Hi all,
I am trying to decide the regression model between Poisson and Negative Binomial.
So far, what I know is that with the presence of overdispersion Negative Binomial is more appropriate than Poisson.
But, the problem is that the over- or under-dispersion is depending on what type of dependent variable I want to use.
My raw data is a count-value: the number of car accidents
But, some researchers normalize the number of car accidents with the number of hours of driving: (# of car accidents/# hours driving)
And, I do think it make sense. Now, again, some of them use it as a percentage and the other use it as just a ratio.
As far as I know, we can also use Poisson not only for a count-value but a continuous value as well. So, my problems are:
1. If I use the dependent variable as a percentage ((# of car accidents/# hours driving)*100), I do have over-dispersion. Should I use Negative Binomial?
2. If I use the dependent variable as a ratio (# of car accidents/# hours driving), I do have under-dispersion. Should I use Poisson?
3. If I use "exposure (# hours driving)", does Poisson or Negative Binomial consider the dependent variable as "Count-value" or "Ratio"?
Could you please share your opinion?
Thank you so much in advance.
I am trying to decide the regression model between Poisson and Negative Binomial.
So far, what I know is that with the presence of overdispersion Negative Binomial is more appropriate than Poisson.
But, the problem is that the over- or under-dispersion is depending on what type of dependent variable I want to use.
My raw data is a count-value: the number of car accidents
But, some researchers normalize the number of car accidents with the number of hours of driving: (# of car accidents/# hours driving)
And, I do think it make sense. Now, again, some of them use it as a percentage and the other use it as just a ratio.
As far as I know, we can also use Poisson not only for a count-value but a continuous value as well. So, my problems are:
1. If I use the dependent variable as a percentage ((# of car accidents/# hours driving)*100), I do have over-dispersion. Should I use Negative Binomial?
2. If I use the dependent variable as a ratio (# of car accidents/# hours driving), I do have under-dispersion. Should I use Poisson?
3. If I use "exposure (# hours driving)", does Poisson or Negative Binomial consider the dependent variable as "Count-value" or "Ratio"?
Could you please share your opinion?
Thank you so much in advance.
