Population choice and variable selection in the estimation and application of risk models

This paper shows that risk adjustment models based on demographic and employment variables are not easily transferable from one population to the next, and that administrative variables are not useful in predicting medical expenditures. We found statistically significant differences between models built on populations of employees from a single employer enrolled in two different health plans, and between models built on populations of enrollees from a single health plan employed at two different companies. Employment-based variables (e.g., length of employment) had little predictive power in any of these risk models. We conclude that policymakers should be careful in applying risk models across populations, and that future versions of risk models for use within large employers need not include employment-based variables.