Confounder, Mediator, or Prognostic Variable

Why Covariate Roles Matter

A regression model can include many variables that help describe or predict an outcome, but that does not make them interchangeable. When the aim is to understand the association between an exposure and an outcome, a covariate’s causal role matters: it may be a confounder, an intermediate step in a causal pathway, or simply a predictor of the outcome. Those roles are defined relative to a particular exposure and outcome, not by a variable’s name or by how strongly it predicts.

Consider a question about whether a treatment is associated with recovery time. Age, treatment adherence, and a baseline laboratory measurement might all help predict recovery time. Yet age could influence both treatment choice and recovery; adherence could occur after treatment and lie on a pathway to recovery; and the laboratory measurement could predict recovery without affecting treatment choice. Treating all three as equivalent “controls” would ignore important differences in their timing and causal relationships.

Definitions Based on a Causal Diagram

A directed acyclic graph (DAG) represents hypothesized causal relationships using variables as nodes and arrows as causal directions. A directed path follows the direction of its arrows. For an exposure \(X\) and outcome \(Y\), a backdoor path is a path from \(X\) to \(Y\) that begins with an arrow pointing into \(X\). Earlier material introduced how such paths can create confounding and how adjustment sets are assessed. Here, the focus is on identifying the roles of individual variables.

These labels depend on the question. A variable may be prognostic and also a confounder: a common cause of exposure and outcome will often predict the outcome. “Prognostic” does not mean “not a confounder.” The important distinction is that predicting \(Y\) alone is not enough to establish that a variable causes, or is associated through a causal path with, \(X\).

Timing helps organize the possibilities. A variable measured before the exposure cannot be a mediator of that exposure if it truly precedes it, but it is not automatically a confounder. It might predict the outcome without influencing exposure. A variable measured after exposure could be a mediator, but timing alone does not prove that it is: subject-matter knowledge is needed to justify the arrows. Measurement dates also need to reflect when the underlying processes occurred, not merely when records were entered.

Regression Prediction Is Not a Causal Diagram

For a numerical outcome, a regression model with exposure \(X\) and covariate \(Z\) might describe the conditional mean as follows:

This equation says how the model represents the mean outcome at given values of \(X_i\) and \(Z_i\). It does not, by itself, establish arrows between the variables. For the equation to serve as an adequate conditional-mean summary, its functional form must be reasonable for the variables being modeled; interpreting a coefficient as causal would require additional causal assumptions and justification. In particular, a nonzero \(\beta_Z\) can reflect a predictive relationship with \(Y\) without \(Z\) being a cause of \(X\) or a source of a backdoor path.

The same caution applies to observed association. If \(Z\) and \(X\) are associated in the data, that does not prove \(Z \rightarrow X\). The association might arise because of other causes, selection, or chance. A causal diagram should be based on timing and subject-matter knowledge, not drawn by turning every statistical association into an arrow.

A Prognostic Variable Need Not Confound

A baseline variable can be useful for predicting an outcome without being a cause of exposure. For example, a biomarker might help predict recovery, while treatment assignment is randomized and therefore unrelated to the biomarker. In a diagram with \(Z \rightarrow Y\) and a separate causal arrow \(X \rightarrow Y\), the biomarker is prognostic but does not form a backdoor path from \(X\) to \(Y\).

Recognizing a Mediator by Its Place in Time and Path

A mediator occurs after exposure in the causal sequence and is affected by the exposure. For instance, treatment may change adherence, which may then influence recovery. The arrows \(X \rightarrow M\) and \(M \rightarrow Y\) make adherence an intermediate variable on a pathway from treatment to recovery. Adherence may strongly predict recovery, but its defining feature as a mediator is its position on that pathway—not its predictive strength.

This classification requires care. A variable recorded after exposure is not necessarily a mediator; it might be unrelated to the exposure, or its causal relation to the outcome may be different from the one proposed. Conversely, an intermediate process can be measured later even if it began earlier. Establishing the role requires a defensible timeline and substantive knowledge of how the variables arise.

A Short Path-Based Justification

The distinction between a prognostic variable and a confounder follows directly from the diagram. A variable can predict \(Y\) because an arrow reaches \(Y\), while having no causal connection to \(X\). In that case, it does not supply the common-cause path \(X \leftarrow C \rightarrow Y\). By contrast, when the common-cause arrows are present, the path from \(X\) to \(Y\) starts with an arrow into \(X\), so it is a backdoor path.

These statements concern the proposed causal structure, not merely the observed regression output. If the structure is uncertain, classification should be reported as a reasoned hypothesis rather than as a fact established by a coefficient or a p-value.

How to Classify a Candidate Covariate

Common Mistakes and Why the Distinction Matters

• “It predicts the outcome, so it is a confounder.” This reasoning considers only the link to \(Y\). A confounder in the common-cause case must also have a causal role in exposure assignment; an outcome predictor alone may be prognostic only.
• “It was measured before treatment, so it is a confounder.” Prior timing rules out its being a mediator of that treatment, but does not show that it causes or is otherwise connected to treatment. It may be prognostic only, irrelevant to both, or a confounder.
• “It is associated with treatment, so it causes treatment.” Statistical association does not determine causal direction. A diagram needs evidence and subject-matter justification.
• “Every covariate has one exclusive label.” A common cause of exposure and outcome can also predict the outcome, so it can be both confounding and prognostic. Labels summarize different features of its role.
• “A post-exposure variable is always a mediator.” Being measured later does not establish that exposure caused the variable or that the variable affects the outcome. Those arrows need justification.

The practical value of the distinction is that covariate selection is not simply a search for variables with strong outcome associations. A variable’s causal position helps determine what an adjusted association represents. For any proposed adjustment, describe the target exposure–outcome question, explain the relevant arrows, and distinguish observed prediction from causal evidence. This makes regression interpretations more transparent without treating a fitted model as proof of a causal diagram.