Predicting pharmacy choice for managed care network design.

Abstract

Background: Pharmacy type selection is a key component of medication access and use. Prior studies have commonly used logistic regression to examine pharmacy type choice, but this approach may not fully capture complex relationships among patient characteristics.

Objective: To examine patient characteristics associated with pharmacy type selection and to compare the performance of traditional logistic regression with ensemble machine learning models for predicting pharmacy type selection.

Methods: We conducted a cross-sectional analysis of adults participating in the 2021 National Consumer Survey on Medication Experience (NCSME-PR; n = 1,502). The outcome was the pharmacy type selected when a prescription medication was needed, categorized as chain, independent, supermarket, mass merchandise, mail-order, or clinic-based pharmacy. Predictors included patient characteristics defined by the Andersen Behavioral Model of Health Services Use, encompassing predisposing, enabling, and need factors. We compared logistic regression with random forest and extreme gradient boosting (XGBoost) models using 5-fold cross-validation and held-out test data. Model discrimination was assessed using the area under the receiver operating characteristic curve.

Results: Across pharmacy types, ensemble models demonstrated higher discrimination than logistic regression, with XGBoost achieving the highest area under the receiver operating characteristic curve values. Prior mail-order pharmacy use, number of chronic conditions, income, and US region were consistently associated with the selection of pharmacy type. Ensemble models captured nonlinear patterns in these associations that were not fully reflected in logistic regression models.

Conclusions: Pharmacy type selection varies by patient characteristics and pharmacy type, and predictive performance differs by analytic approach. Comparative modeling indicates that conclusions about pharmacy choice may depend on the modeling framework used. These findings contribute to pharmacy choice research and highlight considerations for future studies using managed care administrative data.