User-created program for posthoc test for Kruskal-Wallis

[KC Culver]

Hi-

First, I want to thank all of the members who volunteer their time to help others become better, stronger Stata users. I have found answers to so many questions on this forum (and its predecessor).

I am using Stata 13/14 (work/home) and have a dataset with about 1,900 cases. I'd like to examine differences among 4 racial/ethnic groups for 26 different 4-point/5-point Likert scale frequency variables (Never-Very Often). I'm currently using the Kruskall-Wallis test since (a) the frequency variables are ordinal and (b) there's great imbalance among the racial groups (97 latino/as vs. 1,600 whites). Once I find a significant difference, I know I need to follow up with posthoc tests. My reading has suggested possibly using Dunn, Mann-Whitney, or Conover-Iman tests as follow-ups.

A previous Statalist thread pointed that author to "Alexis Dinno's user-written Stata command" for posthoc tests, but the included link doesn't work properly (I've create a link to the previous thread).

I would appreciate any thoughts on which posthoc test(s) may be most appropriate as well as where to find the necessary user-created programs. I'm also open to alternative paths to examining group differences if the K-W omnibus test seems inappropriate.

Thanks-
KC

[Joseph Coveney]

search Alexis Dinno

brings up a Stata Viewer window with hyperlinks to her latest Stata download site among others. You can also

net describe conovertest, from(http://alexisdinno.com/stata)

to bring up the package description in the main Results window, but it won't have download hyperlinks, and so you'll need to manually type the net install command.

Also, consider ordered logistic regression, as was mentioned in post you reference. With that, you'll have access to the official Stata contrast command.

[KC Culver]

Thanks for your reply, Joseph. I'm not sure I have the sample size for ologit, which is why I'm exploring alternatives. I'm also not sure why I can't just create dummy variables to use the K-W if I do find significant results among groups, but I'm not familiar with the Conover test, so I'll have to dig my heels into nonparametrics.

[Ted Kaniuka]

Hello, I tried the net describe suggestion above it failed but had success when I searched for conovertest. Came back with an active link.