0

I have germination data in percentages (with a few zeros) and wish to fit a GLM on them to explore how different seed origins and levels of treatment affect them.

My results are not exactly 'counts', but is it ok to use Poisson family (ultimately quasi-poisson)?

I don't want to transform them to proportions in order to use Binomial. If this is the proper thing to do, what will be next: quasi-binomial or negative Binomial?

chl
  • 50,972
  • 18
  • 205
  • 364
john papas
  • 1
  • 1
    I don't quite follow your situation. If you have percentages, why would you want to use Poisson? What does it mean to "transform [percentages] to proportions"? Why don't you *want* to use the binomial dist? Why would a 'next step' be necessary? If you have data in percentages, do you know what the denominators are? If so, you can recover the number of seeds that germinated & the total number of seeds. This means that you could treat your data as binomial, which, as far as I can tell from your description, is the appropriate way to model your data. – gung - Reinstate Monica Nov 05 '12 at 21:14
  • My data are overdispersed, so I suppose I must use a quasi-likelihood. I read this [link](http://stats.stackexchange.com/questions/17918/quasibinomial-vs-negative-binomial-and-hurdles) and it seems that quasibinomial is the appropriate way but I'm not very sure. – john papas Nov 05 '12 at 22:24
  • The transformation from percentages to proportions is pretty minimal, hwy don't you want to do it? That seems right to me, just as it does to @gung . Why use Poisson (or any count model) on data that are not counts? – Peter Flom Nov 05 '12 at 22:38

0 Answers0