[ADMB Users] Poisson GLMM example

[Mark Payne]

Hi,

Thanks for the replies - that answers the question nicely and makes a lot of sense - it's all there, its just not obvious (to me anyway!).

Three more questions

1. I expect there there is probably going to be a nugget effect in my data  - Am I correct in assuming that I would be better to introduce that in the separable likelihood function as well?
2. As I understand it, the likelihood expression that you use in this example is not the same as that I would get from log_density_poisson() - the "-k!" component at the end is missing. Is there any reason that I can't substitute log_density_possion in there, so that I get the full likelihood and therefore can use AIC etc to compare different models?

and the big qn:

3. In "normal" linear-modelling with spatial correlation (e.g. using lme() or gls() in R), I would look for spatial patterns in the residuals, and generate a variogram to test for autocorrelation - in particular the normalized residuals returned by the R function:
   residual(lme.mdl,type="normalized")
are supposed to be completely free of spatial-correlation (if you're doing everything correctly). However, its a bit trickier here, given that we're dealing with a GLM. The question is, in a Poisson GLMM, how do I calculate the appropriate residuals to feed into a variogram?

Best wishes,

Mark


________________________________________
Fra: H. Skaug [hskaug at gmail.com]
Sendt: 14. juni 2012 20:53
Til: Mark Payne
Cc: users at admb-project.org
Emne: Re: [ADMB Users] Poisson GLMM example

Hi,

Adding to Dave's comments:

> In particular, I am confused about the implementation of the actual model in ADMB. As I understand it, you are using the >NORMAL_PRIOR_FUNCTION to impose the condition that the random effects associated with each point are multivariate >normal, with a mean given by the predictor function, and a variance - covariance matrix that is exponentially decaying with the >distance between the variables.

Correct

>Should the normal_prior be specified as a correlation matrix or a vcov matrix?

As Dave says, we recommend to scale the correlation matrix inside the
likelihood. IN this
example this happens in " exp(_log_sigma)*ui".

>
> FUNCTION void evaluate_M(void)
>
>  get_M(a);
>
> However, at no-point can I see where the evaluate_M function is actually called..

It is being called implicitely. The manual at this point "4.5 Gaussian
priors and quadratic penalties"
is not good, and should be improved.

Hans