[R-sig-ME] Unable to fit glmer with family=binomial(link=identity)

David Duffy David.Duffy at qimr.edu.au
Thu Aug 20 05:24:15 CEST 2009

On Thu, 20 Aug 2009, Ken Beath wrote:

> On 20/08/2009, at 1:58 AM, Doran, Harold wrote:
>> I may be missing something here, but why are you using an identity link 
>> with binomial outcomes? If there is a reason why, you might as well just 
>> use lmer since the identity link is for a normal distribution and that is 
>> the distributional assumption used for the errors in that function.
> There are good reasons, as mentioned it results in relative risks rather than 
> odds ratios. It can be made to work for fixed effects models, the main 
> problem is that the linear predictor must always remain within 0 and 1 or the 
> algorithm fails. For random effects models this is impossible, as the linear 
> predictor plus a normally distributed random effect must always include 
> points outside this range, so it is sensibly not allowed by the software.
> Ken

I do have (fortran) MCMC code for this model -- it rejects proposals 
leading to any predicted probabilities falling outside 0-1 at that 
iteration.  In R, perhaps MCMCglmm could be used.  Alternatively, if a 
point estimate is sufficient, then it could be produced from the logit 
link model fitted values from glmer, or I guess a Poisson model.  Finally, 
there is some literature suggesting that the Gaussian LMM for binary data 
might not be too bad, especially if the probabilities are all >0.05.

2c, David Duffy.

| David Duffy (MBBS PhD)                                         ,-_|\
| email: davidD at qimr.edu.au  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v

More information about the R-sig-mixed-models mailing list