A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
A prediction interval estimates where a new individual observation will fall; a confidence interval estimates the mean response. Prediction intervals are wider because they include both uncertainty in estimating the mean and natural variation around the mean.
From your work with inference in linear regression, you know that the fitted line ŷ = β̂₀ + β̂₁x is itself uncertain — it's estimated from data, so it wobbles depending on which sample you draw. A confidence interval for the mean response captures exactly this uncertainty: at a given x value, where might the true population mean μ_Y|x lie? That interval shrinks as sample size grows, because with more data the estimated line stabilizes around the truth.
A prediction interval asks a different and harder question: where will the *next single observation* at that x value land? Even if you knew the regression line perfectly — even with infinite data — individual observations would still scatter around it. That scatter is the irreducible noise term ε, with variance σ². A prediction interval must account for *both* sources of uncertainty: the estimation uncertainty in the mean (which goes to zero as n → ∞) and the irreducible observation-to-observation variance (which does not).
Mathematically, the prediction interval at a given x* is ŷ* ± t* · SE_pred, where SE_pred² = s²(1 + h), with h capturing the leverage of x* and the "1" term being the irreducible variance contribution. The "1 +" is the essential difference: the confidence interval uses SE² = s² · h alone, without the leading 1. Because SE_pred > SE_mean always, prediction intervals are always wider — often substantially so, especially for small samples.
The practical lesson is to match the interval to the question. If you want to know the expected height of all 40-year-old men in a population, use a confidence interval for the mean. If you want to know where one specific 40-year-old man's height will fall, use a prediction interval. Confusing them leads to either false precision (using a confidence interval when you need a prediction interval) or unnecessary alarm (the opposite direction). The confidence interval tells you about the center of a distribution; the prediction interval tells you about the distribution itself.
As x* moves away from x̄ (the center of your data), both interval types widen — leverage h increases the farther you extrapolate. But the prediction interval widens more slowly in relative terms because the "1" dominates when h is small. Near the center of the data, prediction intervals are roughly twice as wide as confidence intervals; far into extrapolation territory, both balloon together. This is why extrapolation with a prediction interval makes caution concrete: you can literally see how much uncertainty you're projecting onto a single future observation.