Context: Statistical models employed in analysing patient-level cost and effectiveness data need to be flexible enough to adjust for any imbalanced covariates, account for correlations between key parameters, and accommodate potential skewed distributions of costs and/or effects. We compare prominent statistical models for cost-effectiveness analysis alongside randomised controlled trials (RCTs) and covariate adjustment to assess their performance and accuracy using data from a large RCT.
Method: Seemingly unrelated regressions, linear regression of net monetary benefits, and Bayesian generalized linear models with various distributional assumptions were used to analyse data from the TASMINH2 trial. Each model adjusted for covariates prognostic of costs and outcomes.
Results: Cost-effectiveness results were notably sensitive to model choice. Models assuming normally distributed costs and effects provided a poor fit to the data, and potentially misleading inference. Allowing for a beta distribution captured the true incremental difference in effects and changed the decision as to which treatment is preferable.
Conclusions: Our findings suggest that Bayesian generalized linear models which allow for non-normality in estimation offer an attractive tool for researchers undertaking cost-effectiveness analyses. The flexibility provided by such methods allows the researcher to analyse patient-level data which are not necessarily normally distributed, while at the same time it enables assessing the effect of various baseline covariates on cost-effectiveness results.
Keywords: Bayesian regression methods; Cost-effectiveness analysis; Covariate adjustment; Net monetary benefits; Regression methods; Seemingly unrelated regressions.