Writing About Numbers
P14.communication.04 · Audience: guest, it-ml, language-pro · Prerequisites: The One-Slide Answer
The chart was honest. The interval was shown. Then the caption underneath read "the new onboarding flow increased retention by 12%", and every downstream reader — the exec who quoted it, the deck that reused it, the press release that followed — inherited a causal claim the data never made. The study was observational: teams that adopted the new flow also happened to be the more engaged teams. Retention and the new flow moved together; nothing showed one caused the other. No number was wrong. The word "increased" did the damage. This module is about the last surface in the track, the one most people forget is a surface at all — the prose around the numbers, and the discipline of never letting a sentence claim more than the data behind it supports.
An analyst who would never truncate an axis or hide an interval will, in the same breath, write a sentence that claims far more than the chart beneath it. Prose feels informal, a wrapper around the "real" quantitative work — but it is the part most people actually read and quote, and a single loaded word travels further and mutates faster than any chart. The written claim is the finding as it will be remembered and repeated, so it has to be at least as disciplined as the analysis it summarises.
ⓘ Concept: Write only the claim the data supports — no more
Why it matters — A chart can be scrupulously honest and its caption still overclaim, because the sentence is doing separate work: it asserts what the numbers mean, and meaning is where the overreach hides. Readers retain the sentence, not the confidence interval; they quote 'retention increased 12%', not 'retention was 12% higher among adopters in an observational sample'. So the written claim, not the chart, is what propagates — and if it says more than the evidence, the surplus propagates too, hardening into fact through repetition. The discipline is to make the strongest claim the data actually licenses and stop exactly there: describe what you observed, scope it to where you observed it, and refuse the extra clause that would make it sound bigger. A finding written honestly is a finding that survives being quoted.
You have reached the end of the communication track and the close of the Data Visualization & Communication pillar (P14). From the one-slide answer, through dashboards that earn their tiles and the honest presentation of uncertainty, to writing about numbers without overclaiming — you now hold the last mile of data work: turning a correct, honestly-encoded result into a picture and a sentence an audience reads correctly and a decision it can actually bear. Step back to the P14 pillar overview to see the chart-craft and communication tracks together, or follow the cross-links into Time Series & Forecasting (P7) and Causal Inference (P8), where the intervals you present and the causal claims you are allowed to write are actually built.
Going deeper (technical) — the confounder, and why randomisation breaks it
Interview-grade notes for the technical (it-ml) track. The framing above is audience-broad; this is the mechanism underneath the writing rule, and being able to state it is what separates "I know not to say caused" from understanding why.
A confounder is a common cause of both the treatment and the outcome, and it is what makes a correlation uninformative about cause. If engaged teams are both more likely to adopt the new flow and more likely to retain, then adoption and retention will correlate even if the flow does nothing — the engagement drives both. The observed association is a mix of any real effect of the flow and the spurious path through the confounder, and from the correlation alone you cannot separate the two. This is why "control for confounders" is a whole research programme: you can adjust for the confounders you can measure and name, but never prove you have caught them all.
Randomisation is powerful because it severs the treatment from every confounder at once — measured or not. When a coin flip decides who gets the treatment, the treatment is by construction independent of engagement, tenure, prior retention and every other lurking variable, so the two groups differ only by chance and by the treatment. Any outcome gap that survives is attributable to the treatment. That is the entire logic of the A/B test, and it is why a causal verb is honest after a proper randomised experiment and dishonest after an observational study of the same size and cleanliness.
When randomisation is impossible, quasi-experimental designs approximate it — each on an assumption a reader should see. Difference-in-differences, regression discontinuity, instrumental variables and natural experiments each manufacture a comparison that stands in for randomisation under a stated condition (parallel trends, a sharp cutoff, a valid instrument). They can license causal language, but only as far as their assumption holds — so honest writing states the design and its key assumption, letting the reader judge the causal claim rather than take the verb on faith. The full treatment of these designs is Causal Inference (P8); the writing rule is only their shadow: no causal verb without a design that earns it.
Ask the mentor about this module
Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.
Try it yourself
A scratch console for this page's ideas — ungraded, nothing you run here is recorded.
Scratch console
A scratch console with the scientific stack (pandas, numpy, scikit-learn). Runs on the server — no network, resource-limited and measured.
Output appears here.