Wilcoxon/Mann-Whitney tests and weighting

[Kristin]

Hi,

I'm a new user to Stata and need it for my grad project. I need to know if Stata will allow for me to use weights for the Wilcoxon signed rank test and Mann-Whitney U test?

I'm trying to expand my results out to the general population but my demographics are a bit off. I need to weight the gender and ethnicity of my sample but I also have to use the Wilcoxon and Mann-Whitney tests. Is this possible? Can you explain to me how this can be done?

Thanks so much!

[Joe Canner]

Kristin,

The short answer is no. The help for ranksum (which does Wilcoxon/Mann-Whitney) indicates that weights are not permitted. There is a median command that does a median comparison test that uses weights.

A quick online search suggests that the reason weights are not permitted is that a weighted version of Wilcoxon/Mann-Whitney has not been completely worked out. If you want to go it alone, you could do expand weight and create a data set with one observation for everyone in the general population and then use ranksum, but you should proceed with caution. Perhaps others here can comment on the statistical validity of such an approach.

Regards,
Joe

[Nick Cox]

I'd recommend strongly that you look at Roger Newson's somersd (SJ).

For large enough sample sizes, any difference will qualify as significant at conventional levels. One of several merits of somersd is that it will steer you towards more fruitful analyses.