Predicting substance use behaviors: A multi–data set mega-analysis.

Predicting substance use behaviors is a critical public health challenge . Although self-report survey data have been central to understanding substance use, their longitudinal predictive utility has yet to be quantified. We conducted a “mega”-analysis of 44,862 participants, integrating 197 harmonized self-report predictors across demographic, behavioral, familial, and psychosocial domains. Specifically, we employed both linear regression and machine learning approaches to evaluate the extent to which alcohol , tobacco , and cannabis use could be predicted. Models that handled high-dimensional, correlated predictor sets substantially outperformed unregularized linear regression, reflecting the known limitations of fitting unregularized regression to large, collinear predictor sets rather than a general advantage of machine learning. Random forest and regularized regression (Elastic Net) yielded comparable predictive performance ( R 2 ≈ 14% at baseline and ≈4% for longitudinal change). In exploratory leave-one-data-set-out analyses, models did not generalize to independent held-out data sets ( R 2 ≈ 0), indicating that predictive associations were largely context-specific. Externalizing behaviors predicted a significant proportion of substance use variability , particularly at baseline, whereas the predictive power of demographic and familial factors became apparent in models predicting change over time. Combining large-scale harmonization with computational modeling clarifies both the potential and limits of forecasting substance use trajectories. General Scientific Summary This study examined how well self-report survey data can predict substance use behaviors across seven large longitudinal data sets ( N = 44,862). Behavioral, demographic, and family factors predicted within–data set substance use at modest levels ( R 2 ≈ 14% cross-sectionally; R 2 ≈ 4% longitudinally). However, predictive models did not generalize to independent held-out data sets ( R 2 ≈ 0 in leave-one-data-set-out analyses), indicating that predictive relationships were highly context-specific. Regularized linear regression and machine learning performed comparably. These findings clarify the current limits of predicting substance use from commonly available survey measures and highlight the substantial challenges that remain before such models could support individual-level risk identification .

Authors

Publication Details

Journal
Journal of Psychopathology and Clinical Science
Published
2026-11-19
DOI
https://doi.org/10.1037/abn0001167
Primary Topic
Substance Abuse Treatment and Outcomes
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Predicting substance use behaviors: A multi–data set mega-analysis.

Catharine E. Fairbairn, Dahyeon Kang, Nigel Bosch, Lan Jiang et al.
Journal of Psychopathology and Clinical Science
Substance Abuse Treatment and Outcomes
article

Predicting substance use behaviors: A multi–data set mega-analysis.

Catharine E. Fairbairn, Dahyeon Kang, Nigel Bosch, Lan Jiang, Eddie Caumiant
article en

Abstract

Predicting substance use behaviors is a critical public health challenge . Although self-report survey data have been central to understanding substance use, their longitudinal predictive utility has yet to be quantified. We conducted a “mega”-analysis of 44,862 participants, integrating 197 harmonized self-report predictors across demographic, behavioral, familial, and psychosocial domains. Specifically, we employed both linear regression and machine learning approaches to evaluate the extent to which alcohol , tobacco , and cannabis use could be predicted. Models that handled high-dimensional, correlated predictor sets substantially outperformed unregularized linear regression, reflecting the known limitations of fitting unregularized regression to large, collinear predictor sets rather than a general advantage of machine learning. Random forest and regularized regression (Elastic Net) yielded comparable predictive performance ( R 2 ≈ 14% at baseline and ≈4% for longitudinal change). In exploratory leave-one-data-set-out analyses, models did not generalize to independent held-out data sets ( R 2 ≈ 0), indicating that predictive associations were largely context-specific. Externalizing behaviors predicted a significant proportion of substance use variability , particularly at baseline, whereas the predictive power of demographic and familial factors became apparent in models predicting change over time. Combining large-scale harmonization with computational modeling clarifies both the potential and limits of forecasting substance use trajectories. General Scientific Summary This study examined how well self-report survey data can predict substance use behaviors across seven large longitudinal data sets ( N = 44,862). Behavioral, demographic, and family factors predicted within–data set substance use at modest levels ( R 2 ≈ 14% cross-sectionally; R 2 ≈ 4% longitudinally). However, predictive models did not generalize to independent held-out data sets ( R 2 ≈ 0 in leave-one-data-set-out analyses), indicating that predictive relationships were highly context-specific. Regularized linear regression and machine learning performed comparably. These findings clarify the current limits of predicting substance use from commonly available survey measures and highlight the substantial challenges that remain before such models could support individual-level risk identification .

Journal of Psychopathology and Clinical Science
Substance Abuse Treatment and Outcomes
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon