Department of Statistics, University of Washington - Director of Statistical Consulting
This talk follows conveniently upon last week's seminar presentation by Peter Hoff of models for certain data matrices based on the singular value decomposition (SVD). The basis of the analysis we present here is the SVD of a matrix of covariances between a vector of 80 measures representing the midline shape of the corpus callosum from MR brain images and a behavioral profile represented by 230 scores from a five-hour neuropsychological examination.
I will begin with some background on Fetal Alcohol Syndrome (FAS) and Fetal Alcohol Spectrum Disorders (FASD) and diagram of the scope of research studies being carried out at the UW's Fetal Alcohol and Drug Unit (FADU). The particular study of interest here involves 177 adolescents and adults of both sexes, 117 of whom have diagnoses of FASD (FAS or FAE).
I will present our analytical strategy based on the computation of composite scores for features (dimensions) of variation in the midline shape of the corpus callosum covarying with composite neuropsychological scores. The coefficients of these composite scores derive from an SVD of the cross-covariance matrix between the two domains. I will explain a number of details of the statistical data processing, including our approach to making inference about the dimensionality of the SVD model. The composite scores for the group of FASD subjects are considerably more variable than the unexposed, a fact that enables discrimination. The callosal part of this correlation pattern supports a (quadratic) discrimination of FASD from unexposed with cross-validated sensitivity 0.76 and specificity 0.79. The associated neuropsychological dimensions "split" the classic behavioral phenotype into two factors, one for motor function (together with an IQ confound) and the other for an IQ free aspect of executive function (EF). This research is directed toward the context of adolescents and adults without the â€œface of FASâ€ who are referred for possible diagnosis because of FASD-like behavior. Our results suggest that a behavioral screen, the combination of lower IQ with abnormal response timing (fast or slow) on certain executive function tasks, followed by callosal image measures may support a classification of FASD vs. unaffected cases with promising sensitivity and specificity.