Substance use often begins during adolescence, a developmental period when curiosity, risk-taking, and social influences increase while the brain is still maturing. Initiation of alcohol, nicotine, or cannabis use at earlier ages is consistently associated with a higher risk of later substance use disorders and other adverse outcomes, making prevention and early identification of risk a major public health priority [1].

Risk for initiation is multifactorial and spans multiple domains, including individual behavior and mental health, family environment and parenting, peer context, school experiences, and neighborhood and socioeconomic conditions. These influences can change over time, and their timing may matter. For example, changes in peer exposure or parental monitoring during a specific developmental window may be more predictive of near-term initiation than baseline-only measures. Large longitudinal cohorts therefore provide an opportunity to study both whether initiation occurs and when it occurs, using modeling frameworks that respect time ordering [2, 3]. The Adolescent Brain Cognitive Development (ABCD) Study^® is particularly well suited for this work because it is a large, diverse, multi-site longitudinal cohort that follows youth from late childhood into adolescence with rich repeated assessments across psychosocial,environmental, and health domains [4, 5].

Methodologically, two complementary goals are important. First, we need a principled way to screen thousands of time-varying predictors to identify candidate directional relationships that respect temporal precedence (past → future). A common approach is Granger-style analysis, where a predictor is considered informative if its past values improve the prediction of a later outcome, with directionality arising from time ordering rather than simultaneity [6, 7]. In high-dimensional settings, sparse models such as elastic-net logistic regression can perform variable selection while handling correlated predictors, and stability selection (repeated resampling or subsampling) can help distinguish robust signals from chance selections [8, 9, 10, 11, 12]. Second, after identifying robust candidate edges, we want effect-size estimates that adjust for confounding from other past variables. Double/debiased machine learning (DML) provides a framework for estimating low-dimensional effects while flexibly modeling high-dimensional nuisance relationships with machine learning and using cross-fitting to reduce overfitting bias [13, 14, 15, 16].

In this work, we combine these ideas into a two-stage pipeline:

1. graph discovery using lagged predictors, sparse logistic regression, and subject-level bootstrap stability to propose directional edges; and
2. effect estimation for stable edges using DML-style partialling-out with cross-fitting, with machine-learning nuisance models and cluster-robust uncertainty to account for repeated measures within individuals (Figure 1). Together, this approach aims to provide an interpretable causal interface that highlights which time-ordered factors are most consistently linked to subsequent initiation risk, and estimates their adjusted associations under standard causal assumptions.