Erin Kirwin, Mike Paulden, Chris McCabe, Jeff Round, Matt Sutton, Rachel Meacock

Preprint on SSRN.

Decision-makers often determine if technologies are good value for money and should therefore be adopted through value-based decision rules that compare cost-effectiveness analysis results to a threshold value. This decision rule assumes that decision-makers are indifferent between interventions with the same expected value but different underlying uncertainty. Such indifference is unlikely to hold in practice.

We propose a risk-based price and accompanying decision rules to address this limitation. Risk is characterized as the independent per-patient expected value of perfect information (iEVPI), a modification of standard EVPI. The iEVPI estimates the expected value of net benefit losses caused by uncertainty related to a technology, independent of the uncertainty related to alternative treatments. ‘Payer risk tolerance’ is the maximum per-patient risk of making wrong decisions that payers are willing to accept, expressed in monetary terms. The risk-based price is the price at which the iEVPI is equal to the payer risk tolerance. The risk-based decision rules are as follows: (i) a technology is acceptable for adoption if the incremental net benefit of the technology is greater than or equal to zero, and if the iEVPI is less than or equal to the payer risk tolerance, and (ii) the optimal technology has the greatest expected net benefit at the lower of the sponsor submitted or risk-based price at a given cost-effectiveness threshold value.

We demonstrate both risk-averse and risk-neutral payers prefer risk-based pricing outcomes. Risk-based pricing improves incentives for evidence development. Its implementation would increase health system net benefits.

Jeff Round, Erin Kirwin, Sasha van Katwyck, Christopher McCabe

The COVID-19 pandemic has increased public awareness of the influence of epidemiological and economic decision models on public policy decisions. Alongside this is an increased scrutiny on the development, analysis, and reporting of decision models, and how they inform public policy makers. Important technical and ethical questions are raised, bearing on the legitimate role of modelling in public policy decision making.

Decision models for public policy do not exist in a social vacuum. The scope of a model commissioned to inform public decision making is determined by the needs of the socially legitimate decision maker. Model developers may advise, but decision makers are accountable for setting the decision problem, outcomes considered, and policy decisions made.

We consider two challenges in modelling for pandemic policy response. First, the scope of the decision problem is not always made sufficiently explicit by decision makers, leaving modellers having to guess which policy options should be considered, and/or which outcomes should be used to evaluate them. Second, there is rarely sufficient transparency to ensure the public can see what is included in models, which limits the public’s opportunity to advocate to decision makers for the prioritization of specific outcomes and challenge the model results. The effect of these challenges is to weaken the evidence on which decisions are based and the public accountability of those making the decisions.

To address these challenges, we extend the recently described Directed Acyclic Graphs With Omitted Objects Displayed (DAMWOOD) approach to decision analytic models. We apply this approach to a previously published COVID-19 vaccine optimisation model. The enhanced diagrams illustrate the ways in which it is possible to improve communication of model assumptions. The diagrams make explicit which outcomes are omitted from model outputs and provide information on the expected impact of the omissions on aggregated model results.

We discuss the usefulness of DAMWOOD to communicate model structure and results, to frame the decision problem, and the potential value to decision makers of using DAMWOOD to communicate about models and as the basis of decision making.