# How to correct for gender in Mann-Whitney U test?

My statistical knowledge is very limited. I have two groups (patients and controls) and a measured concentration value for each indivdual. Not normally distributed: Mann-Whitney
How do I correct for gender in Mann-Whitney? (I want to know if the difference between patients and controls is due to difference in gender distribution, since males have higher values than females).

You cannot correct for gender in Mann-Whitney. Mann-Whitney is a location test for two groups, and that's all.

There are at least two options here: 1) Stratify by gender. That is, analyze the men and women separately.

2) Do some sort of regression, perhaps OLS or, given your use of Mann-Whitney, perhaps quantile regression, with "concentration" as the dependent variable and two independent variables: Gender and group (patient vs. control)

• Is there any other non-parametric test that allows me to correct for gender? – Camilla Mar 15 '13 at 11:12
• "Nonparametric" covers a wide range of things, including types of regression. I don't know of anything like the Mann Whitney U that corrects for another factor - it's almost part of the definition: Mann Whitney U is simple precisely because it doesn't allow such things. – Peter Flom - Reinstate Monica Mar 15 '13 at 11:49
• Can I perform a Kruskal-Wallis with 4 groups? (male+patient, male+control, female+patient, female+control) – Camilla Mar 15 '13 at 14:17
• regression is ANOVA. – Peter Flom - Reinstate Monica Mar 15 '13 at 19:40

The closest I think you could get with some existing test, is if your data came from a randomized complete block design, but if I recall correctly, this requires people to be randomly assigned to the genders (blocks): this may be rather hard :). If you could, however, you could use a Mack-Skillings test.

You state that your statistical knowledge is limited, so the only true solution I have to offer may be somewhat out of your league: you can build a probabilistic index model (see http://onlinelibrary.wiley.com/doi/10.1111/j.1467-9868.2011.01020.x/full ). These can be used to extend the classical distributionfree tests (like Mann-Whitney) in a similar fashion as you would do through a GLM (a colleague of mine should have a paper published on that any day now), but it would lead me too far to try and explain everything around that theory here.

So: unless you're willing to go the PIM-route, I don't think there is a solution, currently, for distributionfree tests.

• Thank you for your answer, unfortunately it sounds to complicated for me. Can I perform a mack-Skillings test in SPSS? – Camilla Mar 15 '13 at 11:33