Centering Predictors in Quadratic Models

Why Center a Predictor?

In a quadratic regression model, the predictor appears both as a first-degree term and as a squared term. If \(x\) measures age in years, for example, the model includes both \(x\) and \(x^2\). The value \(x=0\) may have little practical relevance, yet the uncentered intercept describes the model at that value. Centering lets us express the same quadratic form relative to a reference value that matters in the setting, such as age 40.

Centering means subtracting a chosen constant from each predictor value. In a quadratic model, the resulting difference must be used in both the linear and squared terms. This changes how the intercept and linear coefficient are expressed. It does not change the units of the outcome or the underlying predictor, and the reference value must be stated in the predictor’s original units.

The Model and the Centered Predictor

Suppose \(Y_i\) is a quantitative response and \(x_i\) is a quantitative predictor for observation \(i\). A quadratic regression model for the conditional mean is

E(Y_i\mid x_i)=\beta_0+\beta_1x_i+\beta_2x_i^2.

Here, \(\beta_0\), \(\beta_1\), and \(\beta_2\) are model coefficients. A customary error form is \(Y_i=\beta_0+\beta_1x_i+\beta_2x_i^2+\varepsilon_i\), where the errors have conditional mean zero. For the usual regression model, errors are also taken to be independent with constant variance; normal errors are needed for some exact small-sample inference procedures, but not for the centering algebra here. The transformation below is an algebraic change of how the quadratic mean is written.

Definition: Choose a reference value \(c\), measured in the same units as \(x\), and define the centered predictor \(z_i=x_i-c\). The centered quadratic model writes the linear and squared terms using \(z_i\): \(E(Y_i\mid x_i)=\alpha_0+\alpha_1z_i+\alpha_2z_i^2\). The reference value \(c\) is the value of the original predictor at which the centered predictor equals zero.

The key is to use \(z_i\) in both terms, giving \(z_i\) and \(z_i^2\). Merely replacing \(x_i\) by \(x_i-c\) in the first-degree term while leaving the old \(x_i^2\) term in place would define a different expression, not the centered version of this quadratic model.

How the Coefficients Translate

Since \(z=x-c\), the original predictor can be written as \(x=z+c\). Substitute this expression into the original quadratic mean and expand the squared term:

\begin{aligned} \beta_0+\beta_1x+\beta_2x^2 &=\beta_0+\beta_1(z+c)+\beta_2(z+c)^2\\ &=\beta_0+\beta_1z+\beta_1c+\beta_2(z^2+2cz+c^2)\\ &=(\beta_0+\beta_1c+\beta_2c^2) +(\beta_1+2\beta_2c)z+\beta_2z^2. \end{aligned}

Comparing this expression with the centered model gives the coefficient translation:

\alpha_0=\beta_0+\beta_1c+\beta_2c^2,\qquad \alpha_1=\beta_1+2\beta_2c,\qquad \alpha_2=\beta_2.

Thus the intercept and linear coefficient depend on the chosen reference value. The quadratic coefficient is unchanged by this translation. In the original model, the predicted mean at \(x=c\) is \(\beta_0+\beta_1c+\beta_2c^2\); that is the centered intercept \(\alpha_0\), because \(z=0\) at the reference. The new linear coefficient \(\alpha_1\) is the rate of change of the quadratic mean with respect to \(x\) at \(x=c\): the original slope is \(\beta_1+2\beta_2x\), which at \(x=c\) equals \(\beta_1+2\beta_2c\). These are descriptions of the coefficients’ relation to the reference point, not claims that the relationship is linear over the whole predictor range.

Key result: When \(z=x-c\) is used in both terms of a quadratic model, the intercept becomes the model’s predicted mean at \(x=c\), and the linear coefficient becomes the original quadratic slope at \(x=c\). The quadratic coefficient stays the same.

Worked Examples

Worked Example: Centering Hours at a Relevant Reference

Suppose a hypothetical model relates a training outcome \(Y\), measured in points, to practice time \(x\), measured in hours:

E(Y\mid x)=12+1.8x-0.04x^2.

Find the centered model using \(c=10\) hours. Then describe how its intercept and linear coefficient relate to that reference.

Define \(z=x-10\), so \(c=10\), \(\beta_0=12\), \(\beta_1=1.8\), and \(\beta_2=-0.04\). Apply the translation formulas:

\begin{aligned} \alpha_0 &=12+(1.8)(10)+(-0.04)(10^2)\\ &=12+18-4=26,\\ \alpha_1 &=1.8+2(-0.04)(10)\\ &=1.8-0.8=1.0,\\ \alpha_2&=-0.04. \end{aligned}

Therefore, the centered model is \(E(Y\mid x)=26+1.0z-0.04z^2\), with \(z=x-10\). The intercept is 26 points, the model’s predicted mean at 10 hours of practice. The linear coefficient is 1.0 point per hour, the slope of the quadratic mean at 10 hours. The quadratic coefficient remains \(-0.04\) points per hour squared.

Worked Example: Centering Age at 35 Years

Consider a hypothetical model for an assessment score \(Y\), in points, using age \(x\), in years:

E(Y\mid x)=42+1.2x-0.015x^2.

Re-express the model using age 35 as the reference value, and calculate the new intercept and linear coefficient.

Set \(z=x-35\). Then

\begin{aligned} \alpha_0 &=42+(1.2)(35)+(-0.015)(35^2)\\ &=42+42-18.375=65.625,\\ \alpha_1 &=1.2+2(-0.015)(35)\\ &=1.2-1.05=0.15,\\ \alpha_2&=-0.015. \end{aligned}

The centered expression is \(E(Y\mid x)=65.625+0.15z-0.015z^2\), where \(z=x-35\). At age 35, the model’s predicted mean score is 65.625 points. The slope there is 0.15 points per year. For this example, age 35 is a reference that gives the intercept and linear coefficient direct meaning in the stated context, unlike the uncentered intercept, which corresponds to age zero.

Worked Example: Changing from One Reference Value to Another

A hypothetical centered model uses \(z=x-10\), where \(x\) is practice time in hours, and has the form

E(Y\mid x)=26+1.0z-0.04z^2,

with \(Y\) measured in points. Rewrite the model using 14 hours, rather than 10 hours, as the reference value.

Let \(w=x-14\). Since \(z=x-10=(x-14)+4=w+4\), substitute \(w+4\) for \(z\):

\begin{aligned} E(Y\mid x) &=26+(w+4)-0.04(w+4)^2\\ &=26+w+4-0.04(w^2+8w+16)\\ &=30+w-0.04w^2-0.32w-0.64\\ &=29.36+0.68w-0.04w^2. \end{aligned}

With the reference moved to 14 hours, the intercept is 29.36 points and the linear coefficient is 0.68 points per hour. The intercept is the model’s predicted mean at 14 hours, and the linear coefficient is the slope there. The new reference value changes both of those coefficients, while the quadratic coefficient remains \(-0.04\) points per hour squared.

Choosing a Reference Value

A useful reference value is one that makes the intercept or local slope easier to discuss in the setting. It may be a scientifically meaningful value, a common value in the observed data, or a value chosen to make zero correspond to a practical comparison point. The reference should be identified with its units; “centered at 35” is incomplete if readers do not know whether 35 means years, kilograms, or another measurement.

Centering does not make an unsuitable quadratic model suitable, and it does not establish that a relationship is causal. It is a way to express the same model form around a more useful point. The choice should therefore serve the interpretation of the model rather than be treated as evidence about the relationship.

Common Mistakes

Centering only the first-degree term. Replacing \(x\) by \(x-c\) in the linear term but leaving \(x^2\) unchanged is not the centered quadratic expression. Use \((x-c)^2\) for the squared term as well.
Keeping the old intercept. After centering, the intercept is not generally \(\beta_0\). It is \(\beta_0+\beta_1c+\beta_2c^2\), the predicted mean at the reference value.
Assuming the linear coefficient is unchanged. The new linear coefficient is \(\beta_1+2\beta_2c\), not usually \(\beta_1\). It represents the slope at the chosen reference, not one constant slope across all values of \(x\).
Using the wrong units or reference. The value \(c\) must be in the same units and scale as \(x\). If the predictor is recorded in hours, a reference of 10 means 10 hours, not 10 minutes.

Check Your Understanding

Use the coefficient translation formulas and keep the reference value and units explicit.

For \(E(Y\mid x)=5+2x+0.1x^2\), center \(x\) at \(c=3\). What are the centered intercept, linear coefficient, and quadratic coefficient?
In a quadratic model, what predictor value does the centered intercept correspond to?
Why must the squared term be written as \((x-c)^2\) when centering a quadratic predictor?
If \(x\) is measured in years, what must be true of the units used for the reference value \(c\)?
Explain why the centered linear coefficient should not be described as the slope at every predictor value.