PURPOSE: To compare two approaches to modeling the impact of multiple health conditions on health utilities: a direct model for summary scores versus modeling individual domain scores which are then combined into summary scores.

METHODS: We use data collected between 1998 and 2004 from the Medicare Health Outcomes Survey. These data include 15 self reported health conditions, age, sex, education, income, and the SF-36 version 1. The SF-6D preference-based summary score was calculated form the SF-36. We compare four models to predict SF-6D scores: additive, multiplicative, censored least absolute deviations (CLAD), and domain-based. The domain-based method uses ordinal logistic regression to model SF-6D domains and domain scores are combined to generate an SF-6D summary score. The accuracy (bias, mean square error, mean absolute deviation) of predictions from these models are assessed using a reserved sample.

RESULTS: We use 103,484 observations to generate models and 102,945 observations for model testing. The domain-based and multiplicative models have the smallest bias for predictions of the mean. CLAD has the smallest bias for predictions on the median for groups with 0 or 1 health condition, all four models are equivalent for groups with 2 to 7 health conditions, and the domain-based and multiplicative models show the smallest bias in groups with 8 or more health conditions. All four models perform similarly on mean squared error and mean absolute deviations in groups with less than 7 health conditions. For groups with 7 or more health conditions, the domain-based and multiplicative models perform best.

CONCLUSIONS: The domain-based and multiplicative models perform better than the additive or CLAD models. The multiplicative model is easier to implement and interpret than the domain-based model and is recommended for modeling the impact of multiple health conditions on SF-6D summary scores in these data. Future work should evaluate the performance of these models in healthier populations.