Abstract

Functional brain connections are the set of statistical relationships between neural activity in different parts of the brain; these are typically estimated from neuroimaging data such as functional MRI . In multi-subject studies, many are interested in identifying the individual functional brain connections or patterns of connections that are different between two groups of subjects or across a clinical population. Popular approaches to this problem include estimating a network for each subject, and then assuming the subject networks are fixed, conducting inference over network metrics. These approaches, however, fail to account for the variability and network estimation error associated with estimating each subject’s brain network, thus resulting in incorrect inferences.

In this talk, we study this problem using Markov Networks as the model for brain connectivity. Statistically, our problem can be described as conducting large scale inference over network edges or groups of edges post graphical model selection, part of a novel statistical paradigm we call Population Post Selection Inference (popPSI). We show that for this popPSI problem, current approaches in neuroimaging have both low statistical power and highly inflated false positive rates. We then develop a new procedure which we term R^3, standing for resampling, random effects, and random penalization. Our approach uses the correct two-level random effects model to account for network variability within a subject (due to network estimation) as well as between subjects. Through simulation studies we show that our method solves many of the problems associated with existing techniques, yielding substantial improvements in terms of both error control and statistical power. We conclude our talk by applying our methods in two case studies – a color-sequence synesthesia study and a neurofibromatosis one study.

Joint work with Manjari Narayan and Steffie Tomson.