Browsing by Author "Hill, Graham"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Open Access
    Bayesian Calibration for Logit Model Microsimulations: Case for PECAS SD in San Diego
    (2017) Hill, Graham; Hunt, John Douglas; Kattan, Lina; Dann, Markus; Wang, Xin
    Bayesian inference is a versatile method for incorporating new information into a model while still respecting existing knowledge. One application of Bayesian inference is the calibration of models that are controlled by a large number of parameters, but where the data usable for calibration is incomplete or unreliable. Microsimulation models of urban development fit both of these criteria, but calibrating them is further complicated by their non-determinism. I investigated a calibration method called Bayesian Expected Value Calibration, which is designed to overcome non-determinism while incorporating existing knowledge, in the context of the PECAS Space Development model of the San Diego area. The test consisted of creating synthetic data using known behavioural parameters, calibrating the Space Development model to targets derived from the synthetic data and with priors reflecting imperfect existing knowledge, and assessing how closely the calibrated parameters matched the true values. I found that BEVC was generally effective at converging towards the true values of the parameters, and often received meaningful contributions from both the prior knowledge and the new observations under a range of plausible conditions. As would be expected from Bayesian theory, increasing the number of observations or the amount of useful prior knowledge improved the accuracy of the calibration. The method was robust under reasonable levels of human fallibility in creating the priors, and only suffered from significant loss of accuracy under extreme assumptions. However, more sophisticated methods of objectively determining the weights to assign to the data sources did not significantly improve calibration accuracy.