Prediction intervals and confidence intervals are both used with regression models, but they answer different questions. A confidence interval estimates where the average response is likely to be for a given x value. A prediction interval estimates where one new individual data point is likely to fall for that same x value.
This distinction matters because predicting an individual outcome is more uncertain than estimating an average.
Understanding Statistics: Prediction Intervals vs Confidence Intervals
A regression line is built from a sample, not from every possible case. Its slope and intercept would change a little if a different sample were collected. This creates uncertainty about the line itself.
Near the centre of the observed x values, the data usually give the line strong support. Far from that centre, there is less support, so interval widths grow. This is why the bands around a regression line often curve outward at both ends.
Using a model beyond the range of the data is called extrapolation. An interval can still be calculated there, but it may give a false sense of security because the real relationship could change outside the measured range.
The two intervals share one source of uncertainty. Both depend on how accurately the regression line estimates the typical response. A prediction interval has an extra source.
Even if the true average were known perfectly, real individuals would not all land exactly on that average. Their differences from the line are called residuals. For example, a model may relate hours of study to test score.
Students who study the same number of hours can have different scores because of sleep, prior knowledge, stress, test conditions, or simple chance. The prediction interval must allow for this natural spread. This extra spread does not disappear just because the sample becomes very large.
The choice of interval depends on the decision being made. A school researcher estimating the average score for all students who study five hours would use a confidence interval. A teacher trying to estimate the score of one particular student would need a prediction interval.
In business, a company might estimate average delivery time for a route, then separately plan for the arrival time of one package. In health studies, researchers may estimate an average outcome for patients with similar measurements, while doctors must remember that one patient can differ greatly from that average. Treating an average estimate as a guarantee for an individual is a common and serious mistake.
Students should pay close attention to the level attached to an interval, often ninety five percent. In repeated studies using the same method, about ninety five out of one hundred such intervals would contain the target value. It does not mean there is a ninety five percent chance that a fixed target moves in or out of one already calculated interval.
The method has the long run success rate. Interval results also rely on model assumptions.
Residuals should show no clear pattern across x values, their spread should be reasonably stable, and unusual points should be checked. A narrow interval is useful only when the model fits the data reasonably well and the data represent the situation where the prediction will be used.
Key Facts
- A confidence interval for the mean response estimates E(y | x0), the average y value at a chosen x0.
- A prediction interval estimates ynew, one future individual response at a chosen x0.
- In simple linear regression, the fitted value is yhat = b0 + b1x.
- A typical confidence interval has the form yhat0 ± t*SEmean.
- A typical prediction interval has the form yhat0 ± t*SEpred, where SEpred is larger than SEmean.
- Prediction intervals are wider because they include both uncertainty in the estimated mean and random scatter of individual data points.
Vocabulary
- Confidence interval
- A range of plausible values for a population parameter, such as the mean response at a given x value.
- Prediction interval
- A range of plausible values for one future individual observation at a given x value.
- Mean response
- The average value of the response variable y for all individuals with the same x value.
- Residual
- The difference between an observed value and the value predicted by the regression model, written as residual = y - yhat.
- Standard error
- A measure of the typical uncertainty in an estimate or prediction.
Common Mistakes to Avoid
- Using a confidence interval to predict a single future value is wrong because a confidence interval only describes uncertainty in the mean response, not individual scatter.
- Thinking a 95% interval means there is a 95% chance the specific computed interval contains the value is wrong in frequentist statistics because the interval is random before sampling, while the true parameter or future value is fixed after the data are collected.
- Assuming the two intervals have the same width is wrong because prediction intervals include extra variation from individual outcomes around the regression line.
- Ignoring the x value where the interval is made is wrong because intervals usually get wider farther from the center of the observed x values.
Practice Questions
- 1 A regression model predicts yhat = 12 + 3x. At x = 4, the standard error for the mean response is 2 and the critical value is t* = 2.1. Find the 95% confidence interval for the mean response.
- 2 For the same model, at x = 4 the standard error for predicting an individual future value is 5 and t* = 2.1. Find the 95% prediction interval for an individual response.
- 3 A scientist wants to estimate the average test score of all students who study 6 hours, while a teacher wants to predict one new student's score after studying 6 hours. Which interval should each person use, and why?