4

I am trying to test a hypothesis for my Masters Thesis.

There are 3 conditions (I will name them X, Y, Z), in each of which the data isn't normally distributed, and measures the improvement for each condition. I ran a Friedman test as a repeated measure, and in post-hoc I ran a Wilcoxon test with the Bonferroni adjustment.

My problem is that I want to compare only X vs Y and X vs Z (and not Y vs Z).

My question is this: for the Bonferroni correction do I need to divide by 2 or 3?

On one hand, there are 3 comparisons done in the SPSS test, but on the other I'm only interested in 2 of them.

Does this even count as a repeated measure?

Thank you!

Izzy
  • 41
  • 4
  • We need more information for the "repeated measure" part of the question. Is this one group of subjects measured three times, or three different groups of subjects each measured once? – TPM Aug 05 '19 at 18:22
  • 1
    Note that "[repeated measure](https://en.wikipedia.org/wiki/Repeated_measures_design)" has a technical meaning in statistics, representing multiple measurements on the same individual under different conditions or over time. I think that what you're asking about is really [multiple comparisons](https://en.wikipedia.org/wiki/Multiple_comparisons_problem). If I'm correct, please change the tag and edit your question so that future visitors to this page won't be confused. If there is a repeated measures issue, please address the comment from @TPM. – EdM Aug 05 '19 at 18:31
  • @TPM This is one group of subjects. They each go through all 3 treatments (conditions) on different sessions, and are measured each time. – Izzy Aug 07 '19 at 07:57

1 Answers1

7

To the best of my knowledge, the Bonferroni correction is based on the number of tests you actually perform, not on the total number of pairwise tests that you could perform. So if you set out to perform two pairwise comparisons, you should divide by two.

ZombiePlan37
  • 145
  • 11
  • 1
    This is correct. One point to consider is that if the p-value for the YZ comparison turns out to be significant, it still shouldn't be considered significant. It should be as if you never calculated that p-value, since its calculation is just a matter of the software doing it by default (and this most likely can be changed). – Dave Aug 05 '19 at 18:27
  • Also you should make sure it is Bonferroni (or Holm) you are interested and want to control family-wise error rate (FWER) rather than BHY (Benjamini-Hochberg-Yakutieli) which controls false discovery rate (FDR). – NBF Aug 15 '19 at 11:01