Solved
New Contributor
Posts: 2

Dunn's Test

Is there a way to do a Dunn's Test following the Kruskal-Wallis in SAS?  If so, what is the procedure that should be used?

Thanks,

Sarah

Accepted Solutions
Solution
‎03-26-2013 03:18 PM
Valued Guide
Posts: 684

Re: Dunn's Test

There are different macros that have been written to do the Dunn's multiple comparisons. A google search gives many hits.  The book Pharmaceutical Statistics Using SAS: A Practical Guide (by Dmitrienko et al.) describes a good one (chapter 6). The data examples and macros for this book are all available at:

http://support.sas.com/publishing/bbu/zip/60622.zip

You will have to learn about using macros, if you haven't used them before.

All Replies
Solution
‎03-26-2013 03:18 PM
Valued Guide
Posts: 684

Re: Dunn's Test

There are different macros that have been written to do the Dunn's multiple comparisons. A google search gives many hits.  The book Pharmaceutical Statistics Using SAS: A Practical Guide (by Dmitrienko et al.) describes a good one (chapter 6). The data examples and macros for this book are all available at:

http://support.sas.com/publishing/bbu/zip/60622.zip

You will have to learn about using macros, if you haven't used them before.

New Contributor
Posts: 2

Thanks so much!

Contributor
Posts: 57

Re: Dunn's Test

Hi everybody,

Are there any alternatives of the Dunn test as post-hoc test of the Kruskall-Wallis test? What do you think about the following:

. Perform Dunnett test on ranked data?

. Perform multiple Mann-Withney test with Bonferonni adjustment?

Posts: 2,116

Re: Dunn's Test

How much data?  Ranks are asymptotically normal, so any of the linear models multiple comparisons procedures could be used with a reasonable amount of data.

Bonferonni can be used with any test, but is inherently conservative.

Doc Muhlbaier

Duke

Contributor
Posts: 57

Re: Dunn's Test

Hi Doc,

Thanks for your answer. I will have 30 subjects in each group, with 4 or 5 treatment groups. Is, Dunnett on ranked data is a good methodology?

Best regards,

Posts: 2,655

Re: Dunn's Test

I would go ahead with the Dunnett adjustment on the analysis of the ranked data.  You have (excluding ties) 120 to 150 values, so the adjustment should be OK, asymptotically.  And if your original data were something like a ratio of two normals (for example, body weight change divided by food intake), then I would recommend this method even more.

Steve Denham

Contributor
Posts: 57

Hi Steve,