Negative binomial model + freqencies

[David Nash]

Dear Evan,

Thanks for the new release - I was particularly excited to see that you have included an option for "negative binomial" models - but have so far not been able to get tis to work (I always get a message saying "an error occurred while running the regression"). 

What I am trying to do is compare the number of a parasites in ant colonies of three different species, which can (within-species) be described reasonably well by a negative-binomial distribution, as most ant colonies have zero parasites, some have one, and a very small number have many (up to 100). I wonder if this is because Wizzard is not accepting zero values for the variable, since the analysis does run if I add one to the value of every parasite count (but that the, of course, changes the means and would no longer correspond to a negative inomial distribution).

Since the negative binomial is a two-parameter distribution, I am also wondering how I can see the (estimated?) dispersion parameter - It looks as though only the means are given in the model interface, the prediction models and XLS output - Or maybe I'm missing something?

Finally, for this type of model, it would be very useful to be able to use a frequency variable to specify how often a particular outcome was observed, rather than having to have one line for each case.

 - David

[Evan Miller]

Hi David,

Wizard ought to accept zero values; if you're able to send me the file
off-list, I can dig into this issue some more and try to figure out
why the regression is failing. Nonlinear regression involves as much
art as science, and sometimes parameters get twisted up in the
iterative procedure.

Currently the dispersion parameter is not displayed or exported, but I
can add this feature to a future version. Now that I think about it,
the dispersion parameter will be useful for displaying a CDF for the
prediction rather than just the mean prediction.

It is possible to use a frequency variable in conjunction with
negative binomial models (or any model for that matter) -- just use
Table > Apply Frequency Weight and select a numeric column with
integer values representing the counts.

I'm glad you like the new version, and I hope we can get these issues
worked out soon!

Evan

> --
> You received this message because you are subscribed to the Google Groups
> "Wizard User Group" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to wizard-users...@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.
>
>



--
Evan Miller
http://www.evanmiller.org/

[Devon McDonald]

Dear Evan,

I haven't been able to get a negative binomial model to work in Wizard, with the same error message appearing. Downloaded Wizard 2 to try but receiving the same error.

Any suggestions?

Also, separately, is it possible to use a log binomial model in Wizard to estimate risk ratios for a common (not rare as in logistic regression) dichotomous outcome?

Thanks,

Devon