A paper can report an estimate that looks useful while leaving substantial uncertainty about the size of the effect. A confidence interval helps you see that uncertainty. It should change the sentence you write about the result.
For a difference in average scores, zero represents no difference. For a ratio, one is the corresponding reference value. That distinction matters before you ask whether an interval crosses the no-effect value. Cochrane recommends interpreting the estimate and interval rather than relying on a significance label alone. Cochrane Handbook, Chapter 15.
This article uses invented data and a fictional score. The numbers are for interpretation practice, not recommendations about a real outcome measure.
Define what a positive number means
Suppose a fictional trial compares two learning-support approaches in a rehabilitation programme. Its outcome is a 0–100 self-management score, where higher is better. The reported effect is the additional-support mean minus the comparison mean at three months.
An estimate of +4 therefore favours additional support on that score. If you reverse the subtraction order, the sign changes. Always write the direction beside the estimate. A minus sign is not automatically bad, particularly when a scale measures problems or symptoms.
Now consider three independent hypothetical results:
| Result | Estimated mean difference | 95% confidence interval |
|---|---|---|
| A | +4 points | −2 to +10 |
| B | +4 points | +3 to +5 |
| C | +1 point | +0.2 to +1.8 |
The examples are deliberately constructed to isolate interpretation. They are not three estimates from the same trial.
Read the whole range
Result A has a positive estimate but a wide interval. It includes a small disadvantage, no difference and a larger advantage. Writing “additional support does not work” discards most of that information. Writing “additional support improves scores by four points” without the interval sounds more settled than the evidence warrants.
A more useful sentence is: “The estimated difference favoured additional support by four points, with substantial uncertainty ranging from a two-point disadvantage to a ten-point advantage.”
Result B has the same estimate but much greater precision. The central number has not changed. What has changed is how tightly this analysis estimates it. Before declaring the effect worthwhile, you still need to know what a few points mean on that instrument and what the support costs or burdens people with.
Result C has an interval entirely above zero, yet the estimated difference is only one point. A precise small effect and an important effect are not the same judgement. The word “significant” can conceal that distinction if you do not explain the scale.
Use an explicit importance threshold as a discussion aid
For this teaching exercise, suppose the discussion group sets three points as a provisional difference worth considering. This is an invented threshold, not a validated clinical cutoff.
Result A leaves open differences both below and above that threshold. Result B sits at or above it. Result C remains below it throughout the interval. This comparison helps the group see why the same statistical question does not resolve every practical decision.
Then ask where a real threshold would come from. It might require research on the instrument, the relevant population and the perspectives of people receiving care. Do not choose a threshold after seeing the result just to make the study look favourable.
The exercise is also a reminder to look beyond the chosen score. A modest benefit with a large workload, access burden or risk might lead to a different discussion from the same benefit with little burden. The confidence interval contains neither those values nor that implementation decision.
What 95% does and does not describe
Under the statistical model and sampling assumptions, a procedure for constructing 95% confidence intervals is designed to include the true parameter in 95% of repeated applications. It does not mean 95% of individual participants improved, and it is not a prediction interval for individual outcomes. The confidence level also does not certify that the study has no bias. Cochrane explanation of confidence intervals.
The endpoints are not hard boundaries beyond which effects are impossible. Nor does the interval assign equal probability to every value inside it. A frequentist confidence interval describes the behaviour of an estimation procedure; it is not a probability distribution for this study’s fixed parameter.
Imagine two students reading Result B. One says, “Ninety-five percent of participants improved by three to five points.” The other says, “The interval describes uncertainty around the estimated group difference.” The second student has identified the relevant quantity. Neither can infer the full distribution of individual responses from that interval alone.
Check the outcome and the analysis
A narrow interval around a poorly chosen outcome can be precise about something that is not your question. A narrow interval from a biased analysis can also be misleading. Read how participants were selected, how missing data were handled and which outcome was specified before the results were known.
Avoid comparing intervals from different outcomes as though they shared a common scale. An interval two points wide on one questionnaire is not automatically more useful than an interval ten points wide on another. The units and meaning of the instrument belong in your interpretation.
Try the ratio version
An invented study reports a risk ratio of 0.80 with a 95% confidence interval from 0.60 to 1.10 for an undesirable event. Which conclusion is most faithful?
“An estimated lower risk, with uncertainty that includes no difference and some increase in risk” preserves the range. “No benefit exists” and “risk is definitely 20% lower” both overstate what the result establishes. A real appraisal would also need absolute risks and study quality.
For your next paper, write three lines: the quantity being estimated, the range of effects compatible with the reported interval under the model, and the practical information still missing. This short record makes your uncertainty specific enough to discuss.
Sources and further reading
- Cochrane Handbook Chapter 15 — Interpreting results — Confidence intervals and interpretation beyond a significance threshold.