EXPECTED VALUE OF SAMPLE INFORMATION FOR CORRELATED DATA: A PRACTICAL APPROACH

   Purpose: Expected value of sample information (EVSI) is a key concept in Bayesian decision theory and illustrates the advantage of the decision-analytic framework over traditional statistical power calculations. We propose a simple and practical framework to compute measures of EVSI.       Methods:   Our approach entails three steps: (1) conduct a probabilistic sensitivity analysis (PSA), (2) regress the model parameters on the incremental net benefit (INB) of the new intervention compared to the standard of care using linear regression metamodeling (LRM), and (3) compute EVSI using the unit normal loss integral (UNLI) method.  The key concept in our approach is that EVSI relies on the correct estimation of the variance of the INB posterior to collecting new data from a new sample (n).  We achieved this goal using LRM, which assumes a linear relationship between the INB and the uncertain model parameters.  Then we adopted the UNLI as a parametric approach to compute EVSI from the fraction of the INB variance explained by the parameters of interest.  We illustrate our approach using a previously published decision model, which compared a new treatment to the standard of care for treating a serious condition.  The new treatment is effective, but associated with additional risk of a critical event.  The uncertain model parameters are (1) the probability of the critical event with the standard care (pC), the probability of side-effects following the new treatment (pSE), the number of quality-adjusted years after a critical event (QE), and the odds ratio of the efficacy of the new treatment compared to standard care (OR).       Results:   The Figure shows the EVSI calculated using our approach and those from the published model for various sample sizes (n).  The EVSI is shown for individual model parameters and for all the parameters combined.       Conclusion:   Our results closely predicted the results from the published model.  While PSA, LRM and UNLI are rooted in simulation studies, they have never been combined in this capacity to express EVSI.  In addition, our approach avoids complex mathematical notations, requires one PSA, and allows for correlation among model parameters.