Permutation-Based Yet Computationally Parsimonious FDR Point and Confidence Interval Estimators

• Activity Number: 673
• Type: Contributed
• Date/Time: Thursday, August 5, 2010 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Joshua Millstein*+ and Dmitri Volfson and John R. Lamb and Hongyue Dai and Stephen H. Friend and Eric E. Schadt and Jonas Bergh
• Companies: Sage Bionetworks and iPierian and Pfizer Inc. and Merck & Co., Inc. and Sage Bionetworks and Pacific Biosciences and Karolinska Institute
• Keywords: multiple testing ; multiple comparisons ; resampling ; exact tests

Abstract

False Discovery Rate (FDR) methods are especially common in genomic studies that include tens of thousands of tests or more. With such an essential role played in the interpretation of results, it is of paramount importance to know the uncertainty in the FDR estimate. In the case of a non-parametric permutation-based estimator, it is also important to know the number of permutations required to provide a reliable estimate. We describe novel non-parametric FDR point and confidence interval estimators. The CI estimator applies a binomial distribution to the count of positive tests resulting in an estimator that does not require additional replicate randomly permuted data sets and is therefore computationally efficient. We applied this approach to gene expression microarray data of breast cancers from non-responder patients to identify genes predictive of breast cancer survival.