Statistics Seminar: Matteo Sesia

Event Date

Location
Mathematical Sciences Building 1147 (Colloquium Room)

SPEAKER: Matteo Sesia, PhD Candidate, Statistics, Stanford University

TITLE: “Developments in knockoff generation and genetic mapping”

ABSTRACT: Modern scientific studies often require the identification of a subset of relevant explanatory variables from a large set of candidates, in the attempt to understand a phenomenon of interest. Knockoffs offer a general framework for this task, making it possible to rigorously test the conditional importance of each variable while controlling the false discovery rate in finite samples, without strong parametric assumptions.

In this talk, we present some developments of this methodology that enable practical applications, specifically considering problems in which either the distribution of the explanatory variables is completely unknown, or is understood to be well-approximated by a hidden Markov model. An interesting example of the latter scenario arises in genome-wide association studies, for which we develop a knockoff-based method for the genetic study of complex phenotypes. Through simulations and analyses of several phenotypes in the UK Biobank data, we demonstrate that this leads to more precise and numerous discoveries than state-of-the-art alternatives, at comparable computational cost.