Challenging false discovery rate: a partition test based on p values in human case-control association studies

Abstract

Background/aims: We consider the situation that multiple genetic variants are underlying a heritable trait and assume that each contributes to the trait only to a small degree. The aim is to develop a statistical test for disease association of these multiple variants.

Methods: We expect that p values resulting from a genome-wide case-control association analysis will fall into two classes: those reflecting true association and those occurring randomly in the interval from 0 to 1. We develop a partition test to find the set of smallest p values deviating most from the number of p values expected under randomness.

Results: Power calculations demonstrate the superiority of our partition test over conventional SNP-by-SNP analyses. Applications of the partition test to six published datasets show that our test is particularly suitable when multiple SNPs appear to contribute to a trait, and furnished more significant results than a well-known procedure to estimate the false discovery rate.

Conclusions: Our partition test also furnishes an estimate of the number of functional SNPs underlying disease and can be highly significant, while single-locus tests may be far from significant.