Predicting cost of care using self-reported health status data

BACKGROUND We examined whether self-reported employee health status data can improve the performance of administrative data-based models for predicting future high health costs, and develop a predictive model for predicting new high cost individuals. METHODS This retrospective cohort study used data from 8,917 Safeway employees self-insured by Safeway during 2008 and 2009. We created models using step-wise multivariable logistic regression starting with health services use data, then socio-demographic data, and finally adding the self-reported health status data to the model. RESULTS Adding self-reported health data to the baseline model that included only administrative data (health services use and demographic variables; c-statistic = 0.63) increased the model" predictive power (c-statistic = 0.70). Risk factors associated with being a new high cost individual in 2009 were: 1) had one or more ED visits in 2008 (adjusted OR: 1.87, 95 % CI: 1.52, 2.30), 2) had one or more hospitalizations in 2008 (adjusted OR: 1.95, 95 % CI: 1.38, 2.77), 3) being female (adjusted OR: 1.34, 95 % CI: 1.16, 1.55), 4) increasing age (compared with age 18-35, adjusted OR for 36-49 years: 1.28; 95 % CI: 1.03, 1.60; adjusted OR for 50-64 years: 1.92, 95 % CI: 1.55, 2.39; adjusted OR for 65+ years: 3.75, 95 % CI: 2.67, 2.23), 5) the presence of self-reported depression (adjusted OR: 1.53, 95 % CI: 1.29, 1.81), 6) chronic pain (adjusted OR: 2.22, 95 % CI: 1.81, 2.72), 7) diabetes (adjusted OR: 1.73, 95 % CI: 1.35, 2.23), 8) high blood pressure (adjusted OR: 1.42, 95 % CI: 1.21, 1.67), and 9) above average BMI (adjusted OR: 1.20, 95 % CI: 1.04, 1.38). DISCUSSION The comparison of the models between the full sample and the sample without theprevious high cost members indicated significant differences in the predictors. This has importantimplications for models using only the health service use (administrative data) given that the past high costis significantly correlated with future high cost and often drive the predictive models. CONCLUSIONS Self-reported health data improved the ability of our model to identify individuals at risk for being high cost beyond what was possible with administrative data alone.