Centering Baseline Outcomes in Interaction Models

Why Center the Baseline Outcome?

In a randomized study, the treatment effect may vary with participants’ baseline outcome values. An interaction model describes that variation, but its treatment coefficient depends on where the baseline scale is set to zero. If zero is unusual or impossible for the study population, that coefficient may be mathematically correct yet hard to interpret. Centering lets us place zero at a meaningful baseline value instead.

Centering is a change of reference point, not a change to the underlying fitted relationship. With the interaction retained, the model predicts the same outcome for every treatment and baseline value before and after centering. What changes is the interpretation of particular coefficients: after centering, the treatment coefficient is the estimated treatment contrast at the chosen baseline value.

The Interaction Model and Its Assumptions

Let $Y$ be a continuous follow-up outcome, $X$ a continuous outcome measured at baseline, and $T$ an indicator of randomized assignment, coded $T=0$ for control and $T=1$ for treatment. Consider the linear interaction model

Y_i=\beta_0+\beta_1T_i+\beta_2X_i+\beta_3(T_iX_i)+\varepsilon_i.

Here, the model describes the conditional mean of the follow-up outcome for a given treatment assignment and baseline value. The coefficients allow the relationship between baseline and follow-up to differ by treatment group. For this interpretation, assume the conditional mean is correctly represented by this linear form over the baseline values of interest, the error has conditional mean zero given $T$ and $X$, and the observations are independent. If the usual classical standard errors are to be used, an additional equal-variance assumption is needed; that assumption is not needed for the algebraic change of reference point described here.

Because treatment was randomized, $T$ represents assignment rather than treatment received. The baseline outcome must have been measured before assignment or treatment. To interpret a treatment contrast at a particular baseline value, that value should also be meaningful for the study and supported by baseline observations in both groups; predictions far outside the observed range require extrapolation.

In the uncentered model, the treatment contrast at baseline value $X=x$ is $\beta_1+\beta_3x$. Thus, $\beta_1$ is the contrast at $X=0$, not necessarily an overall or average treatment effect. The interaction coefficient $\beta_3$ describes how that contrast changes for each one-unit increase in the baseline outcome.

Definition: To center the baseline outcome at a chosen reference value $c$, define $X_{c}=X-c$. The centered interaction model is $$ Y_i=\gamma_0+\gamma_1T_i+\gamma_2X_{c,i}+\gamma_3(T_iX_{c,i})+\varepsilon_i. $$ At $X=c$, the centered baseline value is zero, so $\gamma_1$ is the model’s treatment contrast at that baseline value.

How the Coefficients Change

The centered model is a re-expression of the original model. Substitute $X=X_c+c$ into the uncentered equation and collect terms:

\begin{aligned} Y &=\beta_0+\beta_1T+\beta_2(X_c+c)+\beta_3T(X_c+c)+\varepsilon\\ &=(\beta_0+\beta_2c)+(\beta_1+\beta_3c)T+\beta_2X_c+\beta_3(TX_c)+\varepsilon. \end{aligned}

Therefore, the coefficients in the centered version are

\gamma_0=\beta_0+\beta_2c,\qquad \gamma_1=\beta_1+\beta_3c,\qquad \gamma_2=\beta_2,\qquad \gamma_3=\beta_3.

This calculation also explains the interpretation. At $X=c$, $X_c=0$, and the model’s predicted mean is $\gamma_0$ for control and $\gamma_0+\gamma_1$ for treatment. Their difference is $\gamma_1$. The treatment effect at $X=x$, expressed using the centered model, is $\gamma_1+\gamma_3(x-c)$.

Key result: Centering the baseline outcome changes the intercept and treatment coefficient, but leaves the baseline slope and interaction coefficient unchanged. If the same terms are retained and the coefficients are transformed accordingly, fitted values for every observation are unchanged.

Worked Example: Centering at a Baseline Score of 50

Suppose a hypothetical randomized study measures a follow-up score and a baseline score, both in points. A fitted model using the uncentered baseline score has estimates $\beta_0=40$, $\beta_1=5$, $\beta_2=0.60$, and $\beta_3=-0.20$. The baseline scores of interest include 50, so center at $c=50$. Find and interpret the centered coefficients.

Define $X_{50}=X-50$. Applying the coefficient transformations gives

$\gamma_0=40+(0.60)(50)=70$ follow-up points; $\gamma_1=5+(-0.20)(50)=-5$ points; $\gamma_2=0.60$; and $\gamma_3=-0.20$.

The centered model is therefore

\widehat{Y}=70-5T+0.60X_{50}-0.20(TX_{50}).

At baseline score 50, $X_{50}=0$. The predicted control mean is 70 points, and the predicted treatment mean is $70-5=65$ points. The treatment-minus-control contrast at that baseline score is $-5$ points. This does not say that the contrast is $-5$ at every baseline score: each additional baseline point changes the contrast by $-0.20$ follow-up points.

Check that predictions agree at baseline score 60. In the original model, control is $40+(0.60)(60)=76$, and treatment is $40+5+(0.60)(60)+(-0.20)(60)=69$. In the centered model, $X_{50}=10$: control is $70+(0.60)(10)=76$, and treatment is $70-5+(0.60)(10)+(-0.20)(10)=69$. The two parameterizations give the same predictions.

Choosing and Interpreting the Reference Value

A useful center $c$ is a baseline value that gives the treatment coefficient a clear meaning. It might be a clinically meaningful score, a standard threshold, or a typical value such as the sample mean. The best choice depends on the question and on the scale. For example, centering a score at a value that represents the study’s target population can make the treatment coefficient directly answer a relevant conditional question.

The center does not have to be zero on the original scale, and centering does not change the units. If $X$ is measured in points, then $X_c$ is still measured in points. The intercept is the predicted control mean at $X=c$, while the treatment coefficient is the treatment-minus-control contrast at $X=c$. The interaction coefficient remains the change in that contrast per additional baseline point.

A sample mean is a common reference value, but it does not turn the treatment coefficient into an average treatment effect across the sample. It gives the contrast at the mean baseline value. In a linear interaction model, the average predicted contrast over observed baseline values equals the contrast evaluated at their mean, because the treatment contrast is linear in $X$. That equality depends on the stated model and on averaging over the same baseline values; it does not mean that centering has removed treatment-effect variation.

Worked Example: Centering at a Clinically Meaningful Score

In a hypothetical trial, a baseline score is measured in points, and 60 points is a meaningful reference value. The fitted uncentered model is

\widehat{Y}=22+3T+0.75X-0.10(TX).

Set $X_{60}=X-60$. The transformed intercept is $22+(0.75)(60)=67$. The transformed treatment coefficient is $3+(-0.10)(60)=-3$. The baseline slope and interaction coefficient stay $0.75$ and $-0.10$, respectively. Thus,

\widehat{Y}=67-3T+0.75X_{60}-0.10(TX_{60}).

At baseline score 60, the predicted control mean is 67 outcome units and the predicted treatment mean is 64 outcome units. The estimated treatment-minus-control contrast at that reference score is $-3$ outcome units. For participants whose baseline score is one point higher, the model’s treatment contrast is 0.10 outcome units lower.

The slopes also have a direct interpretation: for control, the predicted follow-up outcome increases by 0.75 units per baseline point; for treatment, it increases by $0.75-0.10=0.65$ units per baseline point. These descriptions concern the fitted conditional means, not a claim that changing a participant’s baseline score would cause a change in their follow-up outcome.

Changing the Center Again

A different reference point changes the numerical treatment coefficient again, while preserving the fitted relationship. For a shift from center $c$ to a new center $d$, let $X_d=X-d$. Starting from a model centered at $c$, note that $X_c=X_d+(d-c)$. Substitution shows that the new intercept is $\gamma_0+\gamma_2(d-c)$, and the new treatment coefficient is $\gamma_1+\gamma_3(d-c)$. The baseline and interaction coefficients remain $\gamma_2$ and $\gamma_3$.

Worked Example: Moving the Center from 50 to 55

Return to the first hypothetical model, centered at 50, with $\gamma_0=70$, $\gamma_1=-5$, $\gamma_2=0.60$, and $\gamma_3=-0.20$. Move the center to 55, so $d-c=5$. The new intercept is $70+(0.60)(5)=73$, and the new treatment coefficient is $-5+(-0.20)(5)=-6$. The other coefficients remain $0.60$ and $-0.20$.

The model centered at 55 is

\widehat{Y}=73-6T+0.60X_{55}-0.20(TX_{55}).

At baseline score 60, $X_{55}=5$. The control prediction is $73+(0.60)(5)=76$; the treatment prediction is $73-6+(0.60)(5)+(-0.20)(5)=69$. These are the same predictions obtained using the center of 50. The treatment coefficient changed from $-5$ to $-6$ because it now describes the contrast at 55 rather than at 50.

Common Mistakes and Practical Checks

Calling the centered treatment coefficient an overall effect. It is the treatment contrast at the selected reference value. With an interaction, contrasts at other baseline values may differ.
Centering but dropping the interaction. Removing the treatment-by-baseline term changes the model; it is not the same re-expression. To preserve fitted values from an interaction model, retain the interaction and transform its coefficients consistently.
Assuming centering changes the predictions. The coefficient values change, but the linear predictor does not when the transformation is done correctly. Checking predicted values at one or more baseline scores is a useful arithmetic check.
Choosing a reference value far outside the data. The transformed coefficient can still be calculated, but its interpretation may rely on extrapolation. Prefer a meaningful value within the range where both treatment groups have baseline observations.
Confusing centering with standardizing. Centering subtracts a constant; it does not divide by a standard deviation or change the unit of the baseline outcome.

Centering is especially useful when the original zero has no practical meaning, when a particular baseline value is central to the scientific question, or when coefficient tables need a more interpretable treatment contrast. It cannot make a misspecified relationship linear, establish that effects are constant, or replace a considered choice of baseline value. Its value is clarity: it makes the treatment coefficient answer an explicitly chosen conditional question while leaving the model’s fitted values intact.

Check Your Understanding

Use the interaction model and coefficient transformations to answer these questions.

In an uncentered model, what baseline value does the treatment coefficient describe?
If the baseline outcome is centered at $c$, write the formula for the treatment coefficient in the centered model using the uncentered coefficients.
A fitted interaction model has $\beta_1=4$ and $\beta_3=-0.15$. What is the treatment contrast at $X=20$?
When the center changes but the interaction model is retained, which fitted quantities stay the same and which coefficient interpretations change?
Why can centering at a value outside the observed baseline range produce an interpretation that is difficult to support?