Case Study 2 — A Report Reorganized Top-Down

A composite, anonymized example. "Dana Whitfield," a data scientist, is a recurring vehicle in this book (you met the churn problem in Chapter 2). Fictional but realistic; no real company or data.

The setup

Dana spent three weeks analyzing why customers were canceling a subscription product. She built a model, found the drivers, and wrote them up for the VP of Marketing, Renée, who would decide what to fund next quarter. Renée is sharp, busy, and not technical—she will give this report about three minutes. Dana's first draft was thorough, accurate, clearly written at the sentence level, and organized exactly the way the work happened. Here is its skeleton.

❌ Before (bottom-up — the order the analysis happened):

Subject: Customer Churn Analysis — Q2

1. Introduction. Background on the churn problem and why it was studied. 2. Data and Methods. Pulled 18 months of subscription data; engineered 40 features; trained and compared three models (logistic regression, random forest, gradient boosting); selected gradient boosting by cross-validated AUC of 0.83; used SHAP values to interpret feature importance. 3. Results. Feature importances, partial-dependence plots, a confusion matrix, model diagnostics. Buried in the middle: the three biggest churn drivers are (a) no logins in the first 14 days after signup, (b) a failed payment that wasn't recovered, and (c) never having used the core "reports" feature. 4. Discussion. Caveats, limitations, comparison to prior quarter. 5. Recommendations. Three suggested interventions, at the very end: an onboarding email sequence targeting first-14-day inactivity, a dunning/payment-recovery flow, and an in-app nudge toward the reports feature.

Every section is competent. And for this reader, it is organized backward. Renée has to travel through a model-selection methodology she doesn't care about and a results section full of SHAP plots before she reaches, at the very bottom, the only thing she needed: what's driving churn and what should we do about it. She will not get there in three minutes. She'll skim section 2, glaze at the plots in section 3, and either email Dana to ask "so what should we do?"—or, worse, set it aside.

The problem is not clarity (the sentences are fine) and not the analysis (it's solid). The problem is structure: it's bottom-up, it buries the conclusion, and it leads with the part the audience cares about least. This is the §4.2 default failure—Dana organized the report the way she discovered the findings, not the way Renée needs to consume them.

Diagnosis by reverse outline

Run the §4.4 / Project Checkpoint diagnostic. Extract one line per section's actual point, in order:

  1. We studied churn.
  2. We built and selected a model (methodology).
  3. Here are model diagnostics… (and, buried) the three churn drivers.
  4. Caveats.
  5. Three things we recommend doing.

Read top to bottom. The most decision-relevant points—the drivers (line 3, buried) and the recommendations (line 5, last)—are at the bottom of the document. The reader's #1 need is the writer's #5 delivery. That inversion is the whole problem, and the reverse outline makes it obvious in seconds.

The reorganization

Reorder top-down. Lead with the recommendation and the drivers; demote the methodology to support; keep everything, move almost everything.

✅ After (top-down — the order the reader needs):

Subject: Churn is driven by 3 fixable problems — recommend funding all three (Q2 analysis)

Executive summary (read this if nothing else). Three problems drive most of our churn, and all three are addressable: (1) new users who don't log in within their first two weeks, (2) failed payments we never recover, and (3) users who never try our core "reports" feature. We recommend funding three matching interventions next quarter—an onboarding sequence, a payment-recovery flow, and an in-app nudge to reports. Estimated reach: these three factors are present in ~70% of churned accounts. Details and methodology below.

## The three churn drivers (in order of impact) ### 1. No logins in the first 14 days — what it is, how big, the fix. ### 2. Unrecovered failed payments — … ### 3. Never used the "reports" feature — …

## Recommended interventions and what we'd need Each driver → a specific intervention → rough cost/effort. (A short table.)

## How we know (methodology) 18 months of data, 40 features, three models compared, gradient boosting selected (AUC 0.83), SHAP for interpretation. Diagnostics and plots in the appendix.

## Caveats Limitations and what we're less sure of.

Appendix: Feature importances, partial-dependence plots, confusion matrix, model comparison.

What changed structurally—each tied to the chapter:

  • Conclusion first (§4.3, inverted pyramid). The executive summary states the drivers, the recommendation, and a number (70%) that quantifies why it matters—readable entirely on its own. Renée can act on the summary alone; the rest is there if she wants it.
  • Methodology demoted, not deleted (§4.2). The model-selection story moves from section 2 to a "How we know" section near the bottom and an appendix. A reader who trusts Dana skips it; a skeptical or technical reader can verify it. Nothing is lost; the order changed.
  • Informative headers (§4.5). "The three churn drivers (in order of impact)" and "Recommended interventions" replace "Results" and "Recommendations." A scanner navigating by headers gets the whole story from the headers alone.
  • Parallel structure (§4.7). The three drivers are presented in an identical shape—what it is → how big → the fix—so Renée learns the pattern once and reads all three effortlessly, and they're ordered by impact.
  • The subject line carries the bottom line. Reading only "Churn is driven by 3 fixable problems — recommend funding all three" already conveys the ask.

The lesson

Same analysis. Same facts. Same sentence-level clarity. Opposite usability, produced entirely by reordering. This is theme 5 in everyday clothes: the bottom-up version served Dana's experience of doing the work; the top-down version serves Renée's need to decide. And it connects to Chapter 2—audience determined which fact leads. Because the reader is a non-technical executive who will spend three minutes, the drivers and recommendation go at the top and the AUC near the bottom. For a fellow data scientist reviewing the model, the methodology might lead instead. The information doesn't change; its order is a function of who's reading.

💡 Tip: Before sending any report, run the reverse outline and ask the §4.3 question: if my reader reads only the first paragraph, do they have what they came for? If not, you wrote it bottom-up—flip it.

Related: Chapter 4 §4.2–§4.3 · Case Study 1 · Exercises M5