT = (,7T , t, 12, XT13, ITt - 1)'. Thenwe have thefootnote on page 65 of hispaper. Therefore, we must findtherestricted MLE (RMLE) A = RTr, (3.1) of x, say *. We proposeto obtain* directly by maxiwheretheelements ofthematrix R = {Rij} are givenby mizing thelikelihood function thefollowing equations.For 1 h h-' b and 1 c i _ t we b 3 have Lac:fJ H )nhv (3.5) subjectto (3.4). In (3.5) theXhvare givenin terms of w (3.2) by(3.1). Denotetherestricted maximum ofL byL*. The constraint set (3.4) is linear in the1T'sand theobjective R3h,i = Phi(I)Phi 9 function logeL can be easilycheckedto be concavein andfor ii j< tifk = it - i(i + 1)/2+jwehave the uT's. The resulting nonlinear programming problem is thus well structured and can be solvedquiteeconom(Phi MPhj(2 + Ph/'IPhi 2) R3h-2,k (3.3) ically on a computer one ofthecommonly using available = -R3h,k . = R3h-I,k algorithms. To findb, the totalnumber of subsamples necessary 3.4 Propertiesof * to estimate thet marginal proportions {fi} and (t) bivariate proportions consider an extreme case (and a The RMLE * is biasedin smallsamples {7rri,}, butis asympmostfavorable (as nh~-Xo, Vh) unbiased.The asymptotic one from thestatistician's viewpoint) in totically varwhich theP valuescan be choseneither equal to zeroor iance-covariance matrix of *T(whichis also the exact
R3h1-,i = Phi (2
-

R3h-2,2 = Phi( '(l

- Phi

h=I v=O

Phi'

),

Tamhane: Randomized Response Techniques

919

variance-covariance matrix of theiUMLE of w) is given known vectorn. Thus,forimplementation purposes one oftheinformation bytheinverse matrix 5; we givebelow mustuse someprior estimate of n. an expression forthe elements of the upperleftt x t Becauseoftheabovedifficulties, we provide only some principal submatrix of 5. For 1 c i,j - t we have heuristic guidelines forthe choice of {Phi('I}. It can be readily verified thatifeach Phi(' = Ilt,thenmatrix R in 5ij = E{2 log L/atirtj} (3.1) becomesa deficient column rankmatrix and hence b 3 forfixed h and 1,thePhi") ,a is notestimable. Therefore, = shouldbe chosenas faraway (in either Enh direction) from Y (111h,)(8ah,1aUYaXh,1aUj)v=O h= I Ilt as possible, subject to somerespondent jeopardy con= 1 for1 - h straint and the constraint that = Ph I P) The remaining elements of5, which wouldinvolve axhvl _ b and 1 = 1, 2. In fact,forlarget, thelength of the can be obtained in an analogous manner. The atij terms, questionnaires can be cuV down for different subsamples various derivatives can be evaluated easilybyusing (3.1). by choosing setsof statements. If PhiCO= 0 fordifferent Expressions forthe variances and covariances of the theresearcher is equally interested in all the attributes, RMLE's of the 0's, say 0's, can be obtained from those thePhi/( shouldbe chosensymmetrically as faras posof the 1T's.Large-sample hypothesis testing concerning sible.For t = 2, sucha symmetric choice is provided by the 0's can be carried out by usingtheexpressions for + p11(2) = 1; subject to this restriction, and P1l(I) PI1(I) their asymptotic variances, withX replacedby its con(2) maybe chosenas farawayfrom A as thejeopardy X = R*. Expressions sistent estimate forthe (asymp- Pjl constraint permits. The choice will depend ontheaverage totic)variancesfor t = 2 are not givenhere but are educational and social sophistication of the population. obtainable from theauthor. A pilotsurvey should be carried outto testdifferent randomizing devices(different {Phi('}), as wellas theques3.5 Testof Independence tionnaire itself. Firstwe note thattesting pairwise independence between theoriginal attributes, sayAi andAj, is equivalent 4. A MEASURE OF RESPONDENT JEOPARDY to testingpairwiseindependence betweenthe corresponding modified attributes. we shall conTherefore, We shallconsider in competition twotechniques with sidertheproblem oftesting between independence pairs the M technique developedhere:the W technique and of modified attributes. the S technique. For a faircomparison betweenthese Suppose thatit is desiredto testthe hypothesis Hi: it is necessary to keep some measure techniques of the set i;. We can jeopardyof respondent's 'i. = lirj forall pairs(i,j) in a certain fixed.In thefollowing privacy use thegeneralized likelihood ratiomethod to testthis section we developsucha measure. 'hypothesis as follows. themaximum ofthelikeCompute lihoodfunction L in (3.5) subjectto thefollowing con- 4.1 Definition of the Jeopardy Function straints on theiT's and Warner Leysieffer (1976) and Lanke (1976) have I ? wi Vi, such for twodifferent developed approaches constructing measures. Herewe shallextend max(O, iT + iTj - 1) ' mij ' min(Qi,iTj) V(i,j) (EJ onlytheLeysieffer-Warattributes: The nerapproach to thecase oft 2 2 sensitive V(i,j) E i. Wij =i7rij in the same manner, Lanke approachcan be extended (3.6) butbecause of lack of space we do notdo so here;the conleads to thesamechoiceofdesign Lanke approach Denotethecorresponding maximum value ofL by Lg*. stants as the Leysieffer-Warner fordifferent techniques Then underHq asymptotically -2 log,(Lg*/L*)has a approach. distribution with chi-squared f degreesoffreedom (df), Considerthe 2' mutually exclusiveand collectively wheref is thenumber ofpairsin theset i. exhaustive is divided groupsintowhichthe population on thepossessionor nonpossession of differdepending entattributes, and denotethesegroups 3.6 Choice of {Phi/} byAIA2 . . . A,, an appropriateis obvious. Consider, say, the group A1A2 ... A,. By ofthe"optimal"(for Thedetermination on the re- usingBayes' theorem and subjectto suitableconstraints criterion in thesame manner as Leysieffer spondent jeopardy;see Sec. 5.1) choice of the design and Warner (1976),we can show thata measureof inbe- formation probabilities problem {Ph,('} appearsto be a difficult fromresponsev in favorof AIA2 resulting for the ... A, against of the expressions cause of the complexities (AIA2 ... A,)c is given by variance-covariances of {*} and thenumber asymptotic of designparameters thatcan be manipulated. It should g(v;A1A2. . . A,) (4.1) be pointedout thateven if expressions for"optimal" = P(v | A1A2 . . . A,)IP(v | (A1A2 . . . A,)c). theywoulddependon theun{Ph1(0} can be obtained,
A,cA2 . .. A,, . . . , A .cA2c ... A,c where the notation

920

Journal of the American Statistical Association, December

1981

as jeopardizing v can be regarded Thustheresponse with 4.3 Equating the Jeopardy Functionsforthe respect to the group AIA2 . . . A, (and not jeopardizing Competing Techniques Ourapproach herewillbe to first equatethejeopardy groups forthecompeting different functions forthefour A, or(AA2 ... A,)c if g(v; AIA2 ... A,) = 1. Now to design constants, equivalent techniques and obtaintheir oftheworst ofan geta measure jeopardyoftheprivacy theS technique. that is, their P valuesandthe, valuefor individualin group AIA2 . . . A, we definethejeopardy yieldedby the Clearly,the values of designconstants as functionforthatgroup in general be consistent. foursets of equationswillnot individuals guarding the We shallfollow theconvention of g(AiA2 ... A,) = maxg(v; AIA2 ... A,). (4.2) is, controlling g(A1A2) in themostsensitive group, that v The nextstep in our approachwill foreach technique. a measure of its perforeach technique in be to compute The jeopardyfunctions forothergroupsare defined We constants. formance based on thesevaluesofdesign an identical manner. to be thetrace The designconstants of each RR technique shouldbe have takenthemeasureof performance matrix of theesof theasymptotic variance-covariance chosenso that thejeopardy function valuesfordifferent forthevarivector.For t = 2, theexpressions groupsdo notexceed some prespecified upperbounds. timator are thethree techniques *2, and*r12 using We note here thatthesejeopardyfunction values will ancesof r1, the from to be givenherebutare obtainable 0's (in contrast dependin general to the too lengthy on theunknown comareusedinthenumerical Theseexpressions case oft = 1). Therefore, somea priori guessesat values author. out in Section5. parisons carried of 0's willbe necessary them. to compute with we see that Equating gw(A1A2) gM(AIA2), withrespect to either A IA2 . . . 1; and notjeopardizing 4.2 Jeopardy Functionsforthe Competing Techniques
PM
=

with respect to (AIA2

. .

. At)C) if g(v; AIA2

. .

. A,) >

{0

12

gw(AIA2)}

/
-

[{012*gw(A1A2)}1/2 + (1

012)1/2]

(4.3)

Usingthe definitions (4.1) and (4.2), we shallderive for PM expressions theexpressions for thejeopardy functions associated with ifgw(AIA2)' (1 - 012)/012*. Similar be obtained can and by equating g(AICA2), g(AIA2c), the W, S, and M techniques fort = 2. Here we shall buttheseare not fortheW and M techniques, consider in- g(AlcA2C) onlythefollowing specialcase of practical cannot the M technique be noted that given here. It should terest. (The general case witht ' 2 is quitestraightforthe W match at low the S technique) (and also technique wardbutalgebraically messyand is henceomitted.) For be two techniques would that is, the levels of gw(A1A2); the W technique we take P1 = P2 = Pw (say) wherePw in matched terms values for the of their jeopardy A1A2' loss of generality. For the S technique we > i without than (1 - 012)/012*. onlyifgw(AIA2)is notsmaller take PI = P1 = Ps (say) and PI = P2 = P (say). For the group and we obtain Next, equating gw(A1A2) gs(A1A2), M technique we takeP11(1)= 1 - P1 (2) = PM(say)where PM > 2 without loss ofgenerality. Ps = ,(2Pw - 1)/[(1- PW) + P(2Pw - 1)] (4.4) deavailablewith Thus we have a class of S techniques 1 - 01 - 02 + 012.The expressions forthejeopardy signconstants classwe (4.4). Fromthis (Ps, 0) satisfying functions are givenin Table 1; thederivations of these can makean optimal that combination choicebyselecting are obtainable expressions from theauthor. the trace of the (asymptotic) ,3) whichminimizes (Ps, Table 1. ExpressionsforJeopardy Functions
W Technique g(A1A2) PW2(1 - 012)/{PwQw(1 - 012 - 012') + Qw2012*} S Technique (Ps + QsO)2(1 - 012)/{QsP(Ps + QsO) x (1 - 012 - 012*) + Qs2p2012*' g(AicA2) Pw2(1 - 02 + 012)/{PwQw(012+ 012*) + Qw2(01 - 012)) (Ps + QsP)(Ps + QsY)(1 - 02 + 012)
+

Define additionalnotationas follows: Qw = 1 - Pw, Qs = 1 - PS, QM = 1 - PM, y = 1 - ,B,and 012* =

M Technique PM2(1 - 012/QM2012*

+ 012*) (1 - 02 + 012)/PMQM(012 012)

{QsY(Ps

+ QSP)012

+ Qs2py(01

+ QsP(Ps + QsY)012*} g(A1A2) PW2(1 - 01 + 012)/{PwQw(012+ 012') + QW2(02 - 012)) (Ps + QsP)(Ps + Qs'Y)(1 - 01 + 012) {QstY(Ps + QsP)012 + Qs2P'Y(02 - 012) + QsP(Ps + QSY)012*} g(AlCA2C) Pw2(1 - 012')/{PWQW(1- 012 - 012') + QW2012} (Ps + Qsy)2(1 - 012')/{QsY(Ps + QsY) PM2(1 - 012')/QM2012 + 012') (1 - 01 + 012)/PMQM(012

x (1

012 - 012*) + QS2'Y2012}

Tamhane: Randomized

Response Techniques

921

variance-covariance matrix of the associatedestimator When01and 02 are small(<.05), theM technique dom= .7 uniformly vector.The actual analytical minimization problem is inates theS technique with 1B inp andPw messy becauseofthecomplexity ofthecriterion function. in all cases studied.When01 and 02 are moderate (bewe can intuit However, theoptimal choiceof(Ps, O) by tween.05 and .10),theM technique dominates whenever noting thatthecriterion function shouldbe a decreasing p is notnegative or Pw is nottoo large, or both.Finally, function ofPs for fixed 1B andthat from (4.4) themaximum when01and 02 are somewhat large(>.10), theM tech= 1. Thustheoptimal valueofPs is obtained when13 S nique dominates the S technique onlywhenp is suffitechnique is the repeatedapplication of the so-called ciently largeand positive andPw is nottoolargeor both. forcedyes technique(Drane 1975) withPs = (2PW - 1)/ The range ofvaluesof01,02, p, andPw forwhich theM dominates theW technique is evengreater. In Pw and 1 = 1. If it is desiredto have gw( ) and gs(Q) technique equal forall thefour groups, thenwe obtain situations dealingwithtwo sensitive atPs = 2PW manypractical = 1. For thischoice of parameters - 1 and 13 thecriterion tributes, 01 and 02 are in factlikely to be smalland the functions fortheW and S techniques are identical, and correlation between theattributes is likely to be positive therefore thetwotechniques are equivalent; thisextends and large.Furthermore, are nottoo large Pw valuesthat the corresponding resultfor t = 1 by Leysieffer and (usuallyin the rangeof .7 to .75) are morecommonly Warner (1976). used. Thusfortheparameter valuesthat are likely to be inpractice, encountered theM technique doesreasonably 5. COMPARISONOF COMPETINGTECHNIQUES well,although notoptimally well. 5.1 Numerical Results 6. APPLICATION OF THEM TECHNIQUE as the Define thetraceinefficiency ofan RR technique ratio ofthetraceofthe(asymptotic) variance-covariance 6.1 Descriptionof the Application matrix of its estimates for01, 02, and 012 to the correthefeasibility oftheM technique infacethedirect when To determine for sponding quantity response technique to-face a interviews, study involving an actual application boththe techniques use the same samplesize n. This was carried out.It was nottheobjective latter quantity is givenby {01(1 - 01) + 02(1 - 02) + ofthetechnique of this small to thepractical feasibilities study compare 012(l - 012)}In. and RR in of all the discussed performances techniques For numerical 10 (01, 02) combinations comparisons, the that would have reprevious sections; comparison to a wide rangeof theseparameters representing likely resources thanwere in practicewere selected;we take 02 quireda largerstudyand greater be encountered it available to us. was decided to includea However, For each (01, 02) comloss of generality. -' 01 without control who would take the reof direct group subjects three012 values were selected:012 = 0, 02/2, bination and who would a datum interview sponse provide against values of therangeof admissible and 02, thuscovering theperformance of theM technique can be comcoef- which 012. For each (01, 02, 012) the value of correlation with and truthto extent of pared respect cooperation ficient P12 was calculated by usingtheformula fulness of responses.Subjectswererandomly allocated 01)(l - 02) P12 = (012 - 0102)/V0102(1 to thetwogroups as explained below. intheSpring 1980Industrial Psychology (IEForeach (0I, 02, 012) combination theresults correspond- Students at 152 class Northwestern the C22) University provided ing to four Pw values (Pw = .70, .75, .80, .85, which for the Three other students from the subjects study. therangeofPw valuescommonly represent used) were same and trained to out the class were recruited carry hereonlytheresults forPw = .70 calculated, although with the student interviews. Based on discussions counand .80 are given;the results forotherPw values are thefollowing oftheUniversity Clinic, obtainable from theauthor. For eachPw thecorrespond- selorand thestaff two issues were identified as senrelevant, potentially ingvalue ofPM was computed by using(4.3). For theS hard and and correlated: sitive, possibly (a) using drugs are the results fortwo (Ps, I) combinations technique the (b) seeking psychiatric help. Accordingly, following withI = 1 and another combination given:an optimal were prepared foruse in the directreone withI = .7; in either case, thePs value was com- two statements the of course,for and M sponse technique interviews; fortheW techputedfrom (4.4). Of course,theresults M in the statement was the second technique presented to I = .5. Thuswe geta detailed niquecorrespond picture the it with of form the inclusion negative by modifying of the performance of the S techniquefor different "not." The values of thetrace theparenthetical choicesof its designconstants. withdesignconforall threetechniques inefficiencies 1: I presently takeor inthepastsix months Statement in the above manner werecomputed havetaken stants determined at leastone ofthefollowing on a regular drugs and are givenin Table 2. basis,thatis, on theaverage,at leastonce a weekfora or longer; month acid,angeldust,cocaine,heroin, quaa5.2 Discussion of the Results ludes,speed,other drugs in thesame category. Identify whether you belongto thisgroupby saying"Yes" or First,notethatthe "optimal"S technique with13 = 1.0 dominates theother techniques in all cases studied. "'No."

922

Journalof the American Stafistical Association,December 1981

Table 2. Trace Inefficiencies

O1
02 012 P12

Pw

PM

W Technique 62.859 16.579 53.792 14.296 47.193 12.634 48.146 12.904 38.520 10.473 32.341 8.936 34.051 9.386 29.074 8.123 25.565 7.233 26.612 7.530 21.264 6.166 18.075 5.353 24.583 7.026 20.930 6.093 18.438 5.456 19.594 5.783 15.617 4.760 13.396 4.189

S Technique .7 1 = 1.0 29.109 11.705 24.850 10.054 21.747 8.850 22.118 9.070 17.610 7.306 14.750 6.185 15.386 6.535 13.075 5.616 11.439 4.964 11.833 5.198 9.365 4.199 7.875 3.593 10.833 4.824 9.159 4.142 8.007 3.671 8.427 3.919 6.621 3.167 5.583 2.728

M Technique 33.126 12.469 27.248 10.416 22.984 8.901 26.592 10.103 19.633 7.710 15.273 6.141 20.736 7.911 16.265 6.469 13.174 5.417 18.020 6.800 12.042 4.984 8.667 3.800 17.624 6.562 13.019 5.248 10.056 4.305 17.121 5.987 9.920 4.179 6.507 3.038

.05 .05 .05 .05 .05 .05 .10 .10 .10 .10 .10 .10 .15 .15 .15 .15 .15 .15

.025 .025 .025 .05 .05 .05 .05 .05 .05 .10 .10 .10 .075 .075 .075 .15 .15 .15

.0000 .0125 .0250 .0000 .0250 .05 .0000 .0250 .0500 .0000 .0500 .1000 .0000 .0375 .0750 .0000 .0750 .1500

- .0367 .3306 .6980 -.0526 .4737 1.0000 - .0765 .3059 .6882 -.1111 .4444 1.0000 -.1196 .2791 .6778 - .1765 .4118 1.0000

.70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80 .70 .80

.6815 .7766 .6875 .7841 .6937 .7919 .6753 .7690 .6874 .7839 .7000 .8000 .6626 .7539 .6746 .7682 .6870 .7835 .6497 .7388 .6739 .7674 .7000 .8000 .6431 .7312 .6611 .7521 .6800 .7747 .6226 .7083 .6595 .7502 .7000 .8000

42.387 13.743 36.221 11.824 31.731 10.426 32.349 10.672 25.807 8.625 21.663 7.327 22.709 7.725 19.336 6.659 16.953 5.905 17.622 6.170 14.003 5.015 11.832 4.319 16.215 5.742 13.749 4.952 12.061 4.409 12.787 4.696 10.110 3.826 8.594 3.329

the deck well and drawa card at Statement 2: In thelastsix months I have(not)sought was asked to shuffle helpfora mental, emotional, or a psychological problem random thesubject (and notshowitto theinterviewer); (secondstatestatement from a professional suchas a psychiatrist, tothefirst psychologist,was askedto respond or a social worker. (club), Identify whether you belongto this ment) ordiamond ifthecardcameup spade,heart, group by saying "Yes" or "No." The card was reand his (her)responsewas recorded. but was repeated, turned to thedeck and theprocedure The sub- the choices of statements The interviewing procedure was as follows. thistime;thus was reversed ject entered theinterview room.His (her)namewas re- P1l(l) = P12(21 - .75. The sheetof paperand the deck corded,which,it was hoped,would make the subject werethen Next,toassess the totheinterviewer. returned take more seriously the sensitivity of the statements. extent overthedirect of theM technique of preference Then the subjectwas randomly allocatedto one of the responsetechnique, questionwas asked: the following two groups(directresponseinterview or M technique "Supposingforthe moment to thatyourtrueresponse interview). In thecase of thedirect responseinterview either was 'Yes,' wouldyou feel of thetwo statements theprocedure was swift and simple and willnotbe elab- more, less, or equally comfortable withthis indirect interview method orated on here.In thecase oftheM technique method to thedirect as compared ofquestioning a sheetofpaperbearing thetwostatements was handed of questioning?" was reThe responseto thisquestion to therespondent alongwith a deckofcards.The subject cordedand thustheinterview the After was concluded.

Tamhane: Randomized

Response Techniques

923

wereaskedto notedownany responded interview, theinterviewers that they wouldfeelequallycomfortable, and theywouldfeelless comfortable. unusualthings in understanding the in- 5 responded (e.g., difficulty These restructions) thathappened sults showthat during thedegree theinterview. oftruthfulness andcooperation by the respondents can be improved by usingthe M 6.2 Results of the Application technique. in the M technique is the summary Following of the responses obtained Finally,out of 77 respondents group, about5 respondents had somedifficulty following the two by using techniques: the instructions and neededto go over the instructions Direct response: n = 75; No-No = 71, Yes-No = 3, one moretime. No-Yes = 1, Yes-Yes = 0. Fromthis we can conclude application that theM techM technique:n = 77; No-No = 14, Yes-No = 5, No- niqueis feasible in practice and is likely to improve the Yes = 41, Yes-Yes = 17. cooperation on thepartof respondents and thusreduce is neededin explaining Thusfrom thedirect interviews we obtain the the bias. Some care, however, response the instructions the to respondents. estimatesalong with their standarderrors following = (giveninsidethe parentheses): .04 = 01 (.0226), 02 [Received January 1977.RevisedDecember1980.] To obtain theRMLE's ofthe0's (byfirst the obtaining REFERENCES RMLE's ofthe1T's) theM technique interview from data, Reduced Grawe must maximize (3.5) subject to (3.4). Forthispurpose ABADIE, J.,andGUIGOU, J.(1969),"The Generalized dient" (in French),Electricite de France (EDF) Working Paper we usedthegeneralized reduced GRG algorithm HI-69/02. gradient ofAbadieandGuigou Response thefollowingBARKSDALE, W.B. (1971),"New Randomized Techniques (1969),which yielded for Control of in Nonsampling Errors Surveys," unpublished Ph.D. = = estimates: .05195 (.0904), .01300 02 (.0774), 012 01 thesis, University ofNorth Carolina, Chapel Hill,Dept.ofBiostatistics. = .01039(.0562).The asymptotic standard errors ofthe BOURKE, PATRICK D. (1975),"Randomized ResponseDesignsfor Estimation," Report 6, Research Project on Confidenestimates (giveninsidetheparentheses) werecomputed Multivariate in Surveys, University of Stockholm, Dept. of Statistics. theformulas from obtained theinformation tiality by inverting CLICKNER, R.P., andIGLEWICZ, B. (1976), "Warner's Randomized matrix value of the ResponseTechnique: givenin Section3.4. The maximum TheTwo Sensitive Questions Case," Proceedlikelihood function was L* oc(.314555)77. Notethat inthis ingsoftheAmerican Statistical SocialStatistics Association, Section, case theRMLE's are the same as theUMLE's; thatis, 260-263. DEVORE, JAYL. (1977),"A NoteontheRandomized TechResponse theUMLE's satisfy theconstraints (3.4). nique,"Communications in Statistics, A6, 1525-1527. For testing 1 and 2 are uncor- DRANE, W. (1975),"Randomized H; thatthe attributes to MoreThanOne QuesResponse Social of theAmerican Statistical Association, thatis, 012 = 0102, theGRG program related, was again tion,"Proceedings 395-397. Section, runwiththeconstraint set (3.6). This yielded themaxi- Statistics (1976),"On theTheory ofRandomized toTwo SenResponses mumvalue ofthelikelihood function under H;, namely sitive inStatistics-Theory Communications andMethQuestions," The valueofthex2 statistic works out ods, A5, 565-574. L9* oc(0.314479)77. ERIKSSON, S. (1973),"Randomized Interviews forSensitive Questo be .0372.Comparing thiswith theupper critical values tions,"unpublished Ph.D. thesis, Gothenburg University, Dept. of ofthechi-squared distribution with one df,we conclude Statistics. thatthe nullhypothesis of independence cannotbe re- FLIGNER,MICHAEL A., POLICELLO, GEORGE E. II, andSINGH, JAGBIR ofTwo Randomized Sur(1977),"A Comparison Response jected.Thissmallvalueofthex2 statistic is possibly due veyMethods, With Consideration fortheLevel ofRespondent Proto two reasons:(a) we are dealingwithrareattributes tection," in Statistics, Communications A6, 1511-1524. hereandtherefore much larger samplesizes are required GREENBERG, B.G., ABUL-ELA, A.A., SIMMONS, W.R., and HORVITZ, D.G. (1969),"The Unrelated Question Randomized Reto obtain a sufficiently powerful test;(b) in general, any sponseModel: Theoretical Journal Framework," of theAmerican RR technique yieldsa less powerful testcompared with Statistical Association, 64, 520-529. thedirect responsetechnique (assuming, of course,re- HORVITZ, D.G., GREENBERG, B.G., and ABERNATHY, J.R. in Randomized (1975),"RecentDevelopments ResponseDesigns," sponsesare equallytruthful forboththetechniques). in A Survey of Statistical Designand LinearModels,ed. J.N. Srivastava,Amsterdam: North 271-285. Holland, Device for (1976),"Randomized Response:A Data Gathering Sensitive International Statistical Questions," Review, 44, 181-196. in Randomized InFirst,we notethatsomewhat higher estimates of the LANKE, J. (1976),"On theDegreeof Protection International terviews," Statistical Review, 44, 197-203. 0's areobtained with theM technique than thoseobtained LEYSIEFFER, R.W.,andWARNER,S.L. (1976),"Respondent Jeopwiththe direct responsetechnique, although thediffer- ardy andOptimal inRandomized Designs Response Models,"Journal Statistical Association, 71, 649-656. encesare notstatistically Thismight significant. indicate oftheAmerican D., and FEDERER, W.T. (1979),"BlockTotalRethat therespondents tendto be moretruthful with theM RAGHAVARAO, in to theRandomized sponseAs an Alternative ResponseMethod technique interview. To thequestion(asked onlyof in- Surveys," Ser.B, 41,40-45. Journal ofthe RoyalStatistical Society, A Survey Response: Technique dividuals intheM technique whether therespond- WARNER,S.L. (1965),"Randomized group) EvasiveAnswer American StaEliminating Bias," Journal ofthe ent wouldfeelmore,equally,or less comfortable with for tistical Association, 60, 63-69. theM technique with than thedirect response technique, (1971),"The LinearRandomized Response Model,"Journal of theAmerican 43 responded Statistical thatthey wouldfeelmorecomfortable, Association, 66, 884-888. 29

.0133 (.0132), 012

0 (.00).

6.3 Discussion of the Results