Bootstrapping for confidence interval estimation and hypothesis testing for parameters of system dynamics models

Methods for confidence interval estimation and hypothesis testing used in the literature and implemented in system dynamics software packages typically assume that the data are normally distributed, not autocorrelated and not heteroskedastic. In dynamic models these assumptions are often violated. Here we propose the bootstrapping method for confidence interval estimation and hypothesis testing in system dynamics models and provide an overview of the issues involved in applying bootstrapping properly in dynamic models. Bootstrapping is a widely used and robust method but its use in system dynamics models is rare. Bootstrapping can handle violations of the maintained assumptions of traditional methods. It is also valid for small samples whereas traditional methods are valid only asymptotically. Another advantage of bootstrapping is that it is a convenient tool for testing hypotheses including complicated functions of parameters (e.g., multiplication, division of parameters). We provide two examples for illustration, one in discrete time and one in continuous time, including experimental data from the beer distribution game and field data from a study of service quality in the banking industry. Copyright (c) 2007 John Wiley & Sons, Ltd.