Generalisable regression methods for costeffectiveness using copulas

Objectives: Covariate explanation of clinical trial cost and outcome is critical to allow reliable estimates of cost-effectiveness. The ordinary simultaneous equations approach however must specify a bivariate distribution for both cost and health outcome that is not typically a product of the best-fitting marginal distribution of each. This paper advocates estimating costs and outcomes simultaneously using copulas to model conditional dependence. The copula is a function that maps univariate marginal distributions of any to some joint distribution. Methods- Copulas are used to fit the bivariate distribution of the simultaneous model for individual cost and outcome in a clinical trial for hysterectomy. These are used to generate counter-factual outcomes and individual-level incremental net benefits due to each procedure, as well as replicating non-parametric techniques for comparison with standard methods. Results- Parametric results from the use of copulas compared to an ordinary Seemingly Unrelated Regression model show better fit with consistent estimates, allowing for the fact that the data is drawn from an underpowered clinical trial. Results also show that estimated coefficients vary in size, sign and statistical significance in different arms of the clinical trial. Non-parametric results compared with standard cost-effectiveness techniques also show more precise estimates of incremental net benefit. Conclusions- Regression-based approaches to cost-effectiveness have the potential to overcome a lot of the limiting assumptions made using non-parametric approaches. By using known information on covariates we can get more precise estimates of the parameters used in standard cost-effectiveness analysis, more precise posterior information and more precise posterior probabilities. Using copulas generates more precise estimation of conditionally-dependent marginal effects.