0

Hello I am doing some stats on some research I did and was having trouble picking the correct statistical tests to use.

The question I am trying to answer is, I have two cell types (I, II) and two treatments (M, WT). I would like to know how they compare to one another, specifically:

I-WT - control for I
I-M - does this still look like I-M?
II-WT - control for II
II-M - does this look more like II-WT or Is?

To answer this it was recommended that I use a two-way anova, however according to Shapiro-Wilk my data is not normal (p < 10e-06 for all features). Is there a recommended non-parametric test for comparisons between 4 groups?

Connorr.0
  • 3
  • 2
  • 1
    1. When testing normality, did you look at the raw response? Or residuals? 2. see here: https://stats.stackexchange.com/questions/2492/is-normality-testing-essentially-useless/2501#2501 (and some of the other answers) – Glen_b Mar 18 '19 at 23:40
  • You can use quantile regression. – Peter Flom Mar 18 '19 at 23:52

1 Answers1

0

You can use the Kruskall-Wallis test which generalizes the Mann-Witney test to multiple groups.

Unfortunately this only gives you a p-value, not a meaningful effect size. So in this case you might want to do a comparison between group I and its control, as well as a comparison between group II and its control, also.

For the two-group comparisons, if the data are not sufficiently close to being normal distributed that you are happy to use a t-test, you might consider some transformation (for example logarithm), or you can provide confidence intervals for the medians in each group.

Helene Hoegsbro Thygesen
  • 361
  • 1
  • 7
  • 3
    Kruskal-Wallis is one-way, not two-way (which the question seems to ask for). (However, under an assumption of a location shift alternative you can estimate location shifts and even [provide a joint confidence region for them](https://stats.stackexchange.com/a/76080/805)) – Glen_b Mar 18 '19 at 23:39