Predictive model assessment for mapping between measures of health outcomes.

Patient-reported outcomes measures (PROMs), which include those evaluating health-related quality of life, are used extensively in health services research and to inform policy making; in general, PROMs can be distinguished as preference-based and non-preference-based measures, depending on whether the valuation method that is used to derive their index score is consistent with economic theory. Unfortunately, it is quite common for clinical studies to employ different PROMs, thus limiting the comparability of the evidence base that they contribute to. For this reason, national agencies as the National Institute for Health and Care Excellence for England and Wales, while have identified in EuroQoL-5D (EQ-5D) the PROM of choice, are accepting the use of a mapping approach to predict EQ-5D from other PROMs when EQ-5D data have not been collected. We consider the problem of directly predicting EQ-5D responses from ‘Short form 12’ by means of multivariate ordered probit models, and carry out the analysis within a Bayesian framework. In particular, we address the key problem of choosing an appropriate summary measure of agreement between predicted and actual results when analysing PROMs, and focus on scoring rules, with particular attention devoted to the case of discrete ordered data and to the importance of propriety.