Date: Mon, 1 Sep 2003 01:43:56 -0400
Reply-To: Richard Ristow <wrristow@mindspring.com>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Richard Ristow <wrristow@mindspring.com>
Subject: Re: ANOVA or Kruskal-Wallis ANOVA
In-Reply-To: <3F52A805.A61A5A0@jhem.jhmi.edu>
Content-Type: text/plain; charset="us-ascii"; format=flowed
At 09:59 PM 8/31/2003 -0400, Christina Cutshaw wrote:
>I want to determine whether four groups: A, B, C, D (total N=73)
>differ in their number of child-related policy goals. I need to
>decide between using a one-way
>ANOVA to compare means or the Kruskal-Wallis ANOVA to compare mean ranks:
>[Appropriate statistical tests rejected normality of distribution of
>the dependent variable, and failed to find heterogeneity of variance
>between groups.] With this information I decided to use a
>Kruskal-Wallis test and then pair-wise Mann-Whitney tests to look at
>the differences between groups in their number of goals. Did I make
>the correct decision?
I'll step in with a rough-and-ready, aware that there are far more
sophisticated mathematical statisticians on the list (Marta, are you
listening?).
Your decision was certainly a defensible one. However, in all
probability a (parametric) one-way ANOVA would give very similar
results. The parametric methods (grouped under the rubric of general
linear models) tend to be quite forgiving of deviations from normality,
which is good, as most real data has a distribution differing markedly
from the normal. The one exception is extreme values: parametric
methods are very sensitive to these, and non-parametric methods are
indicated where extreme values are known or likely. (The alternative of
rejecting the extreme values can lead to many biases: just what is "extreme"?)
In your case, I'd doubt that the observed maximum of "number of
child-related policy goals" is very large, and either method should
give satisfactory results. In fact, you'll probably find the results of
the two methods quite consonant.
| 
