Modelling of UK General Population Utility: The Aldvmm Approach in Practice

OBJECTIVES: A new NICE DSU report on age and sex-adjusted general population utility was recently published (Hernández Alava et al. 2022). The authors fitted a three-component adjusted limited dependent variable mixture model (ALDVMM) to the latest EQ-5D-3L data from the Health Survey for England (HSE) dataset. While the ALDVMM method was specifically developed to address distributional features of the EQ-5D, this model is significantly more complex than other published linear regression models for general population utility, such as the ordinary least squares (OLS) model by Ara and Brazier (2010). Our research aimed to compare the ALDVMM approach against an OLS approach to assess the practicality and stability for extrapolating general population utility in health economic analyses.

METHODS: We compared the NICE DSU ALDVMM model against an OLS model fitted to the same HSE 2014 dataset. ALDVMM and OLS models were also fitted to all available HSE EQ-5D-3L data from 2003 to 2014, to further assess the performance of the ALVDMM approach in the broader dataset. We also tested the stability of each model for probabilistic sensitivity analysis (PSA) using Monte Carlo simulation.

RESULTS: OLS model mean predicted utilities were similar to the DSU ALDVMM approach when fitted to both the 2014 and broader HSE datasets. Compared to OLS models, the ALVDMM approach showed >1% samples among older age groups in Monte Carlo simulations predicting 20% below the expected utilities. Neither ALDVMM nor OLS models predict utilities outside of -0.594 to 1 range in adult population. While the stability of the ALDVMM approach improved when using the total HSE dataset, the model still showed additional volatility relative to an OLS model.

CONCLUSIONS: OLS models may generate similar results to ALDVMM models when predicting UK general population EQ-5D-3L utility using HSE data, while providing less complexity and higher stability for conducting PSA in health economic models.