Comparison of methods for anchoring latent values on the full health-dead utility scale

A key challenge in ordinal methods is to anchor estimated health utility values onto the full health-dead scale. This study assessed five methods of anchoring.

Data were collected between 2016 and 2020 through two surveys conducted in the Quebec general population, with 1,176 and 908 respondents. Health utilities for the Short-Form 6-Dimension version 2 (SF-6Dv2) were estimated using ranking, composite time trade-off (cTTO), discrete choice experiment without (DCE) and with duration (DCE_TTO) methods. Anchoring was performed using five approaches: the dead state for ranking (Rank), linear mapping for DCE (DCE_Mapping), mean value for the worst health state (DCE_WHS), hybrid modeling (DCE_Hybrid), and duration (DCE_TTO). Conditional logit was used for rank and DCE approaches, while a hybrid model and generalized least squares (GLS) were applied for the DCE_Hybrid and cTTO methods, respectively. Approaches were compared based on the sign and ordering of their coefficients, the mean absolute difference (MAD) between the observed mean cTTO values and the estimates of models, and the overall pattern of their estimations.

A total of 17,200, 59,960, 8,500, and 16,464 observations were included for the DCE_TTO, ranking, cTTO, and DCE methods, respectively. The DCE_Hybrid method achieved the lowest MAD (0.056) from observed cTTO values, while DCE_TTO method had the highest (MAD = 0.423). Ranking and DCE_Mapping tended to underestimate utilities for better health states and overestimate for severe ones, while DCE_WHS and DCE_TTO underestimated health utilities for all health states. Visual comparisons confirmed that predictions from models closely aligned with observed cTTO values for better health states.

Hybrid anchoring method demonstrated better performance in predicting observed cTTO values, whereas DCE_TTO model showed greater deviations and a tendency to underestimate health state values. Hybrid anchoring method also appeared as the most appropriate and reliable approach for anchoring ordinal utility data to the full health-dead scale.