Polynomial Mean Functions and Basis Terms

Introduction

A straight-line regression mean changes at a constant rate as its predictor changes. Some relationships are not well described by a straight line: the mean response may bend across the range of observed predictor values. One way to represent such a pattern is to include powers of the predictor in the mean function.

A polynomial regression mean function is a single expression made by adding an intercept and one or more power terms. Each term contributes to the predicted mean at a given predictor value; the curve comes from their combined values, not from any one term acting alone. This tutorial develops that representation and shows how to calculate and interpret its fitted values.

Polynomial Mean Functions and Basis Terms

The basis terms are functions of the same predictor. For example, if \(x=3\), then \(x^2=9\) and \(x^3=27\). They are not separate predictors measured independently of \(x\); their values are determined by the value of \(x\). The model combines these values with their respective coefficients to produce one conditional mean.

The intercept is the constant term \(\beta_0\). Since all positive powers of \(x\) equal zero when \(x=0\), the mean function gives \(m(0)=\beta_0\). This is an interpretation at \(x=0\), whether or not zero lies in the observed range. When zero is outside that range, the intercept is still part of the mathematical expression, but its value may not describe a practically observed situation.

Each coefficient multiplies one basis term. The term \(\beta_jx^j\) contributes \(\beta_jx^j\) to the mean at the specified value of \(x\). In general, a coefficient on a power term is not the entire change in the mean associated with a one-unit increase in \(x\). The fitted mean at any \(x\) is the sum of all the terms, including the intercept.

Mean Model and Fitted Curve

The mean function describes the average response at each predictor value, not every individual response. A standard way to express the regression model is

Here \(x_i\) is the predictor value for observation \(i\), and \(\varepsilon_i\) is that observation’s deviation from the mean function. The zero conditional-mean assumption ensures that \(E(Y_i\mid x_i)=m(x_i)\). This definition of the mean function does not, by itself, require normally distributed errors or constant error variance. Those conditions matter for particular estimation and inference procedures, not for defining the polynomial form.

In data analysis the coefficients are usually estimated. Write the estimated coefficients as \(\hat{\beta}_0,\hat{\beta}_1,\ldots,\hat{\beta}_d\). The resulting fitted mean function is

For a specified \(x\), substituting its value into every term and adding gives the fitted mean response \(\hat{m}(x)\). Evaluating this expression at many predictor values produces the fitted curve. In a plot, the curve is the combined result of all the terms across those values. A power term is not a separate curve to interpret in isolation.

Why the Terms Form One Curve

A polynomial mean function is a sum of basis terms. Its shape is determined by how the value of each term changes as \(x\) changes, and by how the coefficients weight those terms. The following algebraic form makes the combination explicit:

At any fixed \(x\), the values \(1,x,x^2,\ldots,x^d\) are numbers. The expression multiplies each of those numbers by its coefficient and adds the results. As \(x\) varies, the basis-term values vary, so the sum traces out a curve. This is why a polynomial model can represent a mean pattern more flexible than a straight line, while still specifying one mean at each predictor value.

The model is linear in its coefficients: each coefficient appears as a multiplier and the terms are added together. But when a power higher than one is included, the mean function is generally not a straight-line function of \(x\). These two descriptions are compatible because “linear in the coefficients” concerns how the unknown coefficients enter the expression, while “straight-line” concerns how the mean changes with the predictor.

This representation can also be written using basis functions \(\phi_j(x)=x^j\), so that \(m(x)=\sum_{j=0}^{d}\beta_j\phi_j(x)\). The basis-function notation emphasizes that the model is built from chosen functions of one predictor. In polynomial regression, those functions are powers. It does not mean that \(x^2\) or \(x^3\) is a newly observed outcome, and it does not transform \(Y\).

Degrees, Units, and the Predictor Scale

A model with \(d=1\) uses the constant and first power of \(x\); it is a straight-line mean function. Including higher powers allows additional terms in the function. The degree describes the highest power included with a nonzero coefficient. If a coefficient on the highest listed power is zero, the expression has a lower degree than its written maximum power.

Units help check whether a polynomial expression is being used consistently. If \(Y\) is measured in kilograms and \(x\) in meters, then the coefficient of \(x^j\) must have units of kilograms per meter to the \(j\)th power. Consequently, \(\beta_jx^j\) has units of kilograms, matching the response and every other term in the sum. The intercept has the units of \(Y\).

Changing the numerical units of \(x\) changes the numerical values of its powers and therefore changes the coefficients needed to express the same mean pattern. For instance, measuring a distance in centimeters rather than meters changes the numerical predictor value by a factor of 100. The polynomial terms must be interpreted with the predictor’s stated scale. A coefficient cannot be interpreted meaningfully without knowing the units used for \(x\) and \(Y\).

Common Mistakes and Why the Representation Matters

A frequent mistake is to interpret a coefficient as though it described the whole fitted curve. In a polynomial model, the coefficient \(\beta_j\) weights the particular basis term \(x^j\). At a specific predictor value, the term’s contribution is \(\beta_jx^j\), and the fitted mean is the sum of that contribution and all the others. A coefficient alone is not the model’s prediction at that \(x\).

Another mistake is to treat powers as independently observed predictors. The values \(x\) and \(x^2\) are linked because the second is calculated from the first. They are terms in a chosen mathematical representation, not separate measurements. Similarly, including \(x^2\) changes the mean function for \(Y\); it does not mean that the outcome itself has been squared.

The intercept also requires care. It is exactly the mean function’s value at \(x=0\), but that may be an extrapolation if the observed predictor values are all far from zero. The algebraic meaning is fixed; whether it has a useful interpretation in a particular setting depends on the predictor scale and observed range.

Polynomial terms provide a way to describe curved mean patterns using a sum of familiar functions. This is useful because the fitted mean can be evaluated at specific predictor values and displayed as a curve. The representation alone does not establish that a particular polynomial is appropriate for every data set; it defines the form of the mean function, while evaluating its suitability requires attention to the data and the purpose of the analysis.