This presentation is part of: Poster Session - Public Health; Methodological Advances
REGRESSION ARTIFACTS IN LINEAR MAPPING HEALTH STATUS MEASURE TO UTILITY: EVIDENCE FROM MONTE CARLO SIMULATION
Andrew P. Yu, MA, Yanni F. Yu, MA, and Michael B Nichol, PhD. University of Southern California, Pharmaceutical Economics and Policy, Los Angeles, CA
Purpose: Correct measurement of change in utility is paramount in cost-effectiveness analysis of health intervention. Health status measures, such as the SF-36 are ordinal scales with demonstrated ceiling and floor effects, whereas utility has presumed interval characteristics. Scaling inconsistencies may create regression artifacts (regression to the mean) and underestimate the magnitude of utility change. The purpose of this study was to estimate the extent of such regression artifacts. Methods: Monte Carlo simulation methods were used to determine the presence of regression artifacts and estimate their extent under various conditions. Utility was assumed to be linearly associated with latent trait of each SF-36 domain, according to linear regression models (Nichol, 2001). In this study, continuous latent trait scale used population norm t-scores (mean=50, sd=10), which were then linked to discrete domain scores. Population heterogeneous distribution and SF-36 measurement error were realized by random numbers of normal or uniform distributions. Parameters of interest were baseline utility (from 0 to 1), utility change (0.01, 0.05, 0.1), and measurement error (corresponding reliability 100%, 99%, 90%). Results: Changes in utility scores translated from SF-36 strictly underestimated the true utility change in every scenario examined. Even with 100% measurement reliability, underestimation was 22% for 0.01 utility change, approximately 14% for both 0.05 and 0.1 utility change. With increasing measurement error, underestimation increased. For example, at 90% measurement reliability level, when true utility changed 0.1, average estimated utility scores changed only 0.076, with 24.4% underestimation. In general, the regression artifacts presented the least threat at the 0.50 utility score, whereas underestimation was most severe at both ends, reinforcing the pronounced role of ceiling/floor effects. Conclusion: Regression artifacts seemed to be inevitable when mapping HSM to utility based on linear regression algorithm, and may result in underestimation of utility change. Many factors appeared to be responsible for such artifacts including scaling inconsistencies, ceiling/floor effects, and imperfect measurement reliability.
