How to Interpret a Confidence Interval: A Complete Evidence-Writing Example
Read a confidence interval with its point estimate, outcome, units, comparison, and analysis method. Examine both limits against the null value and any justified importance threshold. An interval containing the null does not prove no effect, and a narrow interval does not remove bias.
Read the estimate, the interval, and the outcome together
A confidence interval helps describe the statistical uncertainty around an estimated quantity. Before interpreting it, identify what was estimated: a mean difference, risk ratio, odds ratio, or another measure. Then keep the outcome, comparison, units, and time point beside the numbers. An interval without those details is not an interpretable finding.
The strict frequentist meaning concerns the method: under its assumptions, a procedure with 95% coverage would contain the fixed true parameter in 95% of repeated applications. It does not assign a 95% probability to the parameter being inside this particular observed interval. Greenland and colleagues explain why model assumptions and the distinction between a procedure and one result matter.
This guide uses a fictional education study. Its outcome is a knowledge score from 0 to 100 one week after a teaching session; higher is better. The estimated adjusted mean difference, new session minus usual session, is 4 points with standard error 3. A prespecified normal approximation is assumed for this worked calculation. These invented results are not clinical evidence or a recommendation about teaching practice.
Extract the information before writing a conclusion
Start with the results table and methods, not just the abstract's adjective. In this example the report supplies an adjusted mean difference, so it would be incorrect to silently describe the estimate as an unadjusted comparison of raw group averages. The model and adjustment variables belong in the extraction note even if the short final paragraph does not list every coefficient.
Also distinguish a standard error from a standard deviation. Here the standard error of 3 describes uncertainty in the estimated difference. It is not a statement that individual learners' knowledge scores vary by only 3 points. Using one in place of the other would answer a different question.
For a real paper, record the confidence level and interval method as reported. A t-based, bootstrap, clustered, or model-based interval may use a different calculation. The arithmetic below is a demonstration of the stated normal approximation, not a universal recipe to replace the source analysis.
| Field | Reported or assumed value | Why it matters |
|---|---|---|
| Outcome and direction | Knowledge score, 0–100; higher is better | Positive differences favor the new session |
| Time point | One week after teaching | Do not relabel it long-term retention |
| Comparison | New session minus usual session | Fixes the meaning of positive and negative signs |
| Estimate | Adjusted mean difference +4 points | Keep adjusted status and score units |
| Standard error | 3 points | Uncertainty in the estimate, not individual variation |
| Interval method | Two-sided 95% normal approximation | Use 1.96 as the stated multiplier |
| Importance threshold | +5 points, stipulated for this teaching example | Not a validated threshold for a real instrument |
Calculate the example and inspect both ends
Under the stated approximation, the margin is 1.96 × 3 = 5.88 points. The lower limit is 4 − 5.88 = −1.88; the upper limit is 4 + 5.88 = 9.88. Report the estimated difference as +4 points, with a 95% confidence interval from −1.88 to +9.88 points.
The lower end permits a small disadvantage for the new session; the upper end permits a larger advantage. Zero is also inside the interval. Saying only 'there was no difference' would discard the positive estimate and the range of effects still compatible with the data under the analysis assumptions.
The example's stipulated threshold is +5 points. The point estimate is below that threshold, but the interval includes effects above it. Therefore the result does not establish that the new session reaches the example's important-benefit threshold, nor does it rule such a benefit out. This is a different question from whether zero lies in the interval.
| Step | Calculation or comparison | Result |
|---|---|---|
| Margin | 1.96 × 3 | 5.88 score points |
| Lower limit | 4 − 5.88 | −1.88 points |
| Upper limit | 4 + 5.88 | +9.88 points |
| Null comparison | Does −1.88 to +9.88 include 0? | Yes |
| Importance comparison | Does it include the stipulated +5-point threshold? | Yes; effects below and above it remain compatible |
| Conclusion | Keep estimate, interval, and threshold separate | The size and direction are not resolved precisely enough to establish important benefit |
Compare precision without turning it into a pass label
Consider three invented intervals around the same +4-point estimate. They illustrate different statistical precision, not three actual studies. With a narrow interval from +3 to +5, the estimate is concentrated near a modest improvement. With −2 to +10, it remains compatible with a small disadvantage, no change, or a larger improvement. With −12 to +20, the unresolved range is much wider.
The width alone does not determine whether evidence is trustworthy. A very precise estimate can still be biased by missing outcomes, confounding, selective reporting, or a poorly chosen analysis. Cochrane distinguishes statistical imprecision from the other domains that influence confidence in a body of evidence.
Do not compare intervals from different scales as though their widths had a common meaning. A width of 2 on a 0–10 scale is not automatically the same practical precision as a width of 2 on a 0–100 scale. First preserve the units and the decision the estimate is meant to inform.
| Illustrative 95% interval | Compatible possibilities | What remains unresolved |
|---|---|---|
| +3 to +5 points | A positive difference near the estimate | Whether the effect clearly exceeds the example’s +5 threshold |
| −2 to +10 points | Small disadvantage through larger advantage | Direction and important-benefit threshold |
| −12 to +20 points | Substantial effects in either direction | A much broader set of interpretations |
| Any of these intervals | Statistical results under a model | Risk of bias and applicability require separate appraisal |
Use the correct null value and avoid changing the confidence level
For a difference, no difference is 0. For a ratio, equal quantities give 1. Thus a mean-difference interval from −2 to +10 includes its null, while a risk-ratio interval from 0.70 to 1.20 includes a different numerical null. Keep the measure name visible so the reader does not apply the wrong reference value.
Changing from a 95% to a 90% interval makes the interval narrower under the same usual construction, but it changes the coverage target. It does not improve the data. Choose and report the confidence level according to the analysis plan, rather than selecting the interval that produces the desired conclusion.
For the fictional +4 estimate with standard error 3, a 90% normal interval using 1.645 is approximately −0.94 to +8.94. A 99% interval using 2.576 is approximately −3.73 to +11.73. The estimate stays +4 throughout. The different endpoints reflect the method's different confidence levels.
| Level and multiplier | Calculation | Approximate interval |
|---|---|---|
| 90%; 1.645 | 4 ± 1.645 × 3 | −0.94 to +8.94 |
| 95%; 1.960 | 4 ± 1.960 × 3 | −1.88 to +9.88 |
| 99%; 2.576 | 4 ± 2.576 × 3 | −3.73 to +11.73 |
Do not substitute an individual or prediction claim
The interval in the example concerns an adjusted average difference. It does not say that 95% of learners gained between −1.88 and +9.88 points, or that one particular learner's improvement will fall in that range. Individual variability requires different information.
Likewise, a confidence interval around a meta-analysis mean is not a prediction interval for the effect in a new setting. A precise pooled average can coexist with meaningful variation among settings. When the question is whether another setting will behave similarly, examine heterogeneity, applicability, and any suitable prediction analysis rather than treating the average's confidence interval as the answer.
These distinctions affect the sentence you write. 'The average difference was estimated imprecisely' is a claim about the analysis. 'Every learner will benefit' is a claim about individual outcomes that the interval never tested. Avoid letting the convenience of one pair of endpoints erase the actual target of inference.
Turn the extraction note into an accurate evidence paragraph
A weak summary of the fictional result would be: 'The intervention did not work because the interval crossed zero.' That wording declares absence of benefit without addressing the upper limit. Another weak version would say: 'There is a 95% chance it improves scores by four points.' That turns both the interval and the point estimate into a probability claim they do not provide.
A more faithful paragraph is: 'In the fictional study, the new teaching session had an estimated adjusted mean knowledge-score difference of +4 points at one week compared with the usual session (95% CI −1.88 to +9.88, on a 0–100 scale). The interval includes a small disadvantage, no difference, and improvements exceeding the example's stipulated +5-point threshold. The result therefore leaves important benefit unresolved under the stated analysis, and its practical use would also require appraisal of bias and applicability.'
For a real assignment, replace every element with the study's actual report and cite the result at its source. If the paper does not report the interval method, do not invent one. If an importance threshold is unavailable, say that its practical meaning has not been established rather than choosing a convenient cutoff after seeing the estimate.
- Keep the outcome, time point, measure, comparison direction, units, and confidence level together.
- Describe both ends of the interval and the point estimate without reducing them to a binary label.
- Identify the source of any importance threshold and distinguish it from the null value.
- Keep individual outcomes, future-setting predictions, and overall evidence certainty separate from this interval.