Interpreting Centered Quadratic Coefficients

Why the Centering Value Matters

In a quadratic regression, the coefficients describe different features of a curved fitted relationship. Centering the predictor does not change the fitted curve, but it gives the intercept and linear coefficient a new, often more useful reference point. The intercept becomes the fitted outcome at the centering value, and the linear coefficient becomes the slope there. The quadratic coefficient retains its value because shifting the predictor does not change the curve's curvature.

These interpretations are useful when the predictor value zero has little practical meaning. For example, if \(x\) is a patient's age in years and the observed ages are near 50, an intercept at age zero may not be informative. Centering age at 50 makes the intercept and linear coefficient describe the fitted outcome and slope at age 50 instead.

The Model and Its Centered Form

Let \(Y_i\) be a quantitative outcome and \(x_i\) a quantitative predictor. A quadratic regression model for the conditional mean is

Here \(\beta_0,\beta_1,\beta_2\) are population coefficients. A fitted equation uses estimated coefficients, often written \(b_0,b_1,b_2\), in place of the population coefficients. The interpretation of the fitted equation describes its estimated mean curve, not necessarily the exact population curve.

For interpreting this fitted relationship, the key model condition is that the conditional mean follows the stated quadratic form; in a regression model written with errors, this corresponds to errors having conditional mean zero given the predictor. Conventional least-squares inference also uses assumptions about the errors, such as independence and constant variance, but no inference procedure is needed for the coefficient interpretations in this tutorial. A fitted association alone does not establish a causal effect.

Choose a fixed centering value \(c\), and define the centered predictor \(z=x-c\). Thus \(x=c+z\), and \(z=0\) is the original predictor value \(x=c\). Write the same fitted mean curve in centered form:

The coefficients \(a_0,a_1,a_2\) refer to this centered expression. They do not have the same meanings as the corresponding coefficients in the uncentered expression. In particular, \(a_0\) is the fitted outcome when \(x=c\), not when \(x=0\).

Why the Coefficients Change as They Do

The link between the two forms follows by substitution. Start with the uncentered fitted equation and replace \(x\) by \(c+z\):

Matching the constant, \(z\), and \(z^2\) terms gives the conversion:

The constant term changes because the fitted curve is now referenced to \(x=c\). The linear term changes because the slope of a curved line depends on where it is evaluated. The quadratic term is unchanged because the \(z^2\) term in the expansion has coefficient \(b_2\); translating \(x\) adds contributions to the constant and linear terms but does not alter that coefficient.

The same meanings can be checked directly from the centered fitted equation. Setting \(x=c\) makes both \(x-c\) and \((x-c)^2\) zero, so \(\widehat{Y}=a_0\). Differentiating with respect to \(x\) gives the fitted slope:

The second derivative is \(2a_2\), so the quadratic coefficient is half the second derivative, just as in the uncentered quadratic form. Its units are outcome units per predictor unit squared. The intercept has outcome units, while the linear coefficient has outcome units per predictor unit.

Interpreting a Centered Equation in Context

The centered equation is often most useful when it is read from the reference value outward. At \(x=c\), the fitted outcome is \(a_0\). For a small change \(h\) from the center, so that \(x=c+h\), the fitted value is

The linear part \(a_1h\) describes the change suggested by the slope at the center; the quadratic part \(a_2h^2\) accounts for the changing slope as the predictor moves away from the center. For a quadratic curve, the slope at \(x=c+h\) is \(a_1+2a_2h\). Thus \(a_1\) is specifically the slope at the center, not a constant slope throughout the predictor range.

The units should be kept in view. If the outcome is measured in milligrams and the predictor in days, \(a_0\) is in milligrams, \(a_1\) is in milligrams per day, and \(a_2\) is in milligrams per day squared. The centered predictor \(x-c\) still has units of days; subtracting two values on the same scale does not make the difference unitless.

Changing the Center Again

A different center gives a different intercept and linear coefficient, even though the underlying fitted curve and quadratic coefficient remain the same. This is not a contradiction: the intercept and linear coefficient answer questions at the newly selected reference value. For a new center \(d\), the intercept is the fitted value at \(x=d\), and the linear coefficient is the slope there.

Common Mistake and Why Interpretation Matters

A frequent error is to interpret the linear coefficient of a centered quadratic model as the slope at every predictor value. For example, in the model centered at 10 hours, saying that every extra hour is associated with exactly 1.6 more score points ignores the quadratic term. The correct statement is that the fitted slope is 1.6 points per hour at 10 hours; at another value \(x\), it is \(1.6+2(-0.08)(x-10)\) points per hour.

Another error is to treat the centered intercept as the fitted value at \(x=0\). In the centered model, the intercept is the prediction where \(x-c=0\), which means \(x=c\). In the study-time example centered at 10, the intercept 66 is the fitted score at 10 hours, not at zero hours. Always identify the center before interpreting the intercept.

Centering is valuable because it makes coefficient interpretations correspond to a meaningful predictor value, often one near the observations. It does not itself improve the fit or change any fitted value, and the unchanged quadratic coefficient does not mean the linear coefficient is also unchanged. Coefficients should be interpreted in the equation in which they appear, with the predictor's units and the chosen reference value stated explicitly.