Defining and characterising structural uncertainty in decision analytic models

An inappropriate structure for a decision analytic model can potentially invalidate estimates of cost-effectiveness and estimates of the value of further research. However, there are often a number of alternative and credible structural assumptions which can be made. Although it is common practice to acknowledge potential limitations in model structure, there is a lack of clarity about methods to characterize the uncertainty surrounding alternative structural assumptions and their contribution to decision uncertainty.

A review of decision models commissioned by the NHS Health Technology Programme was undertaken to identify the types of model uncertainties described in the literature. A second review was undertaken to identify approaches to characterise these uncertainties.

The assessment of structural uncertainty has received little attention in the health economics literature. A common method to characterise structural uncertainty is to compute results for each alternative model specification, and to present alternative results as scenario analyses. It is then left to decision maker to assess the credibility of the alternative structures in interpreting the range of results.

The review of methods to explicitly characterise structural uncertainty identified two methods: 1) model averaging, where alternative models, with different specifications, are built, and their results averaged, using explicit prior distributions often based on expert opinion and 2) Model selection on the basis of prediction performance or goodness of fit. For a number of reasons these methods are neither appropriate nor desirable methods to characterize structural uncertainty in decision analytic models.

When faced with a choice between multiple models, another method can be employed which allows structural uncertainty to be explicitly considered and does not ignore potentially relevant model structures. Uncertainty can be directly characterised (or parameterised) in the model itself. This method is analogous to model averaging on individual or sets of model inputs, but also allows the value of information associated with structural uncertainties to be resolved.