Title: Confounder adjustment in large-scale linear structural models
Authors: Qingyuan Zhao - University of Cambridge (United Kingdom) [presenting]
Yang Song - Stanford University (United States)
Abstract: Consider large-scale studies in which thousands of parameters need to be estimated or tests need to be performed simultaneously. In some of these studies, the usual linear regression can be severely biased by latent confounding factors. Two applied examples will be considered. The first is multiple hypothesis testing in genomic screening, where the confounding factors might include batch effect and unmeasured environmental variables. The second example is evaluation of the performance of mutual funds, where the confounders are any systematic risk factors that are not included in a standard factor model (such as the Fama-French-Carhart four factor model). A two-step procedure based on factor analysis and robust regression will be proposed, and some theoretical guarantees will be given. The statistical method will be applied to a mutual fund return dataset.