Baseline Adjustment In Randomized Trials

Why Adjust for Baseline Measurements?

In a randomized trial, treatment assignment is made by chance, so baseline characteristics are not deliberately chosen to favor one group. Even so, measurements such as a participant’s starting symptom score or blood pressure can help explain follow-up outcomes. A model that includes these baseline measurements can compare treatment groups at common baseline values, while retaining the randomized assignment as the treatment variable.

The central question is what that adjusted contrast means. Under an additive model with a common treatment effect across baseline values, the treatment coefficient is the estimated difference between assigned groups after adjustment. It also equals the difference in fitted mean outcomes when both groups are evaluated over the same baseline-covariate distribution. Randomization gives this contrast a clear target: the group difference due to assignment, rather than a comparison that conditions on a characteristic measured after treatment began.

Definition: Let $Y_i$ be the follow-up outcome for participant $i$, $T_i=1$ if the participant is assigned to treatment and $T_i=0$ if assigned to control, and $X_{ij}$ the participant’s $j$th baseline measurement. A common-slope baseline-adjusted model with $q$ baseline measurements is $$ Y_i=\beta_0+\tau T_i+\sum_{j=1}^{q}\beta_j(X_{ij}-c_j)+\varepsilon_i. $$ Here $c_j$ is a fixed centering value for baseline measurement $j$, $\beta_0$ is the modeled control mean at the centering values, $\beta_j$ is the outcome’s modeled change per unit of that baseline measurement, and $\tau$ is the common adjusted treatment contrast. The error $\varepsilon_i$ represents the difference between the observed outcome and its conditional mean under the model.

Centering changes how the intercept is read, but does not change the treatment contrast in this model. For instance, $c_j$ may be the mean baseline value in the trial. The model assumes a linear relationship between each baseline measurement and the conditional mean follow-up outcome, and that the treatment contrast is the same at every combination of baseline values. The latter is the common-treatment-effect assumption.

The outcome $Y_i$ and its units must be defined at follow-up. The variable $T_i$ represents randomized assignment, not necessarily the treatment a participant ultimately received. Thus the treatment contrast is an assigned-group contrast. Randomization supports a causal interpretation of that assignment contrast, subject to the trial’s design and conduct; the regression model describes how baseline adjustment is used to estimate it.

What the Adjusted Treatment Coefficient Targets

The model’s conditional mean for a participant assigned to control is $\beta_0+\sum_{j=1}^{q}\beta_j(X_{ij}-c_j)$. For a participant with the same baseline measurements assigned to treatment, it is this same expression plus $\tau$. Subtracting the two fitted means cancels the intercept and every baseline term:

\begin{aligned} E(Y_i\mid T_i=1,X_{i1},\ldots,X_{iq}) &=\beta_0+\tau+\sum_{j=1}^{q}\beta_j(X_{ij}-c_j),\\ E(Y_i\mid T_i=0,X_{i1},\ldots,X_{iq}) &=\beta_0+\sum_{j=1}^{q}\beta_j(X_{ij}-c_j),\\ E(Y_i\mid T_i=1,X_i)-E(Y_i\mid T_i=0,X_i)&=\tau. \end{aligned}

Here $X_i$ is shorthand for the full set of baseline measurements for participant $i$. The last equality is not a claim that the two treatment groups have identical observed outcomes. It is a statement about the two model-based conditional means at the same baseline values.

The same result holds when the fitted means are averaged over a shared distribution of baseline values. If $F_X$ denotes that distribution, then the model-based contrast is $\int [E(Y\mid T=1,X=x)-E(Y\mid T=0,X=x)]\,dF_X(x)$. Because the expression in brackets is $\tau$ for every $x$, the integral is $\tau$. The choice of shared distribution affects the two adjusted means separately, but under this common-effect model it does not change their difference.

This gives a short justification for why the treatment coefficient targets the randomized group difference. Write $Y_i(1)$ and $Y_i(0)$ for participant $i$’s potential follow-up outcomes under assignment to treatment and control. Under random assignment, assignment is independent of baseline measurements and potential outcomes. If the conditional mean difference is the same $\tau$ at all baseline values, then averaging that difference over the trial’s baseline distribution still gives $\tau$. In a sample, the fitted coefficient estimates this contrast; it need not equal the observed difference between the two raw sample means.

For standard regression-based inference, observations should also be independent across participants, the conditional mean model should be adequate, and the error variance assumptions used by the chosen standard-error procedure should be appropriate. Normal errors are commonly used for exact small-sample $t$- and $F$-inference. Randomization does not itself guarantee linearity, constant variance, normal errors, or a common treatment effect.

Worked Example: One Baseline Measurement Changes the Contrast

Suppose six participants in a hypothetical randomized trial have the following baseline scores and follow-up scores. Both scores are measured in points.

Assignment	Baseline score $X$	Follow-up score $Y$
Control	40	8
Control	50	10
Control	60	12
Treatment	50	12
Treatment	60	14
Treatment	70	16

The raw follow-up mean is $10$ points for control and $14$ points for treatment, so the unadjusted difference (treatment minus control) is $14-10=4$ points. But the treatment group’s average baseline score is also 10 points higher. Fit the common-slope model $Y_i=\beta_0+\tau T_i+\beta_1(X_i-50)+\varepsilon_i$.

Within each assignment group, an increase of 10 baseline points accompanies an increase of 2 follow-up points, giving a slope of $2/10=0.2$ follow-up points per baseline point. The two groups have the same slope, so the pooled fitted slope is also $0.2$. At baseline score 50, the control fitted mean is 10 points. At that same baseline value, the treatment fitted mean is $12$ points: its observed mean is 14, and adjustment back by 10 baseline points gives $14-(0.2)(60-50)=12$.

Thus the fitted treatment contrast is $$ \hat{\tau}=12-10=2\text{ points}. $$ For a direct check, the model with $\hat{\beta}_0=10$, $\hat{\tau}=2$, and $\hat{\beta}_1=0.2$ predicts control outcomes $8,10,12$ and treatment outcomes $12,14,16$, matching every listed value. In this constructed example the regression fit is exact; the adjusted contrast is 2 points, whereas the raw difference is 4 points. The arithmetic illustrates adjustment, not evidence that adjustment will always reduce a contrast or that baseline differences necessarily caused a raw difference.

The hypothetical design assigns participants at random and records the baseline score before treatment. The displayed outcomes follow a linear relationship with a shared slope and constant treatment contrast, so they are consistent with the specified mean model. With real data, those model features and the assumptions needed for inference would still require assessment.

Using Multiple Baseline Measurements

When several baseline measurements are relevant, include them together in the model. Each coefficient $\beta_j$ describes the modeled association between one baseline measurement and the follow-up outcome, holding the other included baseline measurements and assignment fixed. The coefficient $\tau$ remains the adjusted treatment contrast, provided the model has no treatment-by-baseline interaction.

The next example uses hypothetical fitted coefficients to show how the contrast is obtained. The specific coefficient values are stipulated as the result of fitting the stated model; the calculations interpret that fit rather than reconstruct it from participant-level data.

Worked Example: Two Baseline Measurements

A hypothetical trial randomizes 120 participants, 60 per group. The follow-up outcome is a symptom score in points. The baseline measurements are a symptom score $X_1$, in points, and age $X_2$, in years. Suppose model checks are consistent with a linear additive mean model and a common treatment contrast, and the fitted model is $$ \widehat{Y}=20-3.2T+0.60(X_1-30)-0.08(X_2-50). $$

The coefficient $-3.2$ is the adjusted treatment-minus-control contrast in symptom-score points. For example, at baseline symptom score 34 and age 60 years, the fitted control mean is $20+0.60(34-30)-0.08(60-50)=20+2.4-0.8=21.6$ points. The fitted treatment mean is $20-3.2+2.4-0.8=18.4$ points. Their difference is $18.4-21.6=-3.2$ points.

At baseline symptom score 25 and age 40 years, the same model gives control mean $20+0.60(25-30)-0.08(40-50)=20-3+0.8=17.8$ points and treatment mean $14.6$ points. Again, the difference is $14.6-17.8=-3.2$ points. Under this model, treatment assignment is associated with a 3.2-point lower adjusted mean symptom score at either baseline profile. The sign and units matter: a negative treatment-minus-control contrast indicates a lower modeled score in the treatment-assigned group.

The two fitted baseline slopes are not treatment effects. In particular, the age coefficient is a modeled change of $-0.08$ follow-up points per year, holding the other terms fixed. The treatment contrast comes from comparing fitted means that differ only in $T$, not from interpreting a baseline coefficient on its own.

Standardizing Over a Shared Baseline Distribution

A further way to express the adjusted comparison is to predict outcomes for both assignments at a set of baseline profiles, then average each set of predictions using the same weights. This makes the comparison’s reference distribution explicit. Under the additive common-effect model, the adjusted means can change with the chosen profiles, but the treatment-minus-control difference remains $\tau$.

Worked Example: Averaging Adjusted Means Over Two Profiles

Suppose a hypothetical fitted model for a follow-up score in points is $\widehat{Y}=50+4T+2(X_1-30)+(X_2-50)$, where $X_1$ is baseline score in points and $X_2$ is age in years. To illustrate standardization, use two baseline profiles with equal weights: $(X_1,X_2)=(30,50)$ and $(33,52)$.

At profile 1, the control prediction is $50+2(0)+0=50$ points and the treatment prediction is $50+4+2(0)+0=54$ points. At profile 2, the control prediction is $50+2(3)+2=58$ points and the treatment prediction is $50+4+2(3)+2=62$ points. Averaging with weights $0.5$ and $0.5$, the standardized control mean is $(0.5)(50)+(0.5)(58)=54$ points, and the standardized treatment mean is $(0.5)(54)+(0.5)(62)=58$ points. Their difference is $58-54=4$ points, exactly the fitted treatment coefficient.

The two profiles and equal weights are only an illustrative shared target distribution. A real analysis should identify the covariate distribution used to report adjusted means. In this particular common-effect model, changing that shared distribution changes the adjusted means but not their difference.

When the Interpretation Needs Care

The constant-contrast conclusion depends on the specified model. If the treatment effect varies with a baseline measurement, the model needs a treatment-by-baseline interaction to represent that variation. In an interaction model, a single treatment coefficient is the contrast at the chosen centering values; it is not automatically the average contrast over all baseline profiles. That distinction is important, but the examples above intentionally use the common-effect model.

A second point is that “adjusted” does not mean “the treatment coefficient must equal the raw difference.” In the first example, treatment and control participants had different baseline averages, and the outcome model related baseline score to follow-up score. Consequently the raw and adjusted contrasts differed. Conversely, if baseline measurements are not associated with the outcome, or if the groups’ baseline values happen to be similar, adjustment may make little difference to the estimated contrast.

Keep the assignment variable clear. In a randomized trial, $T_i$ in this model denotes assigned group.
State the outcome and units. A contrast of $-3.2$ points is not a contrast of $-3.2$ percent.
Use a shared baseline reference. Comparing fitted group means at different baseline profiles is not the adjusted treatment contrast defined here.
Check the model assumptions. Randomization does not establish linearity, common slopes, or the error conditions used for standard errors.

Key takeaway: In a randomized trial, a follow-up model containing assignment and pre-treatment baseline measurements estimates a baseline-adjusted assignment contrast. With a correctly specified additive mean model and a common treatment effect, the treatment coefficient is the fitted treatment-minus-control difference at every baseline profile and after averaging over any shared baseline distribution.

Check Your Understanding

Answer each question using the common-slope baseline-adjusted model unless the question says otherwise.

In $Y_i=\beta_0+\tau T_i+\beta_1(X_i-c)+\varepsilon_i$, what does $\tau$ represent, and what are its units?
A model predicts 18 points for control and 15 points for treatment at the same baseline profile. What is the treatment-minus-control contrast?
Why does averaging both fitted group means over the same baseline profiles leave the contrast equal to $\tau$ in an additive common-effect model?
In the model, should $T_i$ indicate randomized assignment or treatment received?
If a treatment-by-baseline interaction is present, why might one treatment coefficient not describe the contrast at every baseline value?