Vegan: Anova.CCA accessing original data using option by=

Steve Pawson <Steve.Pawson <at> scionresearch.com> writes:

My apologies for the delay in responding to your request for further

information I have been travelling for

work since you replied and have only just returned to email contact.

The output from the traceback is as follows
# This is the capscale model that I called

beetlecap <-capscale(log(beetles+1) ~ size + Clearfell + Absolute.Distance+

Distance_from_edge+

clearfell.harvest_area + Canopy.Cover + X500mnative + Litter3 + X500mexotic +

X5000exotic +

Condition(AdjLong + AdjLat + AdjLat.2 + AdjLat.2.long + AdjLong.3), environ,

distance = "bray")

This is the ANOVA by margin option with the error

anova(beetlecap, by="margin")

Error in dimnames(x) <- dn :
  length of 'dimnames' [2] not equal to array extent

Corresponding traceback

traceback()

9: `colnames<-`(`*tmp*`, value = c("CAP1", "CAP0"))
8: capscale(formula = log(beetles + 1) ~ size + Clearfell + Absolute.Distance +
       Distance_from_edge + clearfell.harvest_area + Canopy.Cover +
       X500mnative + Litter3 + X500mexotic + X5000exotic + Condition(AdjLong +
       AdjLat + AdjLat.2 + AdjLat.2.long + AdjLong.3) + Condition(size +
       Clearfell + Absolute.Distance + Distance_from_edge +

clearfell.harvest_area +

       Canopy.Cover + Litter3 + X500mexotic + X5000exotic + AdjLong +
       AdjLat + AdjLat.2 + AdjLat.2.long + AdjLong.3), data = environ,
       distance = "bray")

[...clip...]

Dear Steve Pawson,

With the help of this message I was able to construct an example that gives 
the same error message -- this does not prove that the cause of the problem
is the same, but it is possible.

It may be that your *huge* model has redundant variables that cannot be
analysed in marginal test: the other variables explain all, and the marginal
effect of some variables is zero. With that a high number of variables as you
have, this is very likely. It seems that capscale() cannot cope with this case.

I fixed capscale in http://vegan.r-forge.r-project.org and now it handles
smoothly these redundant variables (skips them in permutation test, and
reports df=0). From your point of view it may be unfortunate that I released
a new version of vegan a couple of hours before checking R-News mail, and
therefore this fix is not yet in the next release, and as we just had a release
we probably (hopefully) will not have a new revision very soon. So your
choices are either to use the vegan version in R-Forge (which must be at 
least r1958) or simplify your model so that you don't have redundant variables.
One way of achieving this is to use command

alias(beetlecap, names = TRUE)

which will list the names of the variables that cannot be analysed. You
can remove these variables without influencing your fitted model, because
they really are redundant variables.

Cheers, Jari Oksanen