Abstract
The causal inference literature has provided a clear formal definition of confounding expressed in terms of counterfactual independence. The literature has not, however, come to any consensus on a formal definition of a "confounder," as it has given priority to the concept of confounding over that of a confounder. We consider a number of candidate definitions arising from various informal statements made in the literature. We evaluate these candidates based on two properties: (i) whether control for all "confounders" suffices to control for "confounding" and (ii) whether each confounder in some context helps eliminate or reduce confounding bias. Only one candidate definition satisfies both properties. We propose that a "confounder" be defined as a pre-exposure covariate $C$ for which there exists a set of other covariates $X$ such that the effect of the exposure on the outcome is unconfounded conditional on $(X, C)$, but for no proper subset of $(X, C)$ is the effect unconfounded. We also provide a conditional analogue and propose the term "surrogate confounder" for variables that reduce but do not eliminate bias.
Introduction
In the study of causal inference, "confounding" is well-understood as the bias that arises when the observed association between an exposure and an outcome differs from the true causal effect due to the "mixing of effects." However, while the state of confounding is formally defined through counterfactuals, the identity of a "confounder"—the specific variable responsible for this state—has remained surprisingly informal.
Traditional definitions, such as a variable being associated with both the exposure and the outcome, are frequently used in epidemiology and statistics. Yet, these informal definitions often fail when applied to complex causal structures. This paper evaluates six candidate definitions to establish a rigorous foundation for what constitutes a confounder.
The Criteria for a Valid Definition
To evaluate the candidate definitions, the authors propose two fundamental properties that any robust definition of a confounder should satisfy:
- Property 1 (Sufficiency): If a researcher controls for the set of all variables meeting the definition of a confounder, all confounding bias should be eliminated.
- Property 2 (Utility): Every variable meeting the definition should, in at least some context, be necessary to help eliminate or reduce confounding bias.
Using the counterfactual (potential outcomes) framework and Directed Acyclic Graphs (DAGs), the authors test whether existing definitions hold up under these requirements.
The Failure of Traditional Definitions
The most common definition of a confounder—a variable associated with the exposure and the outcome (often with the caveat of not being on the causal path)—fails Property 2. The primary culprit is "M-bias." In an M-bias structure, a variable may be associated with both the exposure and the outcome through a "collider" path. Adjusting for such a variable actually introduces bias rather than removing it.
Similarly, defining a confounder simply as a variable residing on a "backdoor path" (Definition 2) fails to satisfy the criteria. In some scenarios, controlling for a variable on a backdoor path can increase bias if other variables on that same path are not also controlled, or if the variable is a collider.
The Proposed Solution: Definition 4
After evaluating the candidates, the authors identify only one definition that satisfies both Property 1 and Property 2. This definition, referred to as Definition 4, states:
A variable $C$ is a confounder if it is a member of some minimally sufficient adjustment set.
In this context, a set of variables is "sufficient" if controlling for them eliminates all confounding. It is "minimally" sufficient if removing any single variable from that set would render it no longer sufficient. By defining a confounder as a member of such a set, we ensure that the variable is essential for identification in at least one specific adjustment strategy.
Surrogate Confounders and Scale Dependence
The authors also introduce the concept of a surrogate confounder. This refers to a variable that is not strictly necessary for identification (it is not part of a minimally sufficient set) but is associated with a true confounder. Adjusting for a surrogate can reduce bias even if it cannot eliminate it entirely.
Furthermore, the paper addresses scale dependence. Some definitions of confounding (Definitions 5 and 6) depend on whether the effect is measured as a risk difference or an odds ratio. The authors argue that a fundamental definition of a confounder should ideally be scale-invariant, further supporting the adoption of Definition 4.
Conclusion and Limitations
Defining a confounder as a member of a minimally sufficient adjustment set bridges the gap between graphical approaches (DAGs) and counterfactual frameworks. However, this formalization comes with practical caveats:
- Un-testability: Whether a variable is a confounder cannot be determined by data alone; it requires subject-matter knowledge to construct the underlying causal model.
- Pre-exposure Focus: The current definitions primarily focus on pre-exposure covariates. While they can be extended to non-descendants of the exposure, complex time-varying confounding requires further refinement.
- Faithfulness: The proofs assume "faithfulness," meaning that the statistical independencies in the data accurately reflect the causal structure without perfectly canceling effects.
By providing a rigorous definition, this work allows researchers to move beyond intuitive heuristics and toward a more precise selection of covariates in causal analysis.
Summary updated