É«½äÖ±²¥

In task-based functional MRI (fMRI), collinearity between task regressors in time series models can reduce statistical power. When collinearity is discovered after data collection, researchers often adjust the model to reduce it. However, some of these adjustments are suboptimal and can introduce bias into parameter estimates. Although this issue applies broadly to task-fMRI studies, I'll illustrate it using data from the Monetary Incentive Delay (MID) task in the Adolescent Brain Cognitive Development (ABCD) study. I'll also introduce a procedure to more directly quantify the impact of collinearity on task-relevant measures: the contrast-based variance inflation factor, or cVIF. Using cVIF, I show that common collinearity-reduction strategies-such as omitting regressors for certain task components, modeling sustained activations as impulses, or ignoring response time variability-can bias contrast estimates. Finally, I propose a "Saturated" model that includes all task components, including response times, aiming to reduce bias while maintaining comparable levels of collinearity, as measured by cVIF.

Zoom URL: