This presentation is part of: Poster Session - CEA: Methods and Applications; Health Services Research
ARE RELATIONSHIPS BETWEEN SF-36 SUBSCALES AND HEALTH UTILITY MERELY LINEAR?
Andrew P. Yu, MA, MS, Yanni F. Yu, MA, and Michael B Nichol, PhD. University of Southern California, Pharmaceutical Economics and Policy, Los Angeles, CA
Purpose: The underlying relationship between health status scales (e.g. SF-36 domains) and utility is unknown. Previous studies assumed SF-36 scales mapped linearly to utility measures. Such an assumption provides a simple algorithm, but may be inconsistent with derived power functions, and may yield poor model fitting, lower predictability and regression artifacts. This study was to examine the nonlinear relationships between SF-36 subscales and HUI2 utility. Methods: Data included 6923 Southern California Kaiser Permanente members who filled both SF-36 and HUI2 in year 1994-1995. Missing values were imputed on item level by MCMC and propensity score method with a missing at random assumption. In order to relax the assumption of linear effect SF-36 subscales on utility, we used restricted cubic spline functions (CSF) with four or five knots for each subscale except for RP, SF and RE, which had less than ten unique values and were kept as ordinal categories. Subscale scores were transformed to population norm t-scores according to the SF-36 scoring algorithm. The heuristic shrinkage estimate was used to test for model overfitting. Nonlinear relationships by CSF were visualized by plotting utility against each subscale when holding other covariates at their medians. Nonlinearity of each subscale as well as total nonlinearity of the model was examined by F-test. The final model included the reduced model with significant nonlinear subscales and interactions between subscales and age. Results: The following subscales presented statistically significant nonlinear relationship with utility: PF (F=9.34, d.f.=3, p<.0001), VT (F=3.50, d.f.=2, p=.0302), MH (F=11.14, d.f.=3, p<.0001), BP (F=42.81, d.f.=2, p<.0001). Total nonlinearity was significant (F=11.76, d.f.=16, p<.0001). These nonlinear relationships were also evidenced on fitted plots. Interaction terms between age subscales were found to be significant and were included in the final model. Even though 55 parameters were presented in the final model, adjusted R2 increased from 0.499 for simple linear model to 0.518 (total R2 from 0.500 to 0.522) and the heuristic shrinkage estimate of the final model revealed no concern for overfitting. Conclusions: Some SF-36 subscales show a nonlinear relationship with utility. Researchers should consider the use of nonlinear models in mapping health status to utility.
