Putting Residual Skills Together
This practice set combines skills from “Calculating a Residual by Hand,” “Sign of a Residual: Over- and Underprediction,” “Reading a Residual Plot for Random Scatter,” and “Using s to Describe Prediction Accuracy.” The goal is to move smoothly among a calculation, an interpretation for an individual case, a description of a plot, and a summary of the model’s typical prediction-error size.
For an observed case, the residual is \(y-\hat{y}\): the observed response minus the response predicted by the fitted line at that case’s predictor value. The sign describes direction; the absolute value describes the size of the vertical difference. As in “Standard Deviation of the Residuals, s,” the residual standard deviation \(s\) summarizes the typical size of residuals around the line, in response-variable units.
Use the question to decide which pieces are needed. A request to interpret one residual does not require a discussion of the entire model. A request to assess a residual plot requires describing the pattern across the horizontal range, not just naming one point. And a value of \(s\) summarizes the model’s typical error size; it does not tell you the direction of any one error.
A Quick Audit for Mixed Questions
A useful way to check a response is to move from the calculation outward. First verify the prediction at the observed \(x\). Then check that the subtraction is observed minus predicted. Finally, match the sign, units, and conclusion to the situation. For plot questions, add a separate check: describe the residuals across the range and compare their spread as well as their positions relative to zero.
Find \(\hat{y}\) at the case’s observed predictor value. Then calculate \(y-\hat{y}\).
A positive residual must mean the observed response exceeds the prediction; a negative residual must mean it is less.
Use the residual’s sign for one case, absolute residuals to compare error sizes, and the plot’s overall structure to assess patterns.
Residuals and \(s\) use the response variable’s units. State what the result means for the named response.
Worked Example: Calculate, Interpret, and Compare Residuals
A fictional delivery service uses distance from its depot \(x\), in kilometers, to predict delivery time \(y\), in minutes. Its fitted line is \(\hat{y}=8+3.2x\). One delivery was 5 km away and took 25.1 minutes. Another was 8 km away and took 30.0 minutes. Calculate and interpret both residuals, then identify which delivery was predicted more accurately.
State. Find each delivery’s residual, explain whether the line overpredicted or underpredicted, and compare the sizes of the two errors.
Plan. Substitute each observed distance into the fitted line. For each case, subtract the predicted time from the observed time. Compare absolute residuals to determine which prediction was closer.
Do. For the 5-km delivery, the predicted time is \(\hat{y}=8+3.2(5)=8+16=24\) minutes. Its residual is \(y-\hat{y}=25.1-24=1.1\) minutes. Because this residual is positive, the observed delivery took longer than predicted: the line underpredicted by 1.1 minutes.
For the 8-km delivery, the predicted time is \(\hat{y}=8+3.2(8)=8+25.6=33.6\) minutes. Its residual is \(y-\hat{y}=30.0-33.6=-3.6\) minutes. Because this residual is negative, the observed delivery took less time than predicted: the line overpredicted by 3.6 minutes.
The absolute residuals are \(|1.1|=1.1\) minutes and \(|-3.6|=3.6\) minutes. The first delivery’s prediction was closer, because its absolute residual is smaller.
Conclude. The model underpredicted the 5-km delivery time by 1.1 minutes and overpredicted the 8-km delivery time by 3.6 minutes. It predicted the first delivery more accurately, based on the smaller absolute residual.
Reading a Plot and Connecting It to a Case
A residual plot gives information about errors across many observations. Each vertical position is a residual, and the horizontal axis is usually the predictor \(x\) or the predicted response \(\hat{y}\). A pattern of points on both sides of zero with a vertical spread that changes across the horizontal range is different from a curve: a fan shape concerns how widely residuals vary, while a curve concerns their organized direction.
When reading a plot, describe both features if they are present. For example, residuals may be above and below zero while their spread increases as \(x\) increases. That suggests changing variability in prediction errors, as discussed in “Fan-Shaped Residual Plots and Changing Spread.” It is not enough to say only that some residuals are positive and some negative.
Worked Example: Read a Widening Residual Plot
A fictional library models checkout wait time \(y\), in minutes, from the number of customers waiting \(x\). A residual plot has \(x\) on the horizontal axis. At smaller queue sizes, the residuals are close to zero, generally between about \(-0.5\) and \(0.5\) minute. At larger queue sizes, the residuals appear both above and below zero, with several between about \(-3\) and \(3\) minutes. One plotted case has \(x=8\) and a residual of \(3.2\) minutes. For that case, the fitted line predicts 11.2 minutes. Describe the plot and interpret the case.
State. Describe the overall residual pattern and explain what the residual of 3.2 minutes means for the case with eight customers waiting.
Plan. Compare the plot’s vertical spread at smaller and larger queue sizes, and note whether residuals occur on both sides of zero. For the individual case, use the residual definition and identify the response units.
Do. The vertical spread is narrower for smaller queue sizes and wider for larger queue sizes. Residuals occur both above and below zero, but their distance from zero tends to increase as the queue gets larger. This is a widening, fan-shaped pattern. For the individual case, the observed wait is the prediction plus the residual: \(y=\hat{y}+\text{residual}=11.2+3.2=14.4\) minutes. The observed wait is 3.2 minutes longer than predicted.
Conclude. The plot suggests that prediction errors vary more in size for larger queues, so the model’s prediction accuracy is less consistent across the range. For the case with eight customers waiting, the line underpredicted the wait by 3.2 minutes.
The plot does not show that every prediction for a large queue is poor: the residuals are on both sides of zero. It shows that their spread is larger there. That distinction keeps the conclusion tied to the visible evidence.
Using Residuals to Understand \(s\)
An individual residual tells you the direction and size of one prediction error. The residual standard deviation \(s\) summarizes the typical size of residuals for the fitted line. In simple linear regression, it is calculated from the sum of squared residuals, SSE, using \(n-2\) degrees of freedom:
As in “Using \(s\) to Describe Prediction Accuracy,” interpret \(s\) in the response variable’s units: predictions typically differ from observed responses by about \(s\) units. This is a summary, not a promise that every residual is within \(s\), and it does not say whether any particular prediction is too high or too low. Read it alongside the individual residuals and the residual plot.
Worked Example: Calculate \(s\) and Read the Residuals Together
A fictional music program uses weekly practice time \(x\), in minutes, to predict a performance score \(y\), in points. Its fitted line is \(\hat{y}=42+0.4x\). The following table gives six observations and their predictions. Use the residuals to calculate \(s\), and briefly describe their pattern.
| Practice time, \(x\) (min) | Observed score, \(y\) (points) | Predicted score, \(\hat{y}\) (points) | Residual, \(y-\hat{y}\) (points) |
|---|---|---|---|
| 20 | 51 | 50 | 1 |
| 40 | 56 | 58 | -2 |
| 60 | 67 | 66 | 1 |
| 80 | 75 | 74 | 1 |
| 100 | 80 | 82 | -2 |
| 120 | 91 | 90 | 1 |
State. Calculate the residual standard deviation for these six observations and interpret it in performance-score points.
Plan. Confirm the residuals by subtracting each prediction from its observed score. Square and add the residuals to find SSE, then use \(n-2=6-2\) in the formula for \(s\). Also consider whether the listed residuals show an obvious organized pattern across practice times.
Do. The residuals are \(1,-2,1,1,-2,1\), each in points. Their squared values sum to \(\text{SSE}=1^2+(-2)^2+1^2+1^2+(-2)^2+1^2=1+4+1+1+4+1=12\) points squared. Therefore,
The residuals are above and below zero across the practice-time range, with no obvious curve or steadily changing spread in these six values. The plot would be needed to judge their arrangement visually, but the listed values do not alone show a clear systematic pattern.
Conclude. The residual standard deviation is about 1.73 performance-score points. In this context, predictions from the line typically differ from observed scores by about 1.73 points. This describes a typical error size, not a guarantee for every student or the direction of any particular error.
Common Mistakes and AP Exam Tips
- Stopping at the number. A residual calculation needs an interpretation if the question asks what it means. State whether the model overpredicted or underpredicted the named response, and use response units.
- Using the sign to compare accuracy. A negative residual is not automatically a worse prediction than a positive one. Compare absolute residuals: the smaller absolute value is the closer prediction.
- Confusing one residual with a plot pattern. A single positive residual means underprediction for one case. It does not establish a trend across the full predictor range.
- Describing only which side of zero points occupy. Also check for curves and changes in vertical spread. A fan shape can show changing error variability even when residuals appear on both sides of zero.
- Treating \(s\) as an individual guarantee. Say that predictions typically differ from observed responses by about \(s\) response units. Do not claim that every residual must be less than \(s\) in magnitude.
- Dropping units or context. Residuals and \(s\) have the response variable’s units, not the predictor’s. If the response is a score, report points; if it is a wait time, report minutes.
A full-credit mixed response is specific about the part being asked. Show the fitted value and subtraction for a residual; translate its sign into the direction of the prediction error; and use absolute values when comparing prediction closeness. For a plot, describe the pattern across the horizontal range and connect it to model errors. For \(s\), report the response units and explain the typical error scale without turning it into a guarantee.
Check Your Understanding
For each item, show the requested calculation or explanation and include context and response units where appropriate.
- A fitted line predicts a 16.8-minute wait for a customer. The actual wait is 14.5 minutes. Calculate the residual and interpret its sign.
- Two cases have residuals of \(-1.4\) points and \(2.1\) points. Which case was predicted more accurately? Explain how you decide.
- A residual plot has residuals on both sides of zero, but their vertical spread grows as the predictor increases. Describe the pattern and what it suggests about prediction errors.
- A regression has \(n=8\) observations and \(\text{SSE}=54\) response-units squared. Calculate \(s\) and state what it means in context, using “response units” if no specific context is given.
- Why does a positive residual for one observation not establish that a regression line underpredicts across the full range?