Exploring the variance contributions of correlated model parameters: A sampling-based approach and its application in traffic simulation models

Analyzing the impacts of model parameters on model outputs is an important but challenging topic for scientific research involving simulation models. Global Sensitivity Analysis (SA) has been recently employed by many transportation researchers for such task, but a proper SA is still not a common practice. In particular, many modelers simply assume that all parameters are uncorrelated in the SA. However, this assumption is often unrealistic for traffic simulation models, in which many parameters are actually correlated, leading to wrong conclusions. In this paper, a sampling-based approach is provided for the SA of correlated parameters. It uses Gaussian copula to link the marginal distributions of individual parameters with their global distributions and correlations, and utilizes the extended Sobol’ formula to estimate the variance-based sensitivity indexes in a Monte Carlo framework. Its application is illustrated using two different car-following models: the Intelligent Driver Model (IDM) and the Wiedemann-74 (W74) model. Results show that this method is able to accurately quantify the sensitivity of all model parameters. As a general method, this approach can be transferred like a standard quantitative SA tool to any traffic model or complex model in the wider scientific community, especially when correlated parameters exist.