Optimality Considerations in Testing
Massive Numbers of Hypotheses

Peter Westfall
Texas Tech University


In the current climate of analysis of massive data sets,
there areconcerns about how well methods "scale up."
By most accounts, classical FWE-controlling methods do
not scale up well to increasing p, while FDR-controlling
methods do.  On the other hand, neither method scales up
well with increasing n for typical point null testing.
If priors and losses are correct, and if computation is
feasible, scaling is not an issue for decision-theoretic
methods. I will investigate FWE-controlling methods and
FDR-controlling methods, as well as a newer proposal of
Brad Efron, from the standpoint of minimizing loss, for
different types of realistic loss functions.

Joint work with Ananda Bandulasiri