Set and . These are orthonormal because . Complete them to an orthonormal basis of . An orthogonal transformation of the standard normal error vector has independent normal random variables as coordinates : its joint probability density remains because the transformation preserves squared length and has Jacobian determinant of absolute value one.
Projection onto these basis directions gives
Therefore the joint distribution of least-squares and variance estimators is
The last law is the chi-squared distribution of a sum of squared independent normal random variables with mean zero and variance one, and independence follows because the three statistics use disjoint independent coordinate sets. When all , omit the slope direction: and are independent, and there is no identifiable slope estimator.
For a new independent tumour grown for ten days in B, use . Its fitted mean is
Write . The prediction interval in a normal linear model is
where and is the indicated Student's t-distribution quantile. The design information needed to evaluate is not supplied, so a fully numerical interval cannot be recovered from the coefficient standard errors alone: the coefficient covariances are also needed.
To prove coverage, write , with independent of the original sample. The error has a normal distribution with mean zero and variance . By the joint distribution of least-squares and variance estimators, and is independent of ; it is also independent of . Hence the pivotal quantity
Consequently , which is exactly the claimed prediction interval coverage. This is repeated-sampling coverage conditional on the design, assuming the model; it need not hold conditional on the particular observed training responses. The extra represents new-observation noise and must not be omitted as in a mean confidence interval.
Write the one-column design as and . Maximizing the normal linear model likelihood first gives . Its minimized residual sum of squares is . For variance , the remaining log likelihood is plus a constant, whose maximum is at
Here almost surely because and the true variance is positive. In the exceptional data case , the likelihood is unbounded as , so there is no positive-variance maximum.
Let , the orthogonal projection matrix onto the design column. Then and . These are orthogonal linear projections of an isotropic multivariate normal distribution, hence independent. An orthonormal coordinate system adapted to makes the residual squared length a sum of independent squared standard normals. Thus the joint distribution of least-squares and variance estimators is the product law
In particular the maximum likelihood variance estimator is biased: ; the unbiased residual estimator uses denominator .