Decomposing a Fitted Value Into Contributions

Why Break a Prediction Into Parts?

A fitted value from a multiple regression equation is a total built from several terms. Seeing only the final number can hide how the equation produced it. Separating the total into the intercept and the predictor-term contributions makes the arithmetic transparent: each term can be checked, its sign can be understood, and the pieces can be added back to verify the prediction.

This decomposition is an accounting identity for an additive equation. It does not mean that each term is a separate observed quantity, nor does it establish that a predictor causes a change in the outcome. Each contribution depends on the fitted coefficient and the predictor value used in the equation.

The Additive Model and Its Contributions

For a response \(Y_i\) and \(k\) numeric predictors \(x_{i1},\ldots,x_{ik}\), an additive regression model can be written as

Here \(\beta_0,\ldots,\beta_k\) are model coefficients and \(\varepsilon_i\) is the error for observation \(i\). The additive conditional-mean form says that the model represents the mean response as an intercept plus a separate coefficient-times-predictor term for each predictor. A standard condition for this interpretation is that the errors have conditional mean zero given the predictors: \(E(\varepsilon_i\mid x_{i1},\ldots,x_{ik})=0\). The arithmetic decomposition below uses the fitted equation itself; it does not require a normal error distribution.

After fitting the model, replace each unknown coefficient by its estimate. For a specified predictor profile, the fitted value is the sum of the estimated intercept and the estimated predictor terms.

The contribution \(\hat{\beta}_jx_{ij}\) is the term as it appears in the fitted equation. It is not the same thing as the slope \(\hat{\beta}_j\): the slope has units of response per predictor unit, while the contribution has units of response. For example, if a slope is measured in score points per study hour, multiplying it by a number of study hours produces a contribution measured in score points.

The intercept contribution is the fitted equation’s value when all predictors equal zero. Whether that profile is realistic is a separate question. Even when zero is not a relevant or observed profile, the intercept remains one of the equation’s algebraic terms. Likewise, a predictor contribution is not necessarily a meaningful “share” of the prediction. Contributions may be negative, and their magnitudes depend on how predictors are measured and where zero is located.

A Reliable Way to Decompose a Fitted Value

Use the fitted equation for the particular predictor profile. Keep the intercept as one contribution, calculate each coefficient-times-value product, and then add the terms. A final check is to compare the sum with the original substituted equation. Keeping units beside the calculations helps catch errors such as multiplying a slope by the wrong predictor or omitting a term.

Positive and Negative Contributions

A contribution’s sign follows from the product of the estimated coefficient and the predictor value. When the predictor value is positive, a positive coefficient gives a positive contribution and a negative coefficient gives a negative contribution. A negative contribution subtracts from the sum; it does not mean the entire fitted value must be negative. The intercept and the other terms may outweigh it.

The predictor’s zero point matters. If the predictor is recorded as a positive distance from a chosen origin, its contribution is calculated from that value and the fitted coefficient. A coefficient describes the fitted change in the response associated with a one-unit increase in that predictor while the other predictors are held fixed; a term contribution instead reports the product used in the fitted-value equation at a particular profile. These are related, but they answer different questions.

Why the Contributions Add Up

No special statistical theorem is needed for the check: the fitted value is defined by substituting predictor values into the fitted equation. Each predictor term is a summand in that equation. Naming those summands as contributions changes how the equation is displayed, not its value.

where \(C_0=\hat{\beta}_0\) and \(C_j=\hat{\beta}_jx_{ij}\) for \(j=1,\ldots,k\). Replacing each \(C_j\) by its definition recovers the original fitted equation exactly. This identity also explains the practical verification: if the contributions do not sum to the directly calculated prediction, then a term was omitted, a value was paired incorrectly, or there was an arithmetic or rounding error.

Rounding and What a Contribution Does Not Say

A contribution can be computed from displayed coefficient estimates, which may already be rounded. If products are rounded too early, their sum may differ slightly from a total calculated with more precise values. Keep extra digits during intermediate calculations when available, add the unrounded terms, and round the final fitted value to a precision appropriate to the outcome. If only rounded coefficients are supplied, say that the resulting prediction is approximate.

Contributions also should not be presented automatically as percentages of a fitted value. A ratio such as \(C_j/\hat{y}\) can be unstable or misleading, especially when contributions have different signs or offset each other. In general, state contributions in the response’s units and report the complete prediction. A term’s size reflects both its coefficient and the predictor value, and it depends on the units and zero point used for that predictor.

Common Mistake

A second practical error is to report only the products and leave out the intercept. Unless the model was explicitly fitted without an intercept, the intercept is part of the fitted value and must be included in the sum. Another is to treat a coefficient itself as a contribution: a slope in points per hour is not yet a term in points; it must be multiplied by the predictor value in the equation.