Centering Both Predictors in an Interaction

Why Center Both Predictors?

In an interaction model, the association between one predictor and the outcome depends on the value of the other predictor. As a result, a main-effect coefficient does not describe one universal slope: it describes a slope at a particular value of the other predictor. When the predictors have not been centered, that particular value is zero. If zero is not meaningful or is outside the observed range, the main-effect interpretations may be hard to use.

Centering both predictors replaces each with its difference from a chosen reference value. The model still describes the same fitted outcomes, but the intercept and main effects now refer to those reference values. This tutorial focuses on reading that numerical interaction model: its intercept describes the outcome at both references, and each main effect describes a slope at the other predictor’s reference.

The Model and Its Assumptions

Let \(x_1\) and \(x_2\) be two quantitative predictors, and let \(Y\) be a quantitative outcome. An interaction model with both main effects and their product can be written as

Here, \(E(Y\mid x_1,x_2)\) is the model’s conditional mean outcome at the stated predictor values. In a fitted model, use estimated coefficients \(b_0,b_1,b_2,b_3\) to calculate the fitted outcome \(\hat y\). The product term \(x_1x_2\) allows the slope for one predictor to vary with the other.

The model assumes that the conditional mean has the stated form over the range where it is used, including a product interaction and no additional terms. For usual regression inference, errors are taken to have conditional mean zero; independence and constant error variance are also commonly assumed for standard error calculations, with normal errors used for exact small-sample \(t\)-procedures. Centering is an algebraic change of predictor coordinates; it does not make these assumptions true. The interpretations below describe conditional associations in the model, not causal effects.

Choose reference values \(c_1\) for \(x_1\) and \(c_2\) for \(x_2\), and define the centered predictors

Thus \(u=0\) means \(x_1=c_1\), and \(v=0\) means \(x_2=c_2\). The reference values can be chosen for substantive meaning or because they are representative of the data, but they should be stated along with the resulting interpretations.

How the Coefficients Change

Substitute \(x_1=u+c_1\) and \(x_2=v+c_2\) into the original model. Expanding the product gives

This is the same fitted model written in centered coordinates. The calculation just groups terms by the constant, \(u\), \(v\), and \(uv\); it does not alter any predicted outcome for a given pair of original predictor values.

The intercept adds up the model’s terms at the two references. The centered \(x_1\) main effect includes \(b_3c_2\), because its slope is evaluated at \(x_2=c_2\); the centered \(x_2\) main effect includes \(b_3c_1\) for the corresponding reason. The interaction coefficient stays the same.

Reading the Conditional Slopes

The slope interpretations follow directly by collecting the terms involving each predictor. In the original model, hold \(x_2\) fixed. The terms that vary with \(x_1\) are \(b_1x_1+b_3x_1x_2=(b_1+b_3x_2)x_1\), so the slope for \(x_1\) is \(b_1+b_3x_2\). Similarly, holding \(x_1\) fixed gives an \(x_2\) slope of \(b_2+b_3x_1\).

At the reference values, these slopes become \(b_1+b_3c_2=b_1^*\) and \(b_2+b_3c_1=b_2^*\). Equivalently, in the centered model the slope for \(u\) at a specified \(v\) is \(b_1^*+b_3^*v\), and the slope for \(v\) at a specified \(u\) is \(b_2^*+b_3^*u\). Setting the other centered predictor to zero gives the main-effect interpretation. This is why a main effect in an interaction model must always be reported with the other predictor’s value.

Checking the Re-expression

A direct prediction check is a useful way to confirm that the centered coefficients and reference values have been combined correctly. For \(x_1=5\) hours and \(x_2=2\) tests in the first example, the centered values are \(u=0\) and \(v=0\). The centered model predicts 71 points, matching the original model:

Common Mistakes and Why the Reference Values Matter

A frequent error is to read a main effect in the centered model as a slope that applies regardless of the other predictor. In an interaction model, the slope changes with the other predictor, so the main effect is conditional. Another error is to say the intercept is the overall or average outcome. It is the predicted conditional mean at the specific pair of reference values; it equals an overall mean only under additional conditions, not simply because predictors were centered.

Centering both predictors is useful because it can make the intercept and the two main effects easier to communicate. The choices do not remove the interaction, change the predictions, or establish a causal interpretation. Reference values should be identified in the same units as the original predictors, and interpretations should keep the outcome units attached. For instance, a slope measured in score points per study hour is not a percentage change in score.

• Wrong: “The study-hour coefficient is the increase in score for every additional hour, regardless of practice tests.”
• Correction: “It is the increase in predicted score per additional study hour when practice tests equal the stated reference value.”
• Wrong: “The centered intercept is the average score.”
• Correction: “It is the predicted score when both predictors equal their chosen reference values.”
• Wrong: “The interaction coefficient proves that extra study causes a higher score.”
• Correction: “It describes how the model’s conditional association between study hours and predicted score changes with practice-test count; it does not by itself establish causation.”