A Study on the Effects That An Absence of a Pure-Error Component has in Multipopulation Simulation Experiments

摘要

To facilitate the design and analysis of efficient simulation experiments, many authors have advocated the use of correlation induction strategies such as: (a) common random numbers and (b) antithetic variates, across the various runs that comprise the experiment. The objective of such designed experiments is often to estimate a general linear regression model on the basis of a quantitative response variable generated by the simulation model. This regression model is called the metamodel of the experiment. Along with methods of meta-model parameter estimation under these various designed experiments, statistical methods have been developed to perform appropriate tests of hypothesis as well as confidence interval construction. The validity of these methods is contingent upon the presence of a pure-error component (experimental error) in the response. The purpose of this paper is to examine the effects that the absence of a pure-error component in the response has on these statistical analysis procedures and to provide recommendations for ensuring the presence of pure error in the response. In particular, we investigate the effects that the absence of pure error in the response has on the Schruben-Margolin correlation-induction strategy and its recommended statistical analysis methods. We make recommendations for ensuring the presence of pure error in the response. We also provide an example that clearly illustrates these points. The results of this investigation can easily be extended to multipopulation simulation experiments conducted under other correlation induction strategies.