CMStatistics 2021: Start Registration
View Submission - CMStatistics
B1007
Title: Mass univariate modelling for binary-valued neuroimaging data Authors:  Petya Kindalova - University of Oxford (United Kingdom)
Ioannis Kosmidis - University of Warwick and The Alan Turing Institute (United Kingdom)
Thomas Nichols - University of Oxford (United Kingdom) [presenting]
Abstract: While the vast majority of brain image data are continuous, there is growing interest in binary-valued images describing brain lesions in Magnetic Resonance Imaging (MRI). Binary image data can identify the tissue damaged by a stroke infarct, Multiple Sclerosis lesions in white matter, or bright spots simply called white matter hyperintensities (WMH). While various sophisticated analyses can be proposed, a basic mass univariate regression is vital to map out the influence of the explanatory variables in an unbiased manner. However, the base rate of lesion incident is often very low, leading to many voxels with total- or quasi-separation. We propose a comprehensive simulation framework to evaluate methods for this type of data. We use WMH data from the UK Biobank ($N=40,000$) to define a realistic simulation model that accounts for the spatial dependence in the lesions. Generating arbitrary-$N$ data with covariates where ground truth is known, we compare 3 probit regression methods: Maximum Likelihood (ML), Penalised ML (PML) with Jeffreys prior (aka Firth regression), and a Bayesian Spatial Generalized Linear Mixed Model previously proposed. We find that ML often has problems with separation that PML avoids, and for the smallest sample sizes, the Bayesian method has the best MSE. We will discuss extensions for repeated measures mass univariate modelling with penalised GEE.