Incorporating output variance in local sensitivity analysis for stochastic models

The output of stochastic models is a distribution of values, rather than a single value such as in deterministic models. Local sensitivity analyses of such models typically ignore the higher moments of the output distribution and instead use the distribution mean to represent model output. This might be simplistic, since the shape of the distribution might also be sensitive to changes in model parameters. Here, we construct a simple sensitivity index that captures also the shape of the output distribution, by incorporating its variance in addition to its mean. To evaluate its performance, we reconstructed an existing stochastic individual-based model for mosquitofish (Gambusia holbrooki) population. We compared the performance of the new sensitivity index to the standard sensitivity index (∂Y/∂P) that was calculated using the mean of the output distribution, by ranking model parameters according to their impact on the output. Sensitivity analyses using both methods identified different parameters as the most influential on model output, and rankings were inconsistent between methods regardless of the number of simulations used for generating the output distributions. It is shown that the new index indeed captured better the effect of parameters on model output since it accounted for the variance of the output distribution.