Linearization of local probabilistic sensitivity via sample re-weighting

Local probabilistic sensitivity of input variable X with respect to output variable Z is proportional to the derivative of the conditional expectation E(X|z). This paper reports on experience in computing this conditional expectation. Linearized estimates are found to give acceptable performance, but are not generally applicable. A new method of linearization based on re-weighting a Monte Carlo sample is introduced. Results are comparable to the linearized estimates, but this method is more widely applicable. Results generally improve by conditioning on a small window around z.