Read the estimate, interval, and scale together
Suppose an output reports an estimated mean difference of 2.1, with a 95% confidence interval from 0.4 to 3.8. Read it in this order:
- Estimate: 2.1 is the point estimate from this sample.
- Interval: values from 0.4 to 3.8 are reasonably compatible with the data under the model and interval procedure.
- Scale: these values are differences in the original outcome units. They are not ratios or percentages unless the output says so.
- Null value: for a difference, the usual null value is 0. This interval does not include 0.
That last observation connects the interval to a matching two-sided hypothesis test. Under the same assumptions, a 95% interval that excludes the null corresponds to rejection at the 5% level.2 It does not prove that the effect is real, important, correctly modeled, or free from bias. For that distinction, see how to interpret the p-value and statistical versus practical significance.
“Includes zero” and “crosses one” are different checks
The null value depends on the kind of estimate.
| Result type | Usual null value | What an interval containing it means |
|---|---|---|
| Difference in means, slopes, or additive effects | 0 | The data remain compatible with no additive difference under the model. |
| Odds ratio, risk ratio, or other ratio measure | 1 | The data remain compatible with equal odds, risk, or rates under the model. |
| Correlation | 0 | The data remain compatible with no association of the measured kind. |
So an odds-ratio interval of 0.78 to 1.34 crosses 1, not 0. A mean-difference interval of −0.6 to 1.8 includes 0. Applying the wrong null value is a common but consequential reading error.
What width tells you
Interval width is primarily about precision. Holding other features constant, larger samples usually produce narrower intervals because the estimate is less variable across samples. More variable data, fewer observations, stronger clustering, or a higher confidence level usually produce wider intervals.
A wide interval is not automatically “bad data.” It may honestly show that the available data cannot distinguish between materially different values. For example, an estimate may point in one direction while its interval includes both a negligible and a substantial effect. That is a reason to report the uncertainty, not to hide it behind the point estimate.
There is no universal acceptable confidence-interval width. A width of 2 units could be negligible on one scale and decision-changing on another. The responsible question is whether the interval is narrow enough to separate the values that matter in the study’s domain.
Confidence level is not the interval itself
The confidence level describes the procedure. The interval is the particular pair of limits produced from the observed data. If the same sampling process were repeated and a 95% interval calculated each time, about 95% of those intervals would cover the parameter under the method’s assumptions.1
Increasing the confidence level from 90% to 95% or 99% generally widens the interval. You are asking the procedure to cover the parameter more often, so it must cast a wider net. Comparing intervals is meaningful only when you know their confidence levels and construction methods.
What a confidence interval does not tell you
A frequentist confidence interval does not, by itself, tell you:
- the probability that the parameter lies inside this realized interval;
- the probability that the null hypothesis is true;
- whether an effect is practically important;
- whether the model assumptions are credible;
- whether selection, measurement, or publication bias affected the estimate;
- where a future individual observation will fall.
That last job belongs to a prediction interval, which includes uncertainty about a future observation and is therefore normally wider than the confidence interval for a mean response. Bootstrap and credible intervals require their own interpretations because their construction and probability statements differ.
A compact reading rule
Read a confidence interval as: estimate + compatible range + scale + null value + context. If any of those pieces is missing, resist reducing the result to “significant” or “not significant.” The interval is most useful when it shows both direction and uncertainty without pretending that either is absolute.