The question hidden inside the interval
Imagine a clinic estimating how long people wait for an appointment. One reader wants to know the clinic’s average waiting time. Another wants to know how long their own next wait might be. The same dataset can inform both questions, but the uncertainty is different.
A confidence interval for the mean concerns the first question. A prediction interval for a new individual observation concerns the second. Calling both an “expected range” conceals the distinction. It can make a precise estimate of the average sound like a narrow promise about everyone’s experience.
A numerical example
Suppose an invented sample of 100 independent waiting times has a mean of 20 minutes and a standard deviation of 10 minutes. Under a simple normal model, an approximate 95% confidence interval for the population mean is about 18 to 22 minutes. The corresponding prediction interval for one new waiting time is much wider, roughly 0 to 40 minutes.
The figures illustrate the difference between estimating an average and predicting an individual observation. They are not a recommended model for actual waiting times. Real waiting times may be skewed, depend on time of day, or be correlated within clinic sessions. A normal model can even produce impossible negative lower limits under some data patterns, which would signal a mismatch requiring a more suitable approach.
| Interval target | Uncertainty included in this example |
|---|---|
| Population mean | Sampling uncertainty in the estimated average. |
| One future waiting time | Uncertainty in the average plus variation among individual waits. |
| Average of several future waits | Uncertainty in the average plus variation in that future group’s mean. |
NIST’s prediction-limit formula explicitly includes both the size of the original sample and the size of the future set of observations. Predicting one new value is therefore different from predicting the average of many new values.
Why collecting more data does not make everyone alike
If the clinic observes thousands of visits under the same stable conditions, its estimate of the mean can become very precise. Individual visits can still vary considerably. Larger samples reduce uncertainty about the average; they do not remove the variability of the process being observed.
This distinction appears in many settings. A study may estimate average recovery time precisely while individual recovery remains variable. A university may report a stable mean time to completion while students finish over a broad range. A narrow confidence interval around an average does not imply that an individual outcome will fall close to it.
What does 95% mean here?
For a conventional frequentist confidence interval, the confidence level refers to the long-run coverage of the interval-generating procedure under its assumptions. If the process were repeated many times, about 95% of intervals produced by a valid 95% procedure would contain the fixed parameter. It is not a statement that 95% of individual observations lie inside this particular interval.
A prediction interval instead targets future observations, with coverage defined for that prediction procedure. The distinction is about what is being covered. A Bayesian credible interval uses a different probability framework and should be identified as such rather than treated as a synonym for any range around an estimate.
Regression adds a location to the question
A regression model might estimate waiting time for a particular day or patient category. A confidence band around the fitted mean describes uncertainty about the modeled average at those predictor values. A prediction band for an individual outcome includes residual variation as well.
Predictions far outside the range of the observed data introduce another problem. A model can calculate a narrow-looking interval while relying on a relationship that has not been tested in that region. Mathematical output does not by itself validate extrapolation.
What neither interval fixes
Suppose the clinic records only morning visits but presents the interval as applying to every appointment. A large morning sample does not resolve missing afternoon information. Similarly, if staffing changed after data collection, a well-calculated interval for the old process may not describe the new one.
Intervals make a particular kind of uncertainty visible. They do not automatically include every source of uncertainty, such as systematic measurement error, selection bias, or changes in conditions. The model, population, and measurement process remain part of the interpretation.
A clear label therefore does more than say “95%.” It identifies the quantity, the population or future observation, and the conditions under which the interval was calculated. Those details explain why two different ranges can both be correct while answering different questions.
