Centering One Predictor in an Interaction

Why Center One Predictor in an Interaction?

In a model with an interaction, 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 generally describe an overall or average slope: it describes the slope when the interacting predictor equals zero. If zero is outside the observed range or has little practical meaning, that coefficient may be hard to interpret.

Centering one predictor offers a direct solution. We subtract a chosen reference value from that predictor while leaving the other predictor unchanged. The coefficient of the uncentered predictor then describes its conditional slope when the centered predictor is at the chosen reference value. The fitted relationship itself does not change; only the reference point used for interpreting a coefficient changes.

The Model and the Reparameterization

Consider a linear regression model with predictors \(X_1\) and \(X_2\), including their interaction:

Here \(Y\) is the outcome, \(\beta_0,\beta_1,\beta_2,\beta_3\) are model coefficients, and \(\varepsilon\) is an error term. For the conditional-mean interpretations below, assume \(E(\varepsilon\mid X_1,X_2)=0\), so the displayed regression expression gives the conditional mean of \(Y\). The interaction term allows the slope for one predictor to vary with the other. In particular, at a fixed value \(X_2=x_2\), the conditional slope for \(X_1\) is \(\beta_1+\beta_3x_2\).

Substitute \(X_2=Z_2+c\) into the model and collect terms:

This algebra shows the central result. Centering \(X_2\) changes the coefficient of \(X_1\) from \(\beta_1\) to \(\beta_1+c\beta_3\). The coefficient of \(Z_2\) stays \(\beta_2\), and the interaction coefficient stays \(\beta_3\). The intercept changes to \(\beta_0+c\beta_2\).

Interpreting the Changed Coefficient

The same interpretation follows by looking directly at the original model’s conditional slope. Holding \(X_2=x_2\) fixed, the terms involving \(X_1\) are \(\beta_1X_1+\beta_3X_1x_2\), so the slope for \(X_1\) is \(\beta_1+\beta_3x_2\). At the centering value \(x_2=c\), that slope is \(\beta_1+c\beta_3\)—exactly the coefficient of \(X_1\) in the centered model.

The coefficient of \(Z_2\) has a different reference point. In the centered model, the conditional slope for \(Z_2\), holding \(X_1=x_1\) fixed, is \(\beta_2+\beta_3x_1\). Since \(Z_2\) and \(X_2\) differ only by a constant, this is also the slope for \(X_2\) at \(X_1=x_1\). Its main-effect coefficient \(\beta_2\) therefore still represents the slope for \(X_2\) when \(X_1=0\). Centering \(X_2\) does not change that reference point.

These interpretations apply equally to estimated coefficients. If a fitted model has estimates \(b_1\) and \(b_3\), then after centering \(X_2\) at \(c\), the coefficient of \(X_1\) becomes \(b_1+cb_3\). The product \(cb_3\) must use compatible units: if \(X_2\) is measured in years, \(c\) is in years and \(b_3\) is expressed per unit of \(X_2\).

Predictions Stay the Same

Centering is a change of coordinates in the model, not a change to the underlying relationship. To see why, each original predictor value \(X_2\) corresponds to exactly one centered value \(Z_2=X_2-c\). The algebraic substitution above rewrites the same expression for every \(X_1\) and \(X_2\), so the predicted outcome at each pair of predictor values is unchanged.

When the same observations are used and the model retains \(X_1\), \(X_2\) (or \(Z_2\)), and their interaction, the set of fitted relationships is the same. Ordinary least squares therefore gives the same fitted values and residuals under this reparameterization; the corresponding coefficients are expressed in the new form. This conclusion depends on retaining the full interaction model. It does not mean that arbitrary changes to predictors or model terms leave predictions unchanged.

Choosing the Centering Value

The centering value \(c\) should be chosen to make the conditional slope useful to interpret. It might be a meaningful benchmark, such as age 50, or a value near the center of the observed range. If \(c\) is the sample mean of \(X_2\), the new coefficient of \(X_1\) is the fitted conditional slope at that sample mean. It is not automatically an average of individual-specific slopes under every possible interpretation; it is the slope evaluated at the chosen value.

The value should also be within, or at least close to, the range where the data provide information. Centering at a value far outside the observed range can produce a mathematically valid coefficient that represents an extrapolated conditional slope. Centering does not create observations at that value or resolve limitations in the data.

Common Mistake

A frequent misreading is to say, “The \(X_1\) coefficient changed after centering, so the effect of \(X_1\) changed.” That treats the coefficient as if it described one fixed slope across all values of \(X_2\). In an interaction model the slope depends on \(X_2\), and the coefficient is tied to the reference value \(X_2=0\) in the original model or \(X_2=c\) in the centered model.

The correction is to state the conditioning value and units whenever interpreting a main-effect coefficient. For example: “The estimated slope for activity is 0.7 mmHg per activity hour per week at age 50,” rather than simply, “The activity effect is 0.7.” Centering changes neither the conditional slope at a given original value of \(X_2\) nor the model’s predictions. It changes which conditional slope is written as the main-effect coefficient.