Modeling random responding behavior and extreme response style in surveys

Abstract

Likert-type scale are widely used in social and psychological surveys. Some aberrant responding behaviors have been pointed out in literatures. For example, respondents with low motivation attempt to go through the instrument quickly and consequently endorse the given response options randomly. Another case is extreme response style (ERS), which means that respondents’ tendency of using extreme responses would intervene the usage of response categories. The two responding behaviors were considered in this study. We aim to propose a new item response theory (IRT) model for distinguishing random respondents from attentive respondents and account for ERS of attentive respondents to improve the measurement quality of the test. The parameters were estimated with the Markov chain Monte Carlo (MCMC) method, which is available via the free software WinBUGS. The preliminary results showed the estimated parameters in the new model can be recovered very well. The results also indicate that fitting the new model to data without random responses and extreme responses did not yield seriously biased estimations of parameters. In the opposite way, ignoring random responses and extreme responses by fitting standard IRT models resulted in seriously biased estimations on the item slope parameters; and the item difficulty parameters and the threshold parameters were biased as well. The implications and applications of the new model will be illustrated by an empirical study.