This presentation is part of: Poster Session - Utility Theory; Health Economics; Patient & Physician Preferences; Simulation; Technology Assessment
PREDICTING DERMATOLOGICAL UTILITIES: AN EXPLORATION OF CENSORED UTILITY DATA PREDICTION MODELS
Suephy C Chen, MD, MS1, Maya Sternberg, PhD2, Ahmed M. Bayoumi, MD, MSc3, Ingram Olkin, PhD4, Seaver L Soon, MD1, Ponciano Cruz, MD5, Shenara Sexton, MD6, Calvin O McCall, MD6, and Mary K. Goldstein, MD7. (1) Emory University, Department of Dermatology, Atlanta, GA, (2) Emory University, School of Public Health, Biostatistics, Atlanta, GA, (3) St. Michael's Hospital, University of Toronto, Inner City Health Research Unit, Toronto, ON, Canada, (4) Stanford University, Statistics, Palo Alto, CA, (5) University of Texas at Southwestern, Department of Dermatology, Dallas, TX, (6) Emory University, Dermatology, Atlanta, GA, (7) Stanford University, Center for Primary Care and Outcomes Research, Stanford, CA
Difficulties in obtaining public utilities have prompted researchers to develop regression models to predict utilities from health status measures. However, these models are based on general health instruments and may not be relevant for dermatological health states. Moreover, regression models based on linear regression (LR) are only valid only when certain assumptions are met. Violations of these assumptions can lead to biased prediction models. If the population under study rarely records utilities of 0 and 1 it may reasonable to disregard the ceiling or floor effect and use OLS. However, we have demonstrated that for most dermatological health states, utility scores will cluster close to the ceiling of 1. PURPOSE: We explored several different regression methods to identify the best method to predict health utilities from a validated skin-specific health status measure, Skindex. METHODS: We interviewed 250 consecutive patients from general dermatology clinics at Stanford Medical Center (Stanford, CA), Grady Hospital (Atlanta, GA), and Parkland Hospital (Dallas, TX). Subjects completed both Skindex and a time trade-off utility assessment. We randomly divided our data into derivation and validation sets. The derivation data were analyzed using LR, Tobit regression (TR), Least Absolute Deviations (LAD), Least Trimmed Squares (LTS), Least Median Squares (LMS), and Censored LAD (CLAD). Candidate predictor variables included both Skindex and demographic parameters. We chose the same predictor variables for all models. We used the validation dataset to determine the median absolute prediction error (MAPE) (difference of the predicted and actual utility) and interquartile range (IQR) for each model. RESULTS: In our sample, 48.2% report a health utility of 1, with very little dispersion around the median of .9998 (IQR: 1.0 - .9487). The MAPE and IQR were as follows: LR 0.042 (0.075); TR: 0.073 (0.1); LAD: 0.027 (0.083); LTS: 0.039 (0.096); LMS: 0.039 (0.096); CLAD: 0.053 (0.11). CONCLUSIONS: Although the LR model is not appropriate to estimate utilities when censoring is present, alternative regression models are available to alleviate the problems of bias inherent in ignoring ceiling effects. We have found for our dermatology-based population that the LAD regression gave the least difference in predicted and actual utilities. CLAD did not improve on prediction ability. Other investigators creating prediction models in the face of censored utility data should explore these regression methods.
