Data Stories | Correlation Is Not Causation

Correlation Is Not Causation — a 10-section, A4, print-ready handout built on four simulated datasets, with every table a native Word table you can edit.

What is in it

• Three worked cases. Ice-cream sales vs drowning deaths (r = 0.88, collapsing to −0.25 once temperature is held constant); a 600-patient cohort where Drug X looks 9.7 points worse pooled and 5–15 points better in every severity band (Simpson's paradox, standardised to +8.3); and a 200-person × 8-week panel where a −0.50 cross-sectional correlation turns into a one-way time ordering.
• How causation is tested. Randomisation and its limits, plus difference-in-differences, regression discontinuity, instrumental variables, propensity scores, target-trial emulation and negative controls — each with the assumption it asks you to accept — and a Bradford Hill–style checklist.
• Ten common misreadings, each with the sentence to say instead; nine practice questions with a fully worked answer key; a 13-term glossary; and a method note.
• Five figures, running header, and a footer numbering content pages from 1.

Why the numbers hold together: the tables are generated from results.json and the raw CSVs rather than typed by hand, and each headline statistic is recomputed a second way and asserted to match. The underlying data, the three scripts and the figures are supplied so you can rebuild or re-seed everything.

Checks run: A4 geometry (210 × 297 mm), all 24 tables within the 166 mm text column, text proof-read end to end, and each of the five figures inspected as a rendered image and corrected where labels collided. I could not render the DOCX itself to page images in this environment, so pagination is Word's own flow with explicit breaks before each case study and before Practice, Answers and Misreadings.

All data are simulated (seed 20261031) and labelled as such throughout — nothing here should be cited as evidence about the real world.