Predictive modeling of total healthcare costs using pharmacy claims data: a comparison of alternative econometric cost modeling techniques

OBJECTIVE: We sought to evaluate several statistical modeling approaches in predicting prospective total annual health costs (medical plus pharmacy) of health plan participants using Pharmacy Health Dimensions (PHD), a pharmacy claims-based risk index. METHODS: We undertook a 2-year (baseline year/follow-up year) longitudinal analysis of integrated medical and pharmacy claims. Included were plan participants younger than 65 years of age with continuous medical and pharmacy coverage (n = 344,832). PHD drug categories, age, gender, and pharmacy costs were derived across the baseline year. Annual total health costs were calculated for each plan participant in follow-up year. Models examined included ordinary least squares (OLS) regression, log-transformed OLS regression with smearing estimator, and 3 two-part models using OLS regression, log-OLS regression with smearing estimator, and generalized linear modeling (GLM), respectively. A 10% random sample was withheld for model validation, which was assessed via adjusted r, mean absolute prediction error, specificity, and positive predictive value. RESULTS: Most PHD drug categories were significant independent predictors of total costs. Among models tested, the OLS model had the lowest mean absolute prediction error and highest adjusted r. The log-OLS and 2-part log-OLS models did not predict costs accurately as the result of issues of log-scale heteroscedasticity. The 2-part model using GLM had lower adjusted r but similar performance in other assessment measures compared with the OLS or 2-part OLS models. CONCLUSION: The PHD system derived solely from pharmacy claims data can be used to predict future total health costs. Using PHD with a simple OLS model may provide similar predictive accuracy in comparison to more advanced econometric models.