Yudi Pawitan: Searching for and assessing subtypes in a heterogeneous population

Biological heterogeneity is common in many diseases, and it is often the source of failure to get good prognostic models. In an analysis of lung cancer gene expression data, we are investigating a clustering method to search for molecular subtypes for which we can obtain better prognosis. In the process of assessing whether a subgroup is interesting, we often find p-value distributions with substantial shortage of small values. This turns out to be a common occurrence in heterogeneous populations. Ignoring this effect will lead to biased inference, where we would miss potential discoveries by overestimating the false discovery rate. I will describe a model that explains the phenomenon and show how to analyze the p-value distribution in order to get an unbiased inference.