Managing Structural Uncertainty in Health Economic Decision Models

Health economic models are representations of judgements about the relationships between the model's input parameters and the costs and health effects that the model aims to predict. We recognise that we can rarely define with certainty a 'true' model for a particular decision problem. Building an 'incorrect' model will result in an uncertain prediction error, which we denote 'structural uncertainty'. The absence of observations on the total costs and health effects under each decision option limits the use of data driven approaches to managing structural uncertainty, such as model averaging.

We therefore propose a discrepancy based approach in which we make judgements about structural error at the sub-function level within the model and introduce a series of terms to 'correct' the errors. This is deemed to be easier than making meaningful statements about the error at the level of the model output. The specification of discrepancy terms within the model also allows us to use sensitivity analysis methods to determine the relative importance of the different structural uncertainties in driving output and decision uncertainty.

Following the computation of either the main effect index or the partial expected value of perfect information for each discrepancy term, we can review the structure of those parts of the model where structural uncertainty is an important source of model output or decision uncertainty. We interpret the overall expected value of perfect information for all the discrepancy terms as an upper bound on the expected value of model improvement (EVMI).

We illustrate the sub-function discrepancy method in two case studies: a simple decision tree, and a more complex Markov model. Finally, we propose an efficient method for computing the main effect index and the partial expected value of perfect information when inputs and/or discrepancies are correlated.