Estimating the null distribution for conditional inference and genome-scale screening

摘要: In a novel approach to the multiple testing problem, Efron (2004; 2007) formulated estimators of distribution test statistics or nominal p-values under null suitable for modeling data thousands unaffected genes, non-associated single-nucleotide polymorphisms, other biological features. Estimators can improve not only empirical Bayes procedure which it was originally intended, but also many comparison procedures. Such serve as groundwork proposed based on recent frequentist method minimizing posterior expected loss, exemplified with non-additive loss function designed genomic screening rather than validation. The merit estimating is examined from vantage point conditional inference in remainder paper. simulation study genome-scale testing, conditioning observed confidence level estimated an approximate ancillary statistic markedly improved inference. To enable researchers determine whether rely particular decision making, information-theoretic score provided that quantifies benefit conditioning. As sum degree ancillarity and inferential relevance, reflects balance would strike between two conflicting terms. Applications gene expression microarray illustrate methods introduced.