On the small sample behavior of Dirichlet process mixture models for data supported on compact intervals

Abstract

Bayesian nonparametric models provide a general framework for flexible statistical modeling of modern complex data sets. We compare a rate-optimal and rate-suboptimal Bayesian nonparametric model for density estimation for data supported on a compact interval, by means of the analyses of simulated and real data. The results show that rate-optimal models are not uniformly better, across sample sizes, with respect to the way in which the posterior mass concentrates around a true model and that suboptimal models can outperform the optimal ones, even for relatively large sample sizes.