Comparing Tests of Multinormality under Sparse Data Conditions - a Monte Carlo Study
by Alexander von Eye.
Mardia's tests of multivariate skewness and kurtosis and von Eye and Gardiner's and
von Eye and Bogat's sector and overall tests of multinormality are compared under sparse data
conditions. A Monte Carlo study is reported in which five factors were varied: sample size,
number of variables, type of distribution (normal, uniform, log-transformed, inverse Laplace-
transformed, and cube root-transformed), magnitude of correlation among variables, and the
number of segments used for the 2-tests. Results suggest that, even under sparse data conditions,
(1) the Mardia tests are differentially sensitive to the violations they were designed to detect; (2)
the new sector and overall tests are sensitive to all violations included in the simulations; (3) the
effects of the small samples can be seen in an increased random component of the data structure;
and (4) although the overall and the sector tests are still sensitive to a wide range of violations,
the test statistics are not always distributed as chi2, due to the well known inflation of X2. These
results replicate, in part, the results of von Eye's (2005) study, in which larger samples had been
used. However, they also show that sufficiently large samples are needed for valid statistical
decisions about multinormality.
Mardia's skewness and kurtosis tests; von Eye's sector tests; sparse data; sensitivity
Alexander von Eye, email@example.com
READING THE ARTICLE: You can read the article in
portable document (.pdf) format (519538 bytes.)
NOTE: The content of this article is the intellectual property of the authors, who retains all rights to future publication.
This page has been accessed 2482 times since July 24, 2006.
Return to the Home Page.