The Utility of Local and Global Sensitivity Analysis Methods

Sensitivity analysis methods are used to identify measurements most likely to provide inmportant information for model development and predictions. Methods range from computationally demanding Monte Carlo global methods that require thousands to millions of models runs, to very computationally efficient linear (local) methods able to account for interrelations between parameters that involve tens to hundreds of model runs. Some argue that linear methods can provide 70% of the insight accessible through global methods for 1 to 2% of the model runs. Here this claim is explored using results from a simple model and an evaluation of nonlinearities typical of more complex models. Results suggest that, indeed, linear methods can be very informative for many problems, and that difficulties requiring global methods become evident rather rapidly. It is suggested, therefore, that local methods are nearly always a good first step in model analysis. Global methods remain necessary to fully characterize nonlinearity and obtain detailed probability distributions within the range of models and parameter values of concern.