Interactive demo · Statistical Methods for Modern Biology · Chapter 6

Multiple testing: p-value histogram, Benjamini–Hochberg and hypothesis weighting

One test of differential expression per gene: 33,469 p-values from the airway data (dexamethasone versus control, DESeq2). Their histogram is a mixture of a flat null part and a peak of real effects near 0. Estimate the null proportion π0, read off a false discovery rate, run Bonferroni and Benjamini–Hochberg, split the genes by mean count to see why weighting helps, and check by simulation that BH keeps the FDR at π0φ.

View

Threshold α (reject p ≤ α)

λ for π0 = #{p > λ}/(m(1 − λ))

Target FDR φ

Ranks shown

Tests m

Null proportion π0

Effect size (shift of the z-score)

Scoreboard

Try these

BH: sort the p-values, find the largest k with p(k) ≤ φk/m, reject the k smallest. The weighted version (used by IHW) applies BH to pi/wi with weights that average 1; the weights here are the ones R's ihw chose from baseMean.

In the lecture notes