A nonparametric importance sampling estimator for moment independent importance measures

Abstract

Moment independent importance measures have been proposed by E. Borgonovo in order to alleviate some of the drawbacks of variance-based sensibility indices. They have gained increasing attention over the last years but their estimation remains a challenging issue. An effective estimation scheme in the case of correlated inputs, referred to as single-loop method, has been proposed by Wei et al. In this paper we show via simulation that this method may be inaccurate, making for instance 40% error in the simplest possible Gaussian case. We then propose a new estimation scheme which greatly improves the accuracy of the single-loop method, up to a factor 10 in some simple numerical examples. We prove that our estimator is strongly consistent and several simulation results are presented to demonstrate the advantages of the proposed method.