Background: when randomized studies are unavailable, non-randomized studies – despite the potential for confounding – represent the best available evidence. The presence of a dose-response gradient has long been recognized as an important criterion for evaluating a putative causal relationship. GRADE suggests rating up certainty of evidence if a dose-response gradient is present. However, a dose-response effect may be less credible in scenarios in which the exposure is highly correlated with other potentially confounding factors.
Objectives: we introduce a novel approach to evaluate the plausibility of causal inferences drawn from non-randomized studies. We implemented this approach to inform whether to rate up the certainty of evidence for dose-response in a systematic review addressing the association between red meat intake and cardiovascular and cancer outcomes.
Methods: we conducted a systematic review of cohort studies and found a dose-response association between red meat intake and adverse health outcomes. Because consumption of red meat is highly correlated with other dietary characteristics, we undertook an additional systematic review to compare health outcomes associated with adherence to dietary patterns lower versus higher in red meat. We anticipated that if red meat was indeed a primary causal agent, the observed association between red meat and health outcomes would be greater in studies directly addressing red meat versus studies in which red meat was only one component of a dietary pattern. Conversely, if effect estimates for red meat were of similar or smaller magnitude compared to those for dietary patterns, the association between red meat and health outcomes may be explained by confounding.
Results: we found effect estimates for red meat to be smaller than those from dietary patterns, suggesting that the association between diet and health outcomes is not primarily driven by red meat. Given the potential for residual confounding, in our application of GRADE, we decided to not rate up the evidence for dose-response.
Conclusions: comparing the magnitude of association of indices composed of highly correlated and potentially confounding exposures with the outcome of interest to the magnitude of association between the exposure and the outcome directly may be useful to evaluate the plausibility of causal relationships.
Patient or healthcare consumer involvement: members of the public were involved in deciding outcomes of interest for our systematic review. We propose new methods to evaluate plausibility of causal relationships from non-randomized data. This may improve quality of inferences in areas where randomized studies are not feasible, which may improve health outcomes for the public.