Zero inflated models

[Leah Jones]

I have a question concerning zero inflated negative binomial models. I have a model I'm running, If I use negative binomial I have significance on several key variables. I tried running a zero inflated model. All goodness of fit shows a slightly better fit, however, absolutely nothing is significant. If I check marginal probabilities (Long's SPost program), I have the same significances as I did under the regular neg bin model. Can these marginal probabilites be trusted?

[Richard Williams]

Showing code and output could help. Use code tags. See pt 12 of the FAQ.

[Richard Williams]

Also see

http://m.statisticalhorizons.com/?re...nflated-models

[Leah Jones]

Also see

http://m.statisticalhorizons.com/?re...nflated-models

Showing code and output could help. Use code tags. See pt 12 of the FAQ.

Thank you for your response. I left my flash drive at the office, I'll post code tomorrow if I can get in the building. Also thank you for the link, I'll give it a read.

[Ronán Conroy]

106
Rich's link leads to an interesting blog post entitled "Do we really need zero-inflated models?"
From a statistical point of view, the author argues that we don't need them in most cases. However, the post ignores the fundamental principle of modelling, which is that the model must align with the data generation process. Zero inflated models model a situation in which data are generated by two processes (so do two-part models, and they reflect a conceptually different process at work). I would argue against choosing a model on purely statistical grounds. To do so can lead to a failure to model processes that are of interest.

[Richard Williams]

In fairness, Allison does say "Of course, there are certainly situations where a zero-inflated model makes sense from the point of view of theory or common sense." Also he notes that NB is a special case of ZINB so you can do a formal test to compare them. I think he is mostly arguing that you shouldn't automatically go with a ZIP model; as he concludes "So next time you’re thinking about fitting a zero-inflated regression model, first consider whether a conventional negative binomial model might be good enough. Having a lot of zeros doesn’t necessarily mean that you need a zero-inflated model."