Asymptotic control of FWER under Gaussian assumption: application to correlation tests

Sophie Achard (1) , Pierre Borgnat (2) , Irène Gannaz (3, 4)

1 STATIFY - Modèles statistiques bayésiens et des valeurs extrêmes pour données structurées et de grande dimension
2 Phys-ENS - Laboratoire de Physique de l'ENS Lyon
3 PSPM - Probabilités, statistique, physique mathématique
4 ICJ - Institut Camille Jordan

In many applications, hypothesis testing is based on an asymptotic distribution of statistics. The aim of this paper is to clarify and extend multiple correction procedures when the statistics are asymptotically Gaussian. We propose a unified framework to prove their asymptotic behavior which is valid in the case of highly correlated tests. We focus on correlation tests where several test statistics are proposed. All these multiple testing procedures on correlations are shown to control FWER. An extensive simulation study on correlation-based graph estimation highlights finite sample behavior, independence on the sparsity of graphs and dependence on the values of correlations. Empirical evaluation of power provides comparisons of the proposed methods. Finally validation of our procedures is proposed on real dataset of rats brain connectivity measured by fMRI. We confirm our theoretical findings by applying our procedures on a full null hypotheses with data from dead rats. Data on alive rats show the performance of the proposed procedures to correctly identify brain connectivity graphs with controlled errors.

Mots clés: Multiple testing; structure learning; correlation tests; FWER control; cerebral connectivity

Domaines: Statistiques [math.ST]

Sophie Achard, Pierre Borgnat, Irène Gannaz. Asymptotic control of FWER under Gaussian assumption: application to correlation tests. 2020. ⟨hal-02883720⟩