Extending Beyond Logistic Regression: Machine Learning Approaches To Predicting Pharmacy Type Selection
Location
Hager-Lubbers Exhibition Hall
Description
PURPOSE: To demonstrate that advanced machine learning techniques can better predict pharmacy type selection compared to traditional logistic regression and to identify complex non-linear relationships influencing pharmacy choice. SUBJECTS: The study analyzed 1,502 adult respondents from the 2021 National Consumer Survey on the Medication Experience and Pharmacist Role (NCSME-PR) dataset. METHODS AND MATERIALS: The dependent variable was pharmacy type (independent, chain, supermarket/mass merchandise, mail-order, prescription-only), with 16 independent variables based on the Andersen Behavioral Model. We implemented logistic regression, random forest, and XGBoost models, followed by K-means clustering and SHAP analysis for interpretability. ANALYSES: Model performance was compared using Area Under the ROC Curve (AUC). Feature importance was assessed through pseudo-R² for logistic regression, mean decrease in Gini impurity for random forest, and gain for XGBoost. RESULTS: XGBoost models significantly outperformed logistic regression across all pharmacy types (AUC improvements up to 0.31). Key predictors included chronic conditions, geographic region, mail pharmacy use, and income. Patient segmentation identified four distinct clusters with different pharmacy preferences. SHAP analysis revealed threshold effects where patients with 3+ chronic conditions were 2.4 times more likely to select independent pharmacies in the Southern region. CONCLUSIONS: Machine learning captures complex decision-making patterns in pharmacy selection that traditional statistical approaches miss. The identification of distinct patient segments and non-linear relationships provides actionable insights for targeted pharmacy service design and personalized medication management approaches.
Extending Beyond Logistic Regression: Machine Learning Approaches To Predicting Pharmacy Type Selection
Hager-Lubbers Exhibition Hall
PURPOSE: To demonstrate that advanced machine learning techniques can better predict pharmacy type selection compared to traditional logistic regression and to identify complex non-linear relationships influencing pharmacy choice. SUBJECTS: The study analyzed 1,502 adult respondents from the 2021 National Consumer Survey on the Medication Experience and Pharmacist Role (NCSME-PR) dataset. METHODS AND MATERIALS: The dependent variable was pharmacy type (independent, chain, supermarket/mass merchandise, mail-order, prescription-only), with 16 independent variables based on the Andersen Behavioral Model. We implemented logistic regression, random forest, and XGBoost models, followed by K-means clustering and SHAP analysis for interpretability. ANALYSES: Model performance was compared using Area Under the ROC Curve (AUC). Feature importance was assessed through pseudo-R² for logistic regression, mean decrease in Gini impurity for random forest, and gain for XGBoost. RESULTS: XGBoost models significantly outperformed logistic regression across all pharmacy types (AUC improvements up to 0.31). Key predictors included chronic conditions, geographic region, mail pharmacy use, and income. Patient segmentation identified four distinct clusters with different pharmacy preferences. SHAP analysis revealed threshold effects where patients with 3+ chronic conditions were 2.4 times more likely to select independent pharmacies in the Southern region. CONCLUSIONS: Machine learning captures complex decision-making patterns in pharmacy selection that traditional statistical approaches miss. The identification of distinct patient segments and non-linear relationships provides actionable insights for targeted pharmacy service design and personalized medication management approaches.