Richard Haylock and Anthony O'Hagan
University of Nottingham
Publication details: In Statistics for the Environment 3, Pollution Assessment and Control, 109-128. V. Barnett and K. F. Turkman (eds.). Wiley: Chichester, 1997.
Complex mathematical models are used in many disciplines and notably in environmental modelling. They are often implemented in large computer programs that may take massive computing power to run. Typically, the necessary inputs of the model are not known exactly but are expressed via (often subjective) probability distributions. The user then wishes to know how this uncertainty about the inputs translates into uncertainty about the outputs. This problem is known as uncertainty analysis.
The usual approach is some form of Monte Carlo analysis, sampling values of the inputs, running the program for each and thereby obtaining a sample of outputs. But with large models it may not be possible to do enough runs to get an adequate sample. This paper will present a Bayesian approach to uncertainty analysis which can achieve much better results from dramatically fewer runs of the computer program. Examples are given in the field of radiological protection.