PURPOSE: Uncertain decisions made using a cost-effectiveness threshold applied to each decision problem separately fail to identify the true opportunity costs of displacing other unrelated programmes. The implications for value of information analysis are profound. We demonstrate that the expected value of perfect information (EVPI) for the allocation problem as a whole provides a more appropriate estimate of the returns to further research.

METHODS: A stochastic mathematical programming approach is used to solve the allocation problem. By explicitly distinguishing between parameter uncertainty and variability, the value of acquiring further evidence to inform the allocation problem is derived. The EVPI is expressed in terms of health benefits as the difference between expected benefits with perfect information (no uncertainty) and expected benefits with current information. As no unique shadow price exists to convert these health benefits into monetary terms, the monetary value of information is the reduction in budget required with perfect information to generate the same expected benefits as with current uncertainty. The corresponding threshold is the marginal change in benefits for a unit change in this budget. We compare the EVPIs derived from separate analysis of each decision problem to exactly the same set of decision problems solved simultaneously.

RESULTS: These approaches when applied to identical problems with identical populations and programmes at an entirely equivalent threshold give substantial differences in the value of research. At a budget of £7 million, corresponding to a threshold of £10,000, the EVPI for the allocation problem as a whole is £220,000. The EVPI based on the analysis of each of the decision problems separately substantially overestimated (£800,000) the value of research. EVPI for the allocation problem as a whole provides the correct upper limit because it not only incorporates uncertainty in the choice between treatments for a particular population but also the impact on unrelated treatments that can become possible within the budget constraint. Furthermore, we show that the EVPI for a subset of parameters directly relevant to a programme depends on all the other competing but independent programmes.

CONCLUSIONS: The value of acquiring further evidence to inform an allocation problem must be made in the context of the allocation problem as a whole. Current methods based on the separate analysis of each decision problem may be misleading.