Overview of multiple comparisons choices

Multiple comparisons testing is chosen on two tabs of the analysis parameters dialog.

•The Multiple Comparisons tab specifies the questions you want the multiple comparisons tests to answer. This decision depends on the experimental design and will vary from experiment to experiment.

•The next tab (Options) drills down to choose a test. Those choices tend to be personal or lab preferences, so we put them on the options tab.

These choices should reflect the experimental plan. It is not ok to first look at the data, and then decide which comparisons you want Prism to calculate. If you first look at the data, then you effectively have made all possible comparisons.

How many comparisons?

No multiple comparisons

Multiple comparisons are optional.

Compare each column mean with every other mean

This is probably the most commonly used comparison. Because it makes more comparisons that the other choices, it will have less power to detect differences. You'll choose the exact test on the Options tab, but Tukey's test is used most often.

Compare the mean of each column with the mean of a control column

It is common to only wish to compare each group to a control group, and not to every other group. This reduces the number of comparisons considerably (at least if there are many groups), and so increases the power to detect differences. You'll choose the exact test on the Options tab, but Dunnett's test is used most often.

Compare preselected pairs of columns

Comparing preselected pairs of column means reduces the number of comparisons, and so increases power. But you must have chosen the pairs of means to compare as part of the experimental design and your scientific goals. If you looked at the data first, and then decided which pairs of means to compare, then you  really compared all means. Read about these planned comparisons. You can select up to 40 pairs, but the list may get truncated if you open the file on older versions of Prism).

Test for linear trend

The test for linear trend is a specialized test that only makes sense if the columns are arranged in a natural order (e.g. dose or time) and you want to test whether there is a trend such that the column means tend to increase (or decrease) as you move from left to right across columns. The other multiple comparisons tests pay no attention at all to the order of the data sets. Note that while this choice is on the Multiple Comparisons tab, there really is only one comparison.

Choices Prism doesn't offer

Prism omits two choices that some other programs offer.

•Prism cannot do comparisons (called contrasts) that involve multiple groups -- for example, comparing the grand mean of groups A and B, with the grand mean of groups C, D and E. Scheffe's method is designed to handle these kinds of contrasts, but Prism does not offer it.

•While Prism can test for linear trend between column mean and column order, it cannot test for other trends (quadratic, etc.).