Power Comparison of Statistical Methods for Identifying Differential Expression Genes in Microarray Experiments

• Activity Number: 34
• Type: Contributed
• Date/Time: Sunday, August 7, 2005 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Xiaohua Shu*+ and Xiaohua Zhang and Jianjun Li and Wei Nan and Benjamin I. Sun
• Companies: Temple University and Merck Research Laboratories and Merck Research Laboratories and Merck & Co., Inc. and Merck & Co., Inc.
• Keywords: differentially expressed ; microarrays ; Power ; ROC ; moderated t-statistic ; ANOVA-based t-test

Abstract

To identify differentially expressed genes, the commonly used measures are fold change and t-statistics. However, with small sample size, it is believed t-statistic tends to be highly correlated with the standard error term that appears in its denominator and thus tends to pick up significant findings at a higher rate from among those with low sample variance. To address this concern, moderated t-statistic has been developed by adding value to the denominator of t-test statistic. If there are multiple groups of treatment in a microarray experiment, an alternative is ANOVA-based t-test. In this paper, based on a well-designed CNS microarray experiment, we compared the powers of fold change, regular t-statistic, ANOVA-based t-statistic, SAM t-statistic, and Bayesian moderated t-statistic. The results suggest fold change and SAM t-statistic have the lowest powers in nearly every situation. For true difference, ANOVA-based t-statistic has the biggest power, while regulated t-statistic has nearly the same power as Bayesian moderated t-statistic, especially when the number of replicates is small.