Using directed acyclic graphs to assess open biasing pathways in reviews of observational studies: an example using the impact of residential exposure to animal feeding operations on human health

Background Bias assessment in systematic reviews of observational studies relies on identifying variables that should be controlled as confounders. However, there are no formal approaches to ensuring unnecessary control of colliders is assessed in systematic reviews. In primary observational research, directed acyclic graphs (DAGs) help researchers visualize a hypothesized causal framework for an exposure-outcome relationship and identify variables that must be controlled to estimate the causal effect of the exposure on the outcome, and non-confounding variables, which, if controlled, could introduce bias. Objective The primary objective was to evaluate whether incorporating a DAG into the bias assessment enabled the assessment of open biasing pathways due to uncontrolled confounders or controlled collider variables. The secondary objective was the determine if authors provided the rationale for the variables controlled in the model and reported if the goal was to estimate the direct or total effect of the exposure on the outcome. Methods We examined observational studies investigating a relationship between residential proximity to AFOs and lower respiratory conditions in humans to identify studies that reported incident events and therefore assessed a causal association. We aim to map the variables controlled by the authors onto a DAG to identify any open biasing pathways. We also evaluated if authors reported the rationale for the inclusion of variables selected for control in statistical models. Results Of 36 studies investigating a relationship between residential proximity to AFOs and lower respiratory conditions in humans, 17 reported incident events and were eligible for this study. However, none of the observational studies reported using a DAG to select confounders or avoid control of colliders. We found 8 studies that reported lower respiratory disease. For those 8 studies, we used a previously published DAG to assess if open biasing pathways remained after control for the variables reported by the authors. In six of the eight studies, adjustment for a collider variable opened a biasing pathway. For two of the eight studies, it was unclear if the hypothesized biasing pathways associated with socioeconomic status were controlled by the study design and study population, and so no conclusion was reached. No authors reported whether the goal was to estimate the total or direct effect of exposure on the outcome. Conclusions Systematic reviewers should consider incorporating a DAG into the approach for assessment of bias in reviews of observational studies to ensure a comprehensive assessment of biasing pathways.