06-19-2018 01:10 AM
I am constructing a model of a continuous outcome and mixed predictors using negative binomial regression due to overdispersion. However, I am having trouble figuring out how to compare my various potential models in SAS. Can I use deviance and pearson coefficient as well as the corresponding p-value to compare between different models? For example, the higher the p-value the more adequate the fit? I don't think that makes sense but just trying to think.
I've read a few places that I can do cross validation but is this the only way??
Thanks for any help!
06-20-2018 08:51 AM
If all of your models are negative binomial models and all can be considered nested in (subcases of) a single full model, then you can fit the full model and use CONTRAST statements to test each of your simpler candidate models. See Example 4 of this note. Whether you do this or use AIC or BIC, be sure that the exact same set of observations are used in each model fit. This is often not the case if some of your predictors contain missing values.