An Adjusted Association Is Not Automatically an Effect

Adjustment Describes a Comparison, Not a Cause

Multiple regression lets us compare observations that have the same recorded values of other predictors. This can make an association more informative than a raw comparison, especially when the predictors are related to one another. But the phrase “holding the other predictors fixed” describes a comparison in the data. By itself, it does not tell us what would happen if we intervened and changed the focal predictor.

That distinction matters whenever a regression result is used to discuss a treatment, behavior, policy, or other potential cause. An adjusted coefficient can be useful evidence, but a causal conclusion requires a reason to believe that the comparison represents what would have happened under different values of the focal predictor—not merely that people with different observed predictor values had different outcomes.

What an Adjusted Coefficient Says

Suppose \(Y\) is an outcome, \(X\) is the focal predictor, and \(Z\) is a measured adjustment variable. An additive linear regression model for the conditional mean is

Here \(E(Y\mid X=x,Z=z)\) is the mean outcome among observations with \(X=x\) and \(Z=z\). The model assumes that this conditional mean follows the stated additive linear form. In this model, \(\beta_1\) is the change in the fitted conditional mean associated with a one-unit increase in \(X\), when \(Z\) is held fixed. That is a statement about the model’s conditional comparison. No intervention has yet been described.

Subtract the fitted conditional mean at \(X=x\) from the one at \(X=x+1\), keeping \(Z=z\):

The intercept and the contribution from \(Z\) cancel. This is why the coefficient is called an adjusted slope: it describes the model’s comparison at the same value of \(Z\). The derivation does not show that changing \(X\) would change \(Y\) by \(\beta_1\). The conditional means in the equation are calculated from observed cases; a causal effect concerns outcomes under a change made to \(X\).

What a Causal Effect Means

A causal question asks about outcomes under possible interventions. Let \(Y(x)\) denote the outcome a unit would have if its focal predictor were set to \(x\). For example, \(Y(4)\) could represent a student's exam score if the student were assigned four hours of study under a clearly specified study plan. A causal contrast might be the difference between average outcomes under two settings, such as \(E[Y(4)-Y(3)]\). These potential outcomes are not both observed for the same person: in reality, the person follows one study condition, not both.

To connect an adjusted observed-data comparison to a causal contrast, researchers need substantive assumptions and a design that makes those assumptions plausible. In a simple setting with adjustment variable \(Z\), important conditions include the following:

• Consistency and a well-defined intervention: the intervention represented by setting \(X=x\) has a clear meaning, and a unit observed with \(X=x\) has observed outcome \(Y=Y(x)\). Vague or meaningfully different versions of “one more hour of study” can make this hard to satisfy.
• Conditional exchangeability: within each relevant value of \(Z\), the groups with different \(X\) values are comparable in their potential outcomes. In notation, \(Y(x)\) is independent of \(X\) given \(Z\). This rules out remaining confounding after conditioning on \(Z\); it is an assumption, not something guaranteed by fitting a regression.
• Positivity: at the values of \(Z\) being studied, there is a genuine possibility of observing each relevant value or range of \(X\). If no comparable observations exist, a fixed-\(Z\) comparison must rely on extrapolation.
• Appropriate modeling and measurement: the regression form should represent the conditional mean adequately for the intended comparison, and the relevant variables should be measured well enough for the assumptions to be credible.

The first three conditions explain the causal logic; the last connects that logic to a particular fitted regression. Even if a regression estimates an adjusted association precisely, precision does not establish that these conditions hold.

Why the Assumptions Matter

The link between conditional association and causal effect can be stated directly. Under consistency, among units observed with \(X=x\), their observed outcomes equal their potential outcomes under \(x\). Under conditional exchangeability, conditioning on \(Z=z\) makes the potential outcomes comparable across the observed \(X\) groups. Therefore,

The first equality uses consistency; the second uses conditional exchangeability. Applying the same reasoning at \(x_0\) and \(x_1\), and subtracting, gives

Thus, under these assumptions, the observed conditional mean difference identifies a causal contrast for units with \(Z=z\). If the additive linear model is also suitable, its slope gives the model’s per-unit conditional contrast. Without the causal assumptions, the left side remains an observed-data comparison, while the right side is a causal quantity; the equality has not been established.

Adjustment Cannot Guarantee Comparability

In an observational study, the focal predictor is not randomly assigned. People with different values of \(X\) may differ in factors that also relate to \(Y\). Adjusting for recorded \(Z\) addresses differences represented by \(Z\), but it does not automatically address factors that were unmeasured, poorly measured, or omitted from the model. Consequently, a fitted adjusted slope can remain confounded.

Adjustment can also change the causal question if it includes a variable affected by \(X\). For instance, if a program changes later study habits and those habits affect exam score, controlling for later study habits compares program and control cases at the same value of a possible pathway through which the program works. That is not automatically the program’s total effect. Whether such adjustment is useful depends on the causal question and additional assumptions; “include every available predictor” is not a safe general rule.

Common Mistake: Treating “Adjusted For” as “Caused By”

Wrong reasoning: “The coefficient for study time is positive after adjusting for prior score, so increasing study time causes exam scores to rise.”

The adjustment establishes only that the model’s conditional comparison is positive at fixed prior score. It does not establish that prior score captures every relevant difference between students, that study time is measured without important error, or that changing study time corresponds to a well-defined intervention.

Correction: First describe the result as an adjusted association. Then assess whether the study design and assumptions support a causal interpretation. Random assignment can provide a strong basis for comparability; observational adjustment can also be informative, but only when the adjustment variables and assumptions are defensible. Neither a large coefficient nor a small p-value resolves a causal-design problem.