Classical Statistical Analysis of Simulation-Based Experimental Data

6 Classical Statistical Analysis of Simulation-Based Experimental Data 

Most researchers who have worked with discrete event simulation are familiar with classical statistical analysis. By classical, we mean those tests that deal with assessing differences in means or that perform correlation analysis. Included in these tests are statistic procedmes such as t-tests (paired and unpaired), analysis of variance (univariate and multivariate), factor analysis, linear regression (in its various forms ordinary least squares, LOGIT, PROBIT, and robust regression) and non-parametric tests. 

In most of these tests, the researcher creates an experiment, which is defined in terms of factors (main issnes of interest) and levels (specific values applied to the factors). The result is an experimental design consisting of cells or specific combinations of factors and levels. The interest of the researcher is in determining if there is a significant difference between cell means and if the factors associated with the cells have had a significant impact on the recorded means. 

This approach has several advantages. First, it is well known to most researchers (especially those working with simnlation stndies). These procedures are extensively discussed by Law and Kelton (2000) in their text book on computer simulation. Second, these procedures allow the researcher to assess not only main effects (are there differences in impact between various independent variables) but also interactions. Finally, these procedures are widely implemented in computer statistical packages such as Minitab , SPSS , SYSTAT , IMP , and STATA . 

However, classical statistical analysis, when applied to SCD, does suffer from an important limitation directly attributable to the transient response created by the supply chain dismption itself The dismption introduces a form of data variance. Traditionally, in the case of steady-state simulations, the method proposed for dealing with any transient variance is to delete the data associated with the transient from the dataset. In most cases, this is a reasonable approach since the transient data is typically generated by the simulated system getting to steady-state. 

---