Vuong's test for Poisson, NB, ZIP, an... PreviousNext
Mplus Discussion > Categorical Data Modeling >
 Jeff Greene posted on Thursday, June 18, 2009 - 11:04 am
I am trying to determine the best model for some count data. I know Poisson is nested in Negative Binomial, and ZIP is nested in ZINB so I can use chi-square difference tests to compare those models. But, I've read that Poisson is not nested in ZIP, and NB is not nested in ZINB and that Vuong's test (1989) should be used. Is there a way to get Mplus to produce Vuong's test? If not, is there a way to use information from the output files to calculate Vuong's test?

Thank you!
 Bengt O. Muthen posted on Friday, June 19, 2009 - 10:14 am
Simplest approach would be to use BIC to compare say Poisson to ZIP. You can also try the UG ex 7.25 two-class approach to ZIP to test if 1 class is sufficient using Vuong, that is TECH11. But since TECH11 drops the first class, put the class with restrictions as the second class, not the first as in the example. I have not tried this, however.
 Rob Dvorak posted on Tuesday, July 06, 2010 - 3:33 pm
I have a structural model with 2 ZINB distributions (one as a DV and one as a mediator). Is it possible to get Vuong tests for both of these distributions simultaneously? I don't think this can be done using the two-class approach since I have two separate variables which need to be tested.
 Bengt O. Muthen posted on Wednesday, July 07, 2010 - 10:16 am
It sounds like you want to test the need for zero-inflation, that is, the need for a zero class. And it sounds like you think about running ZINB explicitly as a two-class model, versus a non-inflated model. I don't know how that would work because it is a two-class model with special restrictions. How about just looking at the SEs for the inflation parameters? Or the BIC difference?
 Rob Dvorak posted on Wednesday, July 07, 2010 - 10:41 am
Yes, that's exatly right. I have reported the BIC difference between nb and zinb, but I would like to have Vuong tests to justify the use of zinb over nb. I believe, given the large drop in BIC between the two models, that the zinb model is more appropriate. I would like to report that it isn't possible to simultaneously test for inflation (i.e., Vuong test) of the two different distibutions in the same model. Is this accurate?
 Tihomir Asparouhov posted on Wednesday, July 07, 2010 - 2:03 pm
Use the savedata command
savedata: save=LOGLIKELIHOOD;
to get the log-likelihood values for the zinb and nb models. You can then compute the Voung test as in (see page 16)
 Rob Dvorak posted on Thursday, July 08, 2010 - 5:03 am
@ Tihomir,
This is perfect, exactly what I was looking for. Thanks!
 Rob Dvorak posted on Sunday, November 18, 2012 - 7:35 am
Greetings! I'm trying to obtain the log likelihoods following some multilevel analyses. When I run a multilevel negative binomial the LLs save in the specified file. However, if I switch to zinb (nbi) or hurdle (nbh) the LL column is all zeros. I'm currently working from home on an older version of Mplus (v5.21). Are LLs from nbi & nbh not available in this version... or is it that I am doing something wrong?

Thanks in advance!
 Linda K. Muthen posted on Sunday, November 18, 2012 - 8:54 am
These were available starting with Version 5.1. Please send your input, data, and output to
 Rich Mohn posted on Saturday, August 02, 2014 - 11:39 am
The NYU link posted on July 7, 2010 is no longer valid . . . is there another link to that paper? I am trying to do the same thing.
 Tihomir Asparouhov posted on Monday, August 04, 2014 - 9:00 am

and the reference

Accounting for Excess Zeros and Sample Selection in Poisson and Negative Binomial Regression Models


William H. Greene
Department of Economics
Stern School of Business
New York University
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message