WITHDRAWN Bounding the false discovery rate with pooled controls

Music City Center 

In a multiple testing problem where the number of hypotheses being tested is large, requiring control of the false discovery rate (FDR) means that we will likely only be able to make discoveries if some p-values are extremely small. In some settings, however, only coarse p-values are available due to limited sample size or limited computational budget. By pooling data across multiple tests, we can regain power even in the presence of these constraints---but traditionally, ensuring validity of this type of procedure would require assumptions, such as a Bayesian model, to allow for sharing information across many tests. In this work we show that, with some mild modifications to the procedure, FDR control can hold without any assumptions across the tests.

False discovery rate

Permutation tests