J Clin Epidemiol Vol. 49, No. 5, pp. 545-549, 1996 Copyright 0 1996 Elsevier Science Inc. ELSEVIER 0895s4356/96/$15.00 PI1 SO895-4356(96)00005-4 Respondent-Specific Information from the Randomized Response Interview: Compliance Assessment Brian E. Rittenhouse DEPARTMENT OF PHARMACOLOGY, SCHOOL OF MEDICINE, UNIVERSITY OF MONT&AL, MONT&AL, QUEBEC, CANADA ABSTRACT. In situations in which researchers ask potentially embarrassing questions, respondents may feel uncomfortable with revealing certain behavior. Consequently, response rates or accuracy may be low. The “randomized response interview” (RRI) was developed to enable researchers to better elicit responses to such questions. The technique has clear potential in estimating population proportions engaging in embar- rassing behavior. It does not appear to have been recognized that one may also obtain more respondent- specific information from the application of the RRI. Th is article indicates that while still only probabilistic, respondent-specific information is obtainable from the RRI. J CLIN EPIDEMIOL 49;5:545-549, 1996. KEY WORDS. Compliance, adherence, assessment, interview, randomized responses, estimation INTRODUCTION Survey research accuracy clearly depends among other things on the truthfulness of respondents in answering interviewer questions. In cases in which answers to questions can be embarrassing and in which, as a consequence, responses may not be truthful (or even forthcoming), researchers have used a technique called the “ran- domized response interview” in a sample to estimate population pro- portions of persons engaging in the embarrassing behavior [l-5]. While this technique has many specific formulations, the spirit is simple to convey. The randomized response interview (RRI) allows respondents to insulate themselves from the potential opprobrium of interviewers by virtue of there being only a (specified) probability that the answer given is truthful. A “yes” answer to an embarrassing question is only probabilistically a true “yes” answer. Individual re- sponses are “randomized” according to a protocol by which they are, for example, to answer a question affirmatively (regardless of truth), negatively (regardless of truth), or truthfully based on some probabi- listic event, for example, throws of dice, the number rolled being unobserved by the interviewer (a crucial aspect). Thus, responding in the affirmative to an embarrassing question becomes possible to a respondent because he or she realizes that the interviewer will not know the truthfulness of the specific response. In tests of this method, it has often been shown to yield more accurate results than simple direct questioning about behavior [3,6]. The RR1 has also been shown to be more informative than telephone or self-adminis- tered questionnaires [7,8]. The imperfect knowledge on the part of the interviewer of the truthfulness of the individual response enables respondents (in the- ory) to answer truthfully when the randomization device indicates for them to do so. That this in fact works is supported by some of the research cited above. The obvious cost of this truthfulness is the inability to tell for certain what the actual individual behavior is. Address for correspondence: Brian E. Rittenhouse, Ph.D., DCpartement de pharmacologic, Faculte de mCdecine, UniversitC de Mont&l, C.P. 6128, Succ. Centre-ville, Montreal, Quebec H3C 3J7 Canada. Accepted for publication on 24 May 1995. While it may be obvious that such a buffer may insulate respondents from embarrassment, it is not immediately obvious what value any response has when its truthfulness is unknown. What can be accom- plished is an accurate assessment of the frequency of the behavior in a sample, and, subject to the usual caveats in any research, in the population from which the sample is obtained. The interviewer is in control of the probabilities involved with the RRI. With that knowledge, and the set of responses from the sample, the interviewer can construct an accurate estimate of the proportion of the sample that engages in the behavior of interest. Similarly, a probabilistic estimate of the behavior of any individual can be made. The implicit message from the literature is that this estimate is uniform across the sample (and equal to the proportion of the sample). In many contexts, researchers are not particularly interested in individual behavior but in a sample or population proportion esti- mate of those behaving in particular ways [5]. In other contexts, while the avowed purpose of the RR1 question is to assess population proportions, this may be primarily because it is considered impossi- ble through the RR1 to estimate anything other than uniform indi- vidual probabilities of certain behaviors. It may be that probability estimates for individual behavior are of interest to many researchers, but that estimation of a uniform probability for all individuals in a sample may be considered of minimal value. Rittenhouse [9] has suggested that RR1 procedures may be useful in pharmacy settings, medical practices, or clinical trials to obtain estimates of population proportions of patients who do not properly take their medications, but who may be unwilling to admit to their noncompliance under traditional direct questioning about their be- havior. In such a case, in addition to the potential (primarily re- search) value of population proportion estimates, knowing individ- ual behavior could be of particular interest so that compliance problems may be addressed at individual levels. This article suggests two apparently previously unknown addenda to research on the RRI. The first is that significant information is provided by individuals about their own behavior when responding to such RR1 questions (i.e., one can do better than a uniform proba- bilistic estimate of individual behavior). The second suggestion is