| A new Bootstrap Test for Multivariate One-Sided Hypothesis |
| Consider testing H0: mu1 = mu2 = ... = mup = 0 versus H1: mu1 >0, mu2>0, ..., mup >0 based on a sample in p dimensions. Perlman (1969) obtained the LRT and its null distribution for testing the mean of a p-dimensional multivariate normal distribution. Silvapulle (1997) identifies a fundamental flaw in the LR testing procedure, namely that the LRT can reject the null hypothesis in favor of the alternative even though the true means are "far" from the alternative region. Perlman and Wu (2002) defends LRT by correctly recognizing this "anomalous" behavior as a result of the formulation of the null hypothesis. The present article demonstrates that the "anomaly" remains for non-normal distributions as well and suggests a remedy by reformulating the null hypothesis to be tested. Simulation studies show that the new bootstrap test leads to "correct" decisions for multivariate normal as well as non-normal data. |
|
Abu Minhajuddin Southern Methodist University Student Poster Session |
 |
 |
April 4-5, 2003
Texas A&M University
College Station, TX
Email:
cots@stat.tamu.edu
Fax: (979) 845-3144
Phone: (979) 845-3141