Skip to content

Convergence warning message

9 messages · Christopher David Desjardins, Jackie Wood, Thierry Onkelinx +1 more

Original
Christopher David Desjardins
# 
I am trying to fit a mixed effects binomial model.

The data consists of
- A dependent variable consisting of Bernoulli trials (outcome)
- A time variable (time), which has been mean centered
- An id variable (id)
- A categorical covariate (cat_cov)
- A blocking variable (block) which id is nested in. I realize in the model
below that it should be (1 | id/block) but I am just trying to troubleshoot
my problem at the moment.

When I run the following:

example_data <- read.csv("https://cddesja.github.io/example_data.csv",
header  = T)
example_data$cat_cov <- as.factor(example_data$cat_cov)
example_data$id <- as.factor(example_data$id)
example_data$block <- as.factor(example_data$block)
main_effects <- glmer(outcome ~ 1 + cat_cov + time + I(time^2) + (1 | id),
data = example_data, family = "binomial")

That last line of code gives a warning message:
id), data = example_data, family = "binomial")
Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge with max|grad| = 4.36001 (tol = 0.001, component
1)
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?

I am not exactly sure how to proceed. I know the issue is with cat_cov,
though it's unclear to me why. If I swap out in a different categorical
covariate in the model, not included in that data set, I don't get this
message. I am not running into complete separation with cat_cov.  So, I'm a
little perplexed.

Any advice on what I should do or something I could look at it would be
very helpful.

Thanks,
Chris
Jackie Wood
# 
Hi Chris,

Try checking ?convergence....coincidentally, I was having a similar problem
just yesterday. There are some step by step
instructions for trouble shooting/double checking convergence warnings. For
example, a bit of example code is provided to run your model using a number
of different optimizers. If all optimizers yield similar values, it's
possible that you could be getting false convergence warnings. I'm not sure
if that's the case with your data, but it might be a place to start!

Jacquelyn

On Wed, Mar 16, 2016 at 1:56 PM, Christopher David Desjardins <
cddesjardins at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
Good question.

   I'm afraid that for data sets ~ 100,000 observations or bigger, our 
convergence calculations aren't terribly reliable -- see e.g. the third 
set of figures under https://rpubs.com/bbolker/lme4_convergence ... I 
would follow Jackie's advice ...

        
On 16-03-16 02:24 PM, Jackie Wood wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Christopher David Desjardins
    

    

      
      #
    

  


  

      
Thanks, Jacquelyn and Ben. Jacquelyn, did you mean to attach some code or
just reference the site that Ben did? I had seen Ben's comments on
StackOverflow about potential false convergence messages, so I'll dig a bit
deeper. I just wanted to make sure it wasn't something obvious that I had
overlooked first.

            

      
      
      
    

        
Is that correct?

Chris

        
On Wed, Mar 16, 2016 at 2:35 PM, Ben Bolker <bbolker at gmail.com> wrote:

        

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Jackie Wood
    

    

      
      #
    

  


  

      
Hi Chris,

You can find the example code I was talking about here if you haven't
tracked it down already:

http://127.0.0.1:29918/library/lme4/html/convergence.html

Jackie



On Wed, Mar 16, 2016 at 3:44 PM, Christopher David Desjardins <

        
cddesjardins at gmail.com> wrote:

        

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Thierry Onkelinx
    

    

      
      #
    

  


  

      
Dear Jackie,

127.0.01 points to localhost, which will work only on your computer.

Best regards,
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature
and Forest
team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
Kliniekstraat 25
1070 Anderlecht
Belgium

To call in the statistician after the experiment is done may be no
more than asking him to perform a post-mortem examination: he may be
able to say what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does
not ensure that a reasonable answer can be extracted from a given body
of data. ~ John Tukey


2016-03-16 20:56 GMT+01:00 Jackie Wood <jackiewood7 at gmail.com>:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Christopher David Desjardins
    

    

      
      #
    

  


  

      
Thanks, I'll read through the convergence help page in the lme4 package.
Chris

On Wed, Mar 16, 2016 at 3:12 PM, Thierry Onkelinx <thierry.onkelinx at inbo.be>
wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
https://github.com/lme4/lme4/blob/master/man/convergence.Rd

On Wed, Mar 16, 2016 at 4:12 PM, Thierry Onkelinx

        
<thierry.onkelinx at inbo.be> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Ben Bolker
    

    

      
      #
    

  


  

      
To answer this particular question:

   The importance of accuracy in the approximation of the integral over 
the random effects (for which glmmPQL < Laplace < adaptive Gauss-Hermite 
quadrature) depends on the degree to which the sampling distribution of 
the conditional modes (approx "BLUPs", i.e. the values associated with 
deviations of particular grouping factors from the population average) 
are Gaussian.  In particular, if you have Bernoulli responses and small 
numbers of samples per group (which seems to be the case here -- 1 to 4 
observations per individual if I read your data correctly) then yes, 
this will be a concern ...

        
On 16-03-16 04:33 PM, Christopher David Desjardins wrote: