Diagnosis and exploration of massively univariate neuroimaging models.

The goal of this work is to establish the validity of neuroimaging models and inferences through diagnosis and exploratory data analysis. While model diagnosis and exploration are integral parts of any statistical modeling enterprise, these aspects have been mostly neglected in functional neuroimaging. We present methods that make diagnosis and exploration of neuroimaging data feasible. We use three- and one-dimensional summaries that characterize the model fit and the four-dimensional residuals. The statistical tools are diagnostic summary statistics with tractable null distributions and the dynamic graphical tools which allow the exploration of multiple summaries in both spatial and temporal/interscan aspects, with the ability to quickly jump to spatiotemporal detail. We apply our methods to a fMRI data set, demonstrating their ability to localize subtle artifacts and to discover systematic experimental variation not captured by the model.