Inferring population structure using bayesian clustering programs often requires a priori specification of the number of subpopulations, K, from which the sample has been drawn. Here, we explore the utility of a common bayesian model selection criterion, the Deviance Information Criterion (DIC), for estimating K. We evaluate the accuracy of DIC, as well as other popular approaches, on datasets generated by coalescent simulations under various demographic scenarios. We find that DIC outperforms competing methods in many genetic contexts, validating its application in assessing population structure.