I have a multivariate regression model $\mathbf{Y}$ = $\mathbf{XB}$ + $\mathbf{E}$ where $\mathbf{Y}$ is $n \times m$.

In my case $\mathbf{n < m}$, the number of columns in $\mathbf{Y}$ is greater than the number of rows. Is it still possible to do multivariate hypothesis testing with a model like this (compute Wilks Lambda, Pillai's trace etc.)?

I have been getting errors from some programs and I don't see why. I can't find a mention of this in my multivariate analysis book or online anywhere, maybe I'm not googling the right terms?

If computing these statistics are $\mathbf{not}$ a possibility in this circumstance, could someone explain why not?


Statistics like Wilks lambda, Pillai's trace are certain functions of two covariance matrices, see https://en.wikipedia.org/wiki/Wilks%27s_lambda_distribution and https://en.wikipedia.org/wiki/Multivariate_analysis_of_variance. In this way they generalize, in different ways, the F-ratio in univariate anova, which is a function of two sums of squares.

For the case of Wilks lambda, denote the two matrices as $A, B$. Then $$ \lambda = \frac{\det A}{\det(A+B)} $$ But in the case $n<m$, the matrices are singular, so the determinants are zero ... But each determinant is a product of eigenvalues, so some of this eigenvalues are zero. The other eigenvalues still contain useful information!

Looking into the details (which I will not do here, now), it is possible to see that everything of interest is described by the nonzero eigenvalues. Specifically, Wilks $\lambda$ is a likelihoodratio test, and going through the details in the $n<m$ case, one will see that the test has the same form, with the determinants above replaced with the product of the nonzero eigenvalues. That could be called a generalized determinant, in the spirit of generalized inverses (I don't know if this is an established name). Probably the distribution theory must be checked for this case, and I do not know of any reference containing this generalizations.

For Pillai's trace much of the same will be true.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.