Generalising results from cost-effectiveness analysis: a comparison of methods to account for covariates

OBJECTIVES: Inputs to economic models may depend on patient characteristics which need to be accounted for when estimating cost-effectiveness in a particular target population. A common approach to obtain cost-effectiveness estimates is to plug-in the mean covariate values of the population. However, such estimates correspond to individual patients having the average characteristics of the population. Another approach is to simulate from the joint distribution of covariates in the target population, but this over-estimates uncertainty in the estimates. Instead to obtain estimates specific to a given target population, an integration is required over the joint covariate distribution, however this requires complex computation. This research aims to compare the performance of the different methods. METHODS: The methods are illustrated using a cost-effectiveness model of intravenous immunoglobulin (IVIG) for patients with sepsis. We compare expected incremental net benefit (INB) and probability of IVIG being cost effective for each method: (i) mean of covariates (ii) simulation (iii) integration. RESULTS: For the whole-population analysis, expected INB at £20,000/QALY is equal to -556 when integrating over the covariate distribution and equal to -293 when using the mean covariate value. This results in a 3.5% lower probability of IVIG being cost effective. Simulating from the covariate distribution provides an expected INB at £20,000/QALY of -556, however the overestimation in uncertainty results in a 6.7% lower probability of IVIG being cost-effective than integrating over the covariate distribution. For ICNARC physiology scores representative of the majority of patients, the probability is around 10%-12% lower when integrating over the covariate distribution than when using the mean covariate values. CONCLUSIONS: Using the mean of covariate method introduces bias in estimates of cost-effectiveness. Simulation provides unbiased estimates, but over-estimates uncertainty. Integrating net benefits over the joint covariate distribution is necessary, if it is important to reliably estimate decision uncertainty.