A poll lands: 48 percent of voters back the plan. The headline writes itself. Then someone points to the footnote — margin of error, plus or minus four points. Now the story is "somewhere between 44 and 52 percent," a range that includes both a comfortable win and a dead heat. That range is a confidence interval. It is one of the most useful ideas in data journalism, and one of the most routinely misread.
What a confidence interval actually is
Most numbers journalists care about describe a group too large to measure completely: every voter in a country, every patient in a trial, every household in a city. So we measure a sample and estimate.
A confidence interval puts honest boundaries around that estimate. Instead of "the average commute is 34 minutes," you get "the average commute is 34 minutes, and the data are consistent with anything from 32 to 36."
The "95%" attached to most intervals describes the method, not your particular result. Repeat the same study many times and roughly 95 out of 100 of those intervals would contain the true value. Your one interval either does or doesn't — you just can't tell which. This is subtle, and worth being careful about, because the intuitive reading ("there's a 95% chance the truth is in here") isn't quite what the number means in the standard statistical framework. Some analyses use a Bayesian approach, where the plain-English reading is closer to correct. If the distinction matters for your story, ask which kind of interval you're looking at.
The width is the story, not a footnote
A narrow interval means the estimate is precise. A wide one means the data leave a lot of room. Reporters often treat width as a technicality to bury, when it is frequently the most newsworthy part.
Suppose a trial suggests a treatment cuts the risk of a complication, with an interval running from a 2 percent reduction to a 40 percent reduction. Every value in that range is plausible given the data. Writing "the treatment reduces risk" is defensible. Writing "the treatment cuts risk by 40 percent" picks the most dramatic end of the range and pretends the rest doesn't exist.
Now suppose the interval runs from a 10 percent increase in risk to a 15 percent decrease. It crosses zero, the point of no effect. The honest sentence is that the study didn't settle the question. That's a result too, and readers deserve it.
Width usually comes down to three things: how many observations you have, how much the underlying values vary, and how confident you want to be. More data narrows the range. A 99% interval is wider than a 95% one, because you're asking for more certainty.
Four things a confidence interval is not
- Not a description of the data. It doesn't tell you where 95% of individual values fall. That's a different range, usually a much wider one.
- Not a guarantee of quality. A biased sample produces a tight interval around the wrong answer. Precision and accuracy aren't the same thing.
- Not a probability about your specific interval. The 95% is a property of the procedure, not a bet on this one result.
- Not comparable by default. Two intervals can both say "95%" and mean very different things, depending on design, sample and which sources of error were included.
Where you'll meet them in the wild
Polls
The margin of error on a poll is essentially a confidence interval for a single percentage. It usually covers sampling error only — not question wording, not who bothered to answer, not likely-voter screens. And the margin for a difference between two candidates is larger than the margin for either one alone, which is why a two-point lead inside a four-point margin isn't a lead.
Health and medical research
Intervals around risk estimates are standard here, often expressed as ratios. A result whose interval includes 1.0 — the "no difference" point for a ratio — is usually described as not statistically significant. That is not proof of no effect; it may simply mean the study was too small to detect one. None of this is personal medical advice, and readers making health decisions should talk to a clinician.
Economic and social data
The surveys behind unemployment, inflation and crime figures carry sampling error, and many series get revised later. Treat early point estimates as provisional, and check whether an interval is reported at all. Often it isn't.
How to write about them without losing readers
You rarely need the phrase "confidence interval." You do need the idea.
- Lead with the estimate, then the range. "Support sits at 42 percent, though the poll can't distinguish that from anything between 38 and 46."
- Use plain verbs. "Somewhere between," "could be as low as," "the data can't rule out." Accurate and readable at the same time.
- Resist false certainty in both directions. Don't write "proves." Don't write "meaningless" either, just because an interval is wide.
- Round like a human. "Between 32 and 36 minutes" beats "31.7 to 36.4 minutes" in almost every story.
- Keep caveats close to the claim. A note two paragraphs down is fine. A footnote nobody reads is not.
- Ask what's inside the interval. If a source can't say whether it accounts for non-sampling error, that tells you something about the source.
An interval isn't a hedge. It's the most precise honest sentence the data can support.
Four questions to ask before you publish
- What is the interval, and what confidence level goes with it?
- Which sources of uncertainty does it include, and which does it leave out?
- Does it cross a threshold that matters — zero, 1.0, 50 percent, a policy target?
- What would the story say if the true value sat at each end of the range?
A practical habit for your next story
Before filing, write two sentences: the one you'd use if the estimate landed at the low end of the interval, and the one for the high end. If those sentences point in genuinely different directions, you have a story about uncertainty. If they don't, you can report the number plainly — and say so, without hedging the whole piece into mush.
Readers handle ranges better than journalists often assume. Give them the estimate, give them the boundaries, and explain what the space between means. That isn't dumbing anything down. It's the version of the story that survives contact with the next round of data.
