MAPPING STANDARD GAMBLE AND TIME TRADE-OFF UTILITIES TO THE MEDICAL OUTCOMES STUDY-HIV HEALTH SURVEY AND PATIENT SYMPTOMS

Purpose:  We explored regression methods to map Standard Gamble (SG) and Time Trade-Off (TTO) utility scores to either the Medical Outcomes Study-HIV Health Survey (MOS-HIV) or to a symptom index. Our objectives were to identify the best mapping method and to explore whether symptoms or domain scores yielded better models.

Methods:  We analyzed data from a study of directly elicited utilities from HIV-positive participants. Each participant used a computer-assisted tool to rate their own health using the SG and TTO, the MOS-HIV survey, and an HIV-specific symptom index. Predictor variables were MOS-HIV domain scores or 20 symptoms, each classified as present or absent. We compared quantile regression, which predicts median scores, and five methods that predict mean scores:  generalized linear model (GLM) with identity link, GLM with power link, GLM with log link, robust regression, and a two-step model (logistic plus GLM). We assessed model fit based on the mean absolute error (MAE) in predicted utilities, which we evaluated based on ten-fold cross-validation estimates.

Results: Among 251 individuals, the mean SG score was 80.0 (95% confidence interval [CI] 76.8 to 83.1) and the mean TTO score was 77.2 (95% CI 73.7 to 80.7); the corresponding medians were 90.1 (Interquartile range [IQR] 67.0 to 99.95) and 89.2 (IQR 61.5 to 99.9).  For both SG and TTO methods, symptom scores yielded higher mean absolute errors than MOS-HIV domain scores with GLM-based models but lower errors for other methods. SG models generally yielded lower errors than TTO models. For SG utilities, the lowest MAE (17.63) was observed with quantile regression using symptom scores and the lowest MAE (18.07) among models that predicted mean scores was with robust regression using symptom scores. For TTO utilities, the lowest MAE (19.83) was observed with quantile regression using symptom scores and the lowest MAE (20.29) among models that predicted mean scores was with robust regression using symptom scores. The worst performing models were two-step models, with MAEs of 19.73 for SG predicted by symptom scores and 22.75 for TTO scores predicted by symptom scores.

Conclusion:  Quantile regression or robust regression using symptom scores as predictor variables yielded the best fitting models to predict directly measured utilities among people living with HIV. However, all models had relatively poor predictive performance.