B1523
Title: Penalized mixed models to adjust for batch effects and unobserved confounding in high dimensional regression
Authors: Patrick Breheny - University of Iowa (United States) [presenting]
Abstract: There have been several recent developments with respect to the idea of "deconfounding", or adjusting for hidden confounders. In classical scenarios, this is not possible, but in high dimensions, a hidden confounder will typically leave traces upon multiple features. This introduces the possibility of using the correlation structure of the features to control for confounders, even if they are unobserved. Deconfounding ideas are explored in penalized regression models such as the lasso, illustrating how penalized linear mixed models (LMMs) can correct for hidden confounding. In addition, the performance of this method is contrasted with that of principal components (PC)-based methods. Finally, the use of these methods is illustrated in addressing batch effects and population structure in genomic data analyses, and the computational challenges involved in fitting these models efficiently with genomic-scale data are discussed.
