Improving risk adjustment with machine learning: accounting for service-level propensity scores to reduce service-level selection

The hierarchical condition category (HCC) risk adjustment model tends to produce over-predictions of health care expenditures for individuals who need less costly services and under-predictions of health care expenditures for those who need costlier services. This tendency leads health plans to effectuate service-level selection to attract profitable individuals and avoid unprofitable individuals. In this study, we propose an alternative model using machine learning (ML) techniques to reduce service-level selection by accounting for demographic and diagnostic characteristics as well as service-level propensity scores (SPS) that capture each individual’s need for each service (the HCC + SPS model). Using the 2013–2014 Truven MarketScan database, we compare the performance of the HCC model (the HCC-only model) and the HCC + SPS model. We first fit both models with ordinary least squares (OLS) because traditional risk adjustment models rely on OLS. We also fit these models with ridge regression, which is a regularized ML algorithm, in order to examine whether the performance of the HCC + SPS model improves when combined with ML techniques. We evaluate prediction performance at three levels: group-level, tail distribution, and individual-level. We find that the HCC + SPS model more accurately estimated health care expenditures when combined with ridge regression, especially for individuals with high expenditures. However, we found limited improvements when the HCC-only model was used with ridge regression or the HCC + SPS model was used with OLS. Our findings suggest that accounting for SPS in risk adjustment using ML has the potential to reduce service-level selection.