Avoid Misreading Product Coefficients and Main Effects

Why These Coefficients Are Easy to Misread

In a regression model with an interaction, the association between one predictor and the response depends on the value of the other predictor. That dependence is represented by a product term. It is tempting to read the coefficient of that product as if it were the effect of one predictor on the response, or to quote a main-effect coefficient without saying what value of the other predictor it refers to. Both readings can be misleading.

The key is to interpret the fitted equation as a whole. A main-effect coefficient describes a slope at a particular reference value of the other predictor. The product coefficient describes how that slope changes as the other predictor increases. Neither, on its own, gives the slope everywhere.

The Model and What Its Coefficients Represent

Let \(X\) and \(Z\) be quantitative predictors and \(Y\) a quantitative response. A two-predictor linear interaction model for observation \(i\) is

Here \(\beta_0,\beta_1,\beta_2,\beta_3\) are population model coefficients and \(\varepsilon_i\) is the error for observation \(i\). The fitted model uses estimated coefficients, often written \(b_0,b_1,b_2,b_3\), to give a predicted response:

This equation describes the fitted conditional mean of \(Y\) as a function of \(X\) and \(Z\), provided the mean model is appropriate. For usual regression interpretation, assume that the conditional mean is represented by this equation and that errors have mean zero given the predictors. Standard inferential procedures may also require conditions such as independent errors and constant error variance; those conditions do not change the algebraic meanings of the coefficients. The model describes an association, not necessarily a causal effect.

These statements follow directly by holding one predictor fixed. At a fixed value \(Z=z\), collect the terms involving \(X\):

So the fitted slope for \(X\) at \(Z=z\) is \(b_1+b_3z\), not generally \(b_1\). In the other direction, holding \(X=x\) fixed gives a fitted slope for \(Z\) of \(b_2+b_3x\). The intercept \(b_0\) is the fitted response at \(X=0\) and \(Z=0\).

A Short Derivation: Finite Changes at Fixed Predictor Values

The same algebra shows why a product coefficient is not an overall effect. Suppose \(X\) changes from \(x\) to \(x+\Delta x\) while \(Z=z\) stays fixed. The fitted response difference is

Thus, the fitted change depends on both the size of the change in \(X\) and the value of \(Z\). For a one-unit increase in \(X\), the fitted change is \(b_1+b_3z\). Meanwhile, increasing \(Z\) by one unit changes the \(X\)-slope from \(b_1+b_3z\) to \(b_1+b_3(z+1)\). Their difference is

That difference in slopes is the product coefficient’s interpretation. Its units are response units divided by one unit of \(X\) and one unit of \(Z\). It is not the response change for a one-unit increase in either predictor by itself.

Main Effects Depend on the Reference Value

The coefficient \(b_1\) is sometimes called the main effect of \(X\), but that label must not hide its condition: it is the slope for \(X\) at \(Z=0\). Similarly, \(b_2\) is the slope for \(Z\) at \(X=0\). If zero is not a meaningful or relevant value for the other predictor, the corresponding main-effect coefficient may not have a useful standalone interpretation. It remains part of the model, but its reference condition should be stated.

Centering a predictor replaces it with its difference from a chosen reference value. For example, if \(X_c=X-c\), then \(X_c=0\) means that \(X=c\). In a model expressed using centered predictors, the main-effect coefficients are conditional slopes at those chosen reference values. Centering changes the reference point for interpreting coefficients, not the fitted predictions when the model is rewritten consistently.

Reading the Product Coefficient in Its Units

The units provide a useful check on interpretation. If \(Y\) is measured in kilograms, \(X\) in liters, and \(Z\) in kilograms, then \(b_1\) has units of kilograms per liter, while \(b_3\) has units of kilograms per (liter times kilogram). Multiplying \(b_3\) by a change in \(Z\) gives a change in the \(X\)-slope, measured in kilograms per liter.

A product coefficient can be positive or negative. A positive value means that the fitted slope for \(X\) increases as \(Z\) increases; it does not necessarily mean that the fitted slope for \(X\) itself is positive. If \(b_1\) is negative, a positive \(b_3\) might make that slope less negative or eventually positive, depending on the values considered. The conditional slope \(b_1+b_3z\) answers the direction-and-size question at a particular \(z\).

Common Misreadings and a Reliable Check

A common mistake is to report, “The product coefficient is the effect of \(X\) on \(Y\).” That reasoning treats the \(XZ\) term as though its coefficient were the slope for \(X\). The correction is to hold \(Z=z\) fixed and calculate \(b_1+b_3z\). If the question is about a finite change \(\Delta x\), calculate \(\Delta x(b_1+b_3z)\).

Another mistake is to report “\(b_1\) is the effect of \(X\)” without naming the condition \(Z=0\). The correction is to state that \(b_1\) is the fitted \(X\)-slope when \(Z=0\), or to give the conditional slope at another specified value. If the predictors are centered, identify what their zero values represent.

Before interpreting any coefficient in an interaction model, check three things: which predictor is changing, what value the other predictor is held at, and what units the resulting slope or slope change has. These checks keep a coefficient from being mistaken for a different quantity. They also keep the interpretation tied to the fitted association rather than implying causation.