Calculate Finite-Change Effects Across Predictor Values

Why Compare Finite Changes?

In a regression model with an interaction, the fitted outcome change for increasing one predictor can depend on the value of the other predictor. This means that reporting one change without specifying the other predictor's value may leave out an important part of the interpretation. A useful comparison is to calculate the same fixed increase at two values of the other predictor, then see how much the resulting fitted changes differ.

Here, finite change means the difference between fitted outcomes at two specified predictor values. It is an exact difference in the model's fitted conditional means, not an approximation for a very small change. We will calculate it directly from the interaction equation.

The Interaction Model and Its Assumptions

Let \(x_1\) and \(x_2\) be two quantitative predictors and \(Y\) the quantitative outcome. The product-term regression model for observation \(i\) is

Y_i=\beta_0+\beta_1x_{1i}+\beta_2x_{2i}+\beta_3x_{1i}x_{2i}+\varepsilon_i.

The conditional mean, which is the model's fitted target at fixed predictor values, is

E(Y\mid x_1,x_2)=\beta_0+\beta_1x_1+\beta_2x_2+\beta_3x_1x_2.

The coefficients \(\beta_0,\beta_1,\beta_2,\beta_3\) are fixed model parameters. The error term \(\varepsilon_i\) represents the difference between an observed outcome and its conditional mean. For the mean model, we assume \(E(\varepsilon_i\mid x_{1i},x_{2i})=0\), so that the equation describes the conditional mean. The usual regression setting also treats observations as independent. The finite-change algebra below follows from the stated mean equation; it does not require normal errors or a standard error calculation.

Definition: For a fixed increase \(\Delta x_1\) in \(x_1\), holding \(x_2\) at \(c\), the finite-change effect is the fitted conditional mean at \((x_1+\Delta x_1,c)\) minus the fitted conditional mean at \((x_1,c)\). It is expressed in outcome units and can depend on \(c\).

The order of subtraction matters. In this tutorial, an effect for an increase in \(x_1\) is always the fitted outcome after the increase minus the fitted outcome before it. A positive result indicates a higher fitted outcome after the increase; a negative result indicates a lower fitted outcome.

Deriving the Finite-Change Formula

Write the conditional mean as \(m(x_1,x_2)\). Hold \(x_2=c\) fixed and compare the two fitted means. The terms that do not change with \(x_1\) cancel:

\begin{aligned} m(x_1+\Delta x_1,c)-m(x_1,c) &=\big[\beta_0+\beta_1(x_1+\Delta x_1)+\beta_2c +\beta_3(x_1+\Delta x_1)c\big]\\ &\quad-\big[\beta_0+\beta_1x_1+\beta_2c+\beta_3x_1c\big]\\ &=\beta_1\Delta x_1+\beta_3c\Delta x_1\\ &=\Delta x_1(\beta_1+\beta_3c). \end{aligned}

The result does not depend on the starting value \(x_1\). It does depend on the fixed value \(c\) of the other predictor. In particular, for a one-unit increase in \(x_1\), the fitted change is \(\beta_1+\beta_3c\). This is the conditional slope described in the earlier material; multiplying it by \(\Delta x_1\) gives the fitted change for an increase of any specified size.

Result: At \(x_2=c\), increasing \(x_1\) by \(\Delta x_1\) changes the fitted conditional mean by \(\Delta x_1(\beta_1+\beta_3c)\). Comparing this increase at \(x_2=a\) and \(x_2=b\), the effect at \(b\) minus the effect at \(a\) is \(\beta_3\Delta x_1(b-a)\).

To verify the comparison result, subtract the finite-change effect at \(a\) from the one at \(b\):

\begin{aligned} &\Delta x_1(\beta_1+\beta_3b) -\Delta x_1(\beta_1+\beta_3a)\\ &\qquad=\Delta x_1\beta_3(b-a). \end{aligned}

Thus, for a fixed increase in \(x_1\), the difference between its fitted effects at two values of \(x_2\) is determined by the product coefficient \(\beta_3\), the size of the increase, and the distance between the two \(x_2\) values. The main-effect coefficient \(\beta_1\) contributes to each effect, but cancels from their difference.

Worked Example: Comparing a Two-Hour Increase

Suppose a hypothetical model predicts exam score \(Y\), in points, from study time \(x_1\), in hours, and class size \(x_2\), measured in units of 10 students. Its fitted equation is

\hat{Y}=40+2.4x_1+1.0x_2-0.3x_1x_2.

Compare the fitted score change for adding two study hours when class size is 20 students (\(x_2=2\)) and when it is 40 students (\(x_2=4\)). These are hypothetical coefficients and settings.

At \(x_2=2\), the change is

\Delta \hat{Y} =2[2.4+(-0.3)(2)] =2(1.8) =3.6\text{ points}.

At \(x_2=4\), the change is

\Delta \hat{Y} =2[2.4+(-0.3)(4)] =2(1.2) =2.4\text{ points}.

The effect at 40 students minus the effect at 20 students is \(2.4-3.6=-1.2\) points. Directly from the comparison formula, the same result is

\beta_3\Delta x_1(b-a) =(-0.3)(2)(4-2) =-1.2\text{ points}.

Under this fitted model, the predicted gain associated with two additional study hours is 1.2 points smaller at 40 students than at 20 students. This is an interpretation of model-based association, not evidence that changing study time or class size would cause that difference.

Calculating by Subtracting Fitted Outcomes

The formula is convenient, but the definition offers a second route: calculate both fitted outcomes and subtract. This route can help check signs and make clear which values are being compared. The starting value of the predictor may change the two predictions, even though it cancels from their difference in this particular model.

Worked Example: Checking With Two Fitted Predictions

Consider a hypothetical fitted model for a health score \(Y\), measured in points, with \(x_1\) denoting weekly activity hours and \(x_2\) denoting age in decades:

\hat{Y}=50+1.5x_1-2x_2+0.2x_1x_2.

At age 30 (\(x_2=3\)), compare the fitted scores for \(x_1=4\) and \(x_1=7\), an increase of three activity hours. First calculate the two predictions:

\begin{aligned} \hat{Y}(4,3)&=50+1.5(4)-2(3)+0.2(4)(3)\\ &=50+6-6+2.4=52.4\text{ points},\\ \hat{Y}(7,3)&=50+1.5(7)-2(3)+0.2(7)(3)\\ &=50+10.5-6+4.2=58.7\text{ points}. \end{aligned}

Subtracting the first fitted score from the second gives \(58.7-52.4=6.3\) points. The finite-change formula gives the same result:

3[1.5+0.2(3)]=3(2.1)=6.3\text{ points}.

The comparison confirms that, at age 30, a three-hour increase corresponds to a 6.3-point increase in fitted health score. It does not say that every individual would experience a 6.3-point change.

Reversing Which Predictor Changes

The same subtraction works if \(x_2\) changes while \(x_1\) is held fixed. For an increase of \(\Delta x_2\) at \(x_1=d\), the fitted change is \(\Delta x_2(\beta_2+\beta_3d)\). Comparing that increase at \(x_1=a\) and \(x_1=b\), the effect at \(b\) minus the effect at \(a\) is \(\beta_3\Delta x_2(b-a)\). The roles of the predictors switch, but the product coefficient governs the difference in the same way.

Worked Example: Increasing the Other Predictor

Suppose a hypothetical model predicts crop yield \(Y\), in kilograms, from irrigation \(x_1\), in tens of liters, and fertilizer \(x_2\), in kilograms:

\hat{Y}=20+1.2x_1+3x_2+0.5x_1x_2.

Compare the fitted yield change for adding 2 kg of fertilizer when irrigation is \(x_1=1\) (10 liters) and \(x_1=3\) (30 liters). At \(x_1=1\), the increase is

\Delta\hat{Y}=2[3+0.5(1)]=2(3.5)=7\text{ kg}.

At \(x_1=3\), the increase is

\Delta\hat{Y}=2[3+0.5(3)]=2(4.5)=9\text{ kg}.

The effect at 30 liters minus the effect at 10 liters is \(9-7=2\) kg. The comparison formula gives

\beta_3\Delta x_2(b-a) =(0.5)(2)(3-1) =2\text{ kg}.

In this fitted equation, the predicted yield change for adding 2 kg of fertilizer is 2 kg larger at 30 liters of irrigation than at 10 liters. The arithmetic describes predictions from the hypothetical model; it does not establish a causal fertilizer effect.

Reading the Sign and Units

The sign of \(\beta_3\) tells how the effect of one predictor changes as the other predictor increases, but the units and comparison direction still matter. In the formula \(\beta_3\Delta x_1(b-a)\), a positive value means the effect of the specified increase in \(x_1\) is larger at \(x_2=b\) than at \(x_2=a\). A negative value means it is smaller. If \(b<a\), the factor \(b-a\) is negative, so the direction of the comparison reverses.

The units also provide a useful check. If \(Y\) is measured in points, \(x_1\) in hours, and \(x_2\) in tens of students, then \(\beta_3\) has units of points per hour per unit of \(x_2\). Multiplying by an increase in hours and a difference in \(x_2\) units gives points. The comparison is a difference between two outcome changes, so it has outcome units, not units of slope.

A nonzero product coefficient means the fitted finite-change effect varies across values of the other predictor. If \(\beta_3=0\), the formula for the difference is zero for any specified \(\Delta x_1\), \(a\), and \(b\): the fitted finite change in \(x_1\) is then the same at both \(x_2\) values. This statement concerns the model's coefficients and fitted mean; it is not a claim about statistical significance.

Common Pitfall: Confusing the Product Coefficient With the Whole Effect

A common mistake is to say, “The effect of increasing \(x_1\) is \(\beta_3\).” That is not what the product coefficient represents. At \(x_2=c\), the fitted change for increasing \(x_1\) by \(\Delta x_1\) is \(\Delta x_1(\beta_1+\beta_3c)\). The product coefficient controls how that change differs across \(x_2\) values; it is not generally the finite-change effect itself.

Another mistake is to report the comparison without stating which value is subtracted from which. “The effects differ by 1.2 points” omits direction. In the first worked example, the effect at 40 students minus the effect at 20 students is \(-1.2\) points, whereas the effect at 20 students minus the effect at 40 students is \(+1.2\) points. State the order and retain units.

Key takeaway: For a fixed increase \(\Delta x_1\), calculate the fitted change at \(x_2=c\) as \(\Delta x_1(\beta_1+\beta_3c)\). The effect at \(b\) minus the effect at \(a\) is \(\beta_3\Delta x_1(b-a)\). Interpret the result in context, with the comparison order and units stated.

Check Your Understanding

Use the interaction model and the finite-change formulas to answer each question. Be sure to keep track of units and comparison order.

For \(m(x_1,x_2)=10+2x_1+4x_2-0.5x_1x_2\), calculate the fitted change for increasing \(x_1\) by 3 when \(x_2=2\).
In the same model, what is the effect at \(x_2=5\) minus the effect at \(x_2=2\) for that three-unit increase in \(x_1\)?
In words, what does it mean if the effect at \(x_2=b\) minus the effect at \(x_2=a\) is negative?
For a fixed increase in \(x_2\), name the model coefficients and predictor values that determine the difference between its fitted effects at two \(x_1\) values.
Why is it incorrect to interpret \(\beta_3\) alone as the fitted outcome change for an increase in \(x_1\)?