Clarifications on Dunn-Sidak Approach in multcompare.m
10 views (last 30 days)
Show older comments
SungJun Cho on 14 Jul 2021
Commented: Scott MacKenzie on 14 Jul 2021
I am trying to perform the Dunn's test as a non-parametric post hoc multiple comparison test based on the statistical results obtained from the Kruskal-Wallis test.
I want to ask if a 'dunn-sidak' option provided by multcompare.m is actually the Dunn's test. Based on the paper (Dunn, 1964) and other materials, the Dunn's test is known to use z-statistics, but according to the documentation provided here, it seems like 'dunn-sidak' option uses critical values from the t-distribution.
Other software (e.g., R; cf. dunn.test, p.4) explicitly returns z-test statistics from the Dunn test. When I tested with Python, both Python and MATLAB gave equal results for one dataset but different results when conducted using other datasets.
Can anyone clarify whether the Dunn-Sidak approach in MATLAB is actually the Dunn's test and how does t-statistics fit in here?
My question has also been asked in MATLAB Answers and partially discussed in StackExchange, but I feel like it has not been clearly answered, especially for the MATLAB software.
Scott MacKenzie on 14 Jul 2021
Presumably the MATLAB documentation you cite on Multiple Comparisons is correct -- the Dunn-Sidak test, as implemented in multcompare, "uses critical values from the t-distribution". It's possible this is mathematically equivalent to the test given in the original source you cite The formula is given, so perhaps you can compare the two sources to find out.
Find more on Analysis of Variance and Covariance in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!Start Hunting!