Abstract for: Using Synthetic Data to Compute Effective Sample Size for Improved Uncertainty Estimation of System Dynamics Models

The nonlinear, feedback-rich models used in system dynamics often make use of simplified likelihood functions that violate the basic assumptions required for maximum likelihood estimation and related uncertainty quantification techniques. Although the simplified likelihood is a reasonable approximation of the true likelihood surface, in many cases it can lead to overconfidence in parameter estimates. This work develops an approach for understanding and mitigating this overconfidence effect by using error correlations obtained from synthetic data. We present a methodology for utilizing synthetic data to compute effective sample size, which is subsequently used to rescale the simplified likelihood surface to more faithfully reflect the topology of the true payoff landscape.