I'm a doctoral student @WU Vienna and as part of my dissertation project
I'm about to estimate a panel model (the panel is unbalanced with N=328
and T=8) with a non-negative integer count variable (patent
applications) as a DV. I thus did some reading on -xtnbreg- and
-xtpoisson- but I still struggle with some open issues.
I would really appreciate any help with the following two questions:
1) The data shows signs of over-dispersion, as the standard deviation is
about 3 to 4 times larger than the mean. Based on that my first choice
would be to go for -xtnbreg-. Nevertheless, -xtnbreg- doesn't offer
cluster-robust standard errors.
In Cameron/Trivedi (2009) - Mircoeconometrics Using Stata, on page 627
they say: "/The negative binomial has the attraction that, unlike
Poisson, the estimator is designed to explicitly handle overdispersion,
and count data are usually overdispersed. This may lead to improved
efficiency in estimation and a default estimate of the VCE that should
be much closer to the cluster-robust estimate of the VCE, unlike for
Poisson panel commands. At the same time, the Poisson panel estimators
rely on weaker distributional assumptions - essentially, correct
specification of the mean - and it may be more robust to use the Poisson
panel estimators with cluster-robust standard errors."/
If I get them right they recommend using -xtpoisson- with cluster robust
standard errors since the standard errors of the non-robust -xtnbreg-
are expected to be too low and the robust version of -xtpoisson- should
be able to deal with over-dispersion (if I fit a model with -xtnbreg-
and fixed effects I get a significant result with a p-value of 0.067; if
I fit the same model with -xtpoisson-, fixed effects and cluster robust
standard errors the result is insignificant with a p-value of 0.125).
Since I'm rather new to STATA and statistics in general I'd be very
grateful if some of you could comment on Cameron/Trivedi's (2009)
statement/recommendation. Are there any other options?
2) With -xtnbreg- it is possible to calculate bootstrap standard errors.
However, when using this option, I receive the following error message:
"insufficient observations to compute bootstrap standard errors no
results will be saved".
Basically I try to fit the following model:
xtnbreg dv iv1 iv2... iv8 yeardummy1 yeardummy2 ... yeardummy6, fe
As mentioned, this results in the error message described above. If I
remove the yeardummies from the model everything is fine and the
bootstrap option works. Since the data shows a significant negative time
trend I need the yeardummies in the model. Does bootstraping have a
problem with dummy variables or did I make another (naive) mistake?
Assuming the bootstrap option would work, is the cluster-robust
-xtpoisson- still preferable over the bootstrap -xtnbreg-?