Studies of working memory (WM) function have tended to adopt either a within-subject approach, focusing on effects of load manipulations, or a between-subjects approach, focusing on individual differences. This dichotomy extends to WM neuroimaging studies, with different neural correlates being identified for within- and between-subjects variation in WM. Here, we examined this issue in a systematic fashion, leveraging the large-sample Human Connectome Project dataset, to conduct a well-powered, whole-brain analysis of the N-back WM task. We first demonstrate the advantages of parcellation schemes for dimension reduction, in terms of load-related effect sizes. This parcel-based approach is then utilized to directly compare the relationship between load-related (within-subject) and behavioral individual differences (between-subject) effects through both correlational and predictive analyses. The results suggest a strong linkage of within-subject and between-subject variation, with larger load-effects linked to stronger brain-behavior correlations. In frontoparietal cortex no hemispheric biases were found towards one type of variation, but the Dorsal Attention Network did exhibit greater sensitivity to between over within-subjects variation, whereas in the Somatomotor network, the reverse pattern was observed.

Cross-validated predictive modeling capitalizing on this tight relationship between the two effects indicated greater predictive power for load-activated than load-deactivated parcels, while also demonstrating that load-related effect size can serve as an effective guide to feature (i.e., parcel) selection, in maximizing predictive power while maintaining interpretability. Together, the findings demonstrate an important consistency across within- and between-subjects approaches to identifying the neural substrates of WM, which can be effectively harnessed to develop more powerful predictive models.