The quality team is about to recommend A1C testing on every admission because tested patients come back less often. Write the memo that tells them what the data can and cannot support, backed by a notebook you wrote.
A starter notebook (project-starter.ipynb)
with the load-and-recode cell already written, a patient-level frame builder, and a stub for
welch_ttest(a, b) — signature only, the body is yours (see Session 3.1 for the pieces).
Each checkpoint is one observable artifact you build yourself. Checking all six flips this module's mastery to "project complete" in the tracker. No auto-grading — you mark it once the artifact exists.
Your Module 01 sampling-wobble figure redrawn with the CLT's predicted curve on top, for one statistic of your choosing, with predicted vs. measured spread stated.
A solid version: uses σ/√n (or √(p(1−p)/n)) explicitly, by name, in the writeup — not just plotted; states both the predicted and measured numbers side by side; one sentence on why they differ slightly (finite repeats, not a broken formula).
A 95% interval for a statistic from one sample, plus one sentence in your own words saying what the interval promises.
A solid version: any correct phrasing about the procedure (not the specific output); does not attach a probability to this one interval; states the sample size used.
A two-sample Welch t-test on a continuous outcome by a binary split of your choosing (days by A1C status is fine), built by hand, reported as effect + CI + p.
A solid version: the t-statistic is built from the formula (means, sample SDs with ddof=1, √n) — not only called from a library, though cross-checking against scipy.stats.ttest_ind is expected; states H0 and H1; reports the effect and interval in the column's own units, not just the p-value.
A short paragraph stating what your p-value means, containing one wrong reading you explicitly reject.
A solid version: defines your p-value as an area in the null world (per Session 3.2); the rejected reading is a real misreading — "P(H0)", "97% real," or "big effect" — not a strawman you invented to sound rigorous.
r for a pair of columns, one named confounder, a partial or stratified check, and an honest two-sentence read.
A solid version: the partial/stratified check is actually run, not just asserted; the two-sentence read separates "shown" from "not shown"; does not claim causation in either direction.
≤1 page, plain language, no code: the visit-level and patient-level readmission gaps, whether this is evidence that testing causes lower readmission, the alternative explanations, and the A/B test the hospital would need.
A solid version: leads with the answer, not a step-by-step recap; reports the first-visit gap with its interval (not just the larger, visit-level number); names both selection into testing and the care-after-testing difference (medication-change rate) as alternative explanations; specifies what gets randomized, the single metric, persistent assignment, and what result would change the recommendation.
Repeat checkpoints 3 and 6 on a dataset of your own choosing; or re-run the readmission test with a cluster-aware approach of your choosing (e.g. one row per patient with a robust/clustered standard error) and compare it to the simple first-visit answer from Session 5.2 — unverified by this module's research, so you're on your own here, which is the point.