Skip to content

pairwise comparisons in community structure analysis data

2 messages · Penelope_Pooler at nps.gov, Gavin Simpson

#
I have a question that I'm not sure has a right answer, but I would
appreciate any and all opinions, especially if you know of any citations to
back them up.

In the past, when dealing with univariate data, I have always been a
proponent of using Fisher's Protected LSD for pairwise comparisons because
it is the most powerful procedure, and the accuracy of its experimentwise
error rate is based on a sound argument that was subsequently proven to be
true with simulations in Carmer and Swanson (1973 ).  I included the full
reference below.

At the moment, I'm working with community structure data (multivariate
nonparametric) at multiple sites and there is an need to determine which
sites are different from which other sites.  My colleague has used a
Bonferroni correction  for this type of question the past, but I tend to
think that is most likely too conservative.  I'm interested to know if any
of you have dealt with a similar problem and/or if you know if anyone has
done any work on comparing pariwise comparisons procedures for multivariate
nonparametric data.

I've done a preliminary literature search with no success, but am still
looking.

Thanks for your help.

-Penelope




Carmer, S. G. and M. R. Swanson (1973). "An evaluation of ten pairwise
multiple comparison procedures by Monte Carlo methods." Journal of the
American Statistical Association 68 (341): 66-74.

==============================
Penelope S. Pooler
Quantitative Ecologist
National Park Service I&M
Northeast Coastal and Barrier Network

URI Coastal Institute in Kingston
#1 Greenhouse Rd., Rm 205
Kingston, RI 02881

Penelope_Pooler at nps.gov
Ph.: (401) 874-7060
#
On Mon, 2009-08-10 at 16:43 -0400, Penelope_Pooler at nps.gov wrote:
You don't say how you are comparing sites, i.e. to what is the
Bonferroni correction applied? The p-value from ...?

You are also a bit vague about the data structures/layout. If your data
represent a set of sites with multiple samples taken at each of these
sites, then you might want to look at adonis() and betadisper(), both
functions in the vegan package.

adonis() partitions the dissimilarities between sites on the basis of
certain factors - so it is like ANOVA but with multivariate responses
using any dissimilarity and you can think of this analysis as looking at
whether the distances between points within a group (factor, site etc)
are greater than the between groups distances or distances to other
groups.

If adonis() shows a significant difference between sites, this
difference may be due to differences in the spread or variance of the
respective groups, just like in standard ANOVA. betadisper() can be used
to test for this.

Unfortunately, adonis doesn't do pairwise comparisons, but with the
right amount of force could be changed to do so. betadisper does do
pairwise comparisons and even has a TukeyHSD method for it (for the
multiple comparisons), but a user has recently pointed out that both the
parametric p-value and the permutation-derived p-value of the main
betadisper analysis don't have the right Type I error rates - Ho is
accepted too many times when given random groups. This seems to be a
problem with the general method and not with the implementation in
vegan, and is something that I (as the author of betadisper) and Jari
Oksanen are currently looking into (slowly on my part, I'm afraid...).

If the above and Rodney's reply don't address your real problem per se,
perhaps you could elaborate on what tests you are using and wish to
correct for multiple comparisons?

All the best,

G