Kruskal-Wallis & Regression: unequal n, unequal sd & non-normal dep.var.

fxp

New Member

My problem is as follows:

I conducted a simple experiment to check for the effect of gender and gender pairing in the outcome of a game (the ultimatum game, I won't bother you with unnecessary details).

Important for my question to you, is just:
Each participant makes a monetary offer (decimal number between 0 and 10) to another participant and this offer is the dependent variable.
I have four groups:

Males who play with males
Males who play with females
Females who play with males
Females who play with females

Unfortunately the groups are not of equal sizes at all (17, 20, 11, 31), distribution is not normal, and the standard deviation is not the same across groups.

Question 1

I would like to build subgroups, such as:
female vs. male player
female vs. male opponent
mixed gender vs. same gender

--> Can I simply take the means/medians without accounting for the difference in group sizes?!

By first taking the means of each of the four conditions and only then the means of the wanted subgroups I get the weighted averages. But - it's only one number per condition and I cannot make statistical tests with that! (I would like to do a Wilcoxon Mann Whitney test e.g.).

Question 2

To compare the four groups I think I should use the Kruskal Wallis rank test.

--> Do you agree? If yes, is there a post-hoc method to compare the different subgroups? (I'm thinking of something that is like the Bonferroni method for ANOVA). Or can I simply use a Wilcoxon Mann Whitney then?

Question 3

I would like to run a regression (in order to be able to include other variables, such as age). Can I? Or would that lead nowhere, with all the assumptions that my data violate.

What do you think about the idea of creating a dummy variable that is:
0 if offer ≥X
1 if offer <X

and then running a probit regression?

--

I attach two pictures. They are both histograms of dependent variables that I would like to use in my regression. (one of them is the "offer" that I talked about above).

Thank you very much for your thoughts on these problems!

Karabiner

TS Contributor
--> Can I simply take the means/medians without accounting for the difference in group sizes?!
You don't need to do weighting.
If yes, is there a post-hoc method to compare the different subgroups?
U-tests with Bonferroni-corrected alpha is a common procedure.

But a oneway ANOVA with post hoc tests could also be possible,
if heteroscedacity isn't too arge and you don't have outliers.
I would like to run a regression (in order to be able to include other variables, such as age). Can I? Or would that lead nowhere, with all the assumptions that my data violate.
You can build a regression model and then check assumptions
(e.g. normally distributed error terms, homoscedascity).
What do you think about the idea of creating a dummy variable that is:
0 if offer ≥X
1 if offer <X
and then running a probit regression?
Why probit, not binary logistic regression? But why should you want to do it
anyway? Dichotomization will lead to a loss of statistical power.

With kind regards

K.

fxp

New Member
Thank you very (!) much for you very quick and helpful reply!!