Tutorials › AP Statistics › Predicted Change Versus Actual Change

Linear regression models · Tutorial 844 of 1000

Predicted Change Versus Actual Change

See how residuals account for the difference between a regression line’s predicted change and the change observed between data points.

Intermediate 9 min read

What You'll Learn

  • Distinguish a change in predicted response from a change in an observed response.
  • Calculate the change in predicted response for a chosen increase in the predictor.
  • Use residuals to explain why an observed difference may not match the slope.
  • Recognize why comparing two observations does not necessarily describe change within one individual.
  • Avoid interpreting a slope as a guarantee or as proof of a causal effect.

The Slope Describes the Line’s Predictions

In “Interpreting the Slope in Context,” you learned that the slope \(b\) describes how the predicted response \(\hat{y}\) changes when the predictor \(x\) increases. The word predicted matters. The slope describes the regression line’s predictions; it does not promise that every observed response will change by that amount.

For example, a slope of 4 points per practice hour means the line predicts a response that is 4 points higher for each additional hour of practice. It does not mean every student who practices one more hour will score exactly 4 points higher. Individual observations can fall above or below the line, as discussed in “Population Model Versus Sample Regression Line” and “Reading the Equation of a Regression Line.”

Key idea: The slope gives the change in the fitted line’s predicted response, not a guaranteed change in the observed response for any particular case.

The distinction also matters when the predictor increases by more than one unit. If the slope is \(b\) and the predictor increases by \(\Delta x\) units, the predicted response changes by \(b\Delta x\). That is a change in the line’s predictions. The actual responses may differ from those predictions.

$$ \text{change in predicted response} = b\Delta x $$

Recall from “Population Model Versus Sample Regression Line” that the residual for an observation is its actual response minus the response predicted by the line. A positive residual means the observation is above the line; a negative residual means it is below. Because actual responses include their residuals, comparing two observations involves both the change in predictions and the change in residuals.

$$ y=\hat{y}+\text{residual} $$

Subtracting this expression for one observation from the expression for another gives a useful relationship: the difference between their observed responses equals the difference between their predicted responses plus the difference between their residuals. So the observed difference matches the predicted difference only when the two residuals are the same.

$$ y_2-y_1 = (\hat{y}_2-\hat{y}_1) + (\text{residual}_2-\text{residual}_1) $$

This equation helps explain the word “average” in an informal description such as “the slope describes the average predicted change.” The regression line gives a typical predicted response at each \(x\)-value. Individual observations vary around that line. The slope describes how the line’s predictions change, not the exact change for every case.

Predicted Differences and Observed Differences

Suppose a fitted line is \(\hat{y}=a+bx\). At predictor value \(x_1\), the predicted response is \(a+bx_1\). At \(x_2\), the predicted response is \(a+bx_2\). Subtracting the predictions cancels the intercept, leaving \(b(x_2-x_1)\). Thus, the slope determines the predicted difference for any chosen increase in \(x\).

$$ \hat{y}_2-\hat{y}_1 = (a+bx_2)-(a+bx_1) = b(x_2-x_1) $$

But the actual responses are not required to lie on the line. If one observation has a positive residual and another has a negative residual, their observed difference can be smaller than the predicted difference. If their residuals move in the opposite direction, the observed difference can be larger. The observed difference can even have a different direction from the predicted difference.

There is another important caution: two data points with different \(x\)-values are often two different individuals or cases. Comparing their observed responses does not necessarily tell us how one individual would change if that individual’s \(x\)-value changed. A regression slope describes the pattern of predictions across \(x\)-values; by itself, it is not a guarantee about within-person change or evidence that changing \(x\) causes a change in \(y\). As in “Why Correlation Does Not Imply Causation,” a causal claim requires suitable evidence from the study design.

Worked Examples: Predictions Are Not Guarantees

Worked Example: Practice Hours and Test Scores

A fictional school counselor fits the line \(\hat{y}=62+4x\) to describe the association between weekly practice hours, \(x\), and test score, \(y\), in points, for students. What increase in score does the line predict between 2 and 3 hours? Compare that with two students’ observed scores.

The predictor increases from \(x_1=2\) to \(x_2=3\), so \(\Delta x=1\) hour. The slope is \(b=4\) points per hour. The predicted difference is:

$$ b\Delta x=4(3-2)=4\text{ points} $$

The line predicts 70 points at 2 hours and 74 points at 3 hours:

$$ \hat{y}_1=62+4(2)=70, \qquad \hat{y}_2=62+4(3)=74 $$

Suppose a student who practices 2 hours scores 73, while a different student who practices 3 hours scores 69. Their residuals are \(73-70=3\) and \(69-74=-5\) points. The observed difference in their scores is \(69-73=-4\) points. The residuals explain the gap between that observed difference and the line’s predicted difference:

$$ (\hat{y}_2-\hat{y}_1)+(\text{residual}_2-\text{residual}_1) = (74-70)+(-5-3) = 4-8 = -4\text{ points} $$

The regression line predicts a 4-point increase over this one-hour difference, but these two students’ observed scores differ by 4 points in the opposite direction. This does not mean the line’s slope is a guarantee, nor does it tell us what would happen if one student practiced an additional hour: the observations are from two different students.

Worked Example: Fertilizer and Plant Yield

A fictional greenhouse project models plant yield, \(y\), in grams, from fertilizer amount, \(x\), in grams, using \(\hat{y}=18+2.5x\). For two plants, one receives 4 grams and yields 31 grams; another receives 10 grams and yields 37 grams. Compare the predicted and observed differences.

The predictor difference is \(10-4=6\) grams. The line’s predicted change is:

$$ b\Delta x=2.5(10-4)=15\text{ grams} $$

At 4 grams of fertilizer, the predicted yield is \(18+2.5(4)=28\) grams. At 10 grams, it is \(18+2.5(10)=43\) grams. The observed yield at 4 grams is 31 grams, so that plant’s residual is \(31-28=3\) grams. The observed yield at 10 grams is 37 grams, so the other plant’s residual is \(37-43=-6\) grams.

The observed difference is \(37-31=6\) grams. Using the residual relationship gives the same result:

$$ (43-28)+(-6-3) = 15-9 = 6\text{ grams} $$

The fitted line predicts a 15-gram difference between the two fertilizer amounts, while the two plants’ actual yields differ by 6 grams. Their residuals are not the same, so the observed difference does not equal the predicted difference. This comparison describes these two plants; it does not prove that adding fertilizer caused a particular plant’s yield to change by 15 grams.

Worked Example: Manufacturing Output and Cost

A fictional workshop models daily cost, \(y\), in dollars, using the number of items made, \(x\), with the line \(\hat{y}=120+7x\). One day with 10 items has an observed cost of $196, and another day with 12 items has an observed cost of $220. Find the predicted and observed differences, then explain why they differ.

The predictor increases by \(12-10=2\) items. With a slope of 7 dollars per item, the predicted cost difference is:

$$ b\Delta x=7(2)=14\text{ dollars} $$

The predicted costs are \(120+7(10)=190\) dollars and \(120+7(12)=204\) dollars. The first residual is \(196-190=6\) dollars. The second is \(220-204=16\) dollars. The observed cost difference is \(220-196=24\) dollars, which is 10 dollars larger than the predicted difference because the residual increased by 10 dollars.

$$ (204-190)+(16-6) = 14+10 = 24\text{ dollars} $$

For each additional item, the line’s predicted cost increases by $7. Across two items, that makes a predicted increase of $14. The observed costs on these two days differ by $24 because the second day’s cost is farther above its prediction. The slope does not require every two-item difference in observed costs to equal $14.

Worked Example: A Different Direction in the Observations

A fictional water-monitoring team models clarity, \(y\), in centimeters, from time after a filter change, \(x\), in days, using \(\hat{y}=30+1.2x\). At day 5, one sample has clarity of 34 centimeters; at day 6, another sample has clarity of 32 centimeters. Does this contradict the positive slope?

The fitted line predicts an increase of \(1.2(6-5)=1.2\) centimeters between days 5 and 6. The predictions are \(30+1.2(5)=36\) centimeters and \(30+1.2(6)=37.2\) centimeters. The first residual is \(34-36=-2\) centimeters; the second is \(32-37.2=-5.2\) centimeters.

The observed difference is \(32-34=-2\) centimeters, a decrease. The residual difference is \(-5.2-(-2)=-3.2\) centimeters, so:

$$ 1.2+(-3.2)=-2\text{ centimeters} $$

There is no contradiction. The slope describes the line’s predicted direction, while individual observations can depart from that pattern. These samples are not necessarily repeated measurements of the same water under identical conditions, and the fitted association alone does not establish that the filter caused clarity to change.

Common Mistakes and AP Exam Tips

  • Calling the slope an exact change for each case. “Every student gains 4 points for another hour of practice” overstates what the line says. A full-credit interpretation makes clear that the predicted score increases by 4 points per additional practice hour.
  • Confusing a predicted difference with an observed difference. The slope times the predictor change gives the difference between predictions, not necessarily between actual responses. If a question supplies observed responses, calculate their difference separately.
  • Ignoring residuals when explaining a mismatch. The residual is actual response minus predicted response. A difference in residuals accounts for the difference between observed and predicted changes; it is not an unexplained arithmetic error.
  • Treating two cases as one individual over time. If the observations come from different people, plants, days, or locations, their difference is not automatically a change within one individual. Identify what the cases represent before making a claim.
  • Turning a regression pattern into a causal claim. A positive slope does not, by itself, show that increasing the predictor causes the response to increase. Describe the fitted line’s predictions unless the study design supports a causal interpretation.

For a clear AP response, separate the two quantities: state the predicted change as \(b\Delta x\), then state the observed difference if it is requested. If they differ, compare the residuals. Use the variables and units in context, and avoid claiming that the slope guarantees an individual outcome.

Key takeaway: A regression slope describes how the line’s predicted response changes as the predictor changes. Actual responses can differ because observations have residuals, so the slope is not a guaranteed change for each individual.

Check Your Understanding

Use the line and residual ideas from this tutorial to answer each question.

  1. A fitted line is \(\hat{y}=20+3x\). What change in predicted response does it give when \(x\) increases from 4 to 7?
  2. For an observation, \(y=48\) and \(\hat{y}=52\). Find its residual and state whether the point is above or below the line.
  3. A line predicts 10 units at \(x=2\) and 16 units at \(x=5\). The actual responses are 12 and 14 units, respectively. Find the predicted difference, the observed difference, and both residuals.
  4. Why does a positive slope not guarantee that every pair of observations with a larger \(x\)-value will have a larger observed \(y\)-value?
  5. If two data points represent different people, why should their response difference not automatically be described as the change one person would experience?