Page 434 - Contributed Paper Session (CPS) - Volume 2
P. 434
CPS1916 Erica P. et al.
Figure 1: Regression coefficients with and without modelling the error
component in JAGS. A classical and a Berkson ME assumed
Multivariate models analyses are ongoing.
4. Discussion and Conclusion
We implemented Bayesian hierarchical models to account for the presence
of error in measurements of traffic related air pollution. The Bayesian
formulation allows to model several dependency structures in a very flexible
way, as well as to include an additional component for measurement error. In
our work, we applied such methodology to the study of how TRAP
measurements are associated with high-throughput molecular data, namely
metabolic features sampled from the exposed individuals in a randomized
crossover trial. Our application to the Oxford Street II study showed that the
inclusion of a classical error term in the models resulted in corrections of the
regression estimates whose extent and direction was not clear a priori, which
underlines the importance of explicitly modelling the error component rather
than predicting its effect based on prior beliefs. On the other hand, it
confirmed our expectations that including a Berkson measurement error does
not change the estimates quantitatively.
The explicit formulation of such models was possible thanks to the flexible
structure of Bayesian hierarchical models, and it is relatively straightforward to
embed dependency and measurement error correction in the same
hierarchical structure.
423 | I S I W S C 2 0 1 9