The pairwise test should work on the interaction variable.  Perhaps you made a syntax error?  Or maybe there's something funky in the data?

There are some other post-hoc tests, and a method to produce a compact letter display after pairwise tests here:

http://rcompanion.org/handbook/F_08.html

Thanks Salvatore for the link! Funky I don't know. I have ties in my data and they are bounded (>0 and

The pairwise function shouldn't be bothered by ties or bounds, but may not work, if, for example, the values for two treatments are exactly the same.

The adjustment method shouldn't cause the function to fail.  But you can try some different ones, and see if that fixes it.

*    *    *

If your dependent variable is percentage, you might look at logistic regression or beta regression, which may be better approaches theoretically, and will also allow you to do a two-way analysis with the interaction.

The first example at the following link uses logistic regression.  Note that the percentage data has to be derived from counts of successes and failures.  (It can't be animal A is 50% blue, animal B is 25% blue.  It has to be at location A, I found 15 blue animals and 25 non-blue animals.  Or in animal A I counted 15 blue cells and 25 non-blue cells.)

http://rcompanion.org/rcompanion/e_06.html

*    *     *

For me, I think it's okay to use a one-way style test on the interaction of factors.  I'm not so sure about running the test on the one main factor, and then the other main factor, and then the interaction.  It seems like the one test on the interaction with the post-hoc test essentially tells you what you need to know.

Dear Pauline

You can also achieve Kruskal-Walis test to test AB interaction combinations , with an assumption of independent groups, and in case you get a significant test, then use Mann-Whitney test between each pair.

Good Luck

Really great advice on this form but does is anyone aware of the script for running an interaction without combining the data before hand. Something to the extent of kruskal.test( -----------~ ------*-------, data = ----------). Cheers, Kieran

@Kiernan,

For the K-W test, it doesn't get any easier than using the interaction function as the original poster did.

I thought that might have been the case. Thanks so much for confirming