Skip to content

Error in MEEM(object, conLin, control$niterEM) : Singularity in backsolve at level 0, block 1 in LME model

7 messages · romunov, Arivoara Rabarijaona, Ben Bolker

Original
Arivoara Rabarijaona
# 
Hi,
I hope someone can help me. I'm using nlme to fit models.

My dataframe (1785 obs) :
$ id: Factor w/595 levels
$ treatment: Factor w/3 levels
$ provenance: Factor w/16 levels
$ repetition: Factor w/4 levels
$ bloc: Factor w/66 levels # nested to repetition
$ response: num ... (1777 obs and 8 NA) # 3 repeated measures by id
$ status: Factor w/3 levels, "dominant","codominant","suppressed" # there
are 6 provenances without suppressed trees

I want to run a modele like this :

modele <- lme(response  ~ provenance + treatment + provenance:treatment +
status + status:treatment + statuts:provenance,
                   random = ~ 1|repetition/bloc,
                   correlation = corAR1(form = ~ 1|repetition/bloc/id),
                   data, method= "ML", na.action =na.omit)

I get the message :
Error in MEEM(object, conLin, control$niterEM) :   Singularity in backsolve
at level 0, block 1 in LME model

If I run the modele without the interaction statuts:provenance, it works.

Can anyone tell me how to resolve this error ?

Thanks,
Arivoara Rabarijaona
romunov
# 
Have you tried plotting this? My guess is that you will find something
unexpected in the provenance:treatment level combination.

Cheers,
Roman

On Wed, Jul 7, 2021 at 9:03 AM Arivoara Rabarijaona <arivoara at gmail.com>
wrote:

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Arivoara Rabarijaona
    

    

      
      #
    

  


  

      
Thank you,
provenance:treatment is normal, nothing is unexpected
I think the problem is with provenance:status, but I don't know how to
resolve it.
Using lmer, I get the message: fixed-effect model matrix is rank deficient
so dropping 18 columns / coefficients

Ari

Le mer. 7 juil. 2021 ? 09:52, romunov <romunov at gmail.com> a ?crit :

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
You have constructed a model with multicollinear predictors (another 
way to put this is that your model matrix is rank-deficient).  R's 
formula interface usually takes care of discarding redundant columns, 
but when interactions are spelled out explicitly with it can't always 
manage. You might do better expressing the fixed effects component of 
the model as

(provenance + treatment + status)^2
'
As is often stated in this forum, you may have trouble fitting a random 
effect with only four levels (repetition).

   Ben Bolker

        
On 7/7/21 4:01 AM, Arivoara Rabarijaona wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Arivoara Rabarijaona
    

    

      
      #
    

  


  

      
Thank you for your explanation. I really appreciate it.
However, there is no change using nlme (Error in MEEM(object, conLin,
control$niterEM) : Singularity in backsolve at level 0, block 1 in LME
model) and lmer (fixed-effect model matrix is rank deficient so dropping 6
columns / coefficients). My problem is not resolved.
There is a problem with repetition only when I use lmer model.

Arivoara Rabarijaona

Le mer. 7 juil. 2021 ? 16:26, Ben Bolker <bbolker at gmail.com> a ?crit :

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Arivoara Rabarijaona
    

    

      
      #
    

  


  

      
I can see my data in the attached file.

Arivoara Rabarijaona

Le mer. 7 juil. 2021 ? 17:20, Arivoara Rabarijaona <arivoara at gmail.com> a
?crit :

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
If expressing the model in the R-friendliest form doesn't resolve 
the problem, then you almost certainly have a *real* multicollinearity 
problem, which in turn is almost certainly driven by combinations of 
factors that are missing from your data set (e.g. if you have provenance 
A, B, C and treatment a, b, c and the combination {provenance = A, 
treatment = a} doesn't occur in your data set, then your model matrix is 
multicollinear/unidentifiable.

   Some options:

  * fit in lme4 or another package that automatically handles 
multicollinear terms.  Looking at the mixed model comparison table 
<https://docs.google.com/spreadsheets/d/19itelYaVW0U0gtNtRfqh76ZGt1awlamNcJwT71u_5Uk/edit#gid=0>, 
if you want an AR1 model *and* automatic rank deficiency, you might need 
the INLA package (off-CRAN) ...

* You can construct the model matrix manually and drop collinear terms 
yourself: at least one example is given here: 
https://github.com/glmmTMB/glmmTMB/issues/522

* you can expand the two-way interaction manually and build a one-way model.

   I don't know how to interpret "There is a problem with repetition 
only when I use lmer model".

   cheers
    Ben Bolker

        
On 7/7/21 11:37 AM, Arivoara Rabarijaona wrote: