Skip to content

2 level logit, 2 REs, large sample - log likelihood returns "NaN"

4 messages · Daniel Adkins, Ben Pelzer, Dimitris Rizopoulos

Daniel Adkins
# 
Hi,
I am fitting a large (j=50K, i=9K) 2-level logit with random intercept
and age slope and 14 covariates. Model estimates become stable at
nAGQ>=150 (large, I know). Based on simpler models (random intercept
only, random
slope only, ordinary logit, etc) the solution looks sound. However,
all the fit indices return a value of "NaN", which naturally stands
for "not a number". Why is this? This model should yield a scalar log
likelihood, no? Any advice would be appreciated.

Thanks,
Daniel
Daniel Adkins
# 
To clarify, the models were fit using the glmer cmd of the lme4
package. Specifically, the model with scripted as:

 proto <- glmer(hibpe ~ age + b + b_age + h + h_age + female +
female_age + bxf + hxf + numwaves + dead + nodoctor + nohosp
	+ (age| hhidpn),  nAGQ =150, family=binomial, data=hrs_data,
na.action =na.omit, verbose=TRUE)


Best,
Daniel
On Fri, Jun 10, 2011 at 4:08 AM, Daniel Adkins <deadkins at vcu.edu> wrote:

  
    

      
      
    

  


  


      

    

    
4 days later

    

      
      

        


  

        

      Ben Pelzer
    

    

      
      #
    

  


  

      
Hi all,

Is there a simple method to obtain the predicted probabilities based on 
the FIXED part only of a random intercept logistic regression model 
using lmer?
Thanks for any help!

Ben.

  
  


  


      

    

    

      
      

        


  

        

      Dimitris Rizopoulos
    

    

      
      #
    

  


  

      
Well it depends on what exactly do you mean by 'FIXED part only'. If you 
mean the predicted probabilities for the "average" subject (i.e., a 
subject with random effect value zero), then you can simply obtain these 
probabilities from the estimated coefficients of the model.

However, if you mean the marginal/population probabilities, i.e., the 
probabilities averaged over the subjects, then you in fact need,

P(y = 1) = \int P(y = 1 | b) f(b) db,

where f(b) is the normal density for the random effects. One option to 
get those is using sampling with a bit of extra programming.


I hope it helps.

Best,
Dimitris

        
On 6/15/2011 2:47 PM, Ben Pelzer wrote: