Details and restrictions of the Tukey test
Tukey properties
The Tukey test provides the following optional properties:
Property

Default value

SigLevel

5

SigLevelLow

none

MinBase

30

SmallBase

100

Tukey usage
The Tukey test is not suitable for all tables. When you request the test on a table that is structurally unsuitable, UNICOM Intelligence Professional omits the test, leaves the Tukey and p value rows blank, and writes an explanation to the diagnostics data. You must make sure that the data in the table is generally suitable for testing, that the sample size is suitable, and so on.
Reporter displays a message if you define a Tukey test on an unsuitable table or if you change a table that has a Tukey test defined so that it is no longer suitable for the test. When this happens, you can either change the table so that it conforms to the restrictions, or remove the test from the table.
Rows and columns
The test is allowed only when the columns are independent (the columns do not overlap). The test is not allowed on tables where there is overlap between the columns, which includes all grid tables. The test is ignored when any of the test columns are empty, or when a column's base value is smaller than the minimum base value.
The test is only implemented for rows of means. It is not implemented for rows of proportions.
Effective base
The test does not use the effective base. In an unweighted table the test uses the unweighted figures; in a weighted table the test uses the weighted figures.
Other tests
The Tukey test is not allowed on a table that already includes the Column Means and/or Column Proportions tests, because there would be no way to identify to which test the significance letters refer.
Homogeneous subset detection
Homogeneous subset detection is included in the table builder, but the results can only be displayed in the diagnostics file.
The pairwise test for whether two means are significantly different is:
mean2  mean1 > Q(i,j) * thresholdvalue
while the test for whether two means do not belong in the same homogeneous subset is:
mean2  mean1 > Q(h) * thresholdvalue
It is possible for the pairwise output to deem that two means are significantly different, while the homogeneous subset detection deems they belong in the same subset  or conversely, that they are not significantly different but belong in different subsets.
See also