Fig. 2: SHAP summary plot for random forest cross-validation on the best-performing feature domain (“PEER” - perceptions of peer drinking behaviors).

These models included five features: perceived peer drinking amount and frequency among the nominated heaviest drinkers in one’s social group, perceived peer attitudes toward alcohol use and binge drinking, and perceived peer approval of drinking. Each row shows one feature ranked by its average contribution to the models’ predictions, with the percentage indicating the feature’s relative importance across the full model. Each point represents one prediction for a participant in one of the 100 repetitions of the nested cross-validation pipeline. The horizontal axis shows SHAP values, reflecting the impact of a given feature on the model’s output: values to the right indicate a positive contribution toward predicting intervention effectiveness (class 1), while values to the left indicate a contribution toward predicting no intervention effectiveness (class 0). Color denotes the original feature value (green = high, blue = low). This helps show how different feature values influence the direction of model predictions.