Kruskall-Wallis, ? need for ranksum test

Kruskall-Wallis, ? need for ranksum test

by Arti -

Hi - I was hoping to clarify this issue.

For my table I am doing comparisons between a tertile variable and several variables (categorical and continuous). For the continuous variables - I am choosing between ANOVA and Kruskall-Wallis depending on whether or not the outcome variable is normally distributed. In the table, I have included the p-value results of the Kruskall-wallis (and the ANOVA and chi-squared). Is this okay?

In the reply to my prior post, I was advised to carry out the ranksum test if the Kruskall-Wallis test is statistically significant to determine if there is a significant difference between pairs of categories. I've tried reading about this and tried to understand it but have not yet figured it out. Is this necessary for my table? My goal in the table is to show whether or not there is a statistically significant difference between the 3 tertile groups and I was not interested in comparing pairs. 

If the ranksum test is necessary, how would I interpret the output that I get? (I ran it once and was somewhat confused by the output.)

Thank you.

Re: Kruskall-Wallis, ? need for ranksum test

by Sofia Verstraete -

Arti,

You can report the overall p-value from the Kruskal-Wallis test in your table.  The reason why Ellie suggested doing the rank sum test to compare each group is because you should generally look at the differences between each pair of groups to get more information from your association to further support your conclusions.  For example, if you were to run an ANOVA (in a normally distributed sample) to compare three groups, you would probably follow that by pairwise comparisons to see where the actual differences are (or are not).  This is probably even more important when using Kruskal-Wallis (it is a less "sensitive" test because it does not assume normality and may be more likely to report something as not significant).  You can do that by following your Kruskal-Wallis test with comparisons of each pair of groups using either rank sum or Mann-Whitney (the non parametric equivalent of a t-test).  Depending on the results you can add that to your table or just report it in your results section.  For the purpose of this project reporting the result of the Kruskal-Wallis test is enough, but you should do the rest to have a more clear understanding of your data and your conclusions.

Sofia

Re: Kruskall-Wallis, ? need for ranksum test

by Elizabeth Uy-Smith -

Hello Arti, 

For your table, you are correct - you can just show the Kruskall-Wallis test results between the tertile variable and the other variables. You would just report the single p-value for an overall difference in the 3 categories.

I didn't mean to confuse you by introducing the "ranksum" option. If you believe it is important to determine which of the individual categorical pair combinations (1&2 or 2&3 or 1&3 in a tertile variable) in a multi-categorical variable is significant, then would can choose to perform a ranksum test. Analogous to the t-test for normally distributed variables, the Wilcoxon rank-sum (a.k.a. Mann-Whitney) test is used to test 2 categories in variables with non-normal distributions.  

Hope that helps clarify things. Please let us know if you have more questions.

ellie

Re: Kruskall-Wallis, ? need for ranksum test

by Arti -

Thank you very much. That helps a lot. I understand the concept now.

And I changed my graph so I will now use the Wilcoxon rank-sum test since I only have 2 groups and a non-normally distributed variable.

For the table, I will go ahead and present the values I get from the ANOVA and Kruskall Wallis test. Since the output is already 35 pages long, I will make a separate log file with the paired ttest and the ranksum tests for all of the variables and send it as a separate attachment.

Please let me know if it is okay for me to send those stats as a separate attachment.

Thank you!

Re: Kruskall-Wallis, ? need for ranksum test

by Elizabeth Uy-Smith -

Yes, you can send the paired ttest and ranksum tests as a separate attachment.

Thanks,

ellie