[R-sig-ME] Generalized Linear Models

[Ben Bolker]

On 14-01-14 11:24 AM, Agnes Schneider wrote:
> Hi,
> 
> I'm carrying out an analysis on future time expressions in English on
> the basis of a corpora of spoken language. I am using a linear mixed
> effects model (glmer) because when I coded the data I realized that
> there is considerable variability within each conversation concerning
> the choice of future markers. So I have a model consisting of a
> categorical dependent variable (Future time marker WILL or BE GOING
> TO), a number of fixed effects (syntactic, semantic and
> extralinguistic variables), an a random effect which is File. Not all
> of my independent variables show a significant effect on the choice
> of future marker. My questions now are:
> 
> 1. Is the procedure at arriving at a minimal adequate model the same
> as for logistic regression models (glm)? 

  More or less, although there is some debate as to whether one should
try to discard non-significant random-effects terms or not: see Barr et
al 2013 (ref below), and whether one should consider fixed or random
effects first (I am personally uneasy with the concept of "minimal
adequate models" for confirmatory testing in the first place).

2. How do I find out whether
> there is reason to assume overdispersion? 

  See http://glmm.wikidot.com/faq#overdispersion_est and following (note
that overdispersion is unidentifiable in the case of binary data with
unique predictors, and already taken into account in models with
estimated scale parameters [Gaussian, gamma, etc.])

3. How do I find out
> whether my models (both the initial and the final model) have
> predictive power? 

  Don't know exactly what you mean here.  You could look at
http://glmm.wikidot.com/faq#rsquared

4. How do I determine whether interspeaker
> variability (File) is stronger than the fixed effects?

  As a first cut, comparing the magnitude of the standard deviation
estimate to the size of the fixed effects should do (assuming that the
fixed effect predictors are appropriately scaled).  Beyond that, it
would depend exactly what you mean.

> 
> I am grateful for any comment on my questions!! Thanks Agnes
> 
> _______________________________________________ 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> 


@article{barr_random_2013,
	title = {Random effects structure for confirmatory hypothesis testing:
Keep it maximal},
	volume = {68},
	issn = {{0749596X}},
	shorttitle = {Random effects structure for confirmatory hypothesis
testing},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S0749596X12001180},
	doi = {10.1016/j.jml.2012.11.001},
	number = {3},
	urldate = {2013-06-04},
	journal = {Journal of Memory and Language},
	author = {Barr, Dale J. and Levy, Roger and Scheepers, Christoph and
Tily, Harry J.},
	month = apr,
	year = {2013},
	pages = {255--278},
}