Honest answers to embarrassing questions: Detecting cheating in the randomized response model.

Abstract

Surveys and questionnaires are frequently used by psychologists, social scientists, and epidemiologists to collect data about behavior, attitudes, emotions, and so on.However, when asked about sensitive topics such as their sexual behavior or illegal activity, some respondents lie or refuse to answer. The randomized response method was developed to reduce these evasive answer biases by guaranteeing subject privacy. However, the method has been criticized as being susceptible to cheaters, that is, respondents who do not answer as directed by the randomizing device. Here the authors show that by splitting the sample into 2 groups and assigning each group a different randomization probability, it is possible to detect whether significant cheating is occurring and to estimate its extent while simultaneously protecting the identity of cheaters and those who may have engaged in sensitive behaviors. Surveys and questionnaires are standard methodsfor collecting data about behavior, attitudes, emotions, and so forth. The basic assumption of any interview or survey technique is that the respondents are providing honest information. The validity of this assumption is questionable, however, when researchers ask questions that most would be reluctant to answer publicly. Examples of such questions are those that reveal whether the respondent has engaged in an illegal activity, or an activity that is stigmatized by society, or the query may pertain to a behavior of which the respondent is ashamed or about which the respondent feels is simply too personal to confide in someone else. Faced with such a question, some individuals in a sample will refuse to answer or will lie. Either

A stochastic lie detector

“…Finally, comparable to related randomized-response models (Clark & Desharnais, 1998;Cruyff et al, 2008;Ostapczuk et al, 2009a), the stochastic lie detector assumes that response behavior does not depend on features of the randomized-response design. Specifically, it is assumed that the probability to respond honestly (t) is independent of the randomization probabilities p 1 and p 2 .…”

“…To estimate the two unknown parameters of the model (π and t), it is necessary to draw two independent random samples from the population of interest and to apply different randomization probabilities in each sample (p 1 ≠ p 2 ) (Clark & Desharnais, 1998). Specifically, carriers in the first random sample receive the sensitive question with probability p 1 and the negation of the sensitive question with probability 1 -p 1 , whereas carriers in the second random sample receive the sensitive question with probability p 2 and the negation of the sensitive question with probability 1 -p 2 .…”

“…Several extensions of the forced-response variant of the RRT have been proposed that allow for determining the proportion of respondents who fail to adhere to the instructions by always providing a self-protecting response (Böckenholt, Barlas, & van der Heijden, 2009;Böckenholt & van der Heijden, 2007;Clark & Desharnais, 1998;Cruyff, van den Hout, & van der Heijden, 2008;Ostapczuk, Moshagen, Zhao & Musch, 2009a;van den Hout, Böckenholt, & van der Heijden, 2010). However, all of these approaches share the limitation that the true status of nonadherent respondents cannot be determined.…”

E-research – selected issues concerning the Internet attitudes and opinions research

“…It is important to note that while this modified design addresses a particular type of noncompliance, in practice respondents may exhibit other types of noncompliance behavior. This design is a special case of the design proposed by Clark and Desharnais (1998) who also considered assigning different probabilities for randomization device across groups. Below, we contribute to this literature by developing a multivariate regression technique and power analysis for this modified design.…”

Bayesian randomized item response modeling for sensitive measurements

“…The randomized response (RR) technique is a survey technique that can reduce response bias in sensitive survey studies (Clark & Desharnais, 1998;Lamb & Stem, 1978;Lensvelt-Mulders, Hox, van der Heijden, & Maas, 2005;Reinmuth & Geurts, 1975;Warner, 1965). Initially proposed by Warner, the developed randomized response techniques insure confidentiality of individual responses by introducing a chance mechanism in the response process governed by a randomizing device.…”