Target and Conditional Sensitivity Analysis with Emphasis on Dependence Measures

En attente de publicationIn the context of sensitivity analysis of complex phenomena in presence of uncertainty, we motivate and precise the idea of orienting the analysis towards a critical domain of the studied phenomenon. For this, target and conditional sensitivity analyses are defined.We make a brief history of related approaches in the literature, and propose a more general and systematic approach. Nonparametric measures of dependence being well-suited to this approach, we also make a review of available methods and of their use for sensitivity analysis, and clarify some of their properties. Then, we focus our attention on sensitivity indices based on correlation ratio, namely Sobol- indices, and on two dependence measures the kernel quadratic dependence measure also called Hilbert--Schmidt independence criterion and the Csisz-{a}r divergence dependence measure. We propose adapted versions of these tools for target and conditional analysis, by considering transformation of the output using hard or smooth weight functions.Finally, we show on synthetic numerical experiments both the interest of target and conditional sensitivity analysis, and the efficiency of the dependence measures. We also illustrate the relevance of the proposed smooth versions for conditional estimators