Population-weighted Image-on-scalar Regression Analyses of Large Scale Neuroimaging Data

Recent advances in neuroimaging modeling highlight the importance of accounting for subgroup heterogeneity in population-based neuroscience research. When imaging data are available only for a subsample of participants, estimated associations between brain activity and individual characteristics may differ from the corresponding population-average estimand. We develop a population-weighted image-on-scalar regression framework for estimating associations between scalar predictors and high-dimensional neuroimaging outcomes. We apply this framework to functional magnetic resonance imaging (fMRI) data from the Adolescent Brain Cognitive Development (ABCD) Study's n-back working-memory task, focusing on associations between general cognitive ability and working-memory-related brain activation. The ABCD Study provides baseline weights calibrated to external sociodemographic benchmarks for U.S. children aged 9--10 years; however, the imaging analytic subsample differs from the baseline cohort on several measured child and family characteristics, motivating additional subsample weighting. The proposed approach combines existing baseline population weights with imaging-subsample adjustment weights constructed using inverse propensity score weighting. Simulation studies show that weighted and unweighted estimators perform similarly under correctly specified models, while weighting can improve estimation of population-average coefficient functions when relevant interactions are omitted and selection depends on observed variables. In the ABCD application, weighting changes the magnitude, spatial extent, and statistical detection of brain--cognition association maps, with effects depending on covariate adjustment. Our findings indicate that population weighting adjustments can influence estimated associations between brain activity and cognition and underscore the importance of assessing analytic sample composition and evaluating the sensitivity of results to weighting and model specification.