Why a Coefficient Is Not a Predictor's Full Contribution

Introduction

In a regression equation, a coefficient sits beside a predictor and multiplies it. The coefficient tells us how the fitted outcome changes for a one-unit increase in that predictor, with the other predictors held fixed. But when calculating a fitted value for a particular profile, the predictor contributes its coefficient multiplied by its value. Those are related quantities, not the same quantity.

This distinction matters when reading an equation or explaining a prediction. A coefficient of 3 does not automatically add 3 outcome units to every fitted value. If its predictor is 4, the predictor term is 12; if the predictor is 0, the term is 0. The coefficient remains 3 in both cases.

The Coefficient and Its Term

Consider an additive regression model with one outcome \(Y\) and predictors \(X_1,\ldots,X_k\). Its fitted equation is

Here, \(\hat{y}\) is the fitted outcome for a profile with predictor values \(x_1,\ldots,x_k\). The intercept estimate is \(\hat{\beta}_0\), and \(\hat{\beta}_j\) is the estimated coefficient for predictor \(X_j\). The term for \(X_j\) at a particular profile is \(\hat{\beta}_j x_j\). It is this product—not the coefficient alone—that appears as that predictor’s summand in the fitted value.

The word “contribution” here has a precise, limited meaning: it names one summand in the model’s fitted-value calculation. The fitted value is the sum of the intercept and all predictor terms. It is not necessarily the observed outcome, and a term contribution should not be treated as a causal effect or as a stand-alone measure of predictor importance.

For example, if \(\hat{\beta}_1=3\), then increasing \(x_1\) by one unit adds 3 to the fitted value, with other predictors unchanged. At \(x_1=4\), however, the term is \(3(4)=12\). At \(x_1=0\), it is \(3(0)=0\). The coefficient is the same in each calculation; the term depends on the value being multiplied.

A Term Contribution Is Not the Change Between Profiles

A term contribution describes one predictor’s summand at a given value. A change in fitted outcome compares two profiles. For a change in one predictor, the comparison depends on the difference between its two values.

To see why, hold all predictors other than \(X_j\) fixed. Let the original value be \(x_j\), and let the new value be \(x_j+\Delta\). In the original fitted value, the \(X_j\) term is \(\hat{\beta}_j x_j\). In the new fitted value, it is \(\hat{\beta}_j(x_j+\Delta)\). Subtracting the original fitted value from the new one cancels the unchanged terms:

Thus the coefficient gives the fitted change for a one-unit increase, while \(\hat{\beta}_j\Delta\) gives the fitted change for a \(\Delta\)-unit increase. Neither quantity should be confused with the term \(\hat{\beta}_j x_j\) at one profile.

Term Contributions for Indicators and Multiple Predictors

The same multiplication rule applies to a zero-one indicator. If \(d\) is 0 or 1 and its coefficient is \(\hat{\beta}_d\), its term contribution is \(\hat{\beta}_d d\). For the reference category coded 0, the term is zero; for the category coded 1, the term equals the coefficient. Comparing those categories at the same numeric predictor values gives a fitted difference of \(\hat{\beta}_d\). That contrast is a comparison of fitted values, while the term itself is the summand for one profile.

With several predictors, each term is calculated separately and then combined with the intercept. A positive coefficient does not guarantee a positive term if the predictor value is negative; a negative coefficient does not guarantee a negative term if the predictor value is negative. The sign of the product depends on both factors.

What the Calculation Assumes—and What It Does Not Say

The arithmetic above applies to a fitted additive equation: its right-hand side is the intercept plus the displayed coefficient-predictor products. When the equation is used to describe a conditional mean, the model states that the expected outcome at predictor values \(x_1,\ldots,x_k\) is represented by that additive expression. In a common error formulation, the observed outcome is this conditional mean plus an error with conditional mean zero. The fitted value is the model-based prediction; an individual observed outcome may differ from it.

Interpreting \(\hat{\beta}_j\) as a fitted change requires comparing profiles that differ in \(X_j\) while holding the other predictors fixed, as well as using the same fitted equation and predictor coding. The algebra establishes a model-based association, not a causal effect. A term contribution is likewise a property of the equation at the specified profile, not a claim about what would happen if a person or object were changed in the real world.

A term’s numerical size alone is not a universal measure of importance. It depends on the predictor value at which it is calculated and on how that predictor is represented in the equation. In particular, a zero predictor value produces a zero term, but this does not imply that its coefficient is zero or that the predictor is irrelevant to changes in fitted outcomes.

Common Mistake

A frequent error is to read a coefficient of 2.5 as saying, “This predictor adds 2.5 points to the fitted score for every person.” That confuses the per-unit coefficient with the term at a particular value. In the study-time example, the coefficient is 2.5 points per hour; for a student with 4 hours, the term is 10 points. Another error is to call that 10-point term the change from 3 hours to 4 hours. The change is only \(2.5(4-3)=2.5\) points.