Predicting health behavior for underrepresented groups: A case study on adherence to statins (Working Paper)

Abstract

Background: Accurate prediction of patients health behavior over time is crucial for chronic care. Claims data is an important source for model estimation but is limited for underrepresented groups due to small sample sizes. Purpose: To evaluate the potential for Bayesian hierarchical models to estimate predictive models of adherence for underrepresented groups. Methods: We used claims data from patients enrolled in the American Veterans Affairs health system who initiated statins. Our outcome is if those patients continued using statins for at least one year. Other covariates included LDL, total cholesterol, systolic blood pressure, past adherence, age, race, and sex. We develop a group-based framework using a hierarchical Bayesian model that allows all patient groups to borrow statistical strength from each other while retaining their individualized predictive model. We employ 3-fold cross-validation, with 67 percent of the data used for training and 33 percent for testing. We tested the potential benefit of stratifying our analyses by race and sex vs. a one-size-fits-all approach, estimating the area under the curve (AUC). Results: Our dataset demographics information and pharmacy claims for 3250 patients. Representation differed significantly based on race and sex, with the largest group (white men) representing 70 percent of the VA population. The smallest group, black women, comprise 2.5 percent. We used our model to estimate coefficients for each patient group. One-size-fits-all models yield an AUC of at least 57 percent (black women) and at most 79 percent (white women). Our hierarchical Bayesian model showed significant improvement in AUC for all groups, ranging from 3 percent for white women to 25 percent for black women, with the greatest improvements observed in minority groups. Conclusions: Bayesian hierarchical modeling can identify patients at risk of discontinuing treatment, including those from underrepresented groups. The proposed framework has the potential to improve personalized interventions.

Publication
Working Paper