Message du 18/02/13 ? 21h49
De : "Steve Brewer"
A : v_coudrain at voila.fr, r-sig-ecology at r-project.org
Copie ? :
Objet : Re: [R-sig-eco] adonis and temporal changes
Valerie,
If I understand your design correctly, you're doing a repeated measures
analysis, in which isolation is a between-subjects (I.e., between-sites)
effect. Year and the year x isolation interaction are within-subjects
effects. Because repeated measurements on composition are being taken on
the same site in three years, you use strata to restrict the permutation
within each site as if site were were a random block containing the
different years of measurement. Accordingly, there should be two error
terms: site(isolation) to test the isolation main effect, and the
site*year(isolation), which in this case is equivalent to the residual
error, which is the appropriate error term for testing the year effect and
the year x isolation interaction. The test for isolation is wrong because
adonis cannot use more than one error term to test effect and thus is
using the residual error to test all effects. It should use the
site(isolation) term to test the isolation effect, but it does not. Using
the residual error to test the isolation effect amounts to
pseudoreplication. It assumes that the three measurements of composition
in different years on the same site are independent observations. They are
not. Often, however, people are not interested in the between-subjects
effects (in this case, the main effect of isolation). Rather they are
interested in the interaction with time (in this case, isolation x year).
I don't see that you are justified in pooling any term with the error term
just because it is not significant. Again, the problem is
pseudoreplication. You're treating correlated observations as if they were
independent observations. Pooling the isolation x year interaction with
the residual error term artificially inflates your error df even more.
I'm afraid I don't know R well enough to explain how to analyze the
covariate.
J. Stephen Brewer
Professor
Department of Biology
PO Box 1848
University of Mississippi
University, Mississippi 38677-1848
Brewer web page - http://home.olemiss.edu/~jbrewer/
FAX - 662-915-5144
Phone - 662-915-1077
On 2/18/13 1:49 PM, "v_coudrain at voila.fr" wrote:
Dear Steve,
Thank you very much. I do not exactly understand why the test for
isolation will be wrong, would you have some some explanation?
In a linear regression, you cannot assess the effect of single variable
if the interaction (in which your variable is part) is significant. So if
I get a significant result
for the isolation*year effect I should conclude that there is an
interaction between isolation and year. If the interaction is not
significant, should I drop it to get the
correct estimate for the year effect?
I would have an additional question: I have also an environemental
gradient (continuous, one value pro site, constant over the years). Is it
possible to include it?
Best wishes
Valerie
Message du 18/02/13 ? 15h41
De : "Steve Brewer"
A : v_coudrain at voila.fr, r-sig-ecology at r-project.org
Copie ? :
Objet : Re: [R-sig-eco] adonis and temporal changes
Valerie,
Adonis does not define fixed or random effects, and you therefore cannot
define multiple error terms. However, if your model statement looks
something like this - isolation*year + site, strata = site - then you
will
get the correct test for the isolation x year interaction and the
correct
test for the year effect. The test for isolation will be wrong, because
the residual error is used to test all effects, when it is only
appropriate for testing the year effect and the year * isolation
interaction. The isolation between-subjects effect should be tested with
the site effect but is not.
The key point is here to make strata = site and to NOT specify the site-
interactions with isolation or year. In this way, site will be treated
as
a block for the within-subjects effects and thus could be considered a
random effect.
Hope this helps.
J. Stephen Brewer
Professor
Department of Biology
PO Box 1848
University of Mississippi
University, Mississippi 38677-1848
Brewer web page - http://home.olemiss.edu/~jbrewer/
FAX - 662-915-5144
Phone - 662-915-1077
On 2/18/13 8:19 AM, "v_coudrain at voila.fr" wrote:
Thank you for these explanations. If I put strata=site, this means that
for each site my dissimilarity matrix of year 1 and year 2 will be
permuted and the observed
changes compared to these random permutation? Adding site as a fixed
factor then ensure that I am testing changes in time site by site. Am I
correct?
To my design:
I have 30 permanent sites, 10 of each category of isolation (Isolation
factor with 3 levels: 3x10 sites = 30 sites). I conducted the samples
three years in each
site. I have thus 1 sampling (species composition) pro site pro year. I
would like to know how the sampled communities change with time, either
on a site basis,
or at the level of isolation (I may compare multi-site dissimilarity
among isolation levels between years).
I am not really interested in knowing what proportion of differences in
species community is due to space vs time, but I would like to really
focus on the temporal
changes. That's why I think putting site as a fixed effect should be
appropriate. But if you have any suggestion or think this is not
I would be pleased to
have your opinion.
Cheers,
Valerie
On 18/02/2013, at 14:04 PM, Pierre THIRIET wrote:
Dear Val?rie,
If I remember well, your design includes:
Isolation categories: 3 levels
Sites: nested within Isolation categories (10 levels, a total of 30
sites)
How many replicates per site and time?
Time:? how many years you have? Only one sampling per year? Within
sites and years, samples were random or it is always exactly the same
area you
sample (e.g. permanent quadrats)?
for adonis, consider that strata is for constraining permutations,
which is different than terms in the formulae.
Exactly. The 'strata' only influence the permutations and have no
in formula nor effect defined in the formula.
Currently the 'strata' are the only way to constrain the permutations.
However, in the R-Forge version of vegan and in vegan 2.2-0 (to be
released in April) you
can give a permutation matrix as an input to adonis. You can generate
permutation matrix with, say, shuffleSet function of the permute
This allows
generation of restricted permutations for instance for time series.
command vegandocs("permutations") will open up the vignette of the
permute package
for your inspection, and this will give some examples of defining
restricted permutations. At some timeframe we are completely moving to
the permute package,
but you can already use its permutation matrices as input with these
and upcoming versions of vegan from R-Forge.
Cheers, Jari
--
Jari Oksanen, Dept Biology, Univ Oulu, 90014 Finland