The investigators trained separate predictive models in two adolescent cohorts to discriminate youth who would abstain from binge drinking by mid-to-late adolescence. In the Adolescent Brain and Cognitive Development cohort (ABCD), MRI predictors were measured at age 9–10. The SuperLearner ensemble produced poor discrimination in ABCD, with an area under the receiver operating characteristic curve reported as 0.585. By contrast, the model fit in NCANDA, where MRI measures came from older participants (ages 12–17), achieved better discrimination (AUC = 0.695). The authors reported these AUCs directly and used them as the main comparative performance metrics between cohorts. These results indicate higher predictive performance when neuroimaging features are measured later in development rather than in late childhood, according to the study team.
The research team extracted the ten highest-ranked morphometric and diffusion features from each cohort's SuperLearner model to inspect which brain measures carried most predictive weight. Data indicated little overlap in predictor rankings between ABCD and NCANDA. The preprint did not provide a pooled list of consistent top predictors across cohorts; rather, it presented cohort-specific top tens and highlighted the divergence. The authors suggested that sample age at MRI acquisition and methodological heterogeneity likely contributed to the mismatch in feature importance between the two data sets.
After identifying the top ten predictive features per cohort, the study team applied targeted maximum likelihood estimation (TMLE) to assess whether any individual morphometric or microstructural measure exerted a causal influence on later binge drinking onset. The authors reported that TMLE analyses did not reveal evidence of a causal effect for any of the examined features in either cohort. The manuscript states this finding plainly: no discrete macrostructural or microstructural brain feature, measured in late childhood or adolescence in these samples, showed a causal contribution to binge drinking by the specified ages. The preprint notes these null causal results alongside the predictive performance differences described above.
The pipeline combined a SuperLearner ensemble—used for prediction—and targeted causal inference with TMLE to probe causality among top predictors. In ABCD, structural and diffusion MRI were measured at ages 9–10 and the outcome was abstinence from binge drinking by age 16–17. In NCANDA, MRI predictors were measured later (ages 12–17) while the developmental outcome was binge drinking by age 18. The authors fit separate SuperLearner models in each cohort, extracted the top ten predictive features, and then subjected those features to TMLE to assess causal effects. The preprint did not provide further methodological detail in the abstract about algorithm hyperparameters, feature-preprocessing steps, or exact variable definitions used for each MRI metric; those details may be available in the full text or supplementary materials linked on the preprint page.
The authors reported several constraints intrinsic to this investigation. First, predictive performance in the childhood-measured MRI sample (ABCD) was weak, which limits the stability and interpretability of any ranked predictors drawn from that model. Second, there was minimal overlap in top predictive features between cohorts, which complicates any generalisation about specific neuroanatomical markers. Third, the abstract states that TMLE did not produce evidence for causal effects, but the preprint did not report further sensitivity analyses or alternative causal models in the abstract. The authors explicitly commented that using more sophisticated metrics of brain structure and function may yield better insights; they recommended measuring neuroanatomy in adolescence rather than late childhood if the goal is to predict who will later engage in binge drinking.
The manuscript also documented ethical and data-access details: the ABCD protocol received centralised IRB approval at the University of California, San Diego and site IRBs at the 21 ABCD locations; NCANDA received site IRB approvals. De-identified data were used. Researchers can apply for ABCD access through the NIH NDA link provided on the preprint page, and supplementary material plus code links were made available by the authors on the medRxiv entry.
The report ends with clear unanswered questions that remain in the data presented: which imaging-derived metrics (beyond gross morphometry and standard diffusion measures) might carry causal signal, and whether longitudinal imaging across critical adolescent windows would improve both prediction and causal identification. The authors framed these as empirical next steps rather than definitive claims. The preprint is not peer reviewed and the team stated no competing interests.