I'm running a negative binomial regression. I found big differences in the results* if I compute with the glm y x, (nbinomial 1) link(log) and glm y x, (nbinomial ml) link(log) (for Stata) command.
What is the difference between computing with ML** and with one?
I have big differences in the significance level and I would strongly prefer to take the model with nbinomial 1. The differences between the AIC values are 200 (lower values for nbinomial ml).
*The results showed big differences in the significance and standard error level. ** Stata state's that the "Negative binomial parameter estimated via ML and treated as fixed once estimated". It seems that the parameters are estimated via Maximum Likelihood but what does this mean for the other model?
nbreg y x, dispersion(mean)
with SPSS'sGENLIN Y WITH X /MODEL X DISTRIBUTION=NEGBIN(MLE) LINK=LOG.
, and they have always been exactly the same (although I have had situations in which Stata converges where SPSS does not). As always, are you sure you are estimating the same models? (Same variables, same records with no missing data?) SPSS and Stata have potential differences in how an exposure/offset is specified as well. $\endgroup$(nbinomial 1)
means you set the dispersion to an arbitrary constant, 1. I highly doubt setting it to a constant is appropriate, but one hypothetical situation I can imagine is if by past results you have strong reason to believe it is equal to 1. $\endgroup$