From a Predictor Change to an Outcome Change
A regression equation can give a fitted outcome for any specified set of predictor values. Often the question is not the fitted value at one profile, but how the fit changes when one predictor changes. For example, how much does a model’s fitted daily energy use change when the outside temperature rises by five degrees, if the number of occupants stays the same?
For an additive equation, the answer follows directly from the coefficient of the changing predictor. Multiply that coefficient by the amount of predictor change. The sign matters: a positive slope and a negative predictor change produce a negative fitted change, while a negative slope and a negative predictor change produce a positive fitted change.
The Model and the Conditions for the Calculation
Consider an additive regression model with \(p\) predictors. Its fitted equation, using estimated coefficients, is
Here, \(\hat{y}\) is the fitted outcome, \(\hat{\beta}_0\) is the estimated intercept, and \(\hat{\beta}_k\) is the estimated coefficient for predictor \(x_k\). In this equation, each predictor contributes its coefficient multiplied by its value. “Additive” means these contributions are added, with no products or other terms that make one predictor’s contribution depend on another predictor’s value.
To use the change rule, compare two profiles evaluated by the same fitted equation. The changing predictor must be numeric; its change is new value minus old value. All other predictors must have the same values in both profiles. The equation must contain the changing predictor as an additive term, without an interaction involving it. The two profiles should also be meaningful inputs for the model: arithmetic can produce a fitted value outside the range of the observed outcomes, but that does not guarantee the prediction is useful.
These are conditions for the algebraic comparison, not a complete list of assumptions for statistical inference or for using a regression model to make reliable predictions. The change rule calculates a difference between fitted values. It does not, by itself, show that changing the predictor causes the outcome to change.
Why the Other Terms Cancel
Suppose predictor \(x_j\) changes from \(x_{j,\text{old}}\) to \(x_{j,\text{new}}\). Let \(\hat{y}_{\text{old}}\) be the fit at the original profile and \(\hat{y}_{\text{new}}\) the fit at the new profile. Subtract the original fitted value from the new one. The intercept is identical in both calculations, and each unchanged predictor has the same value in both. Those terms cancel, leaving only the contribution from \(x_j\):
This short derivation explains both the calculation and its limits. The fitted change is not necessarily equal to the coefficient alone: it equals the coefficient times the size and direction of the change. It equals the coefficient only when the predictor increases by exactly one unit. Reversing the order of subtraction reverses the sign of the fitted change.
Worked Example: A Rise in Outside Temperature
Suppose a hypothetical fitted model for daily household energy use is \(\hat{y}=40-0.6T+8N\), where \(\hat{y}\) is fitted use in kilowatt-hours per day, \(T\) is outside temperature in degrees Celsius, and \(N\) is the number of occupants. Calculate the change in fitted use when temperature rises from \(10\) to \(15\) degrees Celsius for a household with three occupants.
Check the comparison. Both profiles use the same equation, and \(N=3\) in each. Temperature is a numeric predictor with an additive term, so the calculation applies.
The predictor change is \(15-10=5\) degrees Celsius. The estimated coefficient of temperature is \(-0.6\) kilowatt-hours per day per degree Celsius. Therefore,
As a check, the original fit is \(40-0.6(10)+8(3)=58\) kWh/day. The new fit is \(40-0.6(15)+8(3)=55\) kWh/day, and \(55-58=-3\) kWh/day. According to this fitted equation, the higher temperature corresponds to a fitted daily energy use that is 3.0 kWh lower for a household with three occupants.
Reading the Sign and Size of the Change
The sign of \(\hat{\beta}_j\Delta x_j\) gives the direction of the fitted change under the stated subtraction order. Its magnitude gives the size of that change in the outcome’s units. The predictor can increase or decrease; it is important to use new value minus old value rather than treating every change as positive.
The coefficient describes the fitted change for a one-unit increase in the predictor, with the other predictors fixed. For a change of several units, the fitted change scales proportionally in an additive linear equation. For instance, a slope of \(3\) outcome points per study hour gives a fitted increase of \(3(1.5)=4.5\) points for an increase of \(1.5\) hours. This multiplication does not turn the coefficient into a prediction: the intercept and other predictor terms are needed to calculate a fitted value at a particular profile.
Worked Example: Increasing Study Time
A hypothetical fitted equation for a quiz score is \(\hat{y}=78+3S-2.5A\), where \(S\) is weekly study time in hours and \(A\) is the number of absences. For a student with two absences, calculate the fitted score change when study time rises from 4 to 5.5 hours per week.
The other predictor, absences, stays at \(2\), and study time appears as an additive numeric term. The change in study time is \(5.5-4=1.5\) hours. The coefficient of study time is \(3\) score points per hour, so
The original fitted score is \(78+3(4)-2.5(2)=85\). The new fitted score is \(78+3(5.5)-2.5(2)=89.5\). Their difference is \(89.5-85=4.5\) points, as calculated. The interpretation concerns the model’s fitted scores at these profiles, not a promise that an individual student’s observed score will rise by 4.5 points.
A Decrease Can Produce an Increase
A negative predictor change does not necessarily mean a negative fitted change. If the slope is negative too, their product is positive. This is why writing the subtraction order and the signed predictor change explicitly is safer than relying on verbal phrases such as “a change of three units.”
Worked Example: Moving Closer to a Location
Consider the hypothetical fitted equation \(\hat{y}=900-25D+100B\), where \(\hat{y}\) is fitted monthly rent in dollars, \(D\) is distance from a city center in kilometers, and \(B\) is the number of bedrooms. For a two-bedroom apartment, calculate the fitted rent change when distance decreases from 8 km to 5 km.
The changing predictor is distance, and bedrooms stays fixed at \(2\). The distance change is \(5-8=-3\) km. The distance coefficient is \(-25\) dollars per month per kilometer, giving
Checking with fitted values, at 8 km the fit is \(900-25(8)+100(2)=900\) dollars per month. At 5 km it is \(900-25(5)+100(2)=975\) dollars per month. The fitted change, new minus old, is \(975-900=75\) dollars per month. Under this equation, the closer apartment has the higher fitted rent for the fixed number of bedrooms.
When the Simple Rule Needs More Than One Term
The coefficient-times-change rule as stated relies on an additive term for the changing predictor. If the equation includes an interaction, the effect of changing a predictor can depend on another predictor’s value. For example, in an equation containing both \(x\) and the product \(xz\), changing \(x\) changes both terms. Subtracting the two fitted values then requires accounting for both contributions; using only the coefficient of \(x\) would be incomplete.
Similarly, if several predictors change, the fitted difference includes a contribution from each changed term. For an additive equation, it is the sum of the coefficient-times-change calculations for those predictors. The focused one-predictor comparison is especially useful because it isolates that predictor’s fitted contribution while holding the rest of the profile constant.
Common Mistakes and Why This Calculation Matters
- Reporting the coefficient as the change for any size of shift. A coefficient is the fitted outcome change per one predictor unit. For a shift of \(\Delta x_j\), multiply by that shift: \(\hat{\beta}_j\Delta x_j\).
- Dropping the minus sign on a decrease. If a predictor moves from 8 to 5, its change is \(5-8=-3\), not \(3\). Keep the order consistent with new fit minus old fit.
- Changing another predictor without noticing. The one-predictor rule compares profiles with all remaining predictors fixed. If another value changes, its contribution must also be included.
- Calling a fitted difference an observed or causal difference. The calculation compares model fits. It does not guarantee the observed outcomes differ by that amount, and it does not establish that intervening on the predictor would cause the change.
- Using the slope as a fitted value. A slope is a rate of fitted change per predictor unit. A fitted value at a profile requires substituting the profile into the whole equation.
This calculation matters because it turns a coefficient into a comparison that answers a specific question: how does the model’s fitted outcome differ between two profiles that differ in one numeric predictor? Stating the original and new values, the fixed predictors, the change direction, and the outcome units makes that comparison clear and checkable.
Check Your Understanding
For each question, keep track of the order of subtraction and state the fitted change in context.
- A model has coefficient \(2.4\) outcome units per hour for study time. What fitted change corresponds to an increase from 3 to 5 hours, with all other predictors fixed?
- A fitted equation has coefficient \(-7\) dollars per kilometer for distance. What is the fitted rent change when distance increases from 2 to 6 km, all else fixed?
- For the previous equation, what is the fitted rent change when distance decreases from 6 to 2 km? Explain why its sign differs from the change for an increase.
- Why does the intercept not appear in the formula for the fitted change when only one predictor changes in an additive equation?
- A model includes an interaction between a predictor \(x\) and another predictor \(z\). Why might the coefficient of \(x\) alone not give the fitted change for a change in \(x\)?