CARWH 2016: Advancing Research to Improve Work and Health October 16-18, 2016 Toronto
Français
Strategies to interpret exposure measurement data have evolved into a framework based on a lognormal distribution. Most currently available tools and approaches are based on traditional frequentist statistical procedures, which have been unable to satisfactorily address several methodological challenges. Bayesian methods, which combine prior information with that in the current data set to derive overall inferences, have been proposed but some important extensions remain to be developed, programmed and made more widely available.
We developed a web-based tool aimed at answering the following types of inferential questions: Is a group compliant with the OEL? Are individuals within this group likely to experience higher risk than the group average? What effect does an intervention have on the underlying exposure distribution? These questions were solved numerically using Bayesian models that integrate the treatment of left-censored data with multiple censoring points.
The Web-based tool is available at www.expostats.ca. In the Bayesian models, variance components are estimated through a hierarchical model with individual workers as the grouping variable, and censoring is treated via multiple imputation by declaring the observations missing from a truncated part of the underlying exposure distribution. The models are estimated using the JAGS Gibbs sampler called through the R statistical package, both installed on a Linux machine with the Shiny Web Server.
Future developments include translating current routines into pure internet language code to facilitate its diffusion to the industrial hygiene community and extensive comparison with their frequentist counterparts.