0

I'm using Wilcoxon Rank Sum Test on my data and I'm wondering if there is a better alternative to use and eventually what post-hoc I could do on my data.

Context: The objective is to find the preferred soundtrack for a video animation. 20 participants, 4 possible soundtracks. The video can be presented in 3 different possible velocity levels. So there are in total 12 (4x3) combinations of video/sound presented to each participant. Each combination is presented only once (no repetitions). Each combination is evaluated through 3 questions each with range 0-100

For instance this is a plot of some sample data for 1 question. Thanks! enter image description here

SF1
  • 101
  • 1
    It reads like the same participant rates several combinations, which makes those ratings dependent. This needs to be taken into account, and the Wilcoxon test doesn't do that. I'd probably fit an ANOVA-type linear model with random effect. Chances are you want to respect the ordinal character of the ratings, however this is hard and required assumptions for linear models can be taken as regarding distributional features rather than the measurement level of the data, although this is controversial. – Christian Hennig Dec 01 '21 at 12:26
  • issue is that the distribution of data is not normal for ANOVA.. – SF1 Dec 01 '21 at 12:34
  • 1
    Some of what I wrote here should also apply in this situation: https://stats.stackexchange.com/questions/324795/normality-test-on-small-samples/543393#543393 – Christian Hennig Dec 01 '21 at 12:41
  • thanks, can I ask to clarify what you mean by "you want to respect the ordinal character of the ratings"? – SF1 Dec 01 '21 at 14:05
  • 2
    Some say that the linear model, t-tests and the like should not be used with ordinal data. I don't agree, but there are arguments either way. – Christian Hennig Dec 01 '21 at 14:10
  • Don't see how Wilcoxon S R test would be useful. It is a 2-sample test. If you really need to do nonparametric test, then maybe Kruskal-Wallis or Friedman would be useful. // I don't know what software you are using to make boxplots, but in R you can often make 'notched' boxplots of 2 or more groups. notches in sides of boxes are nonparametric CIs, roughly calibrated for comparing _pairs_ of plots: if notches don't overlap that's an indication group population medians differ. // Permutation tests might be useful. – BruceET Dec 01 '21 at 16:08
  • thanks! yes I'm using R, will look at the notches (didn't know they were CIs). Data comes from 0-100 slider and distribution is not normal. I'm not good enough with stats to justify using parametric tests with the distributions I have. – SF1 Dec 01 '21 at 23:21

0 Answers0