Abstract
OBJECTIVES: Modelling health-state utility values (HSUVs) from clinical data offers a means to conduct retrospective cost-effectiveness analyses using clinical studies that did not collect direct HSUV measures. Such studies can support the efficient allocation of resources in kidney transplantation (KT). We aim to model KT recipients' EQ-5D-3L HSUVs using routinely collected clinical data. METHODS: From a French observational multicentric prospective cohort, we included 2,787 adult recipients of a first or second single renal graft transplanted between January 2014 and December 2021 who completed 5,679 EQ-5D-3L questionnaires post-KT, from which the HSUVs were calculated. Considering two time periods before and after 1-year post-KT, we estimated a linear mixed effect model (LME), a mixed adjusted limited dependent variable mixture model, and beta and two-part beta mixed models. We compared their predictive performances in terms of precision and calibration. RESULTS: In each model, recipient age, female sex, higher body mass index, presence of comorbidities and time spent on dialysis prior to KT were associated with lower HSUVs. The predicted HSUVs increased during the first year post-KT before slowly decreasing afterwards. The two-part beta mixed model resulted in the most precise predictions but showed poor calibration. The LME was associated with better calibration than the other models. CONCLUSIONS: Our study illustrates the importance of estimating longitudinal predictive algorithms to consider possible time variations in HSUVs. We provide an online calculator for predicting the HSUVs of KT recipients over time. Future studies in international cohorts are important to support the external validity of our results.