A Comment on Blanning's “Metamodel for Sensitivity Analysis: The Regression Metamodel in Simulation”

The purpose of this note is to provide some statistical tools to make operational the idea of “metamodels” as discussed by Blanning (Blanning, R. W. 1974. The sources and uses of sensitivity information. Interfaces4 (4, August) 32–38.) in a recent issue of Interfaces. He distinguishes between the simulation model or “decision model” itself and its approximation called the “metamodel”. An example of a simulation model may be a GPSS program for a queuing system, or a FORTRAN program for an inventory system. The purpose of the metamodel is to measure the sensitivity of the output (e.g., mean waiting time, mean inventory cost) to the inputs, the inputs being either decision variables (e.g., number of service stations, reorder point) or environmental variables (e.g., arrival rate, demand). The sensitivity to decision variables is required for finding optimal or sufficient solutions. The sensitivity to the environmental variables is needed to determine the validity range of the proposed solution. If the solution is very sensitive to the values of the environmental variables, then good estimates of these variables are needed, or suboptimal but more robust solutions (i.e., decision rules) may be suggested, etc.; see also Blanning (1974). As a “metamodel” to study these sensitivities, we propose the linear regression model.