Three-level ZINB in Stata

[Anna Geo]

Hi,

I use Stata 16.1 and I need to run a 3-level Zero Inflated Negative Binomial regression. I have seen some previous answers suggest using "gllamm"

Based on the "gllamm" manual in Stata I wrote the following code:

meglm y x1, x2, x3....xn || lev3: || lev2:, family(nbinomial)

After 11 iterations the log likelihood stops changing.

Does anyone know if I am doing something wrong with the code and how I can get around this.

[Leonardo Guizzetti]

Just a couple of thoughts, and no specific solutions I'm afraid. First, -meglm- and -gllamm- are not the same command, the latter being a community-contributed command. Did you really intend to use -gllamm- ? Second, do you really need to use a ZINB model, or would something like a Poisson model with sandwich variance (robust) variance estimation suffice? It seems to me that, depending on the specifics of the model you are fitting and the data you have, the multi-level ZINB model could be unstable to fit.

[Anna Geo]

Hi Leonardo,

I read that -gllamm- would be a good option but I don't know its command lines. I have run 4 different 2-level regressions in MPlus (couldn't run 3-level models there): Poisson, NB, ZIP and ZINB. Based on LL and AIC, ZINB seems to fit the data better when using 2-level models. Also, the outcome is overdispersed (variance>mean) and it does have more zeros than a normal poisson. Given the above I assumed ZINB would be the right option.

I'm quite new in the field and have never tried methods with "sandwich variance", but will give it a try and let you know how it went.

[Joao Santos Silva]

Dear Anna Geo,

I echo what Leonardo said above, and it looks as if you are mixing up a few things. For example, a Poisson distribution can have any proportion of zeros (so your data cannot have more zeros than a Poisson) and overdispersion is rarely a problem for Poisson (and what matters is the conditional overdispersion, not the relation between the mean and variance of the data). To be able help, we would need to know more about your dependent variable (what is it and how is it measured) and about the goals of your regression (what do you want to use it for).

Best wishes,

Joao