USING DIRECTED ACYCLIC GRAPHS TO ASSESS COLLIDER AND CONFOUNDING BIAS IN REVIEWS OF OBSERVATIONAL STUDIES: AN ILLUSTRATIVE EXAMPLE USING THE IMPACT OF RESIDENTIAL EXPOSURE TO ANIMAL FEEDING OPERATIONS ON HUMAN HEALTH

Background Bias assessment in systematic reviews of observational studies commonly includes consideration about whether key confounders were controlled but rarely evaluates whether control strategies may have introduced bias through control of colliders or mediators. In primary observational research, directed acyclic graphs (DAGs) are used to explicitly represent hypothesized causal structures, identify variables required to block non-causal (biasing) pathways, and distinguish confounders from variables that should not be controlled for. How DAG-based reasoning can be systematically incorporated into bias assessment during evidence synthesis remains insufficiently explored. Objective The primary objective of this study was to demonstrate how an existing, population-level DAG can be used as a reference framework to assess potential confounding and control-related bias in a systematic review of observational studies. The secondary objective was to evaluate whether primary studies reported a rationale for covariate selection. Methods We examined a subset of observational studies identified through a living systematic review assessing community health effects associated with residential proximity to animal feeding operations (AFOs). Studies were eligible if they sought to estimate a causal effect. For studies examining lower respiratory outcomes, covariates reported by study authors were mapped onto nodes in a previously published population-level DAG. The DAG was then used to assess whether reported control strategies were theoretically sufficient to block non-causal pathways and whether control for colliders or other inappropriate variables may have occurred. We also reviewed each study for explicit justification of covariate selection. Results Among 36 studies identified in the parent review, 17 reported incident outcomes and met the inclusion criteria. None of the included studies reported using a DAG or an explicit causal framework to guide covariate selection. Eight studies examined lower respiratory outcomes and could be evaluated using the reference DAG. In six of these studies, control for at least one variable identified as a collider in the reference DAG opened a non-causal pathway between exposure and outcome. In the remaining two studies, it was unclear whether potential confounding pathways, particularly those related to socioeconomic factors, were addressed through study design or population characteristics, precluding firm conclusions. Conclusions This study illustrates how DAG-based reasoning, using a pre-existing population-level causal diagram, can complement conventional risk-of-bias approaches in systematic reviews by making assumptions about confounding and inappropriate control explicit. Incorporating DAGs into evidence synthesis may improve transparency in bias assessment and support more interpretable causal reasoning in reviews of observational studies.