A procedure for assessing uncertainty in models

This paper discusses uncertainty in the output calculation of a model due to uncertainty in inputs values. Uncertainty in input values, characterized by suitable probability distributions, propagates through the model to a probability distribution of an output. Our objective in studying uncertainty is to identify a subset of inputs as being important in the sense that fixing them greatly reduces the uncertainty, or variability, in the output. The procedures we propose are demonstrated with an application of the model called MELCOR Accident Consequence Code System (MACCS), described in Helton et al. (1992). The purpose of MACCS is to simulate the impact of severe accidents at nuclear power plants on the surrounding environment. In any particular application of MACCS there are likely to be many possible inputs and outputs of interest. In this paper, attention focuses on a single output and 36 inputs. Our objective is to determine a subset of the 36 model inputs that can be said to be dominant, or important, in the sense that they are the principal contributors to uncertainty in the output.