I am usinng SPSS 24, and found out that post-hoc tests following Kruskal Wallis (non-parametric 1-way ANOVA) automatically produces pairwise comparison (namely it is expectedly performing Mann Whitney test). However, when we directly run Mann Whitney test, the outcome values are different between the two. Could you tell me which computation is a correct method to retain?

Please explain which procedure you are running and what differences you see. The KW post hoc tests would include a multiple-testing significance adjustment, but running MW two-group tests directly, those results would not have that correction.

And, of course, missing data would complicate the comparison.

https://www.sheffield.ac.uk/polopoly_fs/1.714567!/file/stcp-marshall-KruskalSPSS.pdf

On 3-4page on this document, it says it uses Dunn-Bonferroni test for the pairwise test (post hoc) following the KW. So it was not MW. if this is true, my problem is solved.

I am aware of bonferroni correction and its adjusted or not-adjusted p-values. the raw p-values for the pairwise provided by the KW posthoc was different when I select MW (using the same Analyze - non parametric test - independent analysis) as well as legacy mode MW. The latter two are identical; however, the pairwise provided by KW differ. To note, bonferroni corrected values also differ from the latter two methods.

You can see the official details in the Algorithms doc, which you can find here

https://www.ibm.com/docs/SSLVMB_28.0.0/pdf/IBM_SPSS_Statistics_Algorithms.pdf

thank you.

one line of the PDF says:

"The Kruskal-Wallis, Friedman and Kendall, and Cochran tests use the procedure proposed by Dunn (1964) (originally designed for the Kruskal-Wallis test). The procedure uses ranks (or successes for the Cochran test) based on considering all samples rather than just the two involved in a given comparison".

So it would be user's choice whether keep using the Dunn's correction method (with p-value adjusted for bonferroni) or standard MW with manual correction with the bonferroni.

now the case is solved I believe. Thank you.