[R-sig-ME] Post-hoc test and Cook's distance vs. leverageplot for glmer from package lme4

[Ben Bolker]

Mario Garrido <gaiarrido at ...> writes:

> 
> Hi,
> 
> My name's Mario Garrido, a postdoctoral student in Biology. I am relatively
> new with r and despite I find it brilliant I am having some difficulties in
> finding some functions and interpreting the syntaxes.
> 
> At the moment I am working on a GLZM model which fits a Poisson
> distribution. I am having some problems with two issues
> 
> 1.     after calculating Akaike weights, my best model is one including a
> 4-way interaction term. It is the following:
> 
> model7<-glmer(active~ treatment*daytype*time*age+(1|
> indiv),family=poisson(link=log),nAGQ=1)
> 
> Is there any function or package to perform a Post hoc test to know which
> subset of 2- and 3-way interaction terms have more influence on the model?

  It strikes me that this is a somewhat difficult question 
conceptually, as well as computationally.  How are you dealing
with the issues of marginality?  (See Venables "Exegeses on linear
models", available by internet-searching, for a discussion of
marginality ...)  In other words, how do you define what a
2- or 3-way interaction term means?

 *If* you can define what you mean (e.g. if simply setting the
parameters related to a specific lower-level interaction to zero
makes biological or scientific sense), then you could drop the
terms and look at the difference in AIC or log-likelihood, and
use some sort of multiple comparisons to deal with the post-hocness
of it all.

> 
> 2.     In addition, I find how to compute Cook's distance both for the
> function glmer and for the function lmer using the package influence.ME.
> This package also allow to make some graphs but, is there any
> package or function to do it and obtain a plot similar to this:
> https://climateaudit.files.wordpress.com/2012/09/lew_cooks-distance1.png?
> 

  Have you found a way to compute (or define) leverage for a GLMM?
(Maybe that's what you're asking for.) 

  Ben Bolker