Compare the targets directly
| Feature | Confidence interval for mean response | Prediction interval for new observation |
|---|---|---|
| Target | Conditional mean at a predictor value | One future observation at that predictor value |
| Includes mean-estimation uncertainty | Yes | Yes |
| Includes residual variation of a new observation | No | Yes |
| Typical width | Narrower | Wider |
| Common software label | Confidence interval or mean interval | Prediction interval |
This comparison assumes the same fitted model, predictor value, nominal confidence level, and interval method. Under those matched conditions, the prediction interval contains the corresponding mean-response confidence interval.1
Why a future observation adds uncertainty
Imagine estimating the average outcome for all cases with predictor x = 10. The fitted mean is uncertain because the model was estimated from a sample. A confidence interval represents that estimation uncertainty.
A new observation has another source of variation: individual outcomes differ around the conditional mean. The prediction interval adds that residual variation. More data can narrow uncertainty about the mean substantially, while irreducible observation-level spread can keep the prediction interval wide.
That width is not a software defect. It answers the harder question.
How the two intervals appear in output
R’s linear-model prediction method, for example, distinguishes interval="confidence" from interval="prediction" and documents the assumptions behind the resulting bands.2 Other tools may use labels such as “CI for mean,” “mean response,” “individual prediction,” or “forecast interval.”
Check four details before interpreting the bounds:
- the target—mean or new observation;
- the confidence level;
- the predictor values and whether they are inside the observed range;
- whether the model includes residual and parameter uncertainty in the stated way.
A worked reading
Suppose a model gives a fitted mean of 50 at x = 10, with:
- 95% confidence interval for the mean: [47, 53];
- 95% prediction interval for a new observation: [35, 65].
The first interval describes uncertainty about the conditional mean. The second describes a plausible range for one new outcome under the model and therefore includes much more variation. It is incorrect to say 95% of future observations fall inside the first interval.
Confidence interval interpretation covers the repeated-sampling meaning of the confidence level and the usual null-value checks.
What can make both intervals unreliable
Both intervals depend on the fitted model and construction method. Problems can arise from:
- extrapolation beyond the observed predictor range;
- curvature omitted from a linear model;
- changing residual variance;
- dependent or clustered observations;
- influential cases;
- a future population that differs from the training data.
A narrow interval from a misspecified model is not trustworthy merely because it is narrow. Plot residuals and document the prediction setting.
Prediction intervals are not credible intervals
A Bayesian credible interval makes a probability statement about a parameter or predictive quantity under a prior, likelihood, and posterior model. A bootstrap interval is defined by a resampling procedure. These are separate from the frequentist mean-response and observation-level distinction here. Bootstrap and credible intervals handles those interpretations.
Common reading errors
- Using the confidence interval for a mean as a range for individual outcomes.
- Calling a prediction interval “less precise” without noting its different target.
- Comparing intervals with different confidence levels.
- Ignoring whether the prediction is interpolation or extrapolation.
- Assuming a wide prediction interval proves the sample is poor.
A defensible reporting sentence
“At x = 10, the fitted mean was 50 (95% CI for the mean [47, 53]); the 95% prediction interval for one new observation was [35, 65].” Naming both targets makes the width difference interpretable.