Abstract:
|
Multiple testing approaches that assume exchangeability of tests, i.e., only employ the list of p-values from the tests performed, are widely used. However, in many applications, additional statistics besides the p-values are available - we term these "informative covariates" - that are independent of the p-values under the null hypothesis, but somehow informative of per-test power or prior probability. Ignoring such information wastes overall power. We propose Independent Hypothesis Weighting (IHW), a method that derives hypothesis weights from informative covariates in a data-driven manner without overfitting, to be used in conjunction with a weighted version of the Benjamini-Hochberg method. I will discuss theoretical guarantees of the method and will exemplify the method's performance on numerical experiments and applications from genomics and high-throughput biology. In some cases, dramatic power increases are realized. IHW is a practical approach to discovering associations in large datasets.
Software Availability: www.bioconductor.org/packages/IHW
|