Rencontres Statistiques du CEREMADE (Romain Périer, lundi 7 septembre 2026)

La prochaine séance des Rencontres statistiques du CEREMADE aura lieu le lundi 7 septembre 2026 à 13h45 en salle Bbis. Nous aurons le plaisir d'écouter Romain Périer (Laboratoire de Mathématiques d'Orsay), qui nous parlera de

FDR methods and confidence envelopes for false discoveries with heterogeneous data.

In this talk, after introducing the classical multiple testing setting and some examples where heterogeneity in the data can arise, I present two applications using locality in multiple testing via the closure principle.
The first is the construction of confidence bounds on false discoveries, with an emphasis on the Bretagnolle calibration and its optimization under computational cost constraints. It produces a tighter bound than the DKWM inequality on certain regions of $p$-values.
The second is the application of the $e$-closure principle to produce more powerful multiple testing procedures under heterogeneity and arbitrary dependence of the $p$-values, focusing on the BY $e$-closure principle and the vanilla DBY procedure. The (inexact) shortcut provided here offers a trade-off between statistical power and computational cost.