A confidence interval shows the precision of an estimate. Learn how to read its limits and avoid treating an average effect as a promise for every person.
A study reports that a teaching method improved scores by four points, with a 95% confidence interval from one to seven points. Four is the estimated average effect in that study. The interval tells us something about how precisely the study has estimated it. That extra information matters. Another study could report the same four-point estimate with an interval from a three-point loss to an eleven-point gain. The headline number would be identical, but the second study would leave much more uncertainty about the size and even the direction of the effect.
Confidence intervals are often omitted from short summaries because they seem harder to explain than a single number. That makes research look more settled than it is. You do not need to calculate an interval to use it. Start by checking what is being estimated, reading both ends, and asking whether they would lead to different practical decisions. All the numerical examples in this article are invented for explanation.
Suppose researchers compare two teaching methods using a sample of students. They calculate the difference between the groups’ mean test scores. That difference estimates an average effect, but a different sample of students would probably produce a somewhat different number. The mean is calculated by adding the scores and dividing by the number of students. Comparing the two means gives the estimated difference. Even with a fair study design, chance affects which students end up in each sample and the scores they produce.
The confidence interval expresses uncertainty associated with estimating the effect from that sample. It is calculated using a statistical method with assumptions about the data and how they were collected. Those assumptions matter. An interval is not a guarantee that a study has handled every possible error. It describes precision under the method used, so the design and measurement still need to be examined.
Imagine you could repeat the study many times. Each time, you pick a new sample in the same way. Each time, you work out a new range. For a 95% method, about 95 out of 100 such ranges would cover the true mean, if the rules hold. Some would miss it. You do not get a label that tells you which kind your one range is.
So the 95% describes how the method works over many tries. It does not mean 95% of students had score gains inside your range. And it does not, in this kind of method, give a 95% chance that the fixed true mean is inside this one range. That last point takes care to learn. For a first read, keep this idea: the range helps show how precisely the study has pinned down the mean. The Cochrane guide gives more detail on the formal meaning.
Go back to our made-up test-score study. The main estimate is a gain of four points. The range runs from one to seven points. That leaves room for a small gain and a larger one. It does not pin the effect to exactly four points. A headline that gives just four leaves out useful doubt. Now picture a second study. It also gives four as the main estimate, but its range runs from a loss of three points to a gain of eleven. That range leaves much more open. It includes no change at all.
The two middle numbers match. The two studies do not tell us the result with the same precision. The wider range is less settled.
Suppose the new way to study takes twice as long. A gain of one point may not be worth that time. A gain of seven may be worth it to some students. If the ends of the range would lead you to different choices, that is worth knowing. You may need more evidence before choosing. Do not turn above zero into useful for everyone. A small gain can be real and still matter little for your goal. A wide range can include a useful gain without proving it is there.
Put the cost and the time beside the range. Numbers help with a choice when you keep the actual task in view.
For a difference in scores, zero means no difference. A range from minus three to eleven crosses zero. It includes that no-change value. For a ratio, the key value is one. A ratio compares one amount with another by dividing. If two groups have the same risk, their risk ratio is one. So do not look for zero in every result. Read the label first. Is the range in points, minutes, or a ratio? Write the unit beside the numbers if that helps. A range without a label can be easy to read the wrong way. The units are part of the answer.
Suppose a study recruits only students who are already doing well and have plenty of time to practise. A large sample could produce a narrow interval for an effect in that group. The estimate may still not apply to students who need different support or have less study time. Adding more students from the same restricted group does not automatically fix that problem. An interval also cannot fix a poorly scored test or account automatically for students left out of the study. Ask how the data were gathered. A neat result can still rest on poor work.
This is why narrow does not mean true. It means more precise under the method used. The method itself still needs a check.
For the first example, a reasonable summary is that the study estimated an average gain of four points, with a 95% confidence interval from one to seven. Add who took part and how the comparison was made. Avoid saying that every student will gain between one and seven points: an interval for an average effect does not describe each person's result. The OpenStax introduction to confidence intervals provides further worked explanations. When reading a paper, use the interval alongside the study design. A narrow interval can still accompany a biased comparison, while a wide one may honestly reflect how little the available data can settle.
Cochrane / 2024 / Website
Relative/absolute effects and confidence-interval interpretation. Blog examples are invented, not treatment advice. Checked 2026-10-09.
OpenStax / 2023 / Website
Open textbook reference. Worked blog examples are original and labeled as examples. Checked 2026-10-09.
