I would like to know what model diagnostics I should be checking to ensure that a negative binomial (NB) regression for overdispersed data has meet all of the required assumptions. There is a very helpful post here, however this post does not clarify what I should be looking for in the various plots (for example the residual plots). I would appreciate any clarification/confirmation on the following points:

1) What are the residual plots meant to look like? I understand that a negative binomial does not assume homogenous variances, so should I expect patterns in the residual plots?

2) Is it assumed that the residuals of a negative binomial model are normal? I have read conflicting information.

3) Do you check for influential points in the same way as for a ordinary linear regression (for example by using Cooks, dfbetas etc)?

4) I assume that muliticollinearity can still be an issue in a NB regression and should be checked. Is this correct?

5) Are there any other assumptions/diagnostics that I have missed?

I am working with R.

  • $\begingroup$ Negative binomial is count data, search this site for count data modeling $\endgroup$ Sep 17, 2017 at 14:17