OBJECTIVES: Recommendations about reviewing utility evidence for decision-analytic models do not specify how modellers should select utility evidence from included studies. It is unclear whether and how modellers should incorporate their prior belief about the order of severity between health-states. We aimed to (1) select reliable and consistent utility sources for exemplar health-states from the results of a systematic review (2) establish a set of principles for other modellers selecting utility evidence for multiple health-states.

METHODS: A systematic review following published guidelines identified 403 studies assessing utility for seven health-states related to cardiovascular disease (22/04/2021). From UK-relevant studies using the Euroqol 5 Domain (EQ-5D) measure, we selected a set of utilities to produce a plausible severity ordering. We considered risk of bias and applicability. Risk of bias incorporated participant selection, assessment methods and analysis. Applicability incorporated eligibility criteria, data reporting (relevant time-points, central tendency and dispersion) and utility instrument. To mitigate against unreported sources of heterogeneity, we preferred studies ascertaining higher proportions of the target population. When possible, we sourced multiple estimates from a single study to preserve relationships between utilities.

RESULTS: We could select sources with low risk of bias for all seven health-states, and directly relevant evidence for all but one. We identified single studies for closely-related conditions such as cerebrovascular events and acute coronary syndromes. In our base case decision-analytic model, we prioritised plausibility of severity ordering over population ascertainment. A sensitivity analysis prioritised population ascertainment over order plausibility.

CONCLUSIONS: We recommend three general principles for selecting utility evidence to inform multi-state models: minimising risk of bias (including from unreported participant selection criteria), ensuring applicability to the decision problem and respecting within-study relationships. Modellers must make value judgements about how to apply these principles to their particular context, balancing the need for a plausible ordering of utilities.