
use a residual plot to judge whether a linear model is appropriate, identify influential points, and argue whether an association is causal.
Instruction
A high r2 does not mean a line is the right model. The test is the residual plot: residuals (y − ŷ) plotted against x. If the model fits, the residuals scatter randomly around zero with no pattern. A pattern means the line is missing something systematic.
The heart-disease residuals. For the line 282.43 − 6.12x fitted to Missouri’s age-adjusted heart-disease death rates, 1999-2017:
The residuals are positive from 1999 to 2003 (1999: +18.1), negative from 2004 to 2013 (2009: −15.1), and positive again from 2014 to 2017 (2017: +18.8). That U-shape means the relationship is curved: the rate fell fast at first and then leveled off after about 2012 (it was 193.4 in 2012 and 191.1 in 2017). The line cuts through the curve, too low at both ends and too high in the middle, even though r = −0.948.
Why it matters. Extrapolate the line to 2025 (x = 26): 282.43 − 6.12(26) ≈ 123 deaths per 100,000, about a third below the 2017 rate. The residual pattern warns that the decline had already stalled, so this prediction is not credible. The lesson: always plot the residuals before trusting a prediction, and treat extrapolation with suspicion.
Outliers and influential points. An outlier in regression is a point with a large residual. An influential point is one whose removal would change the slope noticeably. Influential points usually have extreme x-values (high leverage), because the line pivots around (x̄, ȳ) and an end point has the most leverage. Test influence directly: fit the line with and without the point and compare. Never delete a point just because it is inconvenient. Delete it only if it is an error, and report it either way.
Correlation is not causation (A1.DS.A.8). A strong association can arise in several ways:
- Causation: changes in x produce changes in y.
- Common response (a lurking variable): both respond to a third variable. St. Louis and Columbia rainfall correlate at r = 0.777 because the same weather systems reach both. Rain in St. Louis does not cause rain in Columbia.
- Confounding: the effect of x is mixed with the effect of another variable, so they cannot be separated.
- Coincidence: with enough variables, some pairs correlate by chance.
“Year” is correlated with heart-disease deaths, but time itself causes nothing. It stands in for everything that changed over those years, and a scatterplot cannot tell you which of those changes mattered. Only a well-designed experiment (Unit 3) can establish cause and effect directly. With observational data, causal arguments need more: a plausible mechanism, consistency across many studies and places, a dose-response relationship, and alternative explanations ruled out.
In the discourse lab you will take a claim of the form “X causes Y” from a news story or from the data we have used, and argue for or against it. Name at least one lurking variable, and say what study would settle the question.
📖 Read for this lesson
The reading opens over the lesson; close it to come back to exactly where you were. The checks below draw on it.
Formative check
Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.
The heart-disease residuals are positive at both ends and negative in the middle. What should you conclude?
Because rainfall in St. Louis and Columbia has r = 0.777, heavy rain in St. Louis causes heavy rain in Columbia.
Random residual scatter
Tap to revealA linear model is reasonable.
Curved residual pattern
Tap to revealTry a nonlinear model or a transformation.
High-leverage point
Tap to revealExtreme x-value; check influence by refitting without it.
Common response
Tap to revealBoth variables respond to a lurking variable.
Record (or write) a 90-second argument: a news headline says a variable ‘is linked to’ a health outcome. Explain what the correlation does and does not show, name two lurking variables, and describe the experiment or study that would test causation. Then compare your argument with the sample response in this lesson’s reading: say where you agree, and qualify or challenge its proposed lurking variable. A written version of the same length is equally acceptable, and you may also read your argument aloud to a parent or guardian and note one objection they raise.
How confident are you that you can judge a linear model with a residual plot and argue carefully about causation?
Practice
Work these on paper or in your notebook, then open Check your answer. Aim for all of Fluency and Application; try at least one Challenge.
Printable version: this unit’s practice workbook (PDF)
Fluency
Build speed and accuracy with the core skill.
- The heart-disease model is predicted rate = 282.43 − 6.12x (x = years since 1999). The actual 2005 rate was 236.8. Find the residual.
Check your answer
Answer: Predicted ≈ 245.70; residual ≈ −8.90.x = 6. - A residual plot shows a random scatter around 0 with no pattern. What does it suggest?
Check your answer
Answer: A linear model is appropriate. - A residual plot is U-shaped. What does it suggest?
Check your answer
Answer: The relationship is curved; a line is the wrong form. - Classify the explanation: ice-cream sales and drownings rise together because both increase in hot weather.
Check your answer
Answer: Common response (a lurking variable, temperature). - Classify: students who choose to take an SAT prep course score higher, but they also tend to be more motivated, so the effects cannot be separated.
Check your answer
Answer: Confounding. - Which points tend to be most influential in a regression: those with extreme x-values or extreme y-values near x̄?
Check your answer
Answer: Extreme x-values (high leverage).
Application
Use the skill in context. Show your reasoning.
- A line fitted to Missouri’s unintentional-injury death rate, 1999-2017 (CDC), has r = 0.809. Its residuals are positive in 1999, negative 2000-2001, positive 2002-2008, negative 2009-2010, positive in 2011, negative 2012-2015, and positive 2016-2017. What should the report conclude about the model?
Check your answer
Answer: The residuals form waves, so a single line is the wrong model even though it shows a clear upward slope. The report should describe the rises and falls by period rather than use the line to predict.A systematic residual pattern overrides r. - Kansas City precipitation against year, 1973-2024, has slope 0.274 inches per decade. Refit without 2001 (53.51 in, a wet year near the middle of the range): the slope becomes 0.242 inches per decade. Refit instead without 1973 (55.31 in, the first year): the slope becomes 0.681 inches per decade. Which point is influential? Explain.
Check your answer
Answer: 1973 is influential: removing it moves the slope from 0.274 to 0.681 inches per decade, because it is an end point with high leverage and a large residual. 2001 has a large residual but a middle x-value, so removing it moves the slope only from 0.274 to 0.242.Points near the middle of the x-range pull the slope least; end points pull it most. - A news story reports that Missouri counties with more hospitals have more deaths each year, and suggests hospitals are dangerous. Name the lurking variable and explain the association.
Check your answer
Answer: Population size. Larger counties have both more hospitals and more deaths; comparing death rates per 100,000 (age-adjusted) would remove most of the association.
Challenge
Stretch problems. Expect to think before you write.
- Missouri’s age-adjusted heart-disease death rate fell from 300.5 in 1999 to 191.1 in 2017, with r = −0.948 against year. Explain why ‘year’ cannot be the cause, and describe what kind of evidence could identify a cause.
Check your answer
Answer: Time is not a mechanism; it stands in for everything that changed over those years (treatment, prevention, behavior, the population). Identifying a cause would need randomized experiments for specific treatments, or observational studies that measure candidate causes, show dose-response and rule out alternatives across many places.
Review
Keep earlier skills sharp.
- For the data 10, 15, 27, 29, 32, 34, 36: find the mean, median, range and interquartile range. (Find the quartiles as the medians of the lower and upper halves, leaving out the overall median when the count is odd.)
Check your answer
Answer: mean = 26.14; median = 29.0; range = 26; IQR = 19.0 (Q1 = 15.0, Q3 = 34.0) - (Unit 1) Convert an SD of 8.40 inches (St. Louis annual precipitation, 1973-2024) to millimeters.
Check your answer
Answer: About 213.2 mm.Multiply by 25.4.
Nineteen Residuals, Three Refits and a Headline Built From Real Rates
Residuals and refits of the least-squares line for Missouri age-adjusted heart-disease death rates 1999-2017, and the correlation of heart-disease and Alzheimer’s rates, from CDC/NCHS dataset bi63-dtpu; retrieved November 10, 2026 · Public domain (works of the United States government)
Why this reading. The discourse lab asks two things of you: judge a linear model from its residual plot, and argue about a “linked to” headline. The lesson shows the residual plot as a picture; here are the nineteen residuals as numbers, so the plot can be drawn by hand and the U-shape counted rather than eyeballed. Then three refits measure influence, a real headline-shaped claim is built from two columns of the Lesson 2.2 table so that the argument has data under it, and a sample response is included for you to answer.
Table 1. Residuals of the line 282.43 − 6.12x
Missouri age-adjusted heart-disease death rate per 100,000 (CDC/NCHS dataset bi63-dtpu). Predicted values use the unrounded slope −6.1205 and intercept 282.427.
| Year | x | Actual | Predicted | Residual |
|---|---|---|---|---|
| 1999 | 0 | 300.5 | 282.43 | +18.1 |
| 2000 | 1 | 286.0 | 276.31 | +9.7 |
| 2001 | 2 | 273.4 | 270.19 | +3.2 |
| 2002 | 3 | 272.7 | 264.07 | +8.6 |
| 2003 | 4 | 264.9 | 257.94 | +7.0 |
| 2004 | 5 | 248.5 | 251.82 | -3.3 |
| 2005 | 6 | 236.8 | 245.70 | -8.9 |
| 2006 | 7 | 228.8 | 239.58 | -10.8 |
| 2007 | 8 | 218.9 | 233.46 | -14.6 |
| 2008 | 9 | 219.8 | 227.34 | -7.5 |
| 2009 | 10 | 206.1 | 221.22 | -15.1 |
| 2010 | 11 | 201.8 | 215.10 | -13.3 |
| 2011 | 12 | 197.4 | 208.98 | -11.6 |
| 2012 | 13 | 193.4 | 202.86 | -9.5 |
| 2013 | 14 | 194.7 | 196.74 | -2.0 |
| 2014 | 15 | 194.7 | 190.62 | +4.1 |
| 2015 | 16 | 197.9 | 184.50 | +13.4 |
| 2016 | 17 | 192.1 | 178.38 | +13.7 |
| 2017 | 18 | 191.1 | 172.26 | +18.8 |
Count the signs: positive for x = 0 to 4 (1999-2003), negative for x = 5 to 14 (2004-2013), positive again for x = 15 to 18 (2014-2017). Five positives, then a run of ten negatives, then four positives. Random scatter does not produce runs like that. The residuals also sum to zero, as least-squares residuals always do: use that as an arithmetic check. Extrapolated to 2025 (x = 26) the line predicts 123.3 per 100,000 against 191.1 in 2017, while the last four residuals (+4.1, +13.4, +13.7, +18.8) show the line was already falling away from the data.
Table 2. Three refits: which point has influence?
| Data used | Intercept | Slope | Change in slope |
|---|---|---|---|
| All 19 years | 282.43 | −6.121 | +0.000 |
| Without 1999 (x = 0) | 278.06 | −5.766 | +0.354 |
| Without 2017 (x = 18) | 284.52 | −6.490 | −0.369 |
| Without 2009 (x = 10, largest negative residual) | 283.02 | −6.092 | +0.028 |
Removing 2017, the last point, steepens the slope by 0.37; removing 1999, the first point, flattens it by 0.35. Removing 2009, which has the largest negative residual of all but sits at x = 10, next to the mean of x, changes the slope by 0.03. Leverage is about position in x, not about the size of the residual. The endpoint years are the influential ones, and the residual plot already told you they are also the years the line fits worst.
A headline with its data: “Fewer heart deaths linked to more Alzheimer’s deaths”
Take the heart-disease and Alzheimer’s columns of the Lesson 2.2 table as 19 paired points. Their correlation is r = −0.933; the least-squares line predicts an Alzheimer’s rate of 53.01 − 0.1239 × (heart-disease rate). The Alzheimer’s rate rose from 15.0 to 32.3 per 100,000 while the heart-disease rate fell from 300.5 to 191.1, and the points lie close to a line. Someone could write that headline and every number in it would be true.
Now argue. Both series are indexed by year, so any two variables that trend over 1999-2017 will correlate, in one direction or the other, whether or not they have anything to do with each other; year is the common response. Consider also some possibilities that you would have to investigate before believing any of them: people who do not die of heart disease in their sixties may live into the ages at which dementia is common (a real mechanism, but not the one the headline implies); death certificates may increasingly name Alzheimer’s as the underlying cause where they once named another condition (a change in recording, not in disease); and age adjustment removes the aging of the population but not changes in diagnosis. Then say what study could test whether reducing heart-disease deaths increases dementia, and why it could not be a randomized experiment.
A sample response to answer
Illustrative response written for this course; it is not real student work and its claims about recording practices are hypotheses, not findings.
“The headline is built on r = −0.93 between Missouri’s heart-disease and Alzheimer’s death rates. That shows the two series moved in opposite directions over 1999-2017, nothing more. My main lurking variable is time: both series are indexed by year, so almost any two trending series would correlate. My second is how deaths are recorded: if Alzheimer’s was named on death certificates more often over those years, its rate could rise with no change in the disease. To test whether preventing heart-disease deaths raises dementia deaths, I would follow similar people with and without heart disease for many years. A randomized experiment is impossible, because nobody could ethically assign heart disease or withhold treatment.”
What to do
- Draw the residual plot from Table 1 and mark the sign runs. Write one sentence on what the U-shape means for a 2025 extrapolation.
- Verify one refit in Table 2 with =SLOPE on 18 rows. Explain the difference between the 2017 and 2009 removals in terms of leverage.
- Record or write the 90-second argument about the Alzheimer’s headline, naming two lurking variables from the paragraphs above or your own, and the study that would be needed. Then compare your argument with the sample response: say where you agree, and qualify or challenge its second lurking variable. If you like, give your argument aloud to a parent or guardian and note one objection they raise.
Sources: CDC/NCHS, NCHS – Leading Causes of Death: United States, dataset bi63-dtpu (data.cdc.gov)