[R] What error distribution should I use?

[Peter Ehlers]

On 2012-03-27 15:11, Ben Bolker wrote:
> Lívia Dorneles Audino<livia.audino<at>  gmail.com>  writes:
>
>>
>> I'm trying to make a glmm to identify the relationship between insect
>> species richness with fragment size, isolation and time (different years).
>> I already tried to analyse it using poisson distribution error, but I
>> always face with the following warning:
>> *glm.fit: fitted probabilities numerically 0 or 1 occurred *
>>
>> This is probably hapenning because my dataset has a lot of zeros. So, what
>> error distribution should I use?
>>
>
>    I know you haven't gotten a lot of help on r-sig-mixed-models (sorry),
> but it would probably be better to post this question there.  The answer
> is that this is a warning, not an error, so it indicates a need for
> caution but not necessarily that anything is wrong.  In this case,
> an internal call to glm.fit() has difficulty when it tries to fit
> a subset of that data that are all-zero or all-one.  It's quite possibly
> OK, provided that you've looked at your results, plotted predicted values,
> etc., and everything seems to make sense.
>

Livia:
  You might also find this quite extensive recent post from Ted Harding
informative:

   https://stat.ethz.ch/pipermail/r-help/2012-March/307352.html

Peter Ehlers