Skip to content

On what basis one of the predictors/fixed factors should be included in a random slope model?

3 messages · Saudi Sadiq, Alex Fine

Original
Saudi Sadiq
# 
I read this paper  'Random effects structure for confirmatory hypothesis
testing: Keep it maximal' by Barr, Levy, Scheepers and Tily (2013) and
enjoyed it. I tried, then, to keep it maximal with a dataset I am analysing
but I am facing a problem (this may look naive). I hope you will help me
solve it. The dataset includes 9 predictors and I am using R. The model I
ran is:

modle1<- glmer(convergence ~ gender + age.group
+education+residence+sound_before+sound_after+ part.of.speech +
grammatical.gender + style + (1|speaker) + (1|item), data = qaaf,
family='binomial', control = glmerControl(optimizer = "bobyqa"), nAGQ = 1)

As you see, I have two random factors (speaker and item).
Q1: On what basis one of the predictors/fixed factors should be included in
a random slope model?
The model with random slopes could be like model2 or model3:

modle2<- glmer(convergence ~ gender + age.group
+education+residence+sound_before+sound_after+ part.of.speech
+ grammatical.gender + style + (1+gender|speaker) + (1+education|item),
data = qaaf,  family='binomial', control = glmerControl(optimizer =
"bobyqa"), nAGQ = 1)

modle3<- glmer(convergence ~ gender + age.group
+education+residence+sound_before+sound_after+ part.of.speech +
grammatical.gender + style + (1+residence|speaker) + (1+residence|item),
data = qaaf,  family='binomial', control = glmerControl(optimizer =
"bobyqa"), nAGQ = 1)

Model2 has gender and education included in the random slope and model3 has
residence alone.
Q2: Should the fixed factor used in a random slope model be the same (as in
model3)?
Q3: If  it is okay for such a model to have  two fixed factors (as in model2)
or just one (as inmodel3), which is better?
Q4: What should be done if the model does not converge with a random slope,
is there a way to make it work? Or can i trus and report the results of the
model even if there is a warning message?

Best regards
Alex Fine
# 
You may want to check out the response to that paper, which was pretty well
received:  http://arxiv.org/abs/1506.04967

Somewhat unconventionally, here is a facebook thread where a subset of the
authors of each paper talked about it:
https://www.facebook.com/alex.b.fine.9/posts/10102858621451518

A couple quick notes:

1.  I think the basic consensus is that, when using LMMs for hypothesis
testing, you must use the random effects to do what virtually every other
framework for hypothesis testing does, which is account for sources of
variation that may lead to Type I error.  So if "gender" is an experimental
manipulation that might vary depending on item, you should include it.

2.  Does it make sense to have a by-speaker random slope for gender?  This
implies that a given speaker could have either gender.  Without getting too
far afield, I doubt this was indeed the case in your design.  In the
terminology of the Barr et al. paper, a by-speaker random slope for gender
is not "justified by the design".

3.  Do not trust or report a model with an error warning.
On Sat, Apr 9, 2016 at 11:35 AM, Saudi Sadiq <ss1272 at york.ac.uk> wrote:

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    
1 day later

    

      
      

        


  

        

      Saudi Sadiq
    

    

      
      #
    

  


  

      
Hi Alex,
Many thanks for your reply and the historical FB thread.
Best

        
On 9 April 2016 at 17:09, Alex Fine <abfine at gmail.com> wrote: