Message/Author 


Hello Drs. Muthen, I am trying to understand how one interprets the negative binomial dispersion parameter in my output. I just ran a very simple model (i.e., one count variable dependent variable regressed on to two continuous varibles) and received a dispersion parameter estimate of 0.114 and a two  tailed pvalue of 0.461. I am familiar with the alpha dispersion estimate you receive from STAT and SAS and if the parameter isn't statistically significant, dispersion is not an issue and you can use a Poisson model. Can the parameter estimate provided in MPLUS output be interpreted in the same way? Thank you! 


Yes. 


With the caveat of the problem of testing a parameter at the border of its admissible space (zero). I use BIC instead to choose between the various count models. See the Marital Affairs example on slides 3941 in our Topic 5 handout on the website, showing the many variations on the count modeling theme available in Mplus. 


Thanks so much! 


Hi Bengt and Linda, Is there a rule of thumb for how large the dispersion statistic should be in ZINB regression? Is it supposed to be close to 1.00 for the NB to be appropriate? Or should I just be judging its significance, as stated above, as a reflection of whether NB is preferable over Poisson? Tnanks. 


It should be greater than zero. You can check the following reference for a maximum value: Hilbe, J. M. (2007). Negative binomial regression. Cambridge, UK: Cambridge University Press. You can look at significance or you can compare the BIC's from the models with and without dispersion. 

Jen posted on Wednesday, July 30, 2014  9:53 am



Hello, I had another question re: the dispersion parameter estimates. When negative binomial items serve as indicators of a latent factor, is the dispersion parameter estimate still relevant? For several of my indicators, the pvalue is close to 1 for items that have significant dispersion when included as manifest variables in the model. Does this just mean that the 'residual' dispersion after modeling the latent factor is nonsignificant? Thanks! 


The dispersion parameters are relevant also for indicators of factors. Perhaps you are right that the factor absorbs some of the unobserved heterogeneity that the dispersion parameters try to capture and therefore make them go insignificant when used as factor indicators. 

Back to top 