Review of post-clustering inference methods
Résumé
In a classical testing problem, statistical hypotheses are defined before observing the data. However, this principle is violated when hypotheses are based on a clustering of the observations, leading to an inflation of the type I error rate. Recently, post-clustering inference methods have been proposed to address this particular type of data-driven hypothesis selection. These methods are divided in two categories: information partitioning and conditional approaches. This article provides a review of these methods, and a numerical comparison on simulated Gaussian data. This contribution sheds some light on the pros and cons of each approach in terms of (i) type I error control and power, (ii) potential dependence on a specific clustering method, and (iii) assumptions on data distribution.