Statistical question
Hi,
Does anyone have experience with or thoughts on Mantel tests?
I feel sort of silly that I finally figured out how to post on this page and this is all I can come up with today but that's what I'm thinking about at the moment.
5 comments:
Hi Jenny,
We have a spatial statistics group this semester (organized by Cathy Collins) and we are currently figuring out how to do Mantel tests in R. We are also trying to figure out if they are useful for addressing spatial questions. Cathy has a lot of references on the subject that she would probably share with you. Do you want to know anything in particular?
Erin,
I have seen them used in several recent bee papers to test for spatial autocorrelation between different habitats or different parks (one matrix of bray-curtis indices and one of distances between sites). There is a program called "vege" in R that has been used for a Mantel test by the authors of one bee study (on a side note, I have some code for this). I spoke with a stats graduate student here at Iowa State and he was leary of using a Mantel test to look at whether there was any spatial correlation between my roadside sites and the species at each, primarily because I had two treatments (weedy vs. restored). He felt that there would be noise due to the treatments that might clog up any signal from spatial correlation, and that although it was already in the literature, he felt that in those instances the authors didn't find any spatial correlation due to noise resulting from combining different habitats (or parks, or whatever) in the same analyses. So we did two Mantel tests, and just looked at correlation within the treatments. But this doesn't addresses the real question, which is, are paired sites (one weedy, one restored) more alike because they are closer together, and we can't agree on an analysis that will. So my ultimate goal is to find out whether a Mantel test that includes all my sites would actually be the best way to look at spatial autocorrelation (which the literature implies), and if others think it is not, to find something that does. Any thoughts you and Cathy (and others) might have would be appreciated.
It is really cool that you have a statistics group!
Jenny,
You can do a partial Mantel test with three matrices. One with the species, one with geographic distance, and one for treatment "distance" (basically, 1's and 0's). It is more complicated to interpret, but seems like it would suit your study. We read Chapter 15 in Scheiner and Gurevitch 1993 [Design and Analysis of Ecological Experiments] that describes these models. But there are problems with them, so your stats person may know more about that.
Sounds like you are writing up your thesis. Are you thinking about torturing yourself with more graduate school???
Erin,
I am delighted by your suggestion. The stats student I have been in contact with doesn't actually know much about Mantel tests, and so will likely also be excited by the reference you provided. Thanks! Do you know the commands for performing such an analysis in R?
Yep, I am trying to write up my thesis. It is coming very slowly (very).
I am also thinking about more graduate school, but haven't taken much action yet and so may wait another year (or more).
Hope you are well.
Hey folks,
I Gtalked this to Jenny. Am I making any sense here?
Man! I'm a little bit too much of a cowboy when it comes to statistics (just do it and ignore a lot of complicating factors). Maybe not the best way to do things. But I'd sure just put in the distance and species matrices and see what happened... Seems to me that it would be CONSERVATIVE because the sites closest together differ in treatment. Then I think you could say that a significant result meant there was spatial correlation definitely. The only problem would be that your power would probably be low because of the proximity of different-treatment sites, so you couldn't say there ISN'T correlation just because you didn't find it. Cowboy that I am, I would run the data and only worry about the treatment effect if I DIDN'T find correlation. Does this make sense? Maybe I should post this on the blog to see what people think who have studied Mantels more than I.
Post a Comment