OBJECTIVES:

Probabilistic Sensitivity Analysis (PSA) results depend on the assumptions about the marginal distributions of model parameters and their joint distribution. Although joint distribution of some parameters are accounted for (Dirichlet for transition probabilities, multivariate normal for regression parameters), in most health economic models investigation of different types of dependence structures is omitted. As a result, PSA results reflect model inputs that are assumed to be independently distributed. This can lead to inaccurate uncertainty analysis results.

METHODS:

We demonstrated how to fit copulas to data we used to populate a health economic model to construct joint distributions. We then sampled parameter values under different assumptions about their joint distribution: First we assumed they are independent, reflecting the current practices; second we assumed that the variables are jointly normally distributed, and finally we used copulas to sample from the marginal distributions. We compared various plots generated under different assumptions to the scatterplot of the original data. Using a health economic cost-effectiveness model, we analysed the PSA results under three different assumptions.

RESULTS:

Joint distributions with independence and multivariate normal assumption do not accurately represent the dependence observed in the data. PSA results indicate that the variability of model outcomes changed. This has implications on the conclusions about the uncertainty of the base case estimate of cost-effectiveness of the treatment.

CONCLUSIONS:

Ignoring the dependence structure between model parameters can lead to inaccurate results and can distort PSA conclusions. Investigation of the joint distribution of parameters should be a routine part of uncertainty analysis. The methodologies for fitting copulas and simulating random variables with different dependence structures are well documented and they should be incorporated in health economic modelling.