have been performed using data from one research center to minimize technical (e.g., scanner hardware, protocols, and diagnostic assessment) and clinical (e.g., age, medication status, disease chronicity, and severity) heterogeneity. It is therefore not clear whether the MVPA results obtained from these monocenter studies generalize well to other centers, which would be required for clinical application16–18. Interestingly, whereas classification accuracies of monocenter studies only tend to increase with larger samples19,20, accuracies for multicenter studies in other psychiatric disorders such as schizophrenia and autism tend to be lower with increasing sample size12,15–17. This paradoxical effect of lower classification accuracy with larger samples has been attributed to larger sample heterogeneity21, which inevitably increases when combining data from different centers. Here, we used data from the Enhancing Neuro-Imaging and Genetics through Meta-Analysis (ENIGMA) OCD consortium, including 4372 participants recruited at 36 research institutes around the world, with a full range of technical and clinical heterogeneity. We assessed the ability of MVPA to distinguish OCD patients from healthy controls using structural neuroimaging data at the individual subject level. We investigated machine learning classification performance in both single-site and multi-site samples using different validation strategies to assess generalizability. Furthermore, the large sample size enables investigation
