Testing minority stress theory (MST) constructs with machine learning: Predicting LGBTQ+ depression
K Basaria, N Mattei, E Igonya, E Otukpa, W Muga, K StojanovskiAbstract
Background
Minority Stress Theory (MST) holds that discrimination links structural stigma to poor mental health in LGBTQ+ populations, but few studies have evaluated MST using predictive approaches. Our objective was to assess whether discrimination variables consistently emerge as strong predictors of depression among LGBTQ+ adults and whether a reduced Minority Stress model achieves similar predictive performance to a full sociodemographic model.
Methods
We analyzed cross-sectional survey data from 2055 LGBTQ+ adults in Kenya and Rwanda. Depression was assessed with the PHQ-9. Logistic regression and 8 theory-driven supervised machine-learning models (e.g., KNN, random forest, SVM) were trained using 10-fold cross-validation with SMOTE within-training folds, comparing full sociodemographic models to models with only Minority Stress-related variables. Performance was evaluated by specificity, recall, F1, accuracy, and AUC. SHAP values assessed feature importance.
Results
Across all modeling approaches, discrimination measures consistently dominated depression prediction, emerging as the most influential features in all models. Reduced Minority Stress models with only discrimination and LGBTQ+ identity variables performed similarly to full models, with no meaningful predictive performance loss. More complex nonlinear models did not outperform logistic regression, indicating that discrimination-related effects were strong, stable, and largely linear across model specifications.
Conclusions
Our findings provide predictive support for MST and highlight discrimination as a central and highly influential predictor of depressive symptoms among LGBTQ+ individuals. Methodologically, the use of machine learning models demonstrates that these relationships are robust across linear and nonlinear model specifications, reinforcing the stability of Minority Stress constructs and illustrating how predictive modeling can evaluate the sufficiency of theory-driven variable sets.