Sensitivity analysis of computer models: World Bank HDM-III model

A problem often arising in engineering applications of computer models is to determine the importance of each data item in the large pool of required input factors. This paper explores a statistical approach for investigating factor sensitivities. The methodology is demonstrated with the HDM-III highway life-cycle cost analysis model. Specifically, the net present value (NPV) of life-cycle costs predicted by the HDM-III model is analyzed, and sensitivities of NPV to the link characterization input factors are investigated. In the statistically designed experiment, combinations of the input factors are chosen using a method called Latin hypercube sampling, which is well suited to the deterministic HDM-III model. Two analyses of the output data are performed, based on a first-order regression approximation and a Gaussian stochastic-process model. For NPV, the factor rankings are similar, but the sensitivities obtained from the two techniques show some marked differences. This demonstrates the greater flexibility of the stochastic-process model in dealing with nonlinearities and factor interactions in complex input-output relationships.