Scheffé Protection Over Every Linear Contrast

Why Protect Every Contrast?

A mean vector has many possible summaries. For a population mean \(\mu\in\mathbb{R}^p\), each nonzero vector \(c\in\mathbb{R}^p\) defines a linear contrast \(c'\mu\). If the direction \(c\) is selected after looking at the data, an interval calibrated for one prespecified direction need not retain its advertised coverage. Scheffé protection addresses the broader demand: construct bounds that cover every linear contrast at once.

The key is to measure the largest error over all directions after standardizing each contrast by its estimated standard error. That maximum is not an intractable search. It equals the square root of the Hotelling quadratic form for the mean, so an exact Gaussian critical value follows from the \(F\) distribution.

Maximizing the Standardized Contrast Error

For any fixed nonzero \(c\), the estimated standard error of \(c'\bar X\) is \(\sqrt{c'Sc/n}\). The standardized estimation error is therefore \[ \frac{c'(\bar X-\mu)}{\sqrt{c'Sc/n}}. \] To protect all contrasts, we need to control the absolute value of this quantity simultaneously over every nonzero direction \(c\).

To prove the result, set \(a=S^{1/2}c\) and \(b=S^{-1/2}\delta\), where \(S^{1/2}\) is the positive definite square root of \(S\). Then \(c'\delta=a'b\), while \(c'Sc=a'a\). The Cauchy–Schwarz inequality gives \[ \frac{|c'\delta|}{\sqrt{c'Sc/n}} =\sqrt{n}\frac{|a'b|}{\sqrt{a'a}} \leq \sqrt{n\,b'b} =\sqrt{n\,\delta'S^{-1}\delta}. \] When \(\delta\ne0\), choose \(a\) proportional to \(b\), which is equivalent to choosing \(c\) proportional to \(S^{-1}\delta\); equality then holds. When \(\delta=0\), every numerator is zero. This proves the result. \(\square\)

The maximizing direction generally depends on the unknown error \(\delta=\bar X-\mu\). That is exactly why the maximum matters: controlling each direction separately without accounting for the search does not automatically control the direction that happens to have the largest error. The result turns the entire search into one quadratic statistic.

The Exact Gaussian Critical Value

Define Hotelling’s statistic for testing a specified mean vector \(\mu\) by \[ T^2=n(\bar X-\mu)'S^{-1}(\bar X-\mu). \] Under the Gaussian assumptions in the setup, its exact finite-sample distribution is \[ \frac{n-p}{p(n-1)}T^2\sim F_{p,n-p}, \] where \(F_{p,n-p}\) denotes an \(F\) random variable with numerator degrees of freedom \(p\) and denominator degrees of freedom \(n-p\). Write \(F_{p,n-p;1-\alpha}\) for its \(1-\alpha\) quantile. Thus the \(1-\alpha\) quantile of \(T^2\) is \[ K_\alpha=\frac{p(n-1)}{n-p}F_{p,n-p;1-\alpha}. \]

For completeness, the distributional statement follows from the standard Gaussian sampling results: \(\sqrt{n}\,\Sigma^{-1/2}(\bar X-\mu)\) is standard multivariate normal, and it is independent of \((n-1)\Sigma^{-1/2}S\Sigma^{-1/2}\), which has a Wishart distribution with identity scale and \(n-1\) degrees of freedom. The standard Hotelling \(T^2\) distribution for this independent normal vector and Wishart matrix gives the displayed \(F\) relation. The condition \(n>p\) ensures the inverse exists.

To prove the theorem, apply the maximum result with \(\delta=\bar X-\mu\). The event that every contrast is covered is equivalent to \[ \sup_{c\ne0}\frac{|c'(\bar X-\mu)|}{\sqrt{c'Sc/n}}\leq\sqrt{K_\alpha}, \] which, by the maximization identity, is equivalent to \(T^2\leq K_\alpha\). By the exact \(F\) distribution, this event has probability \(1-\alpha\). The bounds therefore have the claimed simultaneous coverage. \(\square\)

This derivation also explains the connection to the mean confidence ellipsoid. The condition \(T^2\leq K_\alpha\) describes an ellipsoidal region for \(\mu\); requiring the contrast bounds to hold for every direction is another way of expressing membership in that region. The common Scheffé multiplier is fixed by the ellipsoid’s exact Gaussian calibration, not by treating the infinitely many contrasts as independent tests.

Worked Contrasts

Interpreting the Protection

The word simultaneous refers to the probability of the whole collection of statements. With probability \(1-\alpha\), all the contrast intervals cover their respective true values at once. It does not mean that each interval is formed from an independent procedure, nor does it require listing the contrasts in advance. A contrast may be chosen after examining the data and the simultaneous guarantee still applies, provided it is a linear combination of the \(p\) coordinates covered by the stated setup.

The bound is unchanged if \(c\) is replaced by any nonzero scalar multiple of \(c\): both the estimated contrast and its margin scale by the same absolute factor. Thus a contrast direction, rather than a particular numerical scaling of its coefficient vector, determines the substantive comparison.

Common Mistake

• Using a one-contrast critical value for a contrast chosen after seeing the data. A \(t\)-critical value can calibrate a single prespecified contrast, but it does not by itself guarantee coverage over every possible direction. The correction is to use the Scheffé multiplier \(K_\alpha=p(n-1)F_{p,n-p;1-\alpha}/(n-p)\) when the claim must hold for all nonzero contrasts.
• Dropping the factor \(p\) or confusing the \(F\) quantile with the \(T^2\) quantile. The exact relation is \((n-p)T^2/[p(n-1)]\sim F_{p,n-p}\). Solving it for the cutoff of \(T^2\) gives \(p(n-1)F_{p,n-p;1-\alpha}/(n-p)\).
• Claiming exact finite-sample coverage without Gaussian sampling. The displayed \(F\) calibration is exact under the independent multivariate normal model with \(n>p\). Without those assumptions, it is not justified as an exact finite-sample result.