MerQuR+Statistics Glossary

Statistics Glossary → Prediction interval

Prediction intervalconcept

The range expected, with a given probability, to contain a single new observation (or the true effect of a new study), rather than the mean response.

In regression, two intervals can be given for a value of x. The confidence interval shows the uncertainty of the mean response (the mean of all units with that x); the prediction interval shows where the value of one new unit might fall. The prediction interval includes both the uncertainty of the mean and the spread between individuals: roughly ŷ ± t · √(s² + SE(ŷ)²). It is therefore always wider, and it does not shrink to zero as the sample grows.

In time-series forecasting the interval widens as the horizon lengthens. In meta-analysis, the prediction interval shows where the true effect of a new, similar study might lie: μ̂ ± critical value · √(SE² + τ²); with large heterogeneity (τ²) the interval can include zero even when the average effect is significant. These intervals assume that the model is correct and that the error distribution (usually normal) is appropriate.

An everyday example

A courier company estimates that 10 km deliveries take 30 minutes on average. The 95% confidence interval for that mean is 28-32 minutes, but the 95% prediction interval for a single 10 km delivery tomorrow is 18-42 minutes, because individual deliveries vary far more with traffic and finding the address.

In MerQur

'Predictions + CI/PI and Save to dataset' in Multiple Linear Regression (with the individual prediction interval columns LICI and UICI); the confidence/prediction band plot in Nonlinear Regression; 'Prediction interval (z or t(k − 2))' in Meta-Analysis; forecast intervals in ARIMA, Exponential Smoothing, VAR, GARCH and the Unobserved Components Model; and the predictive interval in Bayesian Linear Regression.

Analyses where this concept appears

Meta-AnalysisMultiple Linear RegressionNonlinear RegressionBayesian Linear RegressionARIMA ForecastExponential Smoothing (Holt-Winters)VAR (Vector Autoregression)GARCH (Volatility Model)Unobserved Components Model (Structural Time Series)

Related concepts

Confidence level and confidence intervalForecast horizon and forecast interval