The goal of many microarray studies is to identify genes that are differentially expressed between two classes or populations. Many data analysts choose to estimate the false discovery rate (FDR) associated with the list of genes declared differentially expressed. Estimating an FDR largely reduces to estimatingπ1, the proportion of differentially expressed genes among all analyzed genes. Estimatingπ1is usually done throughP-values, but computingP-values can be viewed as a nuisance and potentially problematic step. We evaluated methods for estimatingπ1directly from test statistics, circumventing the need to computeP-values. We adapted existing methodology for estimatingπ1fromt- andz-statistics so thatπ1could be estimated from other statistics. We compared the quality of these estimates to estimates generated by two established methods for estimatingπ1fromP-values. Overall, methods varied widely in bias and variability. The least biased and least variable estimates ofπ1, the proportion of differentially expressed genes, were produced by applying the “convest” mixture model method toP-values computed from a pooled permutation null distribution. Estimates computed directly from test statistics rather thanP-values did not reliably perform well.
Signal identification for rare and weak features: higher criticism or false discovery rates?
