A comparison of different techniques for dividing samples into subsamples in preparation for the Bartlett-Kendall Test of equal variances
Included in this study were nine different procedures for dividing samples into subsamples in preparation for the Bartlett-Kendall test for equal variances (unequal sample size situation). First, each procedure was applied in the circumstance in which variances were homogeneous as a means of determining each procedure's ability to control Type I error estimates. Then, four conditions under which the variances were heterogeneous to different degrees were included so as to provide estimates of each procedure's power at a given nominal alpha level. Two small sample conditions, two intermediate sample conditions, and two large sample conditions were included in this study. Three of the conditions involved small absolute differences among sample sizes while the other three conditions involved large absolute differences among the sample sizes. Nominal alpha levels utilized were .05 and .01 while the number of treatment levels was equal to 3 in all situations. The study dealt only with unequal sample sizes, e.g., n1=9, n2=17, n3=23.
All nine of the subdivision alternatives yielded a very small number of error estimates which deviated significantly from the nominal alpha level. The four alternatives which consistently yielded high power estimates were (1) subsample sizes near √ni, lumping leftover elements with the last subsample formed, (2) subsample sizes near √ni, evenly spreading leftover elements over the subsamples already formed, (3) an equal number of subsamples across samples, lumping leftover elements and (4) an equal number of subsamples across samples, evenly spreading leftover elements.
Thesis75b.L295.pdf
7.59 MB
Adobe PDF
b257e664579553aa9787f934ad6250e1