|
Abstract:
|
A major activity in science and engineering is the development of simulation- or math-based computer models of processes. Such models are virtually always incomplete representations of reality and hence will be rejected by formal tests. The models will be used, however, so the statistical challenge is to understand how to do so effectively. Statistical issues that arise include determination of model bias, use of bias-adjusted predictions and accounting for uncertainty in model parameters and inputs. The methodology used is a mix of Bayesian spatial, hierarchical and nonparametric techniques; such techniques seem necessary, as will be shown with a simple example. Difficulties encountered include needing to deal with functional data, the presence of severe confounding and the surprising phenomenon that full Bayesian analysis is often inferior to a modular, or partial, Bayesian analysis.
|
