Estimating False Discovery Rate for High-Dimensional Discrete Data

• Activity Number: 248
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Naomi Altman*+
• Companies: Penn State University
• Keywords: FDR ; multiple comparisons ; pi0 ; multiple testing

Abstract

In high throughput biology and some other fields with high dimensional data, it has become common to do a statistical test for each feature (variable) and then estimate the false discovery rate (FDR) as a function of the p-value or test statistic to determine a suitable significance cut-off. When the percentage of true null hypotheses Pi0 is not close to 100, an important component of the FDR estimator is an estimator of Pi0. Methods developed for continuous data rely heavily on the fact that under the null hypothesis the distribution of p-values is uniform. However, p-values for discrete tests are discrete, with a distribution that often depends on an ancillary such as a table margin which varies among features. This talk discusses a number of methods for estimating Pi0 when the null distribution of p-values is far from uniform. As well, we show that prescreening to remove low-power tests before estimating FDR improves the ability to identify truly non-null hypotheses among the remaining tests.