Estimates and confidence intervals for importance sampling sensitivity analysis

The performance of the importance sampling/likelihood ratio method for sensitivity analysis is poor for large perturbations—the estimated response eventually goes to zero as parameters change in either direction, no matter what is being estimated. Simultaneously, standard confidence intervals shrink to width zero, although the actual variance increases. We discuss importance sampling methods which give acceptable performance for a wider range of perturbations, and discuss confidence intervals which more accurately reflect the accuracy of estimates.