mann-whitney

[abdalla alsmadi]

hello everybody

 

i am using mann-whitney to compare two groups (control & experimental  \ 20 cases each), but there were a difference  between groups on the pre-test, so what do you suggest for analysis

thanks a lot

 

Abdalla Alsmadi
college of graduate studies
Arabian Gulf University - Bahrain
Tel: [phone] ext. 676

[Bruce Weaver]

Why are you using the Wilcoxon-Mann-Whitney test?  It was designed to test the null hypothesis that p(X < Y) = 0.5, where X and Y are values drawn at random from the two distributions.  When you use it as an alternative to the unpaired t-test (i.e., as a test of location), it makes the strong assumption that the two populations are identical apart from a possible shift in location.  Simulations by Zimmerman (2003) and Fagerland (2009) (and probably others) have shown that the WMW test is VERY sensitive to small differences in variance, and Fagerland has also shown it is sensitive to moderate differences in skewness -- and this is with equal sample sizes.  Those same simulations confirm that the unpaired t-test is very robust under the same circumstances.  

Was there random assignment to groups?  What is the dependent variable?  Do you have some reason for not using ANCOVA, with the baseline score as covariate?  See http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1121605/, for example.

HTH.

abdalla alsmadi wrote

hello everybody

i am using mann-whitney to compare two groups (control & experimental  \ 20 cases each), but there were a difference  between groups on the pre-test, so what do you suggest for analysis
thanks a lot

Abdalla Alsmadi
college of graduate studies
Arabian Gulf University - Bahrain
Tel: [phone] ext. 676

[Paige Ransford]

I will be out of the office from February 18, 2013 to February 26, 2013. If you need immediate assistance please call 617-287-5541.

Paige Ransford
Senior Research Associate
The Center for Women in Politics & Public Policy
UMass Boston

[Rich Ulrich]

When there is a difference at Pre, the starting advice is:
Be very cautious about drawing  *any*  conclusions at Post, and
re-think your whole problem.  Also as a start, look at a plot
of the means of the groups, and (probably) look at a scatterplot
of Pre-Post for each of the two groups.  What is going on?
(Is there a ceiling- or basement-effect on how far scores can move?)


Is this an accidental failure of randomization?  If it is systematic,
what can you say about the reasons? 

Sometime, for large groups, a Pre difference can be a small enough
that its observed size won't matter.  For Ns of 20, however, you are
pointing to a moderately big difference.

If the group that was Better at Pre becomes Worse at Post, then
you may have a fairly firm conclusion.  Otherwise, you *can* consider
each of the ways of measure change scores --
  Ignore data at Pre (look at outcome only). 
  Use raw difference scores. 
  Use covariance (regressed-change).
  Use a version of regressed change that takes its expected regression
from other sources of information. 

Most of those presume that you have a good "metric", that is, that you
have good scaling so that you should not even be tempted to look at
rank-transformation of scores (like, MW testing).  If that was really a
problem, then you need to justify your scaling as it is, or else, fix it.

--
Rich Ulrich

---

Date: Sun, 17 Feb 2013 03:42:49 -0800
From: [hidden email]
Subject: mann-whitney
To: [hidden email]

hello everybody

 

i am using mann-whitney to compare two groups (control & experimental  \ 20 cases each), but there were a difference  between groups on the pre-test, so what do you suggest for analysis

thanks a lot

 

Abdalla Alsmadi
college of graduate studies
Arabian Gulf University - Bahrain
Tel: [phone] ext. 676