F Calibration of the Hotelling–Lawley Trace

Why Calibrate the Hotelling–Lawley Trace?

The Hotelling–Lawley trace summarizes the strength of a multivariate hypothesis by adding its generalized roots. That sum is easy to calculate, but under a general MANOVA null its distribution is not, in general, an exact ordinary F distribution. A practical test therefore needs a reference distribution that translates an observed trace into a tail probability.

The standard approximate F calibration supplies that translation. Its scaling depends on the response dimension, the hypothesis degrees of freedom, and the residual degrees of freedom. The calculation is compact, but the dimensions play distinct roles: confusing them can change both the statistic and its reference distribution.

Definition: Let $p$ be the number of response variables, $q$ the numerator degrees of freedom (equivalently, the rank of the hypothesis in the usual full-rank setup), and $v$ the residual degrees of freedom. Let $\lambda_1,\ldots,\lambda_s$ be the nonzero generalized roots of the hypothesis and error matrices, where $s=\min(p,q)$. The Hotelling–Lawley trace is $$ U=\sum_{i=1}^{s}\lambda_i. $$

For the usual central-null calibration, the hypothesis and error sums of squares and cross-products have the standard independent central Wishart structure, and the error matrix must be nonsingular. In the usual multivariate linear model this requires adequate residual degrees of freedom; in particular, $v\ge p$ gives a full-rank error matrix with probability one under a nonsingular covariance model. The approximation is a null reference distribution, not an assertion that the general finite-sample distribution of $U$ is exactly F.

The Standard Approximate F Transformation

Define $$ s=\min(p,q),\qquad m=\frac{|p-q|-1}{2},\qquad n_0=\frac{v-p-1}{2}. $$ The symbol $n_0$ here is an auxiliary dimension parameter, not the sample size $n$. The conventional Hotelling–Lawley approximation maps the observed trace to an F statistic as follows.

Theorem: Under the central MANOVA null and the regularity conditions above, the standard approximate F calibration is $$ F_{\mathrm{HL}}=\frac{(2n_0+s+1)U}{s(2m+s+1)} \quad\text{with reference degrees of freedom}\quad df_1=s(2m+s+1),\qquad df_2=s(2n_0+s+1). $$ The reference is $F_{df_1,df_2}$, and the approximation is used for an upper-tail test.

These expressions simplify substantially. The simplification is useful in hand calculations and makes clear how both degrees of freedom are determined.

Simplify the numerator degrees of freedom.
If $p\le q$, then $s=p$ and $2m+s+1=q-p-1+p+1=q$. If $q\le p$, then $s=q$ and $2m+s+1=p-q-1+q+1=p$. In either case, $s(2m+s+1)=pq$.

Simplify the other dimension factor.
Since $2n_0+s+1=v-p-1+s+1=v-p+s$, the denominator degrees of freedom are $s(v-p+s)$.

Substitute into the statistic.
The coefficient multiplying $U$ is $(v-p+s)/(pq)$, giving the equivalent formula below.

F_{\mathrm{HL}}=\frac{v-p+s}{pq}\,U,\qquad df_1=pq,\qquad df_2=s(v-p+s),\qquad s=\min(p,q).

The algebra above is exact: it only rewrites the standard calibration’s parameters. The approximation lies in using the resulting F reference distribution for the general null law of the trace. In particular, these degrees of freedom should not be interpreted as proving that $F_{\mathrm{HL}}$ has an exact F distribution for arbitrary $p$ and $q$.

Interpretation: For an observed trace $U_{\mathrm{obs}}$, the approximate upper-tail p-value is $$ P\!\left(F_{pq,\;s(v-p+s)}\ge \frac{v-p+s}{pq}U_{\mathrm{obs}}\right). $$ A larger trace produces a larger calibrated F statistic and therefore stronger evidence against the central null.

Worked Applications

Worked Example: Two Responses and Two Hypothesis Degrees of Freedom

Suppose a test has $p=2$ response variables, $q=2$ hypothesis degrees of freedom, and $v=18$ residual degrees of freedom. The observed roots are $0.30$ and $0.10$. Find the trace, calibrated statistic, and reference degrees of freedom.

First add the roots: $$ U=0.30+0.10=0.40. $$ Here $s=\min(2,2)=2$. The simplified degrees of freedom are $$ df_1=pq=2(2)=4,\qquad df_2=s(v-p+s)=2(18-2+2)=36. $$ The multiplier of the trace is $$ \frac{v-p+s}{pq}=\frac{18-2+2}{2(2)}=\frac{18}{4}=4.5. $$ Therefore $$ F_{\mathrm{HL}}=4.5(0.40)=1.80. $$ The approximate reference is $F_{4,36}$, so the upper-tail p-value is $P(F_{4,36}\ge1.80)$. The two roots are added before calibration; the test does not calibrate each root separately.

Worked Example: One Hypothesis Degree of Freedom

Suppose $p=4$, $q=1$, and $v=18$. There is one nonzero root, observed as $0.25$. Calculate the standard calibration and check its degrees of freedom.

Because $q=1$, $s=1$, and the trace is just the single root: $$ U=0.25. $$ The formula gives $$ df_1=pq=4(1)=4,\qquad df_2=s(v-p+s)=1(18-4+1)=15. $$ The multiplier and calibrated statistic are $$ \frac{v-p+s}{pq}=\frac{15}{4}=3.75,\qquad F_{\mathrm{HL}}=3.75(0.25)=0.9375. $$ Thus the reference is $F_{4,15}$. This is also an important check: for a rank-one hypothesis ($q=1$), the usual exact reduction gives $F=(v-p+1)U/p$ with degrees of freedom $p$ and $v-p+1$. Substitution of $q=s=1$ into the general calibration reproduces that exact result. This special exactness does not extend to arbitrary $p$ and $q$.

Worked Example: More Hypothesis Than Response Dimensions

A model has $p=2$ responses, $q=3$ hypothesis degrees of freedom, and $v=16$ residual degrees of freedom. The two observed roots are $0.20$ and $0.05$. Apply the calibration.

There can be at most $\min(2,3)=2$ nonzero roots, so both listed roots contribute: $$ U=0.20+0.05=0.25,\qquad s=2. $$ The degrees of freedom are $$ df_1=pq=2(3)=6,\qquad df_2=s(v-p+s)=2(16-2+2)=32. $$ The statistic is $$ F_{\mathrm{HL}}=\frac{16-2+2}{2(3)}(0.25) =\frac{16}{6}(0.25) =\frac{4}{6} \approx0.667. $$ The approximate upper-tail p-value is $P(F_{6,32}\ge0.667)$. Although $q=3$, there are only two roots because the response dimension is $p=2$. The numerator degrees of freedom nevertheless equal $pq=6$, not the number of nonzero roots.

What the Approximation Does—and Does Not—Say

The trace $U$ aggregates root strength: every nonzero root contributes, so a collection of moderate roots can yield a substantial trace even when no individual root is especially large. The F calibration turns that aggregate into a familiar reference-scale statistic while accounting for the design dimensions. It is useful when a single upper-tail test is wanted and the standard approximation is appropriate for the problem’s dimensions.

The scaling and degrees of freedom are coupled. The multiplier $(v-p+s)/(pq)$ depends on all three dimensions through $s$, while the reference distribution has numerator degrees of freedom $pq$ and denominator degrees of freedom $s(v-p+s)$. Reporting only the transformed statistic without its reference degrees of freedom is therefore incomplete.

Common pitfall: It is tempting to use the number of nonzero roots, $s$, as the numerator degrees of freedom because $U$ sums $s$ values. That is incorrect for this calibration: $df_1=pq$. The root count enters the denominator degrees of freedom and the scale, but it does not replace $pq$. Also, the p-value is an upper-tail probability; a small value of $F_{\mathrm{HL}}$ is not evidence against the null.

The calibration is not a substitute for checking the model and its assumptions. A singular error matrix prevents the ordinary generalized-root calculation, and a departure from the central-null reference setup can affect the interpretation of a null p-value. Nor does the approximation imply that the trace and the other MANOVA criteria are interchangeable: each criterion summarizes the roots differently and has its own calibration.

Key takeaway: Sum the nonzero generalized roots to obtain $U$, set $s=\min(p,q)$, and calculate $F_{\mathrm{HL}}=(v-p+s)U/(pq)$. Compare it with $F_{pq,\;s(v-p+s)}$ using the upper tail, while remembering that this is generally an approximate null calibration.

Check Your Understanding

Use the definitions and calibration above to answer the following.

For $p=3$, $q=2$, and $v=20$, what are $s$, $df_1$, and $df_2$?
If the observed roots are $0.12$, $0.08$, and $0$, which roots contribute to $U$, and what is the trace?
Explain why the numerator degrees of freedom are not generally equal to the number of nonzero roots.
For $q=1$, simplify the statistic and both degrees of freedom. Why is this case a useful check on the formula?
Which tail of the reference F distribution gives the approximate p-value, and why?

F Calibration of the Hotelling–Lawley Trace

What You'll Learn