Bartlett's Chi-Square Approximation for Wilks Lambda

Why Transform Wilks’ Lambda?

Wilks’ lambda compresses the MANOVA root spectrum into a single determinant ratio: small values indicate that the hypothesis matrix is large relative to the error matrix. Its exact central null distribution is available in terms of products of beta variables, but that distribution can be inconvenient for routine calibration. Bartlett’s transformation takes the logarithm of lambda, applies a finite-sample correction, and compares the result with a chi-square distribution.

The correction matters because the unadjusted quantity \(-v\log\Lambda\) is not generally centered at its chi-square reference value in finite samples. The adjustment depends on both the response dimension and the rank of the tested hypothesis. The resulting chi-square calibration is an approximation, not an exact identity.

Definition: Let \(p\) be the number of response variables, \(h\) the rank of the hypothesis matrix (the numerator degrees of freedom), and \(v\) the error degrees of freedom. Under the central MANOVA null, Wilks’ lambda is \(\Lambda=|E|/|E+H|\), where \(E\) and \(H\) are the error and hypothesis sums-of-products matrices. Bartlett’s transformed statistic is

T_B=-\left(v-\frac{p-h+1}{2}\right)\log\Lambda.

Its reference chi-square degrees of freedom are \(ph\). The approximation is used when the multiplier \(v-(p-h+1)/2\) is positive.

Here \(H\) is the hypothesis matrix for the tested restrictions, not a random variable denoting the hypothesis. The statistic concerns the central null law: the hypothesis effect is zero, and the usual MANOVA model conditions apply, including independent multivariate-normal errors with a common positive-definite covariance matrix. The error matrix must be nonsingular; in the standard full-rank setup this requires \(v\ge p\).

Derivation of the Correction and Degrees of Freedom

The exact central null law of Wilks’ lambda can be expressed as a product of independent beta variables. Write \(m=\min(p,h)\), and let \(U_1,\ldots,U_m\) be independent, with \(U_j\sim\operatorname{Beta}(a_j,b_j)\), where \(a_j=(v-p+j)/2\) and \(b_j=(p+h-2j+1)/2\). Then \(\Lambda\) has the same distribution as \(\prod_{j=1}^{m}U_j\). This representation provides a route to the large-\(v\) chi-square approximation; it does not say that the finite-sample distribution is itself chi-square.

Result: For fixed positive integers \(p\) and \(h\), under the central null and as \(v\) tends to infinity,

-\left(v-\frac{p-h+1}{2}\right)\log\Lambda \xrightarrow{d}\chi^2_{ph}.

The correction removes the leading finite-sample bias in the mean of the unscaled log statistic. It does not make the finite-sample null distribution exactly chi-square.

To see the degrees of freedom, first consider the leading behavior of the independent beta factors. A beta variable \(U_j\sim\operatorname{Beta}(a_j,b_j)\) can be represented as \(G_j/(G_j+K_j)\), where \(G_j\) and \(K_j\) are independent gamma variables with unit scale and shape parameters \(a_j\) and \(b_j\). As \(v\) increases, \(a_j\) increases without bound while \(b_j\) is fixed. The gamma law for \(G_j\), rescaled by its shape, concentrates around 1. Consequently, \(a_j(1-U_j)\) converges in distribution to \(K_j\), which has a gamma distribution with shape \(b_j\) and unit scale. Since \(-\log U_j\) is asymptotic to \(1-U_j\), and \(v/a_j\to2\), it follows that \(v(-\log U_j)\) converges in distribution to a gamma variable with shape \(b_j\) and scale 2.

The variables are independent, and the sum of independent gamma variables with a common scale is gamma with shape equal to the sum of the shapes. Therefore \(v(-\log\Lambda)\) converges to a gamma law with shape \(\sum_{j=1}^{m}b_j\) and scale 2. The sum is \(\sum_{j=1}^{m}(p+h-2j+1)/2=ph/2\), so this gamma law is \(\chi^2_{ph}\). The Bartlett multiplier divided by \(v\) tends to 1, so multiplying by the correction does not change this limiting distribution. This proves the stated convergence.

The reason for the particular correction can also be seen from the mean. For \(U_j\sim\operatorname{Beta}(a_j,b_j)\), \(E[-\log U_j]=\psi(a_j+b_j)-\psi(a_j)\), where \(\psi\) is the digamma function. With \(a_j=(v-p+j)/2\), the large-\(v\) expansion is

E[-\log U_j] = \frac{2b_j}{v} + \frac{2b_j(1-b_j+p-j)}{v^2} + O(v^{-3}).

Summing over \(j\), the first-order terms give \(ph/v\). Summing the second-order coefficients gives \(ph(p-h+1)/(2v^2)\). Hence

E[-\log\Lambda] = \frac{ph}{v} + \frac{ph(p-h+1)}{2v^2} + O(v^{-3}).

Multiplication by \(v-(p-h+1)/2\) cancels the displayed order-\(v^{-1}\) contribution to the mean, leaving \(E[T_B]=ph+O(v^{-2})\). This mean adjustment motivates Bartlett’s correction; the chi-square limit itself follows from the independent-beta argument above.

Worked Calibrations

Worked Example: A Three-Response, Two-Degree-of-Freedom Test

Suppose a MANOVA test has \(p=3\) responses, hypothesis rank \(h=2\), error degrees of freedom \(v=20\), and observed Wilks’ lambda \(\Lambda=0.60\). Calculate the Bartlett statistic and its approximate upper-tail probability.

The correction multiplier is \(20-(3-2+1)/2=20-1=19\), and the reference degrees of freedom are \(ph=3(2)=6\). Thus

T_B=-19\log(0.60) =19(0.5108256) \approx9.7057.

For six degrees of freedom, the chi-square survival probability at \(x\) is \(e^{-x/2}[1+x/2+(x/2)^2/2]\). At \(x=9.7057\), \(x/2\approx4.8528\), so

P(\chi^2_6\ge9.7057) \approx e^{-4.8528}\left(1+4.8528+\frac{4.8528^2}{2}\right) \approx0.138.

The approximate upper-tail probability is therefore about \(0.138\). It is a chi-square approximation to the central null tail probability, not an exact p-value from the finite-sample Wilks distribution.

Worked Example: Verifying a Six-Degree-of-Freedom Tail Probability

Suppose the Bartlett-transformed statistic is \(9.5266\) and the reference degrees of freedom are 6. The same survival formula gives \(x/2=4.7633\), and

e^{-4.7633} \left(1+4.7633+\frac{4.7633^2}{2}\right) \approx 0.008537(17.108) \approx0.1461.

Using the unrounded value in the chi-square survival calculation gives approximately \(0.1461\). Thus the approximate upper-tail probability is about \(0.146\), not \(0.073\). Checking the factor \(e^{-x/2}\) and the polynomial together is important: the polynomial is greater than one and cannot be omitted or halved.

Worked Example: A Rank-One Hypothesis

Consider \(p=2\) responses, a rank-one hypothesis (\(h=1\)), \(v=18\), and observed \(\Lambda=0.75\). The correction multiplier is \(18-(2-1+1)/2=17\), and the chi-square reference has \(ph=2\) degrees of freedom. Therefore

T_B=-17\log(0.75) =17(0.2876821) \approx4.8906.

For two degrees of freedom, the chi-square survival probability is \(e^{-x/2}\). Hence the approximate upper-tail probability is \(e^{-4.8906/2}=e^{-2.4453}\approx0.0867\). This is a useful quick calibration, but the exact rank-one reduction to an \(F\) distribution is available in its appropriate setting and should be used instead when an exact finite-sample calibration is required.

Conditions and Finite-Sample Limitations

The chi-square result is an asymptotic statement with \(p\) and \(h\) fixed as the error degrees of freedom \(v\) increase. It does not establish a universal accuracy threshold such as “\(v\) must exceed a particular number.” Accuracy depends on the dimensions and the observed tail probability as well as on \(v\). When \(p\) or \(h\) is large relative to \(v\), the limiting argument can provide a poor description of the finite-sample distribution, especially in extreme tails.

The ordinary MANOVA null calibration also relies on the model: errors are independent across experimental units, multivariate normal, and have a common positive-definite covariance matrix. Strong departures from these conditions can invalidate the nominal null calibration. In particular, a chi-square reference does not repair dependence or make a model robust to substantial nonnormality.

The multiplier must be positive for the transformed statistic to have the intended direction: smaller \(\Lambda\) should produce larger \(T_B\). With some combinations of dimensions and error degrees of freedom, it may fail this check. A nonpositive multiplier is a warning that this approximation is not usable in that configuration; it should not be silently applied or interpreted as a conventional upper-tail test.

Bartlett’s transformation is useful because it offers a compact, interpretable calibration with \(ph\) degrees of freedom and often improves on an uncorrected log-lambda approximation. But it remains an approximation. When the finite-sample central law is tractable, or when the design is poorly suited to a large-\(v\) approximation, use an appropriate exact calibration or another justified method rather than treating the transformed statistic as exactly chi-square.

Common Pitfall

A frequent error is to use \(p\) as the chi-square degrees of freedom, or to omit the factor \(v-(p-h+1)/2\) and report \(-\log\Lambda\) directly against \(\chi^2_{ph}\). The degrees of freedom count the hypothesis restrictions across all response dimensions, giving \(ph\), while the multiplier puts the log determinant ratio on the corresponding scale. Both pieces are essential.

A second practical trap is a tail-probability arithmetic error. For \(\chi^2_6\), the survival formula contains the full factor \(e^{-x/2}\) multiplied by \(1+x/2+(x/2)^2/2\). At \(x=9.5266\), this evaluates to about \(0.1461\); reporting half that value changes the conclusion and is not a rounding difference.

Key takeaway: Under the central MANOVA null, Bartlett’s approximation compares \(-[v-(p-h+1)/2]\log\Lambda\) with \(\chi^2_{ph}\). The transformation has a clear asymptotic basis and a finite-sample mean correction, but it is not an exact law and must be checked against the model and design dimensions.

Check Your Understanding

Use the definitions, derivation, and calibration examples to answer these questions.

A test has \(p=4\), \(h=2\), and \(v=30\). What are the Bartlett multiplier and the chi-square degrees of freedom?
Why does the limiting chi-square distribution have \(ph\) degrees of freedom rather than \(p\) or \(h\) alone?
For \(p=3\), \(h=1\), and \(v=12\), calculate the Bartlett multiplier. Is it positive?
What does the correction change in the large-\(v\) expansion of the mean of \(-\log\Lambda\)?
Name two conditions or design features that can make the chi-square approximation unreliable.