Uncertainty quantification and global sensitivity analysis for economic models

Abstract

Sensitivity analysis assesses the influence of input parameters on the conclusion of a model. Traditional analysis methods - based on evaluating the model at a reference parameter vector and changing one parameter at a time - are local, linear, and usually do not capture interactions among the parameters. By contrast, the global sensitivity analysis that we present summarizes the parameters’ importance over a range of values, fully capturing nonlinearities and identifying interactions. Specifically, we propose Sobol’ indices, which are based on variance decomposition, and exemplify their use with a standard real business cycle model. Standard approaches to variance decomposition require a large number of model evaluations. To overcome this, we present the state-of-the-art approach for calculating Sobol’ indices, which is based on building a polynomial representation of the model from a limited number of evaluations. In addition, we use this polynomial representation to evaluate the univariate effects, which are conditional expectation functions that can be interpreted as a robust impact of a parameter on the model conclusions.