Skip to content
Prev 1028 / 20628 Next

is multicollinearity of fixed effects resolved by random effects

Disclaimer: The following recommendation of a sequence of steps is not
the only one, perhaps not even the best one. It has worked for me in
the past.

First, I assume
- that centering was done with SCALE=FALSE,
- that you checked that linearity is defensible for the relation
between X13C and your three predictors (i.e., that you do not need
quadratic terms for some of them),
- that you used method = "ML" for comparison of models with identical
random but different fixed effect parts and you read up on the some of
the complications associated with such comparisons discussed on this
list. (You can use method="REML" for comparison of models differing
only in the random effects, again there are some qualifications.)

Second, I recommend that now you focus on the fixed-effects part. It
still is a bit random (pun intended).
If taking out the three.factor interaction does no harm, you may want
to remove the non-significant two-factor interactions. After those are
gone, you may want to check whether there is a non-significant main
effect that is not part of an interaction (cMAP?). Then you may want
to take this one out, too. Such a hiearchical dropping out of effects
may lead you to your current favorite:
X13C ~ cMAT * cLAT + TYPE + (1 | SITE) + (1 | SPECIES)
which expands to:
X13C ~ cMAT+ cLAT + cMAT:cLAT+ TYPE + (1 | SITE) + (1 | SPECIES)
Note there are often good arguments for keeping theoretically
interesting effects in the model, even if they are not significant!

Third, you should plot your effects to get a good idea about the
source of the interaction.

Finally, you specify various random-effects parts and try to
understand what they mean (see previous posts to this list), for
example:
(1|SITE)
(1|GENUS)
(1|GENUS) + (1|SITE)
(1|GENUS) + (1|GENUS:SITE)   equivalent to  (1|GENUS/SITE)

Then allow the significant fixed effects to vary for the random effects.

Reinhold Kliegl
On Tue, May 20, 2008 at 1:55 AM, Jordan Mayor <clavulina at gmail.com> wrote: