Simon Vandekar

My interests pertain to statistical issues in medical image analysis such as quantification of uncertainty and replicability and performing high dimensional inference.

My work lately has focused on making accurate probabilistic statements for high-dimensional neuroimaging data such as computing voxel-wise and spatial extent adjusted p-values that describe how unlikely the observed results are under a particular null hypothesis. In imaging this requires dealing with low-rank estimates for complex covariance matrices, issues of normal approximations in high-dimensional data, and model misspecification. I developed several new (semi)parametric bootstrap joint (PBJ) inference procedures, which are available as an R package (pbj) on github. My current research is attempts to accurately estimate the probability that a thresholded statistical image contains all truly activated regions using small sample sizes. This is more akin to constructing confidence intervals than hypothesis testing and affords more power when the null hypothesis is false. I am also interested in the estimation of the proportion of the image with a nonzero effect size, inference for machine learning models of imaging data, topological data analysis, and unsupervised deep learning.