Skip to content
Prev 298813 / 398503 Next

Significance of interaction depends on factor reference level - lmer/AIC model averaging

Andy Robertson <ar313 <at> exeter.ac.uk> writes:
This is all a little overwhelming.  I appreciate that you
are trying to be thorough, but there's an awful lot to look at here ...
I will give comments until the point where I ran out of time.
Note that you have Location in your model twice, once as a fixed
effect and once as a random effect.  This is bound to lead to trouble.
If you use (1|Location1:Socialgroup) and (1|Location1:Socialgroup:Tattoo)
you will get the random effects you want without also incorporating
a random effect of Location1.

  You could specify the fixed effects as
(AGECAT2+Sex+Location1)^2 if you wanted (it would be equivalent
to this specification).
is this from the 'rockchalk' package?  Do you know that it
isn't doing something funny?
What is c.Sex?  "centered sex" (e.g -1 for males and +1 for females?

  In general I think it is a bad idea to model-average sets of models
some of which contain interactions, because (unless the design is
perfectly balanced and the contrasts are set to sum-to-zero contrasts),
the meaning of the main effects changes between models.  In a model
with an interaction (assuming sum-to-zero contrasts), the main effect
represents the average effect across groups using equal weights:
for example the main effect of sex would be the mean of the
male and female predictions.  In the model without an interaction,
the main effect of groups will represent the average across groups
weighting by the number of individuals per group ...
In general you should not test terms involving categorical variables
(e.g. sex:location) by looking at all of the individual parameter
z-values, but by comparing models with and without the term.
This gets harder when you are doing model averaging. In general
I would say that model averaging and information-theoretic approaches
in general are best for *prediction*, while good old-fashioned
frequentist approaches are best for *hypothesis testing*, which
seems to be what you are trying to do ...

 Also note that the summary is giving you the results of Z-tests,
which do not take the finite size of the data set into account.
Note here that you're getting zero variances for the location
and year variances, and almost identical variances for the
other two random effects (which looks a little fishy to me,
but I can't quite say that it's wrong).
Trying to fit a 4-level or even more extremely a 2-level factor
as a random effect is almost guaranteed to give you zero variance
estimates.  I would strongly consider fitting Location and Year
as fixed effects (you can still include social group within 
location and individual within social group as random effects).
(See point above about how to exclude Location from the random
effects.)
???  Do you mean an effect of sex in location C? I don't see where
the interaction with age comes in ...

 Also note that you seem to have changed from "c.Sex" (a continuous
variable, according to the model summary) to "Sex" (a factor with
"Female" as the first level and "Male" as the second).  Is that
responsible for the differences you are seeing?
stopped here ...

  In general it's not surprising that the apparent effect measured
in the way you have parameterized and are measuring it changes
with parameterization.  The parameters mean different things and
are using a different baseline ...

  A lot of this is basic (although not easy) stuff about
parameterization.