negative F value in 2-way permanova

Theory behind mothur
1

I am performing a permanova using an OTU table and I want to examine 2 factors if they affect the composition of my OTUs
each of the factors is affecting significantly the bacterial OTU composition and I get for each one of them positive F values (as someone would expect)
However, the 2-way interaction between the factors is giving me a negative F-value!
I understand that in PERMANOVA it is a pseudo-F but still… what does a negative F value mean?

here is the output:
Two-way PERMANOVA
Permutation N: 9999
Source Sum of sqrs df Mean square F p
Taxonomy 5.4349E08 2 2.7175E08 9.6671 0.0001
Actino 1.4429E08 1 1.4429E08 5.1329 0.0001
Interaction -1.3142E08 2 -6.5711E07 -2.3376 0.0004
Residual 2.8391E09 101 2.811E07
Total 3.3955E09 106

any ideas?


PS: wasnt sure where to post this question, so I apologize in advance if I posted it in the wrong forum... :(
2

I googled “negative f permanova” your duplicate question on 2 other boards were first, this answer was third.

http://grokbase.com/t/r/r-sig-ecology/1089n8dbgh/r-sig-eco-negative-f-value-in-adonis

3

well, I have been obviously looking for other answers than the one in that forum and by the way I am not using R I am using PAST to run the permanova

as for my duplicated/triplicated question, is there a rule of not posting the same question in a different forum?

4

it’s generally not a great idea to spam a bunch of fora. The link I provided is applicable to permanova regardless what software you ran it in

most likely if your data is strongly non-Euclidean

You likely have an underlying gradient that’s relationship to the communities is a step function rather than a gradual change. This is common in communities which makes permanova inappropriate for evaluating that relationship.

5

in which case (if PERMANOVA is non-applicable) what would you use instead?

6

have you plotted your communities in an NMS? this will tell you if your permanova issues are due to outliers.

How to assess communities that are responding to non-linear gradients is a huge topic in macroecology, dive into that lit. I’ve used regression trees, machine learning, linear modeling, etc.