Skip to content

UNSOLVED: Re: lme4 or other open source mixed model package code equivalent to asreml-R

6 messages · John Clark, David Duffy, Ben Bolker

Original
David Duffy
#
On Thu, 17 Nov 2011, John Clark wrote:

            

      
      
      
    

        
You might look at the regress and spatialCovariance packages of David 
Clifford.  The former allows fitting of (Gaussian) mixed models where you 
can specify the structure of the covariance matrix very flexibly (for 
example, I have used it for pedigree data).  The spatialCovariance package 
uses regress to provide more elaborate models than AR1xAR1, but may be 
applicable.  You may have to correspond with the author about applying it 
to your exact problem.

Just 2c, David Duffy.

  
  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
David Duffy <David.Duffy at ...> writes:

            

      
      
      
    

        
http://stackoverflow.com/questions/7961864/lme4-or-other-open-source-r-package-code-equivalent-to-asreml-r

            

      
      
      
    

        
Hmm.  I took a quick look; it is nice to see another implementation
of mixed effects models (you can never have too many, especially when
they're open and can build on each other ...) -- BUT -- it's not 
immediately obvious to me (maybe this is the "correspond with the author"
part?) how to construct this problem so that the variance structure
corresponds to a sum of specified Gaussian values.  In particular, would
we have to use an outer loop to profile over different values of the
scale parameter (or run a 1-dimensional minimum-finder for the
negative log-likelihood) ?

  
  


  


      

    

    

      
      

        


  

        

      David Duffy
    

    

      
      #
    

  


  

      
On Fri, 18 Nov 2011, Ben Bolker wrote:

        

            

      
      
      
    

        
Well, I do not have any experience carrying out spatial modelling, but it 
did seem to me that the covariances arising from the separable AR1xAR1 
model must be have a close resemblance to some of the other standard 
models.  Specifically, page 85 of Haskard's thesis

http://digital.library.adelaide.edu.au/dspace/bitstream/2440/47972/1/02whole.pdf

seems to imply it is equivalent to a anisotropic Matern model.  This is 
offered by the spatialCovariance package, AIUI.

It took me a while to understand that the example dataset is the Wheat2 
data from the nmle package, which Pinheiro and Bates (2000, P 263) analyse 
using spherical and rational quadratic models.  I am unsure if the OP is 
particularly interested in the exact models he chose for his example, or 
whether one can generally match ASREML's functionality, or has data that 
might be comfortably analysed usibg existing lme correlation structures.

Cheers, David.

  
    

      
      
    

  


  


      

    

    
1 day later

    

      
      

        


  

        

      David Duffy
    

    

      
      #
    

  


  

      
On: Chen et al Genet Epidemiol 2011, 35:650-7.

The latest (dead tree) issue of Genetic Epidemiology has a paper using
simulated and real data to compare methods for testing association
between a measured genotype (fixed effect) and a dichotomous outcome in
pedigrees, so there is residual correlation between observations. They use
a) "ordinary" gaussian linear mixed model treating the trait as 0-1 (in
lmekin) b) the binomial-gaussian GLMM using glmer (0.999375-32) c) GEE
in geepack.  Simulated data were produced under a threshold model and
AFAICT [I don't think the paper well-written], a Wald test was used to
assess the fixed effect for all three.

You can read the abstract, at least, online: they prefer GEE.  Their GLMM 
test Type-1 error tends to drift up a little as the trait prevalence 
increases.  They also experienced problems with GLMM when carrying small 
sample simulations.  They did encounter numerical problems with GEE when 
the trair prevalence was low, but for this situation they preferred the
gaussian LMM, as they found this to have OK Type-I error rates, and 
better power than the GLMM (though twice as slow ;)).

The main weakness of course if that they did not report LRTS results, 
although they do mention Hauck-Donner effects as a possible cause of their 
problems.  Another possible one is the generating model, which is 
convenient but different from the logistic-gaussian.  And fitting the LMM 
to binary variables does usually give correct Type I errors, but when a 
true effect is present overestimates the evidence for association in my 
experience.

Cheers, David Duffy.

  
    

      
      
    

  


  


      

    

    
3 days later

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
David Duffy <David.Duffy at ...> writes:

            

      
      
      
    

        
[snip]

            

      
      
      
    

        
Dave Fournier has posted a solution using AD Model Builder at

http://groups.google.com/group/admb-users/browse_thread/
   thread/96bf76c9f578650b

(sorry, URL broken to make Gmane happy) that tackles (almost) this problem (it
actually
does a spatial GLMM -- a Poisson model layered on the AR1 x AR1
Gaussian random field).

  Ben Bolker