Non-parametric test for 3 matched-pairs?

I showed a video to 3 groups, and they answered questions before and after the test.

Control group watched it with no ads Group A watched it with ad A Group B watched it with ad B

The sample is such that the test needs to be non-parametric.

What do I to test whether the differences between the pre-test/post-test scores among the groups is statistically significant?

1 Answer

One possibility is to do a comparison of post-pre differences for the three groups via a Kruskal-Wallis on those differences.

• So compare control<->a, a<->b, AND control<->b? Or can you Kruskal-Wallis all at once? (Using SPSS) Commented Mar 30, 2013 at 1:47
• Kruskal-Wallis is the nonparametric equivalent of one way ANOVA. That is, it tests for differences in mean rank between all three at once. You also have the possibility of post hoc multiple comparisons, and for location-shift alternatives you can get CIs for the relative location differences. If you have specific contrasts in mind you could do a Mann-Whitney on those. Commented Mar 30, 2013 at 1:53
• What is the distributional issue or issues? Heavy skewness? Discreteness? Something else? Commented Mar 30, 2013 at 1:59
• Aaah, gotcha! That makes sense, at least in theory... But running it gives me these results: i.imgur.com/cPIb3AX.png Where it seems like pretest/posttest are just considered 2 random variables; is there a way to make Mr Kruskal and Mr Wallis contrast the changes between the groups? Or am I missing the point completely (seems likely!) Commented Mar 30, 2013 at 1:59
• I can't say I'm certain what analysis the output corresponds to, but it looks to me like you didn't compute the post-test minus pre-test differences before running the Kruskal-Wallis, which was what I was actually suggesting. Commented Mar 30, 2013 at 2:06