Statistical Methods for Large-scale Multiple Testing Problems

Abstract

A large-scale multiple testing problem simultaneously tests thousands or even millions of null hypotheses, and it is widely used in different fields, for example genetics and astronomy. An error rate serves as a measure of the performance of a testing procedure. The use of the family-wise error rate can accommodate any dependence between hypotheses, but it is often overly conservative and has limited detection power.The false discovery rate is more powerful, however not as widely used due to the requirement of independence and other reasons. In this thesis, we develop statistical methods for large-scale multiple testing problems in pharmacovigilance and genetic studies, and adopt the false discovery rate to improve the detection power by tacking mixed challenges.