We've just arxived our paper on efficient computation for the Expected Value of Partial Perfect Information (EVPPI) based on SPDE-INLA. The EVPPI is a decision-theoretic measure of the impact of uncertainty in some of the parameters in a model on the final decision, informed by current evidence. Basically, it measures how much one should be willing to pay to gather new information to reduce uncertainty in the future outcomes. As such, it can be used to prioritise research and thus has been advocated as a very useful tool in health economic evaluation. Trouble is that it can be very difficult (or even impossible) to compute analytically and even a simulation approach may require an impracticable number of simulations.
The paper is part of Anna's PhD and in it we draw heavily on previous work on Gaussian Process (GP) regression to compute the EVPPI done by Mark Strong and colleagues at Sheffield. Our main idea is to express the GP regression model for the estimation of the EVPPI using a "fictional" spatial problem. As we show in the paper, this allows us to make use of clever models which simplify the computations and we can use INLA to obtain the results in a super-fast way.
As we show in the paper, our method can compute the EVPPI in a matter of seconds and this is virtually irrespective of the number of parameters involved. More importantly, there's a very large reduction in the computational time, in comparison to other non-parametric methods. Also, the accuracy of the estimation is very good.
We're implementing this into the next release of BCEA, which we plan on having ready in a matter of a few weeks!