Nuances of information criteria for Bayesian psychometric models [Author Accepted Manuscript]

It is common practice to compare Bayesian psychometric models via information criteria such as DIC and WAIC. Especially because these criteria can be automatically computed by MCMC software, it is easy to ignore the intricacies related to their computation. This often leads researchers to use noisy criteria that may lead to suboptimal analysis decisions. In this paper, we first review different forms of Bayesian information criteria that could be computed for psychometric models. We then consider best practices, highlighting computational pitfalls that can occur even when one is attempting to follow best practices. Finally, we provide recommendations for the metrics’ practical uses. The paper is intended to clarify conflicting recommendations from the literature and to raise awareness about ways that information criteria can behave unexpectedly. This work was made possible through funding from the Institute of Education Sciences, U.S. Department of Education, Grant R305D210044. reviewed acceptedVersion