Interpreting Sphericity Tests And Their Diagnostics

What a Sphericity Test Can Tell You

A covariance matrix is spherical when every direction has the same variance and distinct coordinates have zero covariance. A likelihood-ratio test of sphericity compares that structure with an unrestricted covariance matrix. Interpreting the test requires more than reading a p-value: the eigenvalues show what kind of sample departure drives the statistic, while the sample size and the quality of the scatter estimate affect how much confidence to place in its calibration.

The test is omnibus. It assesses whether the covariance matrix is proportional to the identity, but a rejection does not identify a particular variable, a scientific mechanism, or a unique pattern of dependence. It provides evidence against equal population eigenvalues under the model and calibration; it does not establish that the null is false.

Definition: For a \(p\)-dimensional random vector with covariance matrix \(\Sigma\), the sphericity hypothesis is \(H_0:\Sigma=\sigma^2 I_p\), where \(\sigma^2>0\) and \(I_p\) is the \(p\)-dimensional identity matrix. Under this hypothesis all \(p\) eigenvalues of \(\Sigma\) are equal to \(\sigma^2\). The unrestricted alternative allows any positive-definite covariance matrix.

For the likelihood-ratio test, suppose \(X_1,\ldots,X_n\) are independent observations from a \(p\)-variate normal distribution with unknown mean and positive-definite covariance matrix. Write \(\bar X=n^{-1}\sum_{i=1}^n X_i\), and use the maximum-likelihood scatter estimate \(S=n^{-1}\sum_{i=1}^n (X_i-\bar X)(X_i-\bar X)^\mathsf{T}\). Let its eigenvalues be \(\lambda_1,\ldots,\lambda_p\). When \(S\) is positive definite, the likelihood-ratio statistic is

T=n\log\left(\frac{\left(\frac{1}{p}\sum_{j=1}^p\lambda_j\right)^p}{\prod_{j=1}^p\lambda_j}\right).

Here \(T=-2\log\Lambda\), where \(\Lambda\) is the likelihood ratio. Under the null, with \(p\) fixed and the stated regularity conditions, its large-sample reference distribution is chi-square with \(p(p+1)/2-1\) degrees of freedom. A finite-sample correction may be used where specified for the test; the examples below use the uncorrected statistic and its large-sample reference distribution to illustrate interpretation.

Why the Eigenvalues Matter

The expression inside the logarithm compares the arithmetic mean of the sample eigenvalues with their geometric mean. It is always at least one, and it equals one exactly when the eigenvalues are all equal. Thus \(T\) is nonnegative and increases as the eigenvalues become more unequal, holding \(n\) fixed. The statistic does not depend on the orientation of the eigenvectors: rotations of the observed coordinates leave the eigenvalues, and hence this test statistic, unchanged.

Result: If \(S\) is positive definite, then \(T\geq 0\), with equality if and only if \(\lambda_1=\cdots=\lambda_p\). Multiplying all eigenvalues by the same positive constant does not change \(T\).

Proof. The arithmetic-geometric mean inequality gives \(\frac{1}{p}\sum_{j=1}^p\lambda_j\geq(\prod_{j=1}^p\lambda_j)^{1/p}\). Raising both sides to the power \(p\) shows that the ratio in the definition of \(T\) is at least one. The logarithm is therefore nonnegative. Equality in the arithmetic-geometric mean inequality holds exactly when every \(\lambda_j\) is equal, proving the equality statement. If each eigenvalue is replaced by \(c\lambda_j\), for \(c>0\), the numerator of the ratio is multiplied by \(c^p\), and the denominator is also multiplied by \(c^p\). The ratio and \(T\) are unchanged. \(\square\)

This scale invariance is useful: the test responds to relative dispersion, not to whether measurements are expressed in large or small units. One very small eigenvalue can make the determinant small and the statistic large, even if the other eigenvalues are similar. Conversely, if all eigenvalues are multiplied by the same factor, that change in overall scale alone does not count as evidence against sphericity.

Worked Example: A Moderate Eigenvalue Contrast

Suppose \(p=3\), the sample eigenvalues are \(4,1,1\), and \(n=30\). Compute the test statistic and interpret the pattern.

The arithmetic mean eigenvalue is \((4+1+1)/3=2\), and their product is \(4\). Therefore,

\frac{2^3}{4}=2, \qquad T=30\log(2)\approx 20.794.

The reference degrees of freedom are \(3(4)/2-1=5\). The 5% chi-square critical value with 5 degrees of freedom is approximately \(11.07\), so the uncorrected statistic exceeds that reference threshold. The sample pattern has one eigenvalue larger than the other two, and that dispersion contributes to the rejection. The omnibus test does not, by itself, say which measured combination of variables accounts for this pattern.

Sample Size and the Strength of the Statistic

The factor \(n\) in \(T\) matters. If the eigenvalue shape is held fixed, increasing \(n\) increases the statistic in direct proportion. This is why a small departure from sphericity may be detected in a sufficiently large sample, while a visibly uneven set of eigenvalues in a small sample may not cross the chosen reference threshold. This observation is about the statistic for a fixed eigenvalue shape; in real studies the observed eigenvalues also vary from sample to sample.

A larger statistic is not itself a measure of scientific importance. It combines the observed relative dispersion with sample size. Interpretation should therefore report the eigenvalue pattern and the test result, rather than equating statistical significance with a large or practically important covariance departure.

Worked Example: The Same Eigenvalue Shape at Two Sample Sizes

Consider \(p=4\) and eigenvalues \(2,1,0.5,0.5\). Their arithmetic mean is \(1\), and their product is \(0.5\), so the ratio in the test statistic is \(1^4/0.5=2\). The reference degrees of freedom are \(4(5)/2-1=9\).

For \(n=24\), \(T=24\log(2)\approx16.636\). For \(n=25\), holding the eigenvalue shape fixed, \(T=25\log(2)\approx17.329\). The 5% chi-square critical value for 9 degrees of freedom is approximately \(16.919\). Thus the first statistic is just below that threshold and the second is just above it.

This comparison isolates the sample-size multiplier; it does not claim that adding one observation to a real data set will preserve the eigenvalues. It illustrates why conclusions near a threshold should be reported with the sample size, the calibration used, and the observed eigenvalue pattern.

Near-Singular Scatter: A Diagnostic Warning

The likelihood-ratio expression requires every sample eigenvalue to be positive, so that \(|S|=\prod_j\lambda_j>0\) and \(\log|S|\) is finite. If a sample eigenvalue is close to zero, the determinant becomes small and the statistic can become very large. This can reflect a genuinely narrow direction in the observed scatter, but it can also arise from redundant measurements, strong linear dependence, too few observations relative to dimension, or numerical problems. The test statistic alone cannot distinguish these explanations.

For centered data, the rank of \(S\) is at most \(n-1\). Consequently, if \(n\leq p\), \(S\) cannot be full rank, and the ordinary unrestricted covariance likelihood has no positive-definite sample maximum. Even when \(n>p\), full rank is not guaranteed if the observed vectors lie in a lower-dimensional subspace. In either situation, the standard likelihood-ratio statistic above is not a finite, ordinary test statistic. Replacing a zero eigenvalue by a tiny arbitrary number and proceeding as if nothing changed does not repair the inferential problem.

Worked Example: A Singular Scatter Matrix

Suppose \(p=3\), \(n=3\), and the three centered observations are \((1,0,0)\), \((0,1,0)\), and \((-1,-1,0)\). Their sum is zero, as required for centered observations. The third coordinate is zero for every observation, so all centered vectors lie in a plane.

Using \(S=n^{-1}\sum_i X_iX_i^\mathsf{T}\), the scatter estimate is

S=\frac{1}{3} \begin{pmatrix} 2&1&0\\ 1&2&0\\ 0&0&0 \end{pmatrix}.

Its eigenvalues are \(1,1/3,0\): the upper-left block has eigenvalues \(1\) and \(1/3\), while the third coordinate supplies the zero eigenvalue. Thus \(|S|=0\), and the logarithm of the determinant in the likelihood-ratio calculation is not finite. The correct diagnosis is that the standard test cannot be applied to this scatter matrix as an ordinary positive-definite case. Reporting an infinite statistic and a routine chi-square p-value would not be a valid substitute.

When the scatter matrix is technically positive definite but nearly singular, inspect its eigenvalues and condition number (the ratio of largest to smallest eigenvalue), and check the data for duplicated or nearly redundant measurements and coding errors. Numerical computation should use stable log-determinant methods rather than forming a determinant and taking its logarithm directly. These checks do not change the model or the test; they help determine whether its calculation and usual calibration are credible.

Reading a Rejection Carefully

Under the assumed model and chosen calibration, a small p-value is evidence against \(H_0\). It does not prove that the population covariance is nonspherical: any test can reject when its null hypothesis is true. Nor does rejection identify a cause. The test is invariant to rotations, so it does not single out a particular coordinate system or explain which variables generated the unequal eigenvalues.

A useful interpretation reports at least the dimension, sample size, test statistic, degrees of freedom, calibration, and the sample eigenvalues or a concise summary of their spread. If the statistic is large because one eigenvalue is exceptionally small, that diagnostic deserves attention; if several eigenvalues differ moderately, the pattern is different even though both situations may lead to rejection. A failure to reject also does not establish sphericity: it means the data did not provide sufficient evidence against it at the chosen level using this test.

Common pitfall: “The test rejected, so the population covariance is definitely nonspherical and the first variable caused it.” This reasoning overstates the conclusion twice. A rejection is evidence against the null, not proof that it is false, and the sphericity likelihood-ratio test is an omnibus eigenvalue-dispersion test, not a variable-specific causal analysis. Report the evidence and inspect the eigenvalue pattern without assigning a cause the test cannot identify.

Key takeaway: The sphericity statistic measures relative eigenvalue dispersion and scales with sample size. Its interpretation depends on the model and calibration, and it requires a positive-definite sample scatter matrix. Rejection is evidence against equal population eigenvalues, not proof of a departure or an explanation of its source.

Check Your Understanding

Use the definitions and diagnostics above to answer these questions.

Why does the arithmetic-geometric mean inequality guarantee that the sphericity statistic is nonnegative?
If every sample eigenvalue is multiplied by the same positive constant, what happens to the statistic, and why?
For fixed eigenvalue shape, how does doubling \(n\) affect \(T\)? Why should this not be described as a guaranteed effect of doubling the size of a real data set?
What is the problem with applying the ordinary likelihood-ratio statistic when one sample eigenvalue is zero?
A test rejects sphericity. Name one conclusion that is supported under the model and calibration, and one conclusion that the test alone does not support.