In surveys it is problematic to get valid responses when studying topics that respondents find threatening to answer questions about (Lee, 1993). Researchers analyzing such sensitive topics are more likely to be confronted with higher non-response rates and/or socially desirable answers, which threatens the validity of results.

One of the solutions to this problem is to use a Randomized-Response (RR) technique (Warner, 1965; Kuk, 1990; Chaudhuri & Mukerjee, 1988; Fox, 1986). In this technique the answer of the respondent depends in part on a randomization device (typically the throw of two dice) and because of this the interviewer/researcher can no longer determine what the response of the individual respondent was. Because the properties of the randomization procedure are known, valid inferences about the behavior under study can still be made.

Although RR-techniques are well-known and meta-analyses show that they improve the validity of the results (Edgell et al., 1982; Lensvelt-Mulders et al, 2005), they are not often used. Several reasons are given for this in the literature. First, to get a comparable precision, RR-techniques require larger sample sizes. Second, it seems that researchers erroneously think that RR-techniques do not allow for analyses at the individual level.

The usage of RR-techniques online is in its infancy. On the one hand the anonymity of the internet is expected to increase the willingness to answer sensitive questions, on the other hand the implementation of RR-techniques online is complicated by the fact that online implementations of a randomization device are less likely to be trusted to guarantee anonymity by the respondent.

Design In our study we collected survey-data of about 3,250 respondents (initial wave of 1,000 and second wave of 750 from wave 1 and 2,250 new respondents) of a large Dutch Internet-panel (200,000+ members) who answered questions about three sets of potentially sensitive topics: traffic violations, possession of illegal software and music, and answering behavior in the online panel. Our design allows, among others, for comparisons with respect to the following issues:

• Voluntary versus involuntary randomization: For those respondents who do not mind answering the questions honestly, using the randomization procedure produces unnecessary noise. We compare direct question and standard randomization response techniques with voluntary choice as to the method of elicitation.
• Browser based versus locally executable randomization. In most online applications, RR-techniques simulate the throw of dice in the browser. Obviously, it would be easy for the researcher to monitor or manipulate the throw of the electronic dice, undermining the faith of the respondent in the procedure. We compare direct question and standard randomization response techniques with the case where the randomization device is a downloadable application.
• Determinants of compliance and test-retest reliability. Despite the procedure, respondents sometimes answer in socially desirable ways just to make sure that they do not confess to any kind of sensitive behavior. Using multiple RR-measurements, we identify which kinds of respondents are more likely to exhibit this kind of behavior. Moreover, in an initial wave we asked the same sensitive questions directly to 1,000 respondents, which allows for another way of determining compliance and test-retest reliability of the measurement procedures.

The data are being collected end of September, so we cannot present results at the time of submission of this abstract. The analysis of RR-data has thusfar been restricted to somewhat non-standard software (such as R and GLIM). Besides the issues mentioned above, in our presentation we will briefly elaborate on ways to use standard software to analyze RR-data.