Department Seminar Series: Michael Newton, Putting lots of things in order: rvalues for ranking in large-scale inference

Abstract

Hypothesis testing approaches have dominated high-dimensional inference in genomic applications.  In many contexts, the precision with which individual parameters are estimated varies greatly among parameters, and testing approaches, which are most concerned with type I errors, behave poorly in ranking and selecting the most interesting (largest) individual parameters owing to an imbalance of power.  I present a framework for evaluating different ranking/selection schemes as well as an empirical Bayesian methodology showing theoretical and empirical advantages over available approaches.   This is joint work with Nicholas Henderson; a preprint is available at http://arxiv.org/abs/1312.5776.  Time permitting I will discuss a new approach to building mixture-model components using constrained latent variables.