How do you stop stakeholders reading a 95% confidence interval as a probability about the true value?

## The problem is a sentence, not an audience "There is a 95% probability the true value is between 4.1 and 5.3" is what people naturally say, and it is not what a frequentist interval means: the parameter is fixed, the interval is the thing that varies from sample to sample, and the 95% is the long-run rate at which the *procedure* covers the truth. Leads who try to solve this by explaining harder tend to lose, because they are fighting the most natural reading of a number that looks like a probability. The workable approach treats it as an interface problem. If the default output produces the wrong sentence, fix the default output. ## Set the house phrasing Decide once what a reported result looks like and make it the template everywhere - dashboards, decks, write-ups: - **Estimate and interval together, in the outcome's own units.** Never a bare verdict, never an interval without the estimate it came from. - **Compatibility language.** "Values from 4.1 to 5.3 are the ones compatible with these data" reads naturally and is correct. "We are 95% confident the mean is between 4.1 and 5.3" is acceptable as shorthand once people know what it stands for. - **A line about what is ruled out.** The practically useful content of an interval is often its exclusions: what the data say cannot be happening. - **Method sentence where it fits.** "Produced by a method that covers the true value 95% of the time" costs one clause and is exactly right. A template that is correct by construction removes the need for individual vigilance, which is the only version of this that survives contact with a busy organisation. ## Explain the picture once, properly There is one image that does the teaching: many intervals drawn from many samples, scattered around a single fixed line, with about one in twenty missing it. Show it in onboarding or a short internal session, tie it to the sentence template, and then let the template carry the load. Repeating the correction in every review meeting is expensive, condescending, and does not scale. ## Triage by consequence A lead's real judgment here is knowing when the misreading matters: - **Low stakes.** The interval sits entirely on one side of every threshold anyone cares about. The loose phrasing leads to the same action as the precise one; correcting it publicly buys nothing and spends credibility. - **High stakes.** A decision threshold falls inside the interval. Now the sloppy mental model does damage - it invites treating a value near an endpoint as nearly settled, or reading a range that contains both a meaningful gain and a meaningful loss as a mild positive. Insist on exact wording here, every time. - **Aggregation.** People who think of each interval as a 95% probability statement start chaining them across many reported results, which produces confidently wrong conclusions. Push back on that pattern wherever it appears. ## Guard the assumptions too The stated level is what the method promises *if* the model behind it holds: how the sample was drawn, whether observations are independent, whether an approximation is reasonable at this sample size. A team that says 95% while quietly violating those conditions has a bigger problem than phrasing. Part of the standard should be that anyone quoting a coverage level can name what it rests on, and that pipelines producing intervals are checked - for instance by simulating the whole procedure against a known value and counting how often the interval covers it. ## Handle the request for a single number Someone will ask for the number without the range. Give the estimate, then a plain-language precision statement in the same breath: "about 4.7, and the data cannot distinguish anything between 4.1 and 5.3." Where the decision threshold lands inside the interval, decline to strip the uncertainty and say why - that is the case where the single number is actively misleading rather than merely incomplete. ## What a strong answer sounds like Diagnose the misreading precisely, then propose systems rather than exhortation: a reporting template, one good visual explanation, an explicit triage rule for where precision is worth insisting on, and a check that the nominal level is actually being delivered. A candidate who says only "I would educate stakeholders" has not made a decision; a candidate who says which sentence appears on the dashboard tomorrow has.