Centering Predictors in Interaction Models

Why Reference Values Matter in an Interaction Model

In a regression model with an interaction, the association between one predictor and the outcome depends on the value of another predictor. That makes a lower-order coefficient—the coefficient attached to one predictor by itself—conditional on a particular value of the other predictor. Centering makes that reference value explicit and can make it more useful than zero on the original scale.

For example, zero hours of study might be outside the range that is useful for interpreting a study-time model. If study time is centered at eight hours, then a coefficient attached to a second predictor describes the association at eight hours of study. The important question is not just what the coefficient’s number is, but which predictor values define the context in which it applies.

Definition: Centering a quantitative predictor \(x\) at a reference value \(c_x\) means replacing it in the model with \(x-c_x\). The centered predictor equals zero when the original predictor equals \(c_x\). In an interaction model, centering identifies the reference values at which the intercept and lower-order coefficients are interpreted.

The Centered Interaction Model

Consider two quantitative predictors, \(x\) and \(z\), and an outcome \(Y\). Let \(c_x\) and \(c_z\) be the chosen reference values. A model for the conditional mean of the outcome is

E(Y\mid x,z)=\beta_0+\beta_1(x-c_x)+\beta_2(z-c_z)+\beta_3(x-c_x)(z-c_z).

Here \(\beta_0\) is the intercept, \(\beta_1\) and \(\beta_2\) are the lower-order coefficients, and \(\beta_3\) is the interaction coefficient. The model includes both lower-order terms along with their product. That is the standard form for an interaction between these two predictors.

The interpretations below concern the model’s conditional mean. They assume that this form is an appropriate description of the mean outcome over the predictor values being discussed, and that the predictor units and reference values are clearly defined. Interpretations should generally stay within the range of data used to fit the model; predictions far outside that range may be unreliable. The coefficients describe model-based associations, not necessarily causal effects. Normality is not required just to explain what the coefficients mean.

Key interpretation: At \(x=c_x\) and \(z=c_z\), both centered predictors equal zero. The intercept \(\beta_0\) is the model’s expected outcome at that reference pair. The coefficient \(\beta_1\) describes the change in expected outcome per unit increase in \(x\) when \(z=c_z\). The coefficient \(\beta_2\) describes the change per unit increase in \(z\) when \(x=c_x\).

These are conditional interpretations. Because the model includes an interaction, \(\beta_1\) is not automatically the association between \(x\) and \(Y\) at every value of \(z\). It applies at the stated reference value \(z=c_z\). Likewise, \(\beta_2\) applies at \(x=c_x\). The interaction coefficient describes how the association changes as the other predictor changes; the lower-order coefficients give that association at the chosen reference values.

Worked Example: Study Time and Practice Sessions

Suppose a hypothetical analysis models exam score in points using study time \(x\), in hours, and practice sessions \(z\), in sessions. The predictors are centered at eight study hours and three practice sessions. The fitted mean model is

E(Y\mid x,z)=72+2.5(x-8)+1.2(z-3)+0.4(x-8)(z-3).

At the reference pair, substitute \(x=8\) and \(z=3\):

E(Y\mid 8,3)=72+2.5(8-8)+1.2(3-3)+0.4(8-8)(3-3)=72.

Thus the model predicts an average exam score of 72 points for students with eight study hours and three practice sessions. The intercept is not the prediction for a student with no study hours and no practice sessions; it is the prediction at the centering values.

The coefficient \(2.5\) means that, when practice sessions are held at the reference value of three, one additional study hour is associated with a model-predicted increase of 2.5 exam points. The coefficient \(1.2\) means that, when study time is eight hours, one additional practice session is associated with a model-predicted increase of 1.2 points. These interpretations name both the outcome units and the reference value of the other predictor.

For example, at \(z=3\), moving from \(x=8\) to \(x=9\) changes the model prediction by

2.5(9-8)+0.4(9-8)(3-3)=2.5+0=2.5\text{ points}.

The interaction term contributes zero for this comparison because practice sessions are at their reference value.

Reading Each Coefficient at Its Reference Point

The meanings of the intercept and lower-order coefficients can be checked by setting the centered predictors to zero or by comparing predictions at the reference values. This is a short way to justify the interpretations without treating the lower-order coefficients as universal effects.

Find the reference values.
In the terms \(x-c_x\) and \(z-c_z\), the values \(c_x\) and \(c_z\) are the original-scale values at which the centered predictors equal zero.

Interpret the intercept.
Set \(x=c_x\) and \(z=c_z\). Every term involving a centered predictor becomes zero, leaving the model prediction \(\beta_0\).

Interpret a lower-order coefficient.
For \(\beta_1\), hold \(z=c_z\); for \(\beta_2\), hold \(x=c_x\). At that reference value, the interaction term contributes zero to a one-unit change in the predictor whose coefficient is being interpreted.

For a more explicit justification, compare the model prediction at \(x=c_x+\Delta x\) with the prediction at \(x=c_x\), while holding \(z=c_z\). The difference is

\begin{aligned} &E(Y\mid c_x+\Delta x,c_z)-E(Y\mid c_x,c_z)\\ &=\{\beta_0+\beta_1\Delta x+\beta_2(0)+\beta_3\Delta x(0)\} -\{\beta_0+\beta_1(0)+\beta_2(0)+\beta_3(0)(0)\}\\ &=\beta_1\Delta x. \end{aligned}

For a one-unit increase in \(x\), \(\Delta x=1\), so the difference is \(\beta_1\). The same argument, changing \(z\) while holding \(x=c_x\), gives a prediction difference of \(\beta_2\) for a one-unit increase in \(z\). This calculation explains why the lower-order coefficients are tied to the other predictor’s reference value.

Worked Example: Rehabilitation Sessions and Baseline Symptoms

Suppose a hypothetical model predicts recovery time in days from weekly rehabilitation sessions \(x\) and a baseline symptom score \(z\). The predictors are centered at four sessions per week and a symptom score of 30. The model is

E(Y\mid x,z)=14-0.8(x-4)+0.15(z-30)-0.05(x-4)(z-30).

At four sessions and a symptom score of 30, the centered values are both zero:

E(Y\mid 4,30)=14-0.8(0)+0.15(0)-0.05(0)(0)=14\text{ days}.

The intercept is the model-predicted mean recovery time at that particular reference pair. It is not a general prediction for everyone in the model.

At a symptom score of 30, increasing rehabilitation from four to five sessions per week changes the prediction by

-0.8(5-4)+0.15(0)-0.05(5-4)(0)=-0.8\text{ days}.

Therefore, the lower-order coefficient \(-0.8\) describes a decrease of 0.8 predicted recovery days per additional weekly session when the symptom score is 30. Conversely, at four sessions per week, increasing the symptom score from 30 to 34 changes the prediction by

-0.8(0)+0.15(34-30)-0.05(0)(34-30)=0.60\text{ days}.

The coefficient \(0.15\) therefore describes an increase of 0.15 predicted days per symptom-score point when rehabilitation is four sessions per week. Neither lower-order coefficient should be read as applying unchanged at all values of the other predictor.

Reference Values Need Not Be Means

A reference value can be chosen for practical meaning; it does not have to be the sample mean. A value such as a common operating level or a clinically meaningful threshold may be more useful. The interpretation always follows the values actually used for centering, so those values should be reported with the coefficients.

Worked Example: Commute Time and Urban Density

Suppose a hypothetical model predicts monthly transit spending \(Y\), in dollars, from commute time \(x\), in minutes, and urban density \(z\), measured in thousands of people per square kilometre. The predictors are centered at a 30-minute commute and a density of 5 thousand people per square kilometre:

E(Y\mid x,z)=120+1.8(x-30)-6(z-5)+0.5(x-30)(z-5).

At the reference pair, the predicted monthly spending is

E(Y\mid 30,5)=120+1.8(0)-6(0)+0.5(0)(0)=120\text{ dollars}.

The coefficient \(1.8\) describes a predicted increase of 1.8 dollars per additional commute minute when density is 5 thousand people per square kilometre. The coefficient \(-6\) describes a predicted decrease of 6 dollars per additional thousand people per square kilometre when commute time is 30 minutes. These are reference-specific interpretations, even though the reference values were selected for practical relevance rather than because they are sample means.

For instance, at density \(z=5\), a commute time of 32 rather than 30 minutes changes the prediction by

1.8(32-30)-6(0)+0.5(32-30)(0)=3.6\text{ dollars}.

The interaction contribution is zero at the density reference value. This example also shows why units matter: \(z\) is measured in thousands of people per square kilometre, so the coefficient \(-6\) is per one-unit increase on that scale—not per single person per square kilometre.

Common Mistakes and Why Centering Helps

A frequent mistake is to interpret \(\beta_1\) as the association between \(x\) and \(Y\) for every value of \(z\), simply because it is written beside \(x-c_x\). With an interaction in the model, that is too broad: \(\beta_1\) applies specifically when \(z=c_z\). The matching rule for \(\beta_2\) is that it applies when \(x=c_x\).

Another mistake is to assume that centering both predictors at their sample means makes the intercept equal to the sample mean outcome. In an additive model that equality can hold under the usual fitted-model conditions, but an interaction term requires care. If both predictors are centered at their sample means, their individual sample means are zero, but the sample mean of their product need not be zero. The interaction term can therefore contribute to the average fitted value. Centering at the means makes the intercept the prediction at the pair of mean predictor values; it does not, by itself, guarantee that this prediction equals the mean outcome.

Wrong: “The coefficient of \(x-c_x\) is the effect of \(x\) at every value of \(z\).”
Correction: It is the modelled change per unit of \(x\) when \(z=c_z\); the interaction makes the association depend on \(z\).
Wrong: “The intercept is the outcome at zero on the original predictor scales.”
Correction: The intercept is the prediction at \(x=c_x\) and \(z=c_z\), which may be nonzero values on both original scales.
Wrong: “Centering makes the coefficient an overall or causal effect.”
Correction: Centering changes the reference point for interpretation; it does not remove the interaction or establish causation.

Key takeaway: In a centered interaction model, read the intercept at the pair of centering values. Read each lower-order coefficient as a conditional association, evaluated when the other predictor is at its own centering value. State those reference values and units whenever you interpret the coefficients.

Check Your Understanding

For each question, identify the relevant reference values and explain the coefficient in context.

In a model with predictors \(x\) and \(z\), centered at 10 and 4 respectively, what outcome value does the intercept describe?
In the same model, what does the coefficient of \(x-10\) describe, and what value of \(z\) must be specified?
Why should the coefficient of \(z-4\) not automatically be interpreted as the association between \(z\) and the outcome at every value of \(x\)?
If both predictors are centered at their sample means, why is it not guaranteed that the intercept equals the sample mean outcome in a model with an interaction?
A predictor is centered at 25 rather than at its sample mean. Does that make the coefficient interpretations invalid? Explain what determines their reference values.