Joint propagation of variability and imprecision in assessing the risk of groundwater contamination.

Estimating risks of groundwater contamination often require schemes for representing and propagating uncertainties relative to model input parameters. The most popular method is the Monte Carlo method whereby cumulative probability distributions are randomly sampled in an iterative fashion. The shortcoming of the approach, however, arises when probability distributions are arbitrarily selected in situations where available information is incomplete or imprecise. In such situations, alternative modes of information representation can be used, for example the nested intervals known as "possibility distributions". In practical situations of groundwater risk assessment, it is common that certain model parameters may be represented by single probability distributions (representing variability) because there are data to justify these distributions, while others are more faithfully represented by possibility distributions (representing imprecision) due to the partial nature of available information. This paper applies two recent methods, designed for the joint-propagation of variability and imprecision, to a groundwater contamination risk assessment. Results of the joint-propagation methods are compared to those obtained using both interval analysis and the Monte Carlo method with a hypothesis of stochastic independence between model parameters. The two joint-propagation methods provide results in the form of families of cumulative distributions of the probability of exceeding a certain value of groundwater concentration. These families are delimited by an upper cumulative distribution and a lower distribution respectively called Plausibility and Belief after evidence theory. Slight differences between the results of the two joint-propagation methods are explained by the different assumptions regarding parameter dependencies. Results highlight the point that non-conservative results may be obtained if single cumulative probability distributions are arbitrarily selected for model parameters in the face of imprecise information and the Monte Carlo method is used under the assumption of stochastic independence. The proposed joint-propagation methods provide upper and lower bounds for the probability of exceeding a tolerance threshold. As this may seem impractical in a risk-management context, it is proposed to introduce "a-posteriori subjectivity" (as opposed to the "a-priori subjectivity" introduced by the arbitrary selection of single probability distributions) by defining a single indicator of evidence as a weighted average of Plausibility and Belief, with weights to be defined according to the specific context.