Thomas Nichols, Neuroimaging Statistics, Institute for Digital Healthcare, Warwick
There has been great interest in discovering and understanding the role of genetic variation in neuroimaging phenotypes. Typical imaging genetics studies use a small number of candidate SNPs, a small number of brain regions, or both. In this talk I will consider methods for searching for gene-brain associations over the entire genome and all brain regions. Such an approach presents massive computational and statistical challenges. I\\\'ll discuss two approaches, a mass-univariate approach and a multivariate approach. A mass-univariate model is the standard tool in neuroimaging analysis, but scaling it up for 100,000 SNPs requires a series of computational and statistical innovations. With our method applied to Tensor-Based Morphometry data from the ADNI project, we report the first gene-brain association to survive whole-genome, whole-brain familywise error correction. Our multivariate approach uses a Sparse Reduced Rank Regression (sRRR) to jointly and parsimoniously explain gene-brain associations. Detailed power analyses show that the multivariate approach should have even greater power than the univariate approach.