Linear Mixed Models Minimise False Positive Rate and Enhance Precision of Mass Univariate Vertex-Wise Analyses of Grey-Matter

We evaluated the statistical power, family wise error rate (FWER) and precision of several competing methods that perform mass-univariate vertex-wise analyses of grey-matter (thickness and surface area). In particular, we compared several generalised linear models (GLMs, current state of the art) to linear mixed models (LMMs) that have proven superior in genomics. We used phenotypes simulated from real vertex-wise data and a large sample size (\mathrm{N}=8,662) which may soon become the norm in neuroimaging. No method ensured a \text{FWER} < 5{\%} (at a vertex or cluster level) after applying Bonferroni correction for multiple testing. LMMs should be preferred to GLMs as they minimise the false positive rate and yield smaller clusters of associations. Associations on real phenotypes must be interpreted with caution, and replication may be warranted to conclude about an association.