Hi, I know that the syntax for performing Kruskal-Wallis in STATA is: kwallis (Variable), by (GroupVar).
Could anyone tell me how to make the same if I would like to calculate p-values for the continuous Variable using the GroupVar PLUS other co-variates (e.g. age sex smoke .... )
Thanks
I think the scenario you propose violates the assumptions of the Kruskal-Wallis test, which requires one ordinal dependent variable and one categorical independent variable. If you have a continuous dependent variable and various other independent variables, you are in regression territory. If you are still talking about an ordinal dependent variable, and multiple other independent variables, try recoding the variables into one. This will only work with a large enough sample, of course.
http://www.ats.ucla.edu/stat/stata/whatstat/whatstat.htm
Thanks Richard for your nice answer. I suppose that I didn't explain my case very well then. What I have is an ordinal dependent variable that is re-coded from a continuous one (percentile groups) and an independent categorical variable. So simply I can apply kwallis in STATA under this conditions, but I would like to have the p-value after adjusting for age and sex and smoking for example, would that be possible?
I also did not get what you meant by recoding the independent variables into one! How can I do that for age, sex and smoking IN ADDITION to my categorical independent variable?
Thanks a lot for your help
If you want a relatively complex model with an ordinal dependent variable, the best approach is to use ordinal regression. How easy this will be probably depends largely on the software package you use. I find it pretty straightforward in R, but discussions on this site gave the impression that it gets pretty ugly in SPSS once you introduce dummy variables and post-hoc testing.
Yeah, I have no idea about performing ordinal regression with Stata.
Judging from the original question, Nawar has a continuous (or at least quantitative) dependent variable and a primary explanatory variable that is categorical, but he would also like to include a few covariates in the model. On the face of it, this sounds like a situation where one would consider using OLS linear regression. But for some reason, Nawar thinks he may need to use some variant on the Kruskal-Wallis test.
I wonder if the underlying issue is that Nawar is looking at the distribution of the raw scores, and finding that it is not normal. If so, Nawar, you are looking at the wrong thing. For OLS linear regression, it is the errors, not the raw scores, that are assumed to be normally distributed. So you need to fit the OLS model, save the residuals, and examine them. Bear in mind too that OLS regression is quite robust to non-normality of the errors--which is a good thing, as there is no such thing as a normal distribution in the real world. (See section 2.5 of George Box's famous article, for example, link given below).
If my guess about why you want to use KW is off the mark, please explain why you think you need to use it (or some other "non-parametric" method rather than OLS linear regression.
Cheers,
Bruce
http://mkweb.bcgsc.ca/pointsofsignificance/img/Boxonmaths.pdf
Thank you Bruce. This was very helpful. Indeed, I predicted the residuals and examined them as you suggested and this is exactly what I needed to calculate the median after that. I suppose KW is not anymore necessary because I can use the trend test after all and it is quite robust.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS