Using simulations to explore the conditions under which "true" dose-response relationships are detectable for environmental exposures: polychlorinated biphenyls (PCBs) and birthweight: a case study.

Abstract

In environmental epidemiology, we use systematic reviews to evaluate the evidence of exposure-outcome relationships with an eye towards regulation. Conflicting results across studies thwart consensus on toxicity. In humans, only observational data is available from studies of environmental exposures, complicating the construction of dose-response relationships across the full range of exposure levels. Individual studies often lack the complete range of exposure levels because environmental exposure levels are tied to study settings. Pooling data across populations seems a natural solution, but strong population-dependent confounding may bias dose-response curves. Using the oft-debated association of polychlorinated bi-phenyls and birthweight as a case study, we describe simulations used to investigate the relative impacts of exposure range-dependent power limitations and confounding on our ability to correctly identify an assumed linear dose-response curve across a representative exposure range. While varying levels of confounding minimally biased estimates in our pooled and meta-analyses, we report very low confidence to ascertain a set underlying dose-response relationship in low-exposure cohorts with a narrow exposure distribution, but high ability in high-exposure cohorts with wide exposure distributions. Our simulations suggest that pooling and meta-analysis should be prioritized despite possible differences in confounding structures, particularly when exposure distributions in individual cohorts are limited.