OBJECTIVES: Generating multivariate random variables is essential in disease simulation applications. In this study we examine the implications of alternative approaches to generate marginal distributions and correlation matrix on simulation outcomes. METHODS: We adopt three alternative methods including Cholesky Decomposition (CD), CD with conditional matching, and the NORmal-To-Anything (NORTA) method to generate a hypothetical simulation sample with arbitrary marginal distributions and correlation matrix. As the comparator, we also create an independent and identically distributed (iid) simulation sample. The samples are individually populated in a previously developed type 2 diabetes microsimulation model to predict the major clinical endpoints over 15 years. The endpoints include all-cause mortality, diabetes-related mortality, and major cardiovascular events. We examine the goodness of fit by total deviance, i.e., the aggregated values of the relative difference between the individual predictions with the endpoints observed in the actual data, in the overall and stratified samples. RESULTS: The results show that, the model predications deviate from the observed data with an iid sample. Over 15 years, the model over-predicts all the numbers of endpoint events by 20%, with the total deviance of 0.73, and the over-prediction is particularly more pronounced in the younger patients. With a sample of a constructed multivariate normal distribution using the CD and CD plus conditional matching approach, the deviance is reduced to 0.41 and 0.58 respectively. A further improvement is observed when using the NORTA method, with the deviance of the endpoints between model prediction and actual data further reduced to 0.11. The reduction was mainly contributed by better approximations in the dispersion of the risk factors among patients. CONCLUSIONS: Random sequences generation has important ramification for simulation outcomes. Poorly-defined multivariate distributions may significantly distort the simulation performance. Given its flexibility for both continuous and discrete variables, NORTA method appears to be a preferable approach.