A comparison of statistical models for analyzing episodes-of-care costs for chronic obstructive pulmonary disease exacerbations

Abstract

Accurate predictive models of costs for episodes of healthcare utilization associated with acute and chronic conditions can be used to develop non-fee-for-service provider remuneration systems. We examined the performance of eight predictive models for costs associated with episodes of care for chronic obstructive pulmonary disease (COPD) exacerbations: ordinary least squares (OLS) regression on untransformed costs, OLS regression on log-transformed costs with Duan’s retransformation, OLS regression on log-transformed costs with heteroscedastic retransformation, OLS regression on log-transformed costs with normal retransformation, robust regression, generalized linear model (GLM) with a Poisson distribution and log link function, GLM with a Gamma distribution and identity link function and GLM with a Gamma distribution and log link function. Administrative health data from Saskatchewan, Canada, including hospital records, physician billing claims, prescription drug records and home care service records were linked to identify individuals with diagnosed COPD and measure their episodes of health service utilization and costs. Cross-validation results showed that none of the models consistently resulted in the best prediction; the OLS regression model on log-transformed costs with normal retransformation had the highest R 2, but the OLS model on untransformed costs and the robust regression model had the best prediction accuracy based on root mean square error and mean absolute prediction error, respectively. Based on these findings, we recommend that researchers consider adopting one of these three models for predicting costs of healthcare use in episodes of care, but also emphasize that further comparisons of model performance are warranted.