You are here:
A Bayesian network approach for estimating distribution parameters with detected and non-detected environmental concentration data
Citation:
Carriger, John F., C. Acheson, D. Kleinmaier, AND R. Herrmann. A Bayesian network approach for estimating distribution parameters with detected and non-detected environmental concentration data. SETAC North America 40th Annual Meeting, TorontoC, November 03 - 07, 2019.
Impact/Purpose:
Present a novel approach for creating probability distributions with environmental concentration data that includes detects and non-detects.
Description:
Environmental data often contain non-detects that can bias a data or exposure analysis if mishandled, such as by substituting artificial point values for the non-detected data. Bayesian analysis is infrequently used with environmental concentration or exposure distributions for handling non-detected data but can incorporate more of the uncertainties for data and parameters than conventional methods. An approach that relies on Bayesian networks for handling non-detects was developed to estimate the parameters of probability distributions built with environmental concentration data. Multiple reporting limits can be used to accommodate the non-detected data and uncertainty ranges can be input to detected data, if useful and appropriate. The approach requires identifying initial prior probability distributions for the distribution parameters and the ranges or point estimates for detected or non-detected data. The uncertainties from the data are propagated to the distributions for parameters using Bayes theorem. The output environmental concentration distribution from the posterior probabilities for the parameters then includes uncertainties from both the parameters and concentration data. We briefly demonstrate the creation of a distribution of environmental concentrations and discuss the application of the Bayesian network distributions in temporal Bayesian learning, Bayesian hypothesis testing, benchmark exceedence calculations, and causal risk-based modelling. The Bayesian network approach to constructing an environmental concentration distribution may be more labor-intensive than classical and other Bayesian approaches as it requires prior probability distributions for parameters and additional time to construct the graphical components of the model. But the Bayesian network approach potentially provides a rigorous and flexible way of including uncertainties from non-detected data and distribution parameters.