I have rank data from a survey of Impact Assessments (IA). Each IA is ranked from A to F on a particular aspect (such as quality of introduction). These IAs come from 7 different sectors, but I don't have an equal number from each sector. The sectors and the number of IAs for each are as follows

Mining: 19

Transport: 9

Infrastructure: 10

Waste: 10


Manufacturing: 5

Environmental: 5

I ran 21 Mann-Whit U tests in SPSS to determine if the ranks received by each sector were significantly different from one another. Many of the tests produce a significant difference at a 0.05 significance level.

My question is, with such a big difference in the number number of observations for each IA sector, and in some cases only 5 observations (such as manufacturing and environment), would it be better for me to be to be only using results of a higher confidence interval such as 0.01? Especially considering that some aspects are not applicable to particular IA (approximately 7% of ranks were not applicable)


1 Answer 1


A standard approach would be to use a Kruskal-Wallis test to test for differences among the sectors, and then Dunns Test to see which pairs of sectors are different. This previous answer is relevant: Post-hoc tests after Kruskal-Wallis: Dunn's test or Bonferroni corrected Mann-Whitney tests?

  • $\begingroup$ Thanks I will run Kruskal Wallis instead of pairwise tests $\endgroup$ Commented Jul 31, 2016 at 12:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.