Seminar Robust and Scalable Inference via Sign-Flipping Score Tests from GLMs to Multivariate Mixed Models

Speaker: Livio Finos (University of Padova)

Abstract:

Permutation and sign-flipping tests offer attractive properties, such as exact Type I error control under exchangeability and robustness to distributional assumptions. However, their application to generalised linear models (GLMs) with confounders is limited due to a lack of exchangeability under the null hypothesis. The effective flipscores test (Hemerik et al., 2020, JRSS-B) addresses this issue by flipping the contributions of the efficient score. This enables asymptotically exact inference in GLMs while accounting for covariates. Further refinement through standardised flipscores (De Santis et al., 2025, JASA) improves finite-sample performance and robustness to variance misspecification, heteroscedasticity, and overdispersion. In this seminar, I will present the theoretical basis and practical implementation of flipscores tests and discuss the most recent extension to include multivariate responses and confidence intervals. I will then extend this framework to the challenging setting of multivariate mixed models, where dependence structures and high-dimensional random effects complicate inference. To address these issues, I will introduce a block-wise sign-flipping approach that preserves within-block dependence while enabling valid, resampling-based inference across clusters. This method generalizes flipscores tests to longitudinal, multilevel, and multivariate data, maintaining robustness and scalability. Applications to real data will be discussed, highlighting how flipscores facilitates flexible and reliable inference in complex, real-world datasets.

Organizer: Silvia Cagnone