^aDepartment of Civil, Construction, and Environmental Engineering, North Carolina State University, Raleigh, North Carolina, USA

Correspondence: Dr. H. Christopher Frey, Department of Civil, Construction, and Environmental Engineering, 311 Mann Hall, Campus Box 7908, North Carolina State University, Raleigh, NC 27695-7908, USA. Tel.: +919 515 1155; Fax: +919 515 7908; E-mail: frey@eos.ncsu.edu

Received 4 October 2005; Accepted 2 December 2005; Published online 25 January 2006.

Abstract

Sensitivity analyses of exposure or risk models can help identify the most significant factors to aid in risk management or to prioritize additional research to reduce uncertainty in the estimates. However, sensitivity analysis is challenged by non-linearity, interactions between inputs, and multiple days or time scales. Selected sensitivity analysis methods are evaluated with respect to their applicability to human exposure models with such features using a testbed. The testbed is a simplified version of a US Environmental Protection Agency's Stochastic Human Exposure and Dose Simulation (SHEDS) model. The methods evaluated include the Pearson and Spearman correlation, sample and rank regression, analysis of variance, Fourier amplitude sensitivity test (FAST), and Sobol's method. The first five methods are known as "sampling-based" techniques, wheras the latter two methods are known as "variance-based" techniques. The main objective of the test cases was to identify the main and total contributions of individual inputs to the output variance. Sobol's method and FAST directly quantified these measures of sensitivity. Results show that sensitivity of an input typically changed when evaluated under different time scales (e.g., daily versus monthly). All methods provided similar insights regarding less important inputs; however, Sobol's method and FAST provided more robust insights with respect to sensitivity of important inputs compared to the sampling-based techniques. Thus, the sampling-based methods can be used in a screening step to identify unimportant inputs, followed by application of more computationally intensive refined methods to a smaller set of inputs. The implications of time variation in sensitivity results for risk management are briefly discussed.