Prediction Intervals Versus Confidence Intervals for Means

Two Intervals, Two Different Questions

A fitted regression value gives a point prediction at a specified predictor value. But an interval around that value can answer either of two questions: how uncertain is the mean response at that predictor value, or where might the outcome for one future individual fall? These targets are related, but they are not interchangeable.

A confidence interval for a conditional mean describes uncertainty in the regression line’s estimated mean at a chosen predictor value. A prediction interval describes uncertainty in a future individual response at that value. The prediction interval must account both for uncertainty in the estimated mean and for the individual outcome’s deviation from that mean.

Model and Interval Formulas

The formulas below are for simple linear regression. Assume the model \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), with independent errors having mean zero, common variance \(\sigma^2\), and a normal distribution. The predictor values are treated as fixed. The conditional mean is \(\beta_0+\beta_1x_0\), estimated by the fitted value \(\hat{y}_0=b_0+b_1x_0\). Let \(S_{xx}=\sum_{i=1}^n(x_i-\bar{x})^2\), and let \(s\) be the residual standard deviation. The residual degrees of freedom are \(df=n-2\).

Under these assumptions, the standard error of the estimated conditional mean at \(x_0\) is \(s\sqrt{1/n+(x_0-\bar{x})^2/S_{xx}}\). For a future individual outcome, the standard error of the prediction error is \(s\sqrt{1+1/n+(x_0-\bar{x})^2/S_{xx}}\). A two-sided interval with confidence level \(1-\alpha\) uses the critical value \(t^*=t_{1-\alpha/2,\,n-2}\).

The two intervals share the same center, \(\hat{y}_0\), and the same \(t^*\). The prediction interval has an additional \(1\) inside the square root. That term represents the residual variation among individual outcomes around their conditional mean. Estimating the mean does not require predicting that extra individual deviation; predicting one future outcome does.

Why the Prediction Interval Is Wider

The distinction can be seen by writing the future outcome as \(Y_{\mathrm{new}}=E(Y\mid x_0)+\varepsilon_{\mathrm{new}}\). Its prediction error is \(Y_{\mathrm{new}}-\hat{y}_0\), or \(\varepsilon_{\mathrm{new}}+(E(Y\mid x_0)-\hat{y}_0)\). The first part is the new individual’s error; the second is the estimation error in the fitted mean. Under the stated model, these components are independent, so their variances add.

Replacing \(\sigma\) with \(s\) gives the prediction standard error in the formula. By contrast, the confidence interval for the mean includes only the uncertainty in estimating that mean. The corresponding estimated variances differ by \(s^2\), so the prediction interval is wider whenever \(s>0\). The width difference does not mean the conditional mean itself is less precisely estimated; it reflects the different target.

A Second Comparison at the Center of the Data

At \(x_0=\bar{x}\), the location term \((x_0-\bar{x})^2/S_{xx}\) is zero. This makes the distinction especially clear: the standard error for the mean depends on averaging across the sample, while the prediction standard error still includes the individual residual variation.

Choose the Interval That Matches the Question

The choice is determined by what the interval is intended to cover, not by which interval looks more useful or is narrower. A confidence interval is appropriate for a question about an average response at a specified predictor value. A prediction interval is appropriate for a question about one new case. Both are conditional on the regression model being suitable and the new case having the stated predictor value.

Common Mistakes and a Practical Check

• Using a mean confidence interval for an individual. A narrow interval around the estimated mean does not describe the spread of individual outcomes. Use a prediction interval when the target is one new response.
• Calling a prediction interval an interval for the mean. The prediction interval includes residual variation and is not a more cautious confidence interval for the same target. It answers a different question.
• Thinking both intervals have different centers. For the same \(x_0\), both are centered at \(\hat{y}_0\); they differ in standard error and width.
• Ignoring units and context. State whether an interval is for a mean or an individual outcome, identify the predictor value, and report the response units.