OBJECTIVES: Cost-effectiveness models typically depend on several patient and disease characteristics (i.e., covariates). To estimate results at the population level the model needs to average over the joint covariate distribution at each evaluation of the model. This requires nested numerical integration, which can be computationally challenging. We aim to explore the bias with two commonly used approximations.

METHODS: We compare the results from three different approaches to covariate adjustment in an economic model for intravenous immunoglobulin (IVIG) in severe sepsis: the mean of covariate method obtained by plugging in the average value of each covariate, the predictive distribution method that is based on simulating from the joint covariate distribution with a single simulation for each evaluation of the model in a probabilistic sensitivity analysis (PSA), and averaging over the covariate distribution by integrating over the joint covariate distribution at each iteration of the PSA (gold standard approach).

RESULTS: Cost-effectiveness results depend on disease severity as defined by APACHE 2 and ICNARC scores. The mean of covariate method overestimates the probability of IVIG being cost-effective by up to 14% compared to averaging over the covariate distribution. That is, for a patient with an ICNARC score of 28, IVIG has a 55% probability of being cost-effective under the mean of covariate approach and a 41% probability of being cost-effective under the average over the covariate distribution approach. The predictive distribution method provides cost-effectiveness estimates aligned to averaging over the covariate distribution method but slightly over- or under-estimates the probability of IVIG being cost-effective.

CONCLUSIONS: The mean of covariate method provides biased cost-effectiveness results due to the non-linear structure of the economic model. The predictive distribution method provides unbiased cost-effectiveness results but does not accurately capture the uncertainty in the probabilistic results due to assigning distributions to additional parameters in the PSA.