Shrinkage Estimation of Expression Fold Change As an Alternative to Testing Hypotheses of Equivalent Expression

摘要: Research on analyzing microarray data has focused the problem of identifying differentially expressed genes to neglect how integrate evidence that a gene is with information extent its differential expression. Consequently, researchers currently prioritize for further study either basis volcano plots or, more commonly, according simple estimates fold change after filtering an arbitrary statistical significance threshold. While subjective and informal nature former practice precludes quantification reliability, latter equivalent using hard-threshold estimator expression ratio not known perform well in terms mean-squared error, sum variance squared bias. On two distinct simulation studies from different studies, we systematically compared performance several estimators representing both current shrinkage. We find threshold-based usually worse than maximum-likelihood (MLE) they often far as quantified by estimated risk. By contrast, shrinkage tend or better MLE never much MLE, expected what about However, Bayesian measure based prior few are indicates local false discovery rate (FDR), best studied. Based ability leverage across genes, conclude use local-FDR instead combinations tests non-shrinkage can be substantially improve reliability prioritization at very little risk doing so less reliably.