Normal and negative binomial distributions with aster

[MatM]

Hi aster users,

I have several questions, related to fitting some distributions in aster models.

1) If I want to use a normal distribution with for the last node of my graph, I need to provide a standard deviation. Is it necessary, as for the negative binomial distributions, to write an ad-hoc function to obtain the correct log likelihood from an aster fit before constructing the likelihood profile? I could not find a TR that would answer clarify this point.

2) From what I understand from the exampled, if I chose to use a negative binomial distribution for a given node, I can find the size that maximizes the likelihood for a given model (the full model), and I will keep using this size when I fit other models to test factors with LRT tests. This makes the approximation that size will not vary much. Should I re-optimize the size for each individual model? But if so, the full model and the simplified model would not be considered nested. or would they?

Thanks a lot for your insights,

[geyer]

It is not clear that you want to follow the examples in the tech reports and try to get a maximum likelihood estimate of the shape parameter for negative binomial or truncated negative binomial or for the variance for normal.  The R function glm just uses method of moments to estimate scale for quasi-likelihood models and perhaps we should do the same.  John Stanton-Geddes used method of moments when he used trunctated negative binomial on his data (I'm not sure where those code examples are).  This is a lot simpler than what we did in our tech reports and is maybe more stable too.

I attach an example of using method of moments for normal location