CMStatistics 2015: Start Registration
View Submission - CMStatistics
B1333
Title: Adaptive selection via SLOPE Authors:  Damian Brzyski - Jagiellonian University (Poland) [presenting]
Abstract: Sorted L-One Penalized Estimation (SLOPE) is a relatively new convex optimization procedure which enables the adaptive selection of regressors under sparse high dimensional designs. The method was designed to control the expected proportion of irrelevant regressors among all selected predictors (false discovery rate, FDR). It was shown that SLOPE performs well in situations when explanatory variables are nearly orthogonal. The control over FDR could be however lost in the presence of strong correlations, which usually happens in the context of Genome Wise Association Studies (GWAS). We show that SLOPE can be successfully used in GWAS after introducing the additional clustering step, where clusters correspond to the groups of strongly correlated SNPs. We use computer simulations to show that after this modification SLOPE controls FDR defined at the clusters level and that it compares favorably to other popular methods for GWAS.