To reduce the variance of metamodel coefficients, simulation analysts often advocate the use of correlation induction strategies. Under certain conditions, these strategies have been shown to reduce the variance of metamodel coefficients without producing significant bias in the coefficient estimates. Although these procedures are very useful for estimating metamodels, the application of many statistical analysis techniques is inappropriate unless the analyst is assured that a pure error component is present in the response. Crenshaw and Tew have demonstrated the absence of pure error in experiments in which all random number streams are used to induce correlations. Mihram argues that a pure error component can be maintained by selecting the seeds for at least one random component randomly and non-repetitively for all design points and replications. In this thesis, random components for which seeds are randomly and non-repetitively selected are referred to as pure error generators. This thesis examines the selection of pure error generators in the context of univariate response, replicated simulation experiments. To assess the impact of pure error generator selection, we give the results of an extensive series of Monte-Carlo experiments in which the Schruben-Margolin strategy is applied for each possible pure error generator in each of two simulation models. To determine causes for the differences in pure error generator performance, four pure error generator selection methods are outlined, tested, and compared to the results of the Monte-Carlo experiments. The results strongly suggest the importance of careful pure error generator selection and indicate that the primary difference in their performance is related to their ability to maintain the prescribed correlation structures of the correlation induction strategy. / Master of Science