Structural MRI in the Analytic Multiverse: Consequences on Discovery and Predictive Accuracy

The choice of preprocessing pipeline for T1-weighted structural MRI substantially influences grey-matter measurements, brain-trait associations, and prediction accuracy. Such analytic variability challenges reproducibility and generalizability in large-scale neuroimaging. Same-Data Meta-Analysis (SDMA), a multiverse framework that combines statistical maps derived from multiple analytic workflows, has been proposed to increase robustness of inference. However, SDMA has not been applied to structural MRI nor evaluated with respect to predictive performance and spatial harmonization constraints. We analyzed structural MRI data from 39,955 UK Biobank participants (discovery n = 23,288; replication n = 16,538). Grey-matter maps were derived using three widely used volume-based pipelines (FSLVBM, FSLANAT, CAT12-ENIGMA), which differ in spatial templates, voxel dimensions, and grey-matter masks. For four phenotypes highly correlated with brain measurements (diabetes, high blood pressure, maternal smoking around birth, and alcohol consumption frequency), we conducted Brain-Wide Association Studies (BWAS). Pipeline-specific association maps were combined using two SDMA approaches (Stouffer and generalized least squares). Because pipelines operate in distinct anatomical spaces, association maps were projected across templates prior to combination, requiring the selection of a reference space. We systematically evaluated how projection strategy and choice of reference space influenced regional discovery, replication, and prediction. SDMA increased the number of significant clusters (+18% on average - standard deviation of 7.8- across all reference space) and improved cluster-level replication rates (e.g., from 5% to 21% in CAT12) compared to single-pipeline analyses, identifying additional cerebellar and subcortical regions not consistently detected by individual workflows. However, results varied depending on the chosen reference space, highlighting that projection between templates introduces non-trivial spatial distortions and affects regional detection. Despite improved regional discovery, SDMA did not enhance prediction accuracy when using significant voxels. When applied to Best Linear Unbiased Predictors (BLUP), SDMA partially mitigated cross-pipeline performance loss but remained inferior to single-pipeline training. In contrast, stacking consistently improved prediction accuracy by ∼23%. These findings demonstrate that while SDMA enhances robustness of regional inference across preprocessing pipelines, projection choices critically shape multiverse outcomes and predictive gains remain limited. Our results provide practical guidance for combining structural MRI workflows and clarify methodological trade-offs in large-scale brain mapping.