Cited page

Citations are available only to our active members. Sign up now to cite pages or passages in MLA, APA and Chicago citation styles.

X X

Cited page

Display options
Reset

Empirical Direction in Design and Analysis

By: Norman H. Anderson | Book details

Contents
Look up
Saved work (0)

matching results for page

Page 373
Why can't I print more than one page at a time?
While we understand printed pages are helpful to our users, this limitation is necessary to help protect our publishers' copyrighted material and prevent its unlawful distribution. We are sorry for any inconvenience.

Overall Anova for More Than Two Groups. The Brown-Forsythe and Welch tests are widely recommended as extensions of overall Anova to handle unequal variance with more than two groups. With normal distributions, they keep α under better control and provide more power than regular F.

Nonnormal distributions, however, undercut both tests. Simulations with the normal and four nonnormal symmetric distributions by Lee and Fung (1983) concluded that the Brown-Forsythe and Welch tests “perform poorly except under normality. Under long-tailed [= heavy-tailed symmetric] distributions, they are conservative and have very low power” (p. 137). c

Accordingly, neither test may be generally useful. This conclusion rests in part on the expectation that unequal variance will usually be accompanied by nonnormality.

Trimming. Trimming may have as much potential, perhaps more, with unequal variance as with equal variance. By decreasing the influence of the tail scores, the trimmed variances may be expected to be more equal.

Trimmed versions of the Brown-Forsythe and Welch tests, both available in BMDP, may deserve consideration, as they seemed to perform fairly well in the cited study. As already suggested, however, two-mean comparisons may be generally preferable to overall Anova between all the means.

Repeated Measures. With repeated measures, two-mean comparisons with unequal variance are handled straightforwardly. Because of likely nonsphericity, each two-mean comparison ordinarily requires its own error term. Each such comparison is equivalent to a test of the corresponding difference score, which automatically takes account of unequal variance.


NOTES
12.1.1a
The integer adjustment is not necessary but simplifies the exposition.
12.1.lb
For the curious, this procedure of replacing the trimmed scores by copies is called Winsorizing, after the statistician Charles Winsor; this is why the subscript W is used. Winsorized Anova, using the Winsorized mean in place of the trimmed mean, seems less desirable than trimmed Anova (Barnett & Lewis, 1994, pp. 79–81).

-373-

Select text to:

Select text to:

  • Highlight
  • Cite a passage
  • Look up a word
Learn more Close
Loading One moment ...
of 864
Highlight
Select color
Change color
Delete highlight
Cite this passage
Cite this highlight
View citation

Are you sure you want to delete this highlight?