Background: Different observational studies often provide multivariate analyses that adjust for different variables. There is no current accepted strategy for proceeding with a meta-analysis when this occurs.
Objectives: To describe how the causal directed acyclic graph (DAG) approach to "confounding" can be used to help decide if observational studies addressing the same research question but using different covariates to adjust for confounding should, or should not be combined.
Methods: We will use the Structural Approach to Bias (based on causal DAGs) to explain how following fundamental principles lead to a coherent strategy for deciding whether two different statistical models are expected to lead to the same unbiased estimate of effect.
Results: Using basic epidemiological principles, we demonstrate that confounding requires an understanding of causal relationships between covariates rather than the traditionally cited associational relationships. Based on the possible underlying causal relationship, we show that one can determine if different statistical models are expected to yield the same estimate of effect (biased or unbiased).
Conclusions: The directed acyclic graph approach provides a logical strategy for deciding if results from observational studies with different statistical models can be combined in one meta-analysis.