0
$\begingroup$

When it comes to predicting the number of days or intervals, I read that Poisson regression produces a better fit. In some papers, they also made reference to negative binomial.

Can someone suggest which gives better results and in what situations?

Use case: Trying to predict the "number of days for Harvest"

$\endgroup$
  • $\begingroup$ I think that "which gives better results" is likely to be much too broad for our site - essentially any number of answers can make up a criterion for what thing to do better and locate some applications and point to some particular individual experience or other, so there could be dozens of valid but different answers -- that's not okay, it turns it into a big list of responses and invites anecdotes. You'll need to ask something considerably more specific $\endgroup$ – Glen_b -Reinstate Monica Jul 30 '17 at 4:56
  • $\begingroup$ While "number of days" looks like a count I think you're actually intending to refer to a discretized interval of time. A Poisson or negative binomial model may by happenstance be quite adequate in some situations but as a first attempt I'd be thinking more about something like an exponential or gamma model (noting that discretizing an exponential would correspond to a geometric, which is a particular case of the negative binomial, so it may work okay across a variety of situations). $\endgroup$ – Glen_b -Reinstate Monica Jul 30 '17 at 4:56
  • $\begingroup$ Hi Glen_b, If the question was more broader, that's due to my limited knowledge on the subject. linear regression equation didn't produce satisfactory result with the test data. I will try out the options and post the results in a day or two. Anyway, many thanks for the detailed explanation. $\endgroup$ – Nat_G Aug 2 '17 at 15:38
2
$\begingroup$

(While this isn't as complete an answer as I'd really like to see here, this should have some answer, so I'm pulling my brief comment down to an answer.)

While "number of days" looks like a count I think you're actually intending to refer to a discretized interval of time. A Poisson or negative binomial model may by happenstance be quite adequate in some situations but as a first attempt I'd be thinking more about something like an exponential or gamma model (noting that discretizing an exponential would correspond to a geometric, which is a particular case of the negative binomial, so it may work okay across a variety of situations).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.