Effects of Sample Size and Residual Noise on Interval Width

Why Prediction-Interval Width Changes

A prediction interval for a new outcome has two sources of uncertainty. The fitted regression mean at a predictor value is estimated from a finite sample, so that mean is uncertain. In addition, a new individual outcome can differ from its conditional mean because of its own error. More observations or a better-spread set of predictor values can reduce the first source. They do not, by themselves, remove the second.

This distinction explains why a prediction interval can remain wide even when a fitted mean is estimated precisely. Its width depends on sample size, the residual standard error, and how well the observed predictor values support a prediction at the target location. These factors work through different parts of the interval formula and should not be treated as interchangeable.

The Formula and Its Components

For simple linear regression, let \(x_0\) be the predictor value for the future case, \(\bar{x}\) the sample mean of the observed predictor values, and \(S_{xx}=\sum_{i=1}^{n}(x_i-\bar{x})^2\). The usual two-sided prediction interval for one new outcome at \(x_0\) is

Here, \(\hat{y}_0\) is the fitted value at \(x_0\), \(s\) is the residual standard error, and \(t_{n-2,\,1-\alpha/2}\) is the relevant critical value from a \(t\) distribution with \(n-2\) degrees of freedom. The total width is the upper endpoint minus the lower endpoint:

The formula is for the standard regression prediction interval under a linear mean model with independent errors of mean zero and constant variance; the usual \(t\)-based procedure also assumes normally distributed errors. It treats the new predictor value as fixed. The interval estimates uncertainty for one future outcome, not just uncertainty about the conditional mean.

The predictor-dependent quantity \(h_0=1/n+(x_0-\bar{x})^2/S_{xx}\) is the leverage contribution for this prediction. It is smallest at \(x_0=\bar{x}\), where it equals \(1/n\). Moving the prediction location away from the sample mean increases it, unless a larger \(S_{xx}\) offsets that distance. The term \(S_{xx}\) measures the spread of observed predictor values around their mean.

A Short Justification for the Two Uncertainty Sources

The future outcome can be written as its conditional mean plus a new error. The prediction error is therefore the sum of the error in estimating the mean and the new individual's error. Under the model assumptions, these two contributions are uncorrelated; their variances add. In simple regression, the variance of the estimated mean at \(x_0\) is \(\sigma^2[1/n+(x_0-\bar{x})^2/S_{xx}]\), while the new individual error has variance \(\sigma^2\). Thus the prediction-error variance is

Replacing the unknown \(\sigma\) by the residual standard error \(s\), and using the \(t\)-based critical value, gives the interval formula above. This variance decomposition is the reason the leading \(1\) cannot be dropped when the target is a new individual outcome.

Sample Size, Residual Noise, and Predictor Spread

Increasing \(n\) can reduce \(1/n\), and often increases \(S_{xx}\) when the additional observations extend the range or density of predictor values. Either change can reduce mean-estimation uncertainty. However, “more observations” alone does not guarantee that \(S_{xx}\) increases substantially: if new observations cluster tightly in an unhelpful part of the predictor range, the prediction at a distant \(x_0\) may still have high leverage. The degrees of freedom also affect the \(t\) critical value.

Residual standard error \(s\) multiplies the entire width. If the design and critical value are held fixed, multiplying \(s\) by a factor multiplies the interval width by the same factor. Lower residual noise therefore narrows both the mean-estimation and individual-outcome contributions as measured in outcome units. But the residual standard error is not the same thing as mean-estimation uncertainty: it describes the scale of outcome variation left around the regression line.

Increasing predictor spread can help predictions away from \(\bar{x}\). With the same distance \(x_0-\bar{x}\), a larger \(S_{xx}\) makes \((x_0-\bar{x})^2/S_{xx}\) smaller. The improvement applies to the mean-estimation part; it does not erase the leading \(1\). Also, a change in predictor spread does not necessarily improve every prediction equally: at the center, the distance term is already zero.

What the Width Can and Cannot Tell You

The formula offers a useful way to diagnose why two prediction intervals differ. A smaller \(s\) indicates less residual variation in outcome units. A smaller leverage contribution indicates more precise estimation of the conditional mean at the target location. A larger sample can contribute to the latter, but only in combination with where the predictor values lie. These are distinct features of the data and the model.

Even if the mean-estimation contribution became very small, the width would still include the leading \(1\). With \(s\) and the critical value fixed, the width approaches \(2t\,s\) as the leverage contribution approaches zero. This is a conceptual lower limit under those fixed quantities, not a guarantee that a real interval will attain it. In real analyses, changing the sample or design can also change \(s\), the fitted model, and the critical value.

A prediction interval’s units are the outcome units. If an outcome is measured in milligrams, for example, both its width and residual standard error are in milligrams; predictor spread has squared predictor units and enters only through a dimensionless leverage term. Keeping these units straight helps prevent confusing predictor variability with outcome noise.

Common Mistake: “A Bigger Sample Makes Every Prediction Precise”

The incorrect reasoning is: “The sample size doubled, so the prediction interval must become half as wide.” The formula does not support that conclusion. Only particular components change with \(n\); the outcome-noise term remains, and predictor-location leverage depends on both the distance from \(\bar{x}\) and \(S_{xx}\). The critical value changes too, but not in direct proportion to sample size. A larger sample that adds little useful predictor spread may bring only a modest improvement for a distant prediction.

A related error is to interpret a narrower prediction interval as proof that individual outcomes have become less variable. If the interval narrowed because \(n\) increased or predictor spread improved while \(s\) stayed fixed, the mean was estimated more precisely; the residual outcome variation did not decrease. Always identify which part of the formula changed before explaining a width difference.