Poster Presentation

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

A model-based alternative to cluster-based permutation tests for the analysis of time-resolved data

Ladislas Nalborczyk¹, Paul-Christian Bürkner²; ¹CNRS, ²Technische Universität Dortmund

Presenter: Ladislas Nalborczyk

A central analytic goal in cognitive neuroscience is to determine whether, when, and where neural responses differ across conditions or groups. The standard solution relies on mass-univariate testing with multiple-comparison correction, often through cluster-based permutation methods. Although these methods control error rates effectively, they do not support precise inference about the onset or offset of an effect because inference is made at the cluster level rather than at individual time points. We introduce a model-based alternative based on Bayesian generalised additive multilevel models (BGAMMs), which estimate smooth time-resolved effects while accounting for temporal dependencies and between-participant variability. The method yields posterior odds that an effect is above zero (or above chance) at each time point, enabling direct probabilistic statements about temporal localisation. We benchmarked the approach on simulated EEG data with known ground-truth onset and offset and compared it against raw mass-univariate tests, FDR and FWER corrections, cluster-based permutation tests, threshold-free cluster enhancement, and changepoint detection. Across 10,000 simulated datasets, the BGAMM approach showed the lowest bias and error for both onset and offset estimation. We then applied the method to empirical MEG decoding time courses and evaluated stability across many split-half subsets of participants. The proposed approach produced interpretable clusters and highly stable temporal estimates, while standard methods were either overly lenient or less precise. These results suggest that Bayesian additive multilevel modelling provides a practical and more temporally precise alternative to cluster-based inference for analysing time-resolved data. To facilitate adoption, we implemented the method in an open-source R package: https://github.com/lnalborczyk/neurogam.

Topic Area: Methods, Tools, Theory & Neural Coding