Interpreting Coefficients After Post-Treatment Adjustment

Why the Adjustment Set Changes the Question

In longitudinal studies, a covariate measured after an earlier treatment can be both a consequence of that treatment and a predictor of a later treatment and outcome. Including it in a regression may help describe comparisons among people with the same measured covariate history. But it also changes what the treatment coefficient compares: the coefficient no longer includes outcome differences that operate through the covariate held fixed by the model.

This distinction matters because a regression coefficient is not automatically a total causal effect. Its statistical meaning is determined by the model and the variables held constant; its causal meaning additionally requires assumptions about how treatment and covariates were assigned. In particular, a coefficient after post-treatment adjustment generally answers a conditional question, not the question of what would happen if treatment changed and all its downstream consequences were allowed to unfold.

What a Treatment Coefficient Conditions On

Consider a regression with an additive conditional mean and no treatment-by-covariate interactions:

Here \(A\) is treatment, \(L\) is an included covariate, and \(Y\) is the outcome. Under this model, \(\beta_A\) is the difference in the model’s expected outcome for a one-unit increase in \(A\), holding \(L\) fixed. If \(A\) is binary, it is the modelled difference between \(A=1\) and \(A=0\) at the same value of \(L\). The units are outcome units per treatment unit. In a fitted regression, the estimated coefficient describes this adjusted association in the data, subject to the model specification.

The total effect instead compares outcomes under treatment strategies while allowing downstream variables to take the values they would take under each strategy. For a binary treatment at one time, its population average is \(E[Y(1)-Y(0)]\), where \(Y(a)\) denotes the outcome that would occur under treatment level \(a\). In a longitudinal setting, the strategies specify treatment at multiple times. A coefficient that holds a later covariate fixed does not generally make this total comparison.

Worked Example: A Treatment Effect Transmitted Through a Covariate

Why the Adjusted Coefficient Need Not Be the Total Effect

The example illustrates a general algebraic point. Suppose, for simplicity, that a treatment changes a later covariate by \(\alpha\) units per treatment unit, and that the covariate changes the outcome by \(\beta_L\) units per covariate unit. If the outcome equation also has a treatment term \(\beta_A A\), then substituting the covariate equation into the outcome equation adds the transmitted effect \(\alpha\beta_L\) to the treatment’s total effect.

The coefficient \(\beta_A\) is the contrast holding \(L\) fixed; the substituted coefficient \(\beta_A+\alpha\beta_L\) is the total effect in this simple linear model. This result relies on the equations being correctly specified, on the absence of relevant interactions, and on the stated error and treatment-assignment assumptions. It is not a universal rule for converting adjusted and unadjusted regression coefficients.

With time-varying treatment, the interpretation requires particular care. A later covariate may be affected by an earlier treatment while also predicting a subsequent treatment and the outcome. Adjusting for it conditions on a consequence of prior treatment, so the earlier treatment coefficient excludes pathways through that covariate. Further, the resulting conditional association need not be a clean causal direct effect: common causes of the covariate and outcome, or other consequences of treatment, can make such an interpretation invalid. A regression coefficient alone does not establish the assumptions needed for a causal reading.

Worked Example: A Two-Period Treatment History

Interactions Change the Conditional Contrast

A coefficient is not even a single common conditional contrast when the regression includes a treatment-by-covariate interaction. In that case the treatment contrast depends on the covariate value. The coefficient on treatment alone gives the contrast at the covariate’s reference value, not a universal effect across all covariate levels.

When Can a Causal Interpretation Be Defended?

A fitted coefficient always has a model-based statistical interpretation, but a causal interpretation requires more. At minimum, the treatment and covariate interventions must be well defined, the model must adequately describe the relevant conditional means, and the data must contain support for the treatment comparisons being made. Unmeasured common causes of treatment and outcome can undermine a causal interpretation. With a post-treatment covariate, additional assumptions are needed to justify conditioning on that covariate for the particular causal contrast of interest.

Even when a conditional coefficient does have a causal interpretation under a carefully specified model and assumptions, it is a conditional effect: it describes a comparison with the covariate fixed. That is a different target from a total effect that allows the covariate to respond to treatment. Naming the target before interpreting the coefficient helps prevent these two questions from being conflated.

Common Mistake

Incorrect reasoning: “The regression includes the later covariate, so the treatment coefficient controls confounding and is the treatment’s effect on the outcome.”

Correction: The coefficient compares treatment values conditional on the included covariate values. If the covariate was affected by earlier treatment, that comparison holds part of the treatment history fixed and generally leaves out pathways through it. Whether the coefficient has any causal interpretation also depends on the design, assumptions, and model; inclusion in a regression is not, by itself, proof of confounding control.