2
$\begingroup$

My research question is to test whether two groups have a difference in variance-covariance across multiple measures. However, the data do not follow the multivariate normality assumption required for Box's M and the sample size is relatively modest at 30 per group. Is there an alternative in this case?

I have found some suggestions by Layard, but few implementations in practice and there is some criticism in the literature. I also found this useful summary on visualisation of tests for equality of Covariance Matrices by M. Friendly and M. Sigal.

Improve this question
$\endgroup$
8
  • 2
    $\begingroup$ You also might want to consider if “correlation” if the right kind of dependence structure to consider. Without multivariate normality, you could have a funky but interesting dependence structure. This is related to something called the copula. $\endgroup$
    – Dave
    Commented Jan 15, 2020 at 10:56
  • 1
    $\begingroup$ @Nick I would start by hitting the marginal distributions with the probability integral transform (ecdf(x)(x) in R) and plotting to see what kind of relationship there is. $\endgroup$
    – Dave
    Commented Feb 27, 2020 at 10:40
  • 2
    $\begingroup$ sciencedirect.com/science/article/pii/S0047259X08001474 Optimal tests for homogeneity of covariance, scale, and shape $\endgroup$
    – Nick
    Commented Feb 28, 2020 at 15:51
  • 1
    $\begingroup$ I am reading description of the package ‘ICSNP’ as Tools for Multivariate Nonparametrics. $\endgroup$
    – Nick
    Commented Feb 29, 2020 at 4:28
  • 1
    $\begingroup$ @readbeard, maybe useful for you stat.ethz.ch/pipermail/r-help/2005-September/079210.html $\endgroup$
    – Nick
    Commented Mar 1, 2020 at 3:23

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.