Rescuing missing data in connectome-based predictive modeling
Recent evidence suggests brain-behavior predictions may require very large sample sizes. However, as the sample size increases, the amount of missing data also increases. Conventional methods, like complete-case analysis, discard useful information and shrink the sample size. To address the missing data problem, we investigated rescuing these missing data through imputation. Imputation is the substitution of estimated values for missing data to be used in downstream analyses. We integrated imputation methods into the Connectome-based Predictive Modeling (CPM) framework. Utilizing four open-source datasets--the Human Connectome Project, the Philadelphia Neurodevelopmental Cohort, the UCLA Consortium for Neuropsychiatric Phenomics, and the Healthy Brain Network (HBN)--we validated and compared our framework with different imputation methods against complete-case analysis for both missing connectomes and missing phenotypic measures scenarios. Imputing connectomes exhibited superior prediction performance on real and simulated missing data as compared to complete-case analysis. In addition, we found that imputation accuracy was a good indicator for choosing an imputation method for missing phenotypic measures but not informative for missing connectomes. In a real-world example predicting cognition using the HBN, we rescued 628 individuals through imputation, doubling the complete case sample size and increasing explained variance by 45%. Together, our results suggest that rescuing data with imputation, as opposed to discarding subjects with missing information, improves prediction performance.