Matching in observational research: from the directed acyclic graph perspective

Abstract

Matching is a standard method for selecting research objects regarding the observational research, which controls confounding factors and improves statistical efficiency. However, its role in controlling confounding is not consistent in different observational studies. Matching can eliminate the confounding bias of matching variables in cohort studies, but checking on itself cannot eliminate confounding bias in case-control studies. In matched case-control studies, researchers may not accurately judge whether the variable is a confounder. Sometimes the variables that are not confounders are mistakenly matched. In that case, it will result in overmatching, which will lead to the decline of statistical efficiency or the introduction of unavoidable bias or increase of workload. If the real confounding factors are omitted, it will cause confounding bias. Therefore, researchers should consider what kind of matching variable selection criteria should be formulated. A directed acyclic graph is a visual graphic language that can show the complicated causality among different epidemiological research designs. This article analyzes the role of Matching in different observational research designs from the perspective of the directed acyclic graph, formulates the selection criteria for matching variables in matched case-control studies, and provides some reference suggestions for future epidemiological research design.