Abstract
The causal inference literature has provided a clear formal definition of confounding expressed in terms of counterfactual independence. However, it has not produced a clear formal definition of a confounder, giving priority to the concept of confounding over that of a confounder. This paper considers several candidate definitions and evaluates them based on whether control for all "confounders" suffices to control for "confounding" and whether each confounder helps eliminate or reduce bias. The authors propose that a "confounder" be defined as a pre-exposure covariate $C$ that is a member of a minimally sufficient adjustment set. Variables that reduce bias but do not meet this criteria are termed "surrogate confounders."
Introduction
In the study of causal effects, researchers aim to isolate the impact of an exposure ($A$) on an outcome ($Y$). This task is frequently complicated by "confounding"—a phenomenon where the observed association between $A$ and $Y$ is distorted by other factors. While the causal inference community has established a rigorous definition for the absence of confounding (typically defined as the independence of counterfactual outcomes and the exposure), the definition of the individual variables we call "confounders" has remained surprisingly informal and often contradictory.
This lack of formalization is not merely a semantic hurdle; it has practical implications for how researchers select variables for adjustment. Without a precise definition, practitioners may inadvertently adjust for variables that introduce new biases, such as M-bias or collider-stratification bias.
The Search for a Formal Definition
The authors evaluate six candidate definitions for a "confounder" using the counterfactual (potential outcomes) framework and causal directed acyclic graphs (DAGs). To determine the utility of these definitions, they test them against two "common-sense" properties:
- Property 1 (Sufficiency): If a researcher controls for all variables meeting the definition, confounding is eliminated.
- Property 2A (Necessity/Utility): Every variable meeting the definition must actually help eliminate bias in at least one causal context.
Candidate Definitions Evaluated
The paper examines several common approaches found in the literature:
- The Traditional Definition: A variable associated with both the exposure and the outcome.
- The Backdoor Path Blocker: A variable that lies on a "backdoor path" between $A$ and $Y$.
- The Necessary Variable: A member of every minimally sufficient adjustment set.
- The Minimal Member (Proposed): A member of some minimally sufficient adjustment set.
- The Bias Reducer: A variable whose control reduces bias (often scale-dependent).
- The Change-in-Estimate Variable: A variable that, when adjusted for, changes the effect estimate (collapsibility-based).
Key Insights: Why Most Definitions Fail
Through mathematical proofs (Propositions 1–7), the authors demonstrate that most of these definitions fail to satisfy the two essential properties across all possible causal diagrams.
The Traditional Definition is perhaps the most problematic. In "M-bias" scenarios, a variable can be statistically associated with both the exposure and the outcome, yet adjusting for it creates a non-causal association where none existed before. This violates Property 2A, as the variable does not help eliminate bias; it creates it.
Similarly, definitions based on "Change-in-Estimate" or "Bias Reduction" are flawed because they are scale-dependent. A variable might appear to be a confounder when looking at the risk difference but not when looking at the risk ratio. The authors argue that being a "confounder" should be a structural property of the causal system, not an artifact of the chosen statistical scale.
The Proposed Definition: The Minimal Member
The authors conclude that Definition 4 is the only candidate that consistently satisfies both Property 1 and Property 2A. Under this definition, a confounder is a pre-exposure covariate $C$ for which there exists a set of other covariates $X$ such that the effect of $A$ on $Y$ is unconfounded conditional on $(X, C)$, but is not unconfounded given any proper subset of $(X, C)$.
In simpler terms, a confounder is a necessary component of at least one valid strategy to block all backdoor paths between the exposure and the outcome. This definition provides a rigorous mathematical bridge between the informal language used by practitioners and formal counterfactual theory.
Surrogate Confounders
A significant contribution of this paper is the distinction between "confounders" and "surrogate confounders." A surrogate confounder is a variable that helps reduce bias—perhaps because it is a proxy for an unobserved confounder—but does not meet the formal requirement of being part of a minimally sufficient set. While these variables are useful in practice, the authors argue they should be categorized separately to maintain the structural integrity of the definition of a confounder.
Conclusion and Practical Implications
The primary takeaway for researchers is that identifying confounders requires more than just statistical testing; it requires subject-matter knowledge represented through causal structures. Because it is generally impossible to verify if a variable is a confounder using data alone, researchers must rely on DAGs to identify the minimally sufficient adjustment sets. By formalizing the definition of a confounder, this paper clarifies how we should identify and talk about the variables that threaten the validity of causal claims.
Summary updated