In the new baseline simulations, we made an important improvement to the health plan choice model by including the effect of health status as a variable that affects plan choice. This is important for two reasons. First, health status may be an important factor in predicting plan choice, so the addition of this variable should improve the fit of the choice model, other things equal. Second, it may be that case that sick (healthy) people prefer certain plans, which would drive the premiums up (down). Specifically, if sick people are attracted to traditional plans, it could lead to a "death spiral" of increasing premiums and falling enrollment for the traditional plans. One of our goals in the new simulations is to determine whether the addition of a HDHP as a choice will tend to de-stabilize the market for health insurance.
To account for health status, we used the claims data for contract holders (employees) in the employers we examined in our prior plan choice analysis. For this analysis, we obtained the claims data for the year prior to their possible enrollment in a high deductible health plan. We used the diagnosis code information from these prior year claims records to calculate a set of 34 Adjusted Diagnosis Groups (ADGs) using a methodology developed by Johns Hopkins University researchers.8 Several of these 34 ADGs identify a diagnosis indicating the presence of a chronic condition. With this information we constructed a dummy variable indicating the presence of chronic illness. This variable construction permitted us to develop a medical care cost regression model to predict future medical expenditure of the MEPS population enrolled in each plan type. A description of the results of these prediction regressions (for each plan choice) is presented in the appendix. Using an aggregate measure of health status represented as a binary variable allowed us to create a variable we could map from the MEPS database in order to predict health plan choice.
The final health plan choice model is presented in the appendix. Besides the inclusion of health status, we also interacted premium and cost sharing variables with more demographic variables. These interactions were introduced to account for possible associations not accounted for by the own-price response to premium and cost sharing variables.