Quantitative intersectional research aims to understand how various social positions jointly influence health risk behaviors. Traditional methods like moderated multiple regression, while common, can be unreliable due to data limitations and complexity. Machine learning offers a promising alternative, adept at handling numerous interactions in sparse datasets, yet its application remains limited. This study highlights the utility of group-lasso interaction network (glinternet), an innovative machine learning technique employing hierarchical regularization, to evaluate intersectional disparities in substance use.
Applying glinternet, which excels in variable selection and parameter stabilization for both main and interaction effects, this research examined two-way interactions among gender, sexual orientation, and race in predicting the prevalence of heavy episodic drinking, cannabis use, and cigarette use. The analysis utilized the All of Us Research Program (N = 283,403), a diverse national sample. Findings were validated using holdout cross-validation and benchmarked against logistic regression estimates.
Glinternet demonstrated superior stability in prevalence estimates across discovery and replication samples compared to logistic regression, particularly for underrepresented groups. Elevated prevalence rates for cigarette and cannabis use were observed among sexual minority and White cisgender women relative to heterosexual and non-White women, respectively.
These results suggest that glinternet can enhance traditional moderated multiple regression in intersectional research by improving model efficiency and parameter stability. This machine learning approach provides valuable new avenues for quantifying health disparities across intersectional social positions, offering advanced analytical positions for researchers in the field.
