Comparison of Estimators for Multi-level Randomized Response Data: Evidence from a Case of Sexual Identity

Abstract

In surveys, proportions of individuals self-reporting as bisexual and homosexual are likely to suffer from misreports or refusal to answer when directly questioned. For more reliable information, the randomized response technique is often used. Therefore, to reduce social desirability bias, a multi-level randomized response technique has been used in Taiwan to collect self-reported sexual identity. Using the Taiwan Social Change Survey as a case study, we estimate proportions of individuals aged 18–54 years self-reporting as bisexual and homosexual. Empirical results show individual determinants of truthful response differ across data collection modes and estimation methods. Moreover, estimated proportions of non-heterosexuals in small samples obtained by the maximum likelihood method are less meaningful and accurate than Bayes ones.