Guidelines for optimal estimation of correlation ratios based on Monte Carlo simulation of computer codes

Abstract The computational cost of many computer codes is a burden that obliges users to use the same set of simulations to perform uncertainty and sensitivity analysis. Correlation ratios are a simple and straightforward tool to estimate first order sensitivity indices. Nevertheless, they demand the use of debiasing factors and of appropriate partitions of the support of each input parameter to deliver optimal estimates. The way to estimate these indices depends in fact on the actual values to be estimated. This work contains an exhaustive study developed for providing a set of guidelines about the optimal way to deliver such estimates, which involves the minimum sample sizes needed, the selection of the partition and the use of appropriate debiasing factors.