
investigate the relationship between two quantitative variables with real data, fit and evaluate a linear model, and communicate what the model does and does not show.
Instruction
This task brings together everything in the semester’s statistics: a real question, real data, a model you can evaluate and an argument about what it means.
1. Ask a question about a relationship. You need two quantitative variables measured on the same individuals. Choose one data set; all three earn full credit and all three support every part of this task.
- Data Set 1 — printed here (illustrative). Fifteen school districts. x = percent of adults age 25+ with a bachelor’s degree; y = median household income in thousands of dollars. (12, 40), (15, 42), (18, 43), (21, 63), (24, 62), (27, 55), (30, 59), (33, 78), (36, 79), (39, 88), (42, 87), (45, 83), (48, 93), (51, 85), (54, 97). The association is strong, positive and close to linear, so a least-squares line is the right tool. Note what the individuals are: districts, not people.
- Data Set 2 — printed here (illustrative). A compact car driven at a steady speed on level road. x = speed in miles per hour; y = fuel economy in miles per gallon. (20, 20), (25, 25), (30, 29), (35, 32), (40, 34), (45, 35), (50, 34), (55, 32), (60, 30), (65, 27), (70, 24), (75, 20). The form here is clearly curved, and a linear model is the wrong tool. That is a legitimate finding: fit the line anyway, show that r is near zero while the scatter plot shows a strong relationship, read the arch in the residual plot, and explain in your report why r measures only linear association.
- Data Set 3 — data you gather. Two quantitative variables you download or collect, from a named public source you cite: Census QuickFacts (census.gov/quickfacts) for county income, education, commute time and home values; Missouri Comprehensive Data System (dese.mo.gov) for enrollment and staffing by school district; NOAA Climate at a Glance (ncei.noaa.gov/access/monitoring/climate-at-a-glance) for temperature and precipitation by county; or class data such as hand span and height, collected anonymously with an option to decline.
2. Document the data. Record the source, the table or file name, the year, the date you downloaded it, and exactly what each variable measures with its unit. For county or school data, note what the individuals are: counties or schools, not people. A pattern among counties does not necessarily describe individuals within them.
3. Explore. Make a scatter plot with labeled axes. Describe direction, form, strength and any unusual points. If the form is clearly curved, say so: a linear model is then the wrong tool, and your report should explain why.
4. Model. Fit the least-squares line with a spreadsheet, Desmos or CODAP. Interpret the slope in a sentence with units and the words “predicted” or “on average”. Decide whether the intercept has meaning in your context. Compute r and interpret it.
5. Evaluate. Make a residual plot and describe it. Is there a pattern? Are there large residuals, and what are those individuals? A county with a large positive residual is one the model underestimates; ask why.
6. Use and question the model. Make one prediction inside the data’s range and explain how much to trust it. Then argue carefully about causation: could one variable cause the other? Name confounders, consider reverse direction, and say what study could test a causal claim.
The report. 700-1,000 words, with the scatter plot and line, the residual plot, and the key numbers. Write for a reader who is smart but did not take this class: every figure gets a sentence saying what it shows.
The presentation. Three to four minutes, live or recorded: question, scatter plot with line, what the slope means, how well the model fits, and your causation argument. Expect one question about a specific point on your residual plot.
📖 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.
Residual
Tap to revealActual minus predicted.
Pattern in the residual plot
Tap to revealA sign the linear model is the wrong form.
r near ±1
Tap to revealStrong linear association; the sign gives direction.
Strong correlation from observational data
Tap to revealAssociation, not proof of cause.
Among the counties of a state, higher median income is strongly correlated with a higher percent of adults with a bachelor’s degree. Which conclusion is justified?
How confident are you that you can carry out and present a complete bivariate investigation with a fitted and evaluated linear model?
Unit 4 checkpoint
Do this before you start the task. It covers the whole unit and tells you what to review first.
Of 60 ninth graders, 12 have jobs. What is the conditional relative frequency of having a job given grade 9?
Of 45 students with jobs, 33 are twelfth graders: % (to one decimal).
Equal conditional relative frequencies across rows suggest no association.
The slope of predicted arm span ≈ 1.02 × height − 3.76 means…
A residual is actual minus .
Which r shows the strongest linear association?
r = 0.9 means 90% of the points lie on the line.
A residual plot curves. What should you conclude?
With predicted sleep ≈ 8.85 − 0.64x, the prediction for x = 3 is about hours.
A strong correlation in observational data proves causation.
What establishes cause most convincingly?
Predicting arm span for a 100 cm child from data on 152-183 cm students is…
Submit: Bivariate Data Study (700-1,000 words with plots and model) and presentation
Task: Investigate the relationship between two quantitative variables using real data: describe the association, fit and interpret a least-squares line, evaluate it with residuals and r, make and qualify a prediction, and argue about causation.
Submission workflow
- Stage 1 — Plan. Question, variables with units and what the individuals are, data source and collection or download plan.
- Stage 2 — Peer-reviewed draft. A partner re-computes one residual and checks your slope interpretation and causation language; submit a revision note.
- Stage 3 — Revised final. Scatter plot with line, model, residual plot, r, one qualified prediction, causation argument and sources, in 700-1,000 words.
- Stage 4 — Presentation. Three to four minutes, live or recorded, answering a question about one point on the residual plot.
How it is scored
| Criterion | What proficient looks like |
|---|---|
| Scatter plot and description (A1.DS.A.5) | Labeled plot; direction, form, strength and unusual points described in context. |
| Linear model (A1.DS.A.5, A1.DS.A.5a) | Least-squares line fitted with technology and used for a prediction within the data’s range. |
| Interpretation (A1.DS.A.6) | Slope interpreted with units and appropriate language; intercept’s meaning or lack of meaning explained. |
| Fit (A1.DS.A.5a, A1.DS.A.5b, A1.DS.A.7) | Residual plot analyzed; r computed and interpreted correctly, including its limits. |
| Causation (A1.DS.A.8) | Confounders and direction considered; a study that could test cause described. |
| Communication (MP.3) | Clear written argument with sources; presentation answers questions with reference to the data. |
The Bivariate File: Three Paired Data Sets, Fitted, Checked and Criticized
NOAA NCEI Climate at a Glance city time series, July average maximum temperature, San Diego (USW00023188) and Fresno (USW00093193), 1995–2025, retrieved September 23, 2026; plus the illustrative speed-and-fuel-economy data set printed in this lesson · Public domain (works of the United States government); worked regressions SLS original
Why this reading. The task needs two quantitative variables measured on the same individuals, a documented source, a fitted line, a residual plot, an r, a prediction and a causation argument. Three complete data sets are below. Two are real federal records and one is the lesson’s own illustrative set, and each is worked far enough that you can check your own regression rather than trusting whatever a spreadsheet hands back. None of them behaves perfectly, which is the point: a report that says why a model fits badly is worth more than one that found a tidy line.
Data Set 1 — A weak trend: San Diego Julys against the year
Source. NOAA National Centers for Environmental Information, Climate at a Glance, city time series: July average maximum temperature at San Diego International Airport (station USW00023188), 1995–2025. Retrieved September 23, 2026. x = the year; y = the July average maximum temperature in °F. The individuals are Julys, not people and not days.
(1995, 73.3), (1996, 73.7), (1997, 72.2), (1998, 73.0), (1999, 72.4), (2000, 73.2), (2001, 72.8), (2002, 71.6), (2003, 74.1), (2004, 76.2), (2005, 74.2), (2006, 80.9), (2007, 73.8), (2008, 73.4), (2009, 75.3), (2010, 70.1), (2011, 74.1), (2012, 72.7), (2013, 74.2), (2014, 76.9), (2015, 75.7), (2016, 74.4), (2017, 76.0), (2018, 78.6), (2019, 73.3), (2020, 74.4), (2021, 76.0), (2022, 72.2), (2023, 74.1), (2024, 76.7), (2025, 73.3).
The model. x̄ = 2010, ȳ = 74.284, Sxx = 2480, Sxy = 161.8, Syy = 136.56. Slope = 161.8 ÷ 2480 = 0.0652 °F per year; intercept = −56.85; r = 161.8 ÷ √(2480 × 136.56) = 0.278. The line predicts 73.3 °F for 1995 and 75.3 °F for 2025.
The residuals, and what they are for. The three largest are July 2006 at +6.88, July 2010 at −4.18 and July 2018 at +3.79. Every other residual is under 2.4 in size. A residual plot here shows no curve and no fan — just scatter, with 2006 far above everything. So a line is not the wrong shape; it is simply a weak description. With r = 0.278 the line accounts for well under a tenth of the variation, and a report that presents a trend line here without saying that is overselling it.
What to argue about. Year cannot be caused by temperature, so the usual reverse-direction worry is gone; what remains is whether the station measured the same thing for thirty-one years. Compare this with Fresno, where the same calculation on the same years gives a slope of 0.165 °F per year and r = 0.536. Two stations in one state, two very different-looking trends.
Data Set 2 — A moderate association: two cities, the same Julys
Source. The same two NOAA series, paired by year. x = San Diego July average maximum (°F); y = Fresno July average maximum (°F). The individuals are the thirty-one Julys from 1995 to 2025.
(73.3, 94.7), (73.7, 100.8), (72.2, 95.5), (73.0, 96.7), (72.4, 95.6), (73.2, 93.5), (72.8, 96.0), (71.6, 99.2), (74.1, 101.3), (76.2, 98.5), (74.2, 101.7), (80.9, 103.0), (73.8, 98.2), (73.4, 98.6), (75.3, 100.7), (70.1, 98.5), (74.1, 96.5), (72.7, 97.6), (74.2, 101.0), (76.9, 99.9), (75.7, 96.8), (74.4, 100.0), (76.0, 101.9), (78.6, 102.8), (73.3, 99.0), (74.4, 99.9), (76.0, 103.3), (72.2, 100.8), (74.1, 101.8), (76.7, 105.2), (73.3, 97.0).
The model. x̄ = 74.284, ȳ = 99.226, Sxx = 136.56, Sxy = 102.53, Syy = 235.84. Slope = 0.751, intercept = 43.45, r = 0.571.
Interpreting. Each degree warmer in San Diego’s July is associated with about 0.75 of a degree warmer in Fresno’s. The intercept, 43.45 °F, is the predicted Fresno July for a San Diego July of 0 °F — seventy degrees outside the data, so it means nothing and should be labeled as meaningless rather than left for a reader to trip over.
A prediction inside the data. San Diego at 75.0 °F predicts Fresno at 0.751(75.0) + 43.45 ≈ 99.8 °F. The San Diego values run from 70.1 to 80.9, so this is an interpolation. With r = 0.571, say how loosely to hold it: real Fresno Julys in years when San Diego was near 75 ranged over several degrees.
Large residuals worth naming. 2006 sits 1.19 °F below the line even though it was San Diego’s hottest July on record — the model expected Fresno to be hotter still. 2000 sits 4.91 °F below the line and 2024 sits 4.16 °F above it. A county or a year with a large residual is a question, not a defect: ask what happened in it.
Data Set 3 — The wrong shape: speed and fuel economy
Source. The illustrative data set printed in this lesson: a compact car driven at a steady speed on a level road. These are teaching numbers, not measurements of a real vehicle, and any report using them must say so. x = speed in miles per hour, y = fuel economy in miles per gallon.
(20, 20), (25, 25), (30, 29), (35, 32), (40, 34), (45, 35), (50, 34), (55, 32), (60, 30), (65, 27), (70, 24), (75, 20).
The model, fitted anyway. x̄ = 47.5, ȳ = 28.5, Sxx = 3575, Sxy = −100, Syy = 309. Slope = −100 ÷ 3575 = −0.0280, intercept = 29.83, r = −0.095.
The residuals, in order of speed: −9.27, −4.13, +0.01, +3.15, +5.29, +6.43, +5.57, +3.71, +1.85, −1.01, −3.87, −7.73. Plot them against speed and they form a clean arch: negative at both ends, positive in the middle. That is the residual plot telling you a line is the wrong kind of model, and it is the one message a scatter plot can hide but a residual plot cannot.
What the report should say. Fuel economy and speed are strongly related here — knowing the speed pins the economy down to within a mile or two per gallon — and r = −0.095, nearly zero. There is no contradiction, because r measures linear association only. Reporting “no relationship, r = −0.10″ would be the single worst sentence you could write about this data set, and a spreadsheet will hand you exactly that number without comment.
Documentation your report must carry
For whichever set you use: the issuing agency or the fact that the data are illustrative; the station, series or table identifier; the period covered; the date retrieved; what one value measures and its unit; and what the individuals are. All three sets above have individuals that are not people — Julys, paired Julys, test runs — and every claim in your report has to be a claim about those same things.
Sources: NOAA NCEI, Climate at a Glance: City Time Series (Fresno) (ncei.noaa.gov); NOAA NCEI, Climate at a Glance: City Time Series (San Diego) (ncei.noaa.gov)