Prioritizing hypothesis tests for high throughput data.

Abstract

The advent of high throughput data has led to a massive increase in the number of hypothesis tests conducted in many types of biological studies and a concomitant increase in stringency of significance thresholds. Filtering methods, which use independent information to eliminate less promising tests and thus reduce multiple testing, have been widely and successfully applied. However, key questions remain about how to best apply them: When is filtering beneficial and when is it detrimental? How good does the independent information need to be in order for filtering to be effective? How should one choose the filter cutoff that separates tests that pass the filter from those that don't? We quantify the effect of the quality of the filter information, the filter cutoff and other factors on the effectiveness of the filter and show a number of results: If the filter has a high probability (e.g. 70%) of ranking true positive features highly (e.g. top 10%), then filtering can lead to dramatic increase (e.g. 10-fold) in discovery probability when there is high redundancy in information between hypothesis tests. Filtering is less effective when there is low redundancy between hypothesis tests and its benefit decreases rapidly as the quality of the filter information decreases. Furthermore, the outcome is highly dependent on the choice of filter cutoff. Choosing the cutoff without reference to the data will often lead to a large loss in discovery probability. However, naïve optimization of the cutoff using the data will lead to inflated type I error. We introduce a data-based method for choosing the cutoff that maintains control of the family-wise error rate via a correction factor to the significance threshold. Application of this approach offers as much as a several-fold advantage in discovery probability relative to no filtering, while maintaining type I error control. We also introduce a closely related method of P-value weighting that further improves performance. R code for calculating the correction factor is available at http://www.stat.uga.edu/people/faculty/paul-schliekelman pdschlie@stat.uga.edu Supplementary data are available at Bioinformatics online.