A framework for addressing structural uncertainty in decision models

Abstract

Decision analytic models used for health technology assessment are subject to uncertainties. These uncertainties can be quantified probabilistically, by placing distributions on model parameters and simulating from these to generate estimates of cost-effectiveness. However, many uncertain model choices, often termed structural assumptions, are usually only explored informally by presenting estimates of cost-effectiveness under alternative scenarios. The authors show how 2 recent research proposals represent parts of a framework to formally account for all common structural uncertainties. First, the model is expanded to include parameters that encompass all possible structural choices. Uncertainty can then arise because these parameters are estimated imprecisely from data, for example, a treatment effect of doubtful significance. Uncertainty can also arise if there are no relevant data. If there are relevant data, uncertainty can be addressed by averaging expected costs and effects generated from probabilistic analysis of the models with and without the parameter. The weights used for averaging are related to the predictive ability of each model, assessed against the data. If there are no data, additional parameters can often be informed by eliciting expert beliefs as probability distributions. These ideas are illustrated in decision models for antiplatelet therapies for vascular disease and new biologic drugs for the treatment of active psoriatic arthritis.