RE: Obtaining robust standard errors after xtnbreg with random effects

[DY Kim]

Hello!

I am using negative binomial regression with random effects: xtnbreg crime unemployment poverty region, re vce(robust).
When I included the option of vce(robust), I got the error message:

vcetype 'robust' not allowed
r(198);

Is there anyway I can obtain robust standard errors after xtnbreg with random effects?

Any help would be appreciated.

[Fei Wang]

I would recommend using -xtpoisson, re- with -vce(robust)-. Poisson model is preferred to negative binomial in most cases.

[DY Kim]

I would recommend using -xtpoisson, re- with -vce(robust)-.

Thank you for the advice. However, given that my dependent variable has many zero values, I need to use xtnbreg. Is there any command option to obtain robust standard errors in the context of xtnbreg?

[Fei Wang]

Well I would stick to #2 and strongly recommend -xtpoisson-. We can hardly find a situation where -xtnbreg- is applicable while -xtpoisson- is not qualified. See more discussions here.

[Joao Santos Silva]

Just to add to what Fei Wang said, Poisson regression is compatible with any percentage of zeros, so having many zeros is not a reason to go for the negative binomial. I would strongly recommend xtpoisson with fixed effects.

[DY Kim]

I would strongly recommend xtpoisson with fixed effects.

I am using nbreg with random effects, not fixed effects.

Do you have any references that support using xtpoisson with random effects over nbreg with random effects? Any help would be appreciated.

[Joao Santos Silva]

Do you have to use RE? Using RE in this context is tricky because the model is only valid under very strong assumptions. That is why I suggest Poisson with FE, which is quite robust; see Jeff Wooldridge's 1999 paper "Distribution-Free Estimation of Some Nonlinear Panel Data Models" in the Journal of Econometrics.

[DY Kim]

Do you have to use RE? Using RE in this context is tricky because the model is only valid under very strong assumptions. That is why I suggest Poisson with FE, which is quite robust; see Jeff Wooldridge's 1999 paper "Distribution-Free Estimation of Some Nonlinear Panel Data Models" in the Journal of Econometrics.

Hello Joao,

I have to include time-invariant variables as controls using random effects models. If I use fixed effects models, I will lose all time-invariant variables, which are important for my research. Can Wooldridge's paper be a reference for my case? Thank you for the help.

[Joao Santos Silva]

Then I would perhaps just use Poisson regression; please check the paper.