Detailed statistical methods for computing the Expected Value of Sample Information metric (306637)*Hawre Jalal, University of Pittsburgh
Keywords: value of information, clinical trials, sample size, research prioritization, expected value of sample information, EVSI, methods
Expected Value of Sample Information (EVSI) involves computing the value of additional information from a finite sample (n) using a mathematical model f(.) which is often complex. Computing EVSI involves (1) simulating data collection (D) from n individuals, (2) updating the prior parameters based on the newly generated data (p|D), and (3) calculating the maximum expected net monetary benefit of such data collection experiments over all strategies. EVSI has been underutilized due to the computational burdens of evaluating f(.) within two nested Monte Carlo loops.
This talk presents four recent developments that enable significant reduction in computational time: (1) Strong proposed a low dimensional summary statistic for the data collection, (2) Menzies proposed a reweighting scheme based on the likelihood of observing each simulated dataset, (3) Jalal proposed a Gaussian-Gaussian Bayesian updating approach to estimate the posterior expectation of the parameters, and (4) Heath proposed a modified nested Monte Carlo approach. These methods compute EVSI with reasonable accuracy, can run on standard personal computers, and may increase the adoption of EVSI among researchers.