Skip to content

Likelihood function of lme4

7 messages · Ed Merkle, Yashree Mehta, Ben Bolker

Original
Yashree Mehta
# 
Hi,

Where can I find the density function used for lmer model estimation in
lme4?
I need the log likelihood value for each observation because I have to run
a non-nested test statistic for which I have to compare the log likelihood
value of each observation with that from another model.

Regards,
Yashree
4 days later
Ed Merkle
# 
Yashree,

You might look at llcont.lmerMod() from the merDeriv package.

Ed
Yashree Mehta
# 
Ed,

This function worked perfectly. Thank you!

Regards,
Yashree

On Tue, Apr 20, 2021 at 3:33 PM Merkle, Edgar C. <merklee at missouri.edu>
wrote:

  
  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
I would also point out that Ed is the author of the nonnest2 package, 
which already implements the Vuong test (which may be what you're trying 
to do ??)

        
On 4/20/21 10:12 AM, Yashree Mehta wrote:

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Yashree Mehta
    

    

      
      #
    

  


  

      
Yes, I am trying to run the Vuong test.

The other model with which I am comparing the lme4 model belongs to the
"frontier" class of models. I did use nonnest2 but got the following error:
Error in UseMethod("llcont") :
  no applicable method for 'llcont' applied to an object of class "frontier"

So I decided to calculate the Vuong statistic by constructing it using
observation-level loglikelihoods.

Thank you.

Regards,
Yashree

  
  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
OK.  You should make sure to check that lme4 and frontier are 
computing *commensurate* log-likelihoods (e.g. that the normalization 
constants are consistently included in both packages); if possible (it 
isn't always), it's good to try to fit *identical* models in both 
package to make sure they give the same log-likelihoods. (You can't fit 
an lmer model without a random effect, but you can re-evaluate its 
likelihood with the RE variances set to zero.)

        
On 4/20/21 11:52 AM, Yashree Mehta wrote:

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Yashree Mehta
    

    

      
      #
    

  


  

      
OK. I will make sure.

        
On Tue, Apr 20, 2021 at 6:02 PM Ben Bolker <bbolker at gmail.com> wrote: