Permutation-based methods for One- and Two-sample Problems with High-dimensional Covariates and Multivariate Responses

• Activity Number: 223
• Type: Contributed
• Date/Time: Tuesday, August 10, 2004 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Greg DiRienzo*+
• Companies: Harvard University
• Keywords: genotype sequence ; joint confidence region ; multiple testing ; aimultaneous inference ; sparse data ; variable confidence band

Abstract

Permutation-based resampling methods are proposed for the multiple-testing problem with P>=1 response variables. Simultaneous alpha-level variable confidence bands are obtained for an R-dimensional vector of test statistics. Inference is simultaneously conducted with respect to all P response variables. These methods are utilized to: reduce the dimension of a large set of covariates; test association between covariate patterns and response; test for a treatment group effect on response separately for each of several covariate patterns. We present a simulation study comparing our methods and those of Pollard and vander Laan (2003) in the gene expression setting and show that precise variance-covariance estimation enables our methods to outperform theirs for small sample sizes or skewed response data. Simultaneous P-dimensional confidence regions for statistics of interest are easily obtained for each setting considered. Examples include relating HIV-1 genotype sequence and resistance phenotypes for P drugs, and joint estimation of treatment effect on time to virologic failure and time to off-track within several strata of baseline covariates.