- Linear regression is not advised for count outcomes
- Limited options for interpretation
- Incorrect standard errors / significance tests
- The Poisson is rarely the best fit to the data
- The standard errors can sometimes be drastically off
- Start with negative binomial and skip Poisson since you have to estimate the negative binomial model to test if Poisson is appropriate anyway
- Consider what it is you want to describe
- Predicted rates are almost always of interest
- Predicted probabilities of specific outcomes or of a range of outcomes are also often of interest
- Remember that the models are nonlinear and interactive
- Sometimes effects (especially for the rate) will look relatively linear—but not always
- Zero inflated models do a good job of dealing with “excess” zeros
- These can be very theoretically interesting models, allowing you to model two processes simultaneously
- There are good and reliable statistical tests for choosing the best fitting model
- Consider the balance of parsimony and fit; BIC does this nicely
- countfit / countfit() is useful and saves you a lot of time—performing all of the reasonable tests for you automatically
Back to top