Sleep As a Protective Factor In Adolescent Suicidal Ideation: A Double Machine Learning Approach within Stress and Emotion Regulation Frameworks

Background

Adolescent suicidal ideation is a major public health concern, particularly among youth exposed to psychosocial stressors. Guided by the Stress Process Model, this study conceptualizes bullying, cyberbullying, sexual violence, substance misuse, and academic strain as key risk exposures. The Sleep and Emotion Regulation Framework suggests that adequate sleep supports emotional regulation and reduces vulnerability to distress, whereas insufficient sleep increases psychological risk. Yet few studies have examined whether sleep has an independent protective effect after accounting for co-occurring stressors and complex confounding. Because sleep duration and psychological stressors are often interrelated, conventional approaches may not fully disentangle their independent effects. Moreover, the relationship between sleep duration and suicidal ideation may be nonlinear, as both insufficient and excessive sleep have been linked to poorer mental health. To address this gap, this study applies Double Machine Learning (DML), which adjusts for high-dimensional confounders while reducing bias. We examine (1) the association between sleep duration and suicidal ideation, (2) whether protective effects persist after adjustment, and (3) whether a nonlinear dose-response pattern exists.

Methods

We used data from the 2023 Youth Risk Behavior Survey (N = 17,145). Suicidal ideation was operationalized as a binary indicator based on a self-reported item asking whether respondents had seriously considered attempting suicide during the past 12 months. Continuous sleep duration was the focal predictor. Covariates included sex at birth, grade, race/ethnicity, sexual minority status, school bullying, cyberbullying, forced sexual intercourse, opioid misuse, physical activity, academic grades, and survey year. We compared naive ordinary least squares (OLS), covariate-adjusted OLS, and partial-linear Double Machine Learning. DML adjusted for high-dimensional confounding and reduced misspecification bias. DML stability was evaluated through repeated split estimation, and Shapley Additive exPlanations (SHAP) values were used to assess feature importance.

Results

Across all models, longer sleep duration was associated with lower suicidal ideation. In naive OLS, a one-hour increase in sleep was significantly associated with a 4.9-point decrease in suicidal ideation (theta = -0.0491, 95% [CI] [-0.0512, -0.0469]). In adjusted OLS, the estimate attenuated to 3.9 points but remained significant (theta = -0.0393, 95% CI [-0.0414, -0.0372]). DML produced a nearly identical estimate to adjusted OLS (theta = -0.0393, 95% CI [-0.0416, -0.0370]), indicating that the protective effect of sleep persisted after adjustment for observed confounders. Repeated DML analyses showed high stability, with a mean theta of -0.0393, minimal variation across runs, and complete sign consistency. An exploratory dose-response analysis found a nonlinear relationship, with suicidal ideation risk declining between 4 and 8 hours of sleep and plateauing thereafter. SHAP analyses identified sleep duration, forced sexual intercourse, school bullying, and cyberbullying as leading contributors in predictive models.

Conclusions

Findings suggest that sleep is not merely an individual health behavior but a meaningful protective factor within broader adolescent stress processes. Although the effect size was reduced after controlling for co-occurring social stressors, sleep remained independently protective. The consistency between adjusted OLS and DML further supports robustness across estimation strategies. These findings position healthy sleep as a modifiable intervention target for school- and community-based suicide prevention.