Re: [R] glmmADMB: Generalized Linear Mixed Models using AD Model Builder

From: Spencer Graves <>
Date: Wed 21 Dec 2005 - 03:49:48 EST

          I get upset when software dies and refuses to give me an answer. I'd much rather have a routine give me a wrong answer -- with an error message -- than just an error message. Maybe refuse to print standard errors when the hessian is singular, but at least give me a progress report with the singular hessian. Without that, I have to program "optim" or something else separately to get the answers and the hessian in order to do my own diagnosis -- if I know enough to do that.

	  Just my 0.02 Euros.
	  spencer graves

Roel de Jong wrote:

> Of course it is generally possible to generate datasets for a perfectly 
> well-defined model that are hard to fit, but in this particular case I 
> feel it should be possible. In my observations, glmm.admb is far more 
> numerically stable fitting GLMM's than other software I've seen. Further 
> , I don't think the data I generated come from a model that is 
> overparameterized, severely contaminated with outliers, has no noise, or 
> is nonlinear. But I encourage anyone to run a simulation study with 
> generated data they think are acceptable and compare the robustness of 
> several methods. I leave it at this.
> Best regards,
> 	Roel de Jong
> Berton Gunter wrote:

>>May I interject a comment?
>>>When data is generated from a specified model with reasonable
>>>values, it should be possible to fit such a model successful,
>>>or is this
>>>me being stupid?
>>Let me take a turn at being stupid. Why should this be true? That is, why
>>should it be possible to easily fit a model that is generated ( i.e. using a
>>pseudo-random number generator) from a perfectly well-defined model? For
>>example, I can easily generate simple linear models contaminated with
>>outliers that are quite difficult to fit (e.g. via resistant fitting
>>methods). In nonlinear fitting, it is quite easy to generate data from
>>oevrparameterized models that are quite difficult to fit or whose fit is
>>very sensitive to initial conditions. Remember: the design (for the
>>covariates) at which you fit the data must support the parameterization.
>>The most dramatic examples are probably of simple nonlinear model systems
>>with no noise which produce chaotic results when parameters are in certain
>>ranges. These would be totally impossible to recover from the "data."
>>So I repeat: just because you can generate data from a simple model, why
>>should it be easy to fit the data and recover the model?
>>Bert Gunter
> ______________________________________________
> mailing list
> PLEASE do read the posting guide!

Spencer Graves, PhD
Senior Development Engineer
PDF Solutions, Inc.
333 West San Carlos Street Suite 700
San Jose, CA 95110, USA <>
Tel:  408-938-4420
Fax: 408-280-7915

______________________________________________ mailing list
PLEASE do read the posting guide!
Received on Wed Dec 21 20:06:34 2005

This archive was generated by hypermail 2.1.8 : Wed 21 Dec 2005 - 23:48:31 EST