Interpreting A Treatment Coefficient After Covariate Adjustment

What Does an Adjusted Treatment Coefficient Compare?

Suppose treatment is given at two visits. Let $A_0$ denote treatment at the first visit, $L_1$ a covariate measured later, $A_1$ treatment at the second visit, and $Y$ the outcome. A linear regression of $Y$ on $A_0$, $L_1$, and $A_1$ adjusts the comparison involving $A_1$ for the other regressors. The interpretation is conditional: it compares observations with the same values of $A_0$ and $L_1$, but different values of $A_1$, under the regression model.

That description is not, by itself, a causal claim. In particular, $L_1$ may be affected by $A_0$, and may also influence the later treatment $A_1$ and outcome $Y$. Holding $L_1$ fixed describes a comparison among observations with the same measured covariate value; it does not automatically describe what would happen if treatment were assigned by intervention.

Definition: In the additive linear conditional-mean model $$ E[Y\mid A_0,L_1,A_1] =\beta_0+\beta_0^{(A)}A_0+\beta_1^{(L)}L_1+\beta_1^{(A)}A_1, $$ the coefficient $\beta_1^{(A)}$ is the modeled conditional mean difference in $Y$ for a one-unit increase in $A_1$, holding $A_0$ and $L_1$ fixed. For binary $A_1$, it is the modeled difference between $A_1=1$ and $A_1=0$, at fixed $A_0,L_1$. This interpretation requires the stated conditional-mean model; a coefficient in a fitted regression is an estimate of its parameter.

A Conditional-Mean Model Gives the Interpretation

The distinction between an outcome equation and a conditional-mean model matters. Writing $Y$ as a sum of terms plus an error does not alone establish that the coefficient of $A_1$ is a regression coefficient. To make that conclusion, the error must have conditional mean zero given the regressors, or an equivalent model for the conditional mean must be specified.

For example, suppose

Y=10+2A_0+4L_1+3A_1+\varepsilon, \qquad E[\varepsilon\mid A_0,L_1,A_1]=0.

Taking conditional expectations gives $E[Y\mid A_0,L_1,A_1]=10+2A_0+4L_1+3A_1$. Thus, at any fixed pair $(A_0,L_1)$, the conditional mean at $A_1=1$ exceeds the conditional mean at $A_1=0$ by 3. This is an exact model statement, not a claim that the data necessarily contain both treatment values at every covariate pattern, nor that the contrast is causal.

Worked Example: Holding the Covariate History Fixed

Suppose the conditional mean model above applies, and compare two patients with $A_0=1$ and $L_1=6$, one with $A_1=0$ and one with $A_1=1$. The model predicts

For $A_1=0$, $E[Y\mid 1,6,0]=10+2(1)+4(6)+3(0)=36$.

For $A_1=1$, $E[Y\mid 1,6,1]=10+2(1)+4(6)+3(1)=39$.

The conditional difference is $39-36=3$, which agrees with the $A_1$ coefficient. The comparison holds $A_0$ and $L_1$ fixed. It does not compare treatment histories that may produce different distributions of $L_1$, and it does not by itself estimate the effect of assigning a treatment regime.

The calculation extends directly to any fixed values $a_0,l_1$. Subtracting the two modeled conditional means gives

\begin{aligned} &E[Y\mid A_0=a_0,L_1=l_1,A_1=1] -E[Y\mid A_0=a_0,L_1=l_1,A_1=0]\\ &\qquad = (\beta_0+\beta_0^{(A)}a_0+\beta_1^{(L)}l_1+\beta_1^{(A)}) -(\beta_0+\beta_0^{(A)}a_0+\beta_1^{(L)}l_1) =\beta_1^{(A)}. \end{aligned}

This derivation also exposes an assumption built into the additive model: the conditional difference is the same for every $a_0,l_1$. If treatment effects vary with covariate values, this single common-difference form may not describe the conditional means. The coefficient then need not represent every stratum's difference. The interpretation above is therefore tied to the model form, not just to the fact that a regression was run.

Adjustment Can Change the Association

A coefficient from an adjusted regression need not equal the unadjusted difference in observed outcomes. The following calculation isolates the role of $L_1$. Consider a slice of the population with $A_0=0$. Let $L_1$ and $A_1$ be binary, with $P(L_1=1)=P(L_1=0)=0.5$, and let

P(A_1=1\mid L_1=1)=0.8,\qquad P(A_1=1\mid L_1=0)=0.2, \qquad E[Y\mid L_1,A_1]=5+4L_1+2A_1.

Assume all four $(L_1,A_1)$ combinations have positive probability and the conditional mean model is correct. Within either $L_1$ stratum, the modeled treatment difference is 2. But treatment groups have different covariate distributions, so their unadjusted means need not differ by 2.

Worked Example: Adjusted and Unadjusted Differences

First calculate the joint probabilities. For treated observations, $P(L_1=1,A_1=1)=0.5(0.8)=0.4$ and $P(L_1=0,A_1=1)=0.5(0.2)=0.1$. Hence $P(A_1=1)=0.5$, and $P(L_1=1\mid A_1=1)=0.4/0.5=0.8$.

For untreated observations, $P(L_1=1,A_1=0)=0.5(0.2)=0.1$ and $P(L_1=0,A_1=0)=0.5(0.8)=0.4$. Thus $P(A_1=0)=0.5$, and $P(L_1=1\mid A_1=0)=0.1/0.5=0.2$.

The treated mean is $E[Y\mid A_1=1]=5+4(0.8)+2=10.2$. The untreated mean is $E[Y\mid A_1=0]=5+4(0.2)=5.8$. Their unadjusted difference is $10.2-5.8=4.4$, not 2. The extra 2.4 arises because the treated group contains more observations with $L_1=1$, whose conditional outcome mean is higher by 4.

After adjustment for $L_1$, the conditional means are $5+4L_1$ for $A_1=0$ and $7+4L_1$ for $A_1=1$. Their difference is 2 at both $L_1=0$ and $L_1=1$. The adjusted coefficient therefore represents the common conditional difference in this model, while the crude difference mixes that treatment contrast with the groups' different covariate distributions.

When Is the Regression Coefficient Actually the Conditional Difference?

Two facts justify interpreting an ordinary least-squares coefficient as the parameter in a specified conditional-mean model. The first is the subtraction above: the model's $A_1$ coefficient equals the conditional difference if the model is additive and the conditional mean is correctly specified. The second is that, under standard moment and rank conditions, the population least-squares projection recovers the coefficients of that conditional mean.

Result: Let $X=(1,A_0,L_1,A_1)^\mathsf{T}$, assume $E[\|X\|^2]<\infty$ and $E[Y^2]<\infty$, and suppose $E[Y\mid X]=X^\mathsf{T}\beta$. If $E[XX^\mathsf{T}]$ is positive definite, the unique population least-squares coefficient vector is $\beta$.

Proof. Define the residual $R=Y-X^\mathsf{T}\beta$. The conditional-mean assumption gives $E[R\mid X]=0$. Therefore

E[XR]=E\{X E[R\mid X]\}=0.

The population least-squares objective at a candidate vector $b$ is $Q(b)=E[(Y-X^\mathsf{T}b)^2]$. Its normal equations are $E[XX^\mathsf{T}]b=E[XY]$. Since $Y=X^\mathsf{T}\beta+R$, we have $E[XY]=E[XX^\mathsf{T}]\beta+E[XR]=E[XX^\mathsf{T}]\beta$. Positive definiteness makes $E[XX^\mathsf{T}]$ invertible, so the unique solution is $b=\beta$. The finite second-moment assumptions ensure these expectations exist. $\square$

This result concerns the population linear regression. In a finite sample, an estimated coefficient varies from sample to sample. If the conditional mean is not linear in the included regressors, least squares instead estimates a best linear projection; its coefficient can still be calculated, but it does not generally equal a common conditional treatment difference.

Worked Example: Why an Error Condition Cannot Be Skipped

Consider the equation $Y=5+2L_1+3A_1+\varepsilon$, but suppose $\varepsilon=A_1$. The equation is true algebraically, and its displayed coefficient on $A_1$ is 3. However, substituting the error gives $Y=5+2L_1+4A_1$, so

$E[Y\mid L_1,A_1]=5+2L_1+4A_1$, and the conditional difference between $A_1=1$ and $A_1=0$, holding $L_1$ fixed, is 4, not 3.

The failure is $E[\varepsilon\mid L_1,A_1]=A_1\ne 0$. Thus the coefficient written in an equation is not necessarily the coefficient of the conditional regression. To infer the latter from an error-form equation, a condition such as $E[\varepsilon\mid A_0,L_1,A_1]=0$ is needed. Even that condition establishes a statistical conditional-mean interpretation, not automatically a causal interpretation.

Conditional Comparison Is Not Automatically Causal

In longitudinal settings, $L_1$ is measured after $A_0$. Conditioning on it changes the comparison: patients with the same $L_1$ may have reached that value under different earlier treatment histories. If $L_1$ is affected by $A_0$, an $A_1$ coefficient adjusted for $L_1$ is not, merely by its definition, the total effect of a treatment history. The earlier material on treatment-confounder feedback explains why $L_1$ can be both a predictor of later treatment and part of an earlier treatment's pathway to the outcome.

A causal reading of the conditional contrast requires additional assumptions and a clearly defined estimand. For example, a causal contrast of $A_1$ within observed history $(A_0,L_1)$ would require consistency, positivity of both $A_1$ choices within the relevant histories, and conditional exchangeability of $A_1$ with the potential outcomes given that history. Correct outcome-model specification is also needed to use this regression coefficient as that contrast. These requirements are distinct from the algebraic interpretation of the coefficient.

The alternatives answer different questions. A crude comparison contrasts observed treatment groups with their naturally occurring covariate distributions. An adjusted regression coefficient describes a conditional comparison under its model. A standardized marginal comparison averages conditional predictions over a specified covariate distribution, so its value depends on that target distribution and is not generally just the conditional coefficient. None should be substituted for another without stating the intended comparison.

Diagnostics and Common Pitfalls

Check support. If one treatment value never occurs for a relevant $(A_0,L_1)$ history, the data do not directly support that within-history comparison. A regression may still return a number by extrapolation.
Check the conditional-mean form. Residual patterns against $A_1$, $L_1$, or fitted values can indicate that the additive linear model is inadequate. A common coefficient can conceal differing conditional contrasts.
Do not equate an equation coefficient with a regression coefficient. Verify the conditional-mean-zero error condition or state the conditional-mean model directly.
Do not call adjustment a causal guarantee. Covariate adjustment describes what the model conditions on. It does not establish exchangeability, positivity, or correct specification.
Keep the comparison explicit. Say “the modeled difference in mean outcome for different $A_1$, at fixed $A_0,L_1$,” rather than simply “the effect of treatment.”

The central point is precise but limited: in a correctly specified additive conditional-mean model, the $A_1$ coefficient is a common conditional mean contrast. Whether that contrast answers a causal question depends on the treatment assignment process, the covariate history, and the target being asked about.

Check Your Understanding

Answer each question, then compare with the solutions below.

In an additive conditional-mean model, what values are held fixed when interpreting the coefficient of $A_1$ in a regression on $A_0,L_1,A_1$?
Suppose $E[Y\mid A_0,L_1,A_1]=8+A_0+2L_1+5A_1$. What is the modeled conditional mean difference for changing binary $A_1$ from 0 to 1?
In the numerical example, why was the unadjusted difference 4.4 although the within-$L_1$ difference was 2?
What conditional-mean assumption is sufficient to interpret the coefficient in an error-form outcome equation as the regression coefficient?
Name two reasons why an adjusted treatment coefficient should not automatically be called a causal effect.

Solutions. (1) Hold $A_0$ and $L_1$ fixed. (2) The difference is 5, by subtracting the conditional means at $A_1=0$ and $A_1=1$. (3) Treated observations had $P(L_1=1\mid A_1=1)=0.8$, compared with 0.2 among untreated observations, and $L_1=1$ raised the conditional mean by 4; this contributes $4(0.8-0.2)=2.4$, so $2+2.4=4.4$. (4) A sufficient condition is $E[\varepsilon\mid A_0,L_1,A_1]=0$. (5) For example, exchangeability and positivity may fail, or the conditional-mean model may be misspecified; moreover, conditioning on a post-$A_0$ covariate does not by itself identify a total effect of a treatment history.