2

I currently have a GLM with a poisson family but my dependent variable is not an integer. Although R gives results I get warnings. So I tried with the quasi poisson family and this did not give any warnings. Therefore I am wondering if I can use the quasi poisson to resolve the problem of the non integer dependent variable ( I could only found that it correct for overdispersion)? If quasi poisson is the solution is it possible to find a AIC value as it now gives NA?

Thank you.

jondow0404
  • 21
  • 5
  • 1
    Not the best, poisson is meant for counts or rate data. You are almost modeling it like a rate now, with equal weights. What exactly is your dependent variable? – StupidWolf May 11 '21 at 07:20
  • https://stats.stackexchange.com/a/7076/203662 – polkas May 11 '21 at 07:36
  • "Runs without warnings or errors" is not the same as "gives the correct answer". – Limey May 11 '21 at 08:40
  • I understand that it is still maybe not correct, that is why I try to understand why it does not give warnings and if it then can resolve this problem – jondow0404 May 11 '21 at 09:15

1 Answers1

0

You can use glm with quasipoisson family with (nonnegative) fractional data, but first ask yourself how the fractions arose. Maybe they arose from dividing counts by some area, or population size, or ... to get a rate, in that case look at Poisson rate regression, that is, model the count with an offset (which is log of the ratio denominator).

The Poisson distribution has variance equal mean, but there is no reason that should hold for fractional data. So formally it makes sense to treat it as over (or under)-dispersion. As always run your diagnostics.

kjetil b halvorsen
  • 1,206
  • 2
  • 18
  • 28