Abstract

We propose a new Bayesian nonparametric multivariate testing procedure for comparing several treatments against a control. The test is based on a general model where the distribution of each treatment group can be identical to (or different from) the control group distribution, depending on the value of a latent binary vector. This vector is endowed with a spike-and-slab prior distribution carefully chosen to ensure a multiplicity correction. Group distributions are modeled in a flexible way using a dependent Dirichlet process. Monte Carlo experiments suggest that our proposal performs better than state-of-the-art frequentist alternatives for small sample sizes.