Bugs in Dunn's multiple comparisons test following the Kruskal-Wallis test.

Fixed bugs: Dunn's post test:

• In all versions of Prism up to 4.03 (Windows) and 4.0c (Mac), the Dunn's test is computed incorrectly if the total number of values (in all groups) is 1291 or greater. When the total N is this large, there is an overflow in the calculations and the results are invalid. Note that this threshold (1291) applies to the total number of values in all groups, not the number of rows of data. Fixed in 5.00.
• In Prism 3, avoid entering data with subcolumns for replicates. Fixed in version 4, which averages the subcolumns and does the ANOVA and post tests on the average values.
• Versions of Prism prior to 3.02 calculated the Dunn's post test incorrectly when there were missing values (blank spots) in your table. The blanks would count when the algorithm figured out the significance levels. So if you had lots of blanks, or were right on the borderline of significance levels, Prism would report a P value that is too low. Fixed in 3.02 and 3.0c.
• All versions of Prism up to 4.03 and 4.0c compute Dunn's test incorrectly when the data table is formatted for entry of Mean with SD or SEM (a format that makes no sense for nonparametric tests). Even if the SD or SEM subcolumns are blank, their presence confuses Prism and leads to invalid Dunn's test results. Use Change...column format to reformat the data table to have no subcolumns.
• Versions of Prism up to 4.03 and 4.0c computed Dunn's test incorrectly when some values were excluded. Prism correctly figured out the difference between mean ranks, ignoring the excluded values. But then it takes into account sample size and incorrectly counted all the values in each column, even the excluded ones. Starting with 5.00, Prism correctly ignores the excluded values when computing the n values. Since Prism 4 "thought" the sample size was bigger than it actually was, it sometimes erred in the direction of finding significance incorrectly or reporting ** when it should have reported *. This bug has the most impact with tiny samples, or large samples with a huge fraction of excluded values.