[R-sig-eco] Interactions in MRM

[Leamy, Larry]

Dear Dr. Goslee:

I found some posts about potential interactions in MRM, and I wonder if it is acceptable to construct an interaction term as follows?
leaf1 = MRM(gendist ~ regiondist + envirdist +I(regiondist*envirdist) + moldist, mrank = TRUE, nperm=1000)

(I am using mrank = TRUE, incidentally, because our gendist matrix has a number of 0 values)

If this is not acceptable, what if I simply multiply the elements in the two matrices to generate a separate (interaction) matrix? This is the usual way I have formed interactions in multiple regression models in SAS.

Which (if either) is preferable, and is there any problem with significance testing with an interaction term in the model?

Thanks much for your help.

Best,

Larry Leamy




Dr. Larry Leamy
Department of Biology
University of North Carolina at Charlotte
9201 University City Blvd.
Charlotte, NC 28223
Phone: [phone]
FAX: [phone]
Email: ***@uncc.edu


[[alternative HTML version deleted]]

[Sarah Goslee]

Hi Larry,

Post by Leamy, Larry
I found some posts about potential interactions in MRM, and I wonder if it is acceptable to construct an interaction term as follows?
leaf1 = MRM(gendist ~ regiondist + envirdist +I(regiondist*envirdist) + moldist, mrank = TRUE, nperm=1000)
(I am using mrank = TRUE, incidentally, because our gendist matrix has a number of 0 values)

Nothing has changed since those earlier discussions: MRM will not
process interaction terms. You can use standard regression methods to
get coefficients, but not measures of significance because os the
permutation test needed for properly testing hypotheses based on
dissimilarities.

Post by Leamy, Larry
If this is not acceptable, what if I simply multiply the elements in the two matrices to generate a separate (interaction) matrix? This is the usual way I have formed interactions in multiple regression models in SAS.

I don't know of any testing of the effectiveness of that method, and
cannot recommend it. The extensive literature on partial Mantel tests
should give you some idea that dealing with anything involving
dissimilarities is not statistically straightforward.

That said, if you find some research on the subject or develop your
own rigorous testing I will add your findings into the ecodist
package. :)

Post by Leamy, Larry
Which (if either) is preferable, and is there any problem with significance testing with an interaction term in the model?
Thanks much for your help.
Best,
Larry Leamy