1

I am trying to model a count data and it is suggested that it is better using Poisson regression rather than linear model. I am puzzled that when I compare the coefficient's significance between linear and poisson regression, the significance is higher in linear. Does it mean that linear regression is better than poisson regression?

Or it is matter of dispersion of the data? Since the requirement for poisson is that it should be equidispersed (mean=variance).

How can I check the dispersion of data using the output of the model? and any suggestion to improve the logistic model?

I am really new in logistic regression and any help will be much appreciated

kjetil b halvorsen
  • 77,844
danang crysnanto
  • 41
  • The interpretation of the parameters are different between the two models! See my answer here: http://stats.stackexchange.com/questions/142338/goodness-of-fit-and-which-model-to-choose-linear-regression-or-poisson/142353#142353 – kjetil b halvorsen Nov 16 '16 at 22:23

1 Answers1

1

Since count data is at best approximately normal, major differences between a linear model and a count model likely just points to the linear model being inappropriate even as an approximation. Often a picking a model that accounts for overdispersion (e.g. negative binomial) is a good choice and avoids worries about unaccounted for overdispersion.

Björn
  • 32,022