
investigate the relationship between two quantitative variables from a public dataset, fit and critique a model, and defend conclusions about association and causation.
Instruction
The question. You choose it, with this form: How is [response] related to [explanatory variable] for [cases], and how well does a model describe the relationship?
Datasets you may use (all public federal data; name the source and retrieval date):
- CDC NCHS Leading Causes of Death (data.cdc.gov, dataset bi63-dtpu): age-adjusted death rates by cause, state and year, 1999-2017. For example, the Missouri cancer rate against year, or the unintentional-injury rate against year. The injury rate rose, dipped, rose again and then jumped between 1999 and 2017 (43.2 in 1999, 41.9 in 2000, 50.1 in 2006, 47.0 in 2013, 58.8 in 2017), a pattern a single line will miss.
- NOAA NCEI Global Summary of the Year: temperature or precipitation against year for a Missouri station (St. Louis, Kansas City, Springfield or Columbia), or one station against another.
- Census decennial counts for Missouri (1910-2020), with a nonlinear model.
- Another federal dataset, if a parent or guardian agrees that it is a good choice and you cite it with its retrieval date.
The analysis you must show.
- A scatterplot with labeled axes and units, and a description: direction, form, strength, unusual points.
- The correlation, with an explanation of what it measures and whether it is appropriate given the form.
- A least-squares model: a line, or a transformed exponential if the context and scatterplot suggest it. Give the equation with units, the slope and intercept interpreted in context (or growth rate for an exponential), and r2 interpreted.
- A residual plot and your judgment of the model. If the residuals show a pattern, say what it means and try a better form.
- An influence check: refit without the most extreme x or y point and report the change in slope.
- A causation paragraph: could x cause y? Name at least two lurking variables, and describe the study that could test cause and effect.
- One prediction by interpolation and one by extrapolation, with a judgment of how far each can be trusted.
A model paragraph. “From 1999 to 2017, Missouri’s age-adjusted heart-disease death rate fell by about 6.1 per 100,000 per year on average (ŷ = 282.43 − 6.12x, x = years since 1999, r2 = 0.899). The residual plot, however, is U-shaped: the line under-predicts in 1999-2003 and 2014-2017 and over-predicts in between, so the decline was slowing. The linear model should not be used for years after 2017.”
The defense. Four minutes: your question, your scatterplot and residual plot, your model, and one limitation. Be ready for the question “why this model and not another?” Record yourself, or give the defense aloud to a parent or guardian, and answer the questions listed with the task.
📖 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.
Your residual plot shows a clear curve. What should your report do?
A report can claim that x causes y if r² is above 0.9.
Stage 1 plan: state your question, dataset (with URL and retrieval date), variables with units and the cases you will include. Paste a first scatterplot and write two sentences on the form you expect and why.
How confident are you that you can carry out, report and defend a two-variable investigation with a model I can justify?
Unit 2 checkpoint
Do this before you start the task. It covers the whole unit and tells you what to review first.
Which r indicates the weakest linear association?
If r = 0.777, then 77.7% of the variation in y is explained by x.
With r = 0.8, s_y = 10 and s_x = 4, the least-squares slope is .
The model predicts the death rate falls 6.12 per 100,000 per year. Which is the best interpretation?
A point has actual y = 201.8 and predicted y = 215.1; its residual is .
A residual plot is curved. What should you do?
A strong correlation between two variables shows that one causes the other.
Which point is most likely to be influential?
If log₁₀ P = 6.51 + 0.0025t, the annual growth factor is about .
A model with r = 0.9894 can safely be used to predict 50 years beyond its data.
St. Louis and Columbia rainfall correlate at r = 0.777 because…
The least-squares line always passes through (x̄, ȳ).
Submit: Bivariate Investigation Report (1,300-1,600 words) + four-minute model defense
Task: Investigate the relationship between two quantitative variables from a public federal dataset. Use the hosted tables in this lesson’s reading and the readings for Lessons 1.3, 2.1 and 2.4 (CDC/NCHS death rates, bi63-dtpu; NOAA NCEI Global Summary of the Year; Census counts, each with its retrieval date). That route is complete; the agency portals (data.cdc.gov, ncei.noaa.gov, census.gov) are where to look if you want a different data set, and any data set you find yourself must be cited with its retrieval date. Tools: CODAP, a free spreadsheet (SLOPE, INTERCEPT, CORREL, RSQ, LN/LOG10), or optionally Python in JupyterLite. A hand-drawn scatterplot and residual plot built from a printed table earn full credit.
Submission workflow
- Stage 1 — Plan. Question, dataset and variables with units, the cases included, and a first scatterplot.
- Stage 2 — Self-checked draft. Using the summary table in this lesson’s reading, recompute your own slope from b = r·s_y/s_x and check that your line passes through (x̄, ȳ). Then critique your residual plot and causation paragraph with the checklist in the reading (or ask a parent or guardian to read the draft and note one question); submit your notes and your revision plan.
- Stage 3 — Final report. 1,300-1,600 words with scatterplot, model, residual plot, influence check, causation paragraph and predictions.
- Stage 4 — Model defense. Four minutes, recorded or given aloud to a parent or guardian, answering these questions about your choice of model: why this form, what the residual plot showed, and how far you would trust a prediction.
How it is scored
| Criterion | What proficient looks like |
|---|---|
| Scatterplot and description (A1.DS.A.5) | Direction, form, strength and unusual points described in context. |
| Correlation (A1.DS.A.7, PS.13) | Computed correctly; its meaning and limits explained given the form. |
| Model fitting (PS.12, A1.DS.A.5a, A1.DS.A.5b) | Appropriate least-squares or transformed model with correct equation and units. |
| Interpretation (A1.DS.A.6) | Slope or growth rate, intercept and r² interpreted in context. |
| Residuals and influence (A1.DS.A.5) | Residual plot analyzed; influence tested by refitting. |
| Causation (A1.DS.A.8) | Lurking variables named; the study that could test causation described. |
The Self-Check Table: Summary Statistics for Every Pair the Course Hosts
Regression summaries for the ten causes in CDC/NCHS dataset bi63-dtpu (Missouri, 1999-2017) and for four Missouri NOAA stations against year (1973-2024), with pointers to the hosted station-pair and Census tables; retrieved November 10, 2026 · Public domain (works of the United States government)
Why this reading. Stage 2 of the task has you recompute your own slope from b = r·sy/sx and critique your own residual plot. That check only works if you have the summary statistics for the pair you chose. Here they are for every pair this course hosts: all ten CDC causes against year for Missouri, the six station pairs, the four stations against year, and the census counts. Pick from this menu and anyone who checks your work, a parent or guardian included, is working from the same numbers.
Menu 1. CDC age-adjusted death rates for Missouri against year, 1999-2017
Dataset bi63-dtpu, x = years since 1999 (x̄ = 9, sx = 5.627, n = 19). The full rows for the first six causes are in the Lesson 2.2 reading; the others can be read from the same dataset, or you may take the summary statistics below as given and state that in the report.
| Cause | 1999 | 2017 | ȳ | sy | r | b | a | r² |
|---|---|---|---|---|---|---|---|---|
| Heart disease | 300.5 | 191.1 | 227.34 | 36.33 | −0.948 | −6.121 | 282.43 | 0.899 |
| Cancer | 207.9 | 167.3 | 189.48 | 13.38 | −0.990 | −2.353 | 210.66 | 0.980 |
| Stroke | 65.6 | 41.0 | 49.99 | 9.09 | −0.964 | −1.556 | 64.00 | 0.929 |
| Chronic lower respiratory diseases | 51.7 | 50.7 | 50.20 | 3.06 | 0.539 | 0.293 | 47.57 | 0.290 |
| Alzheimer’s disease | 15.0 | 32.3 | 24.83 | 4.82 | 0.907 | 0.778 | 17.83 | 0.823 |
| Unintentional injuries | 43.2 | 58.8 | 48.36 | 4.19 | 0.809 | 0.603 | 42.94 | 0.654 |
| Diabetes | 26.3 | 21.0 | 22.48 | 2.68 | −0.874 | −0.417 | 26.24 | 0.763 |
| Influenza and pneumonia | 28.3 | 16.7 | 20.93 | 4.06 | −0.935 | −0.674 | 26.99 | 0.874 |
| Suicide | 12.6 | 18.5 | 14.27 | 2.09 | 0.920 | 0.342 | 11.19 | 0.846 |
| Kidney disease | 15.1 | 19.6 | 18.01 | 1.40 | 0.839 | 0.209 | 16.13 | 0.703 |
The table is a lesson in itself. Five causes fall (heart disease, cancer and stroke the most; cancer is the straightest line in the set, r² = 0.980), five rise (Alzheimer’s disease, suicide, kidney disease, unintentional injuries and chronic lower respiratory diseases), and chronic lower respiratory diseases has by far the lowest r² (0.290), which means a line explains little of it. Look beyond r²: heart disease (0.899) and stroke (0.929) have high r² but residuals with a run of ten or more of the same sign, and unintentional injuries rose from 43.2 to 58.8 with a jump after 2015 (50.9, 57.0, 58.8). The dataset does not say why; you would need other sources to investigate. Choosing one of these is allowed, and it makes the residual plot the center of your report rather than an afterthought.
Menu 2. Station against station, or station against year
The Lesson 2.1 reading gives r and the distance for all six station pairs, and the Lesson 1.3 reading has the 52 rows of annual precipitation for 1973-2024. For a station against year, the summary statistics are:
| Station | Mean PRCP (in) | SD (in) | r with year | Slope (in per decade) |
|---|---|---|---|---|
| St. Louis | 41.04 | 8.395 | 0.280 | 1.552 |
| Kansas City | 38.72 | 8.332 | 0.050 | 0.274 |
| Springfield | 45.19 | 7.956 | −0.028 | −0.147 |
| Columbia | 41.32 | 8.271 | 0.060 | 0.327 |
A line fitted to any of these will have r² of at most 0.08, so the lesson’s interpretation of a low r² is the report: the slope is small compared with the year-to-year variation, and the honest finding may be “no clear trend”. St. Louis has the largest slope (1.55 inches per decade, r = 0.28); check whether that holds in the 1950-2024 St. Louis record in the Lesson 1.1 reading.
Menu 3. Census counts
The Lesson 2.4 reading has the twelve Missouri counts, their logarithms, and three fits worked with residuals. A report using them should go beyond reproducing the fits: for example, fit 1940-2020 and test whether any single exponential describes it. The 1910-1970 and 1970-2020 growth factors are 1.0058 and 1.0059 per year.
What to recompute for your own pair (a checklist)
- b from r, sy and sx, and a from the means. The line must pass through (x̄, ȳ): substitute and check.
- The residual signs in x order. A run of five or more of one sign is a pattern to explain, not noise.
- The influence check: slope with the first and with the last x removed. Report both changes.
- Whether the causation paragraph names year as a common response wherever both variables are indexed by year.
Sources: CDC/NCHS, NCHS – Leading Causes of Death: United States, dataset bi63-dtpu (data.cdc.gov); NOAA National Centers for Environmental Information, Global Summary of the Year (GSOY), stations USW00013994, USW00003947, USW00013995, USW00003945 (ncei.noaa.gov); U.S. Census Bureau, 2020 Census Apportionment Results, apportionment.csv (www2.census.gov)