F Calibration of Pillai's Trace

Why Calibrate Pillai’s Trace?

Pillai’s trace summarizes the strength of a multivariate hypothesis across several response directions. To use it for a hypothesis test, we need a reference distribution that turns the observed trace into a tail probability. The classical F calibration provides a convenient approximation: it transforms the trace using the response dimension, the rank of the tested restrictions, and the residual degrees of freedom.

This calibration is not a claim that Pillai’s trace has an exact F distribution in every multivariate design. It is an approximate null calibration under the standard multivariate linear model. Keeping that distinction clear matters: the calculation of the transformed statistic is exact, while interpreting its F tail as a p-value uses an approximation.

Quantities in the Calibration

Suppose a multivariate linear model has $p$ response variables and a tested hypothesis consisting of $q$ independent, estimable restrictions. Let $v$ be the residual degrees of freedom. Write $H$ and $E$ for the hypothesis and error sum-of-squares-and-cross-products matrices. When $E$ is positive definite, let $\lambda_1,\ldots,\lambda_s$ be the nonzero generalized roots of $H$ relative to $E$, where $s=\min(p,q)$ is the maximum possible number of such roots. Pillai’s trace is

V=\sum_{i=1}^{s}\frac{\lambda_i}{1+\lambda_i}.

The generalized roots are nonnegative, so each term lies in $[0,1)$. In particular, $0\leq V<s$. A degenerate hypothesis may have fewer than $s$ nonzero roots, but the calibration’s dimensions still use $p$ and $q$, not the observed number of nonzero roots.

Definition: In the classical Pillai F calibration, $p$ is the number of response variables, $q$ is the rank of the tested restrictions in the design or contrast space, and $v$ is the residual degrees of freedom. Set $s=\min(p,q)$, $m=(|p-q|-1)/2$, and $n=(v-p-1)/2$. The calibrated statistic and its approximate degrees of freedom are $$ F_{\mathrm{P}}=\frac{2n+s+1}{2m+s+1}\frac{V}{s-V}, \qquad \nu_1=s(2m+s+1), \qquad \nu_2=s(2n+s+1). $$ Under the null hypothesis, $F_{\mathrm{P}}$ is compared with an $F_{\nu_1,\nu_2}$ distribution as an approximation.

The rank $q$ counts restrictions, not the rank of $H$. Generically, the hypothesis SSCP matrix has rank $\min(p,q)$, and in a degenerate case its rank can be smaller. For example, a scalar response has $p=1$: the matrix $H$ can have rank at most one even when the design tests $q=2$ restrictions. The degrees of freedom must use $q=2$, the restriction rank.

Reading and Simplifying the Formula

The definitions of $m$ and $n$ give simpler expressions for the degrees of freedom. This is useful both for calculation and for checking that the dimensions have been entered correctly.

Derived result: The numerator and denominator degrees of freedom can equivalently be written $$ \nu_1=s\max(p,q), \qquad \nu_2=s(v-p+s). $$

To verify the first identity, observe that $2m+s+1=|p-q|+s$. If $p\geq q$, then $s=q$, so $|p-q|+s=p-q+q=p=\max(p,q)$. If $q\geq p$, then $s=p$, so $|p-q|+s=q-p+p=q=\max(p,q)$. Thus $\nu_1=s\max(p,q)$. For the second identity, $2n+s+1=v-p-1+s+1=v-p+s$, and multiplying by $s$ gives $\nu_2=s(v-p+s)$.

The coefficient in the statistic is the ratio of these degrees of freedom:

\frac{2n+s+1}{2m+s+1}=\frac{\nu_2}{\nu_1}.

Consequently, the same statistic can be calculated as $F_{\mathrm{P}}=(\nu_2/\nu_1)V/(s-V)$. The equality follows by substituting the definitions of $\nu_1$ and $\nu_2$; their common factor $s$ cancels. This form is a practical arithmetic check: the multiplier must be the denominator degrees of freedom divided by the numerator degrees of freedom.

Derived result: For fixed $p,q,v$, the transformation from $V$ to $F_{\mathrm{P}}$ is strictly increasing on $0\leq V<s$. Indeed, $$ \frac{d}{dV}\left(\frac{\nu_2}{\nu_1}\frac{V}{s-V}\right) =\frac{\nu_2}{\nu_1}\frac{s}{(s-V)^2}>0, $$ because $s,\nu_1,\nu_2$ are positive. Thus a larger trace gives a larger calibrated statistic and a smaller upper-tail p-value.

The derivative establishes the ordering of trace values under the transformation; it does not establish that the F reference distribution is exact. The p-value is the upper-tail probability in that reference distribution evaluated at the observed transformed statistic:

p_{\mathrm{approx}}=P\!\left(F_{\nu_1,\nu_2}\geq F_{\mathrm{P,obs}}\right).

Worked Calculations

Worked Example: Two Restrictions and One Response

Suppose there is one response variable, $p=1$, two tested restrictions, $q=2$, and $v=20$ residual degrees of freedom. The observed generalized root is $\lambda_1=0.5$. Calculate the calibrated statistic and its approximate p-value.

There is one possible root because $s=\min(1,2)=1$. The trace is

V=\frac{0.5}{1+0.5}=\frac{1}{3}.

Here $m=(|1-2|-1)/2=0$ and $n=(20-1-1)/2=9$. Therefore

\nu_1=1(0+1+1)=2, \qquad \nu_2=1(18+1+1)=20.

The multiplier is $\nu_2/\nu_1=20/2=10$, and $s-V=1-1/3=2/3$. Hence

F_{\mathrm{P}}=10\frac{1/3}{2/3}=5.

For $F_{2,20}$, the upper-tail probability at $x$ is $(1+2x/20)^{-10}$. Therefore the calibrated p-value is

p_{\mathrm{approx}}=(1+2(5)/20)^{-10}=1.5^{-10}\approx 0.01734.

The example also illustrates why $q$ must remain the restriction rank: $H$ has rank at most one for a scalar response, but the numerator degrees of freedom here are calculated using two restrictions.

Worked Example: A Two-Response, Two-Restriction Hypothesis

Let $p=q=2$, let the residual degrees of freedom be $v=20$, and suppose the two generalized roots are $0.5$ and $0.25$. Calculate Pillai’s trace, the approximate F statistic, and the upper-tail p-value.

The trace is the sum of the two transformed roots:

V=\frac{0.5}{1.5}+\frac{0.25}{1.25} =\frac{1}{3}+\frac{1}{5} =\frac{8}{15}.

Now $s=2$, $m=(|2-2|-1)/2=-1/2$, and $n=(20-2-1)/2=17/2$. The degrees of freedom are

\nu_1=2(-1+2+1)=4, \qquad \nu_2=2(17+2+1)=40.

Since $s-V=2-8/15=22/15$, the trace ratio is $(8/15)/(22/15)=4/11$. Thus

F_{\mathrm{P}}=\frac{40}{4}\frac{4}{11}=\frac{40}{11}\approx 3.6364.

For $F_{4,40}$, set $z=4F_{\mathrm{P}}/(4F_{\mathrm{P}}+40)$. Substitution gives $z=4/15$. The upper tail is $I_{1-z}(20,2)=(1-z)^{20}(1+20z)$, so

p_{\mathrm{approx}} =\left(\frac{11}{15}\right)^{20}\left(1+\frac{80}{15}\right) =\left(\frac{11}{15}\right)^{20}\frac{19}{3} \approx 0.0128.

The result is an approximate p-value from the F calibration, not a statement that the multivariate null distribution is exactly $F_{4,40}$.

Worked Example: A Rank-One Hypothesis With Two Responses

Suppose $p=2$, $q=1$, and $v=20$. There is one nonzero generalized root, $\lambda_1=2$. Work through the calibration.

Here $s=1$, and the trace is $V=2/(1+2)=2/3$. The parameters are $m=(|2-1|-1)/2=0$ and $n=(20-2-1)/2=17/2$. It follows that

\nu_1=1(0+1+1)=2, \qquad \nu_2=1(17+1+1)=19.

The trace ratio is $(2/3)/(1/3)=2$, so

F_{\mathrm{P}}=\frac{19}{2}(2)=19.

For $F_{2,19}$, the upper-tail probability at $x$ is $(1+2x/19)^{-19/2}$. Consequently,

p_{\mathrm{approx}}=(1+2(19)/19)^{-19/2} =3^{-19/2}\approx 0.0000293.

This rank-one case has a particularly simple tail calculation. More generally, whether a special case has an exact reduction is a separate question from how the general Pillai calibration is computed.

Conditions and Common Pitfalls

The classical approximation is used in the standard fixed-design multivariate linear model under the null hypothesis. The responses are modeled jointly with independent observational errors having a common covariance matrix; the usual derivation assumes multivariate normal errors. The tested restrictions must be estimable, and $q$ is their rank after removing redundant restrictions. The error SSCP matrix must be nonsingular for the ordinary generalized-root formulation; under the standard model this requires adequate residual degrees of freedom (in particular $v\geq p$). The displayed F degrees of freedom must be positive.

These conditions describe the setting for the standard calibration, not a guarantee of exact F sampling behavior. When the model or covariance assumptions are doubtful, or when an exact reference law is required, the approximate p-value should not be presented as exact. Report the trace, the transformed statistic, its two degrees of freedom, and that the calibration is approximate.

Common mistake: Treating $q$, the rank of the tested restrictions, as the rank of $H$, or substituting the latter into the formula. These ranks need not agree: $\operatorname{rank}(H)\leq\min(p,q)$, and it may be smaller still in degenerate cases. Determine $q$ from the estimable restrictions, then calculate $s=\min(p,q)$. A second arithmetic trap is using the wrong exponent in a simple $F_{2,\nu}$ tail: it is $-\nu/2$, so for example $P(F_{2,20}\geq2)=(1+2(2)/20)^{-10}\approx0.1615$, not the same expression raised to $-20$.

Pillai’s calibration is useful because it gives a single, interpretable route from a multivariate trace to an approximate test: calculate the trace from the roots, obtain the dimensions from the design, transform, and evaluate an upper tail. Its reliability depends on respecting that division between the statistic’s algebra and the reference distribution’s approximation.

Check Your Understanding

Use the definitions and calculations above to answer these questions.

In a design with $p=4$, $q=2$, and $v=30$, find $s$, $m$, $n$, $\nu_1$, and $\nu_2$.
Why can the rank of $H$ not be used in place of $q$ when calculating the calibration degrees of freedom?
For fixed $p,q,v$, what does the monotonicity of the trace transformation imply about the ordering of calibrated statistics?
If $V=0.8$, $s=2$, $\nu_1=4$, and $\nu_2=20$, calculate $F_{\mathrm{P}}$.
Why should a reported p-value from the general Pillai F calibration be described as approximate?