[Ludovic Van Cau]

Without the vce option:

Negative binomial regression                    Number of obs     =    462,187
                                                LR chi2(501)      =   39266.96
Dispersion     = mean                           Prob > chi2       =     0.0000
Log likelihood = -851639.86                     Pseudo R2         =     0.0225

-----------------------------------------------------------------------------------
         teamsize |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
------------------+----------------------------------------------------------------
    internetdummy |  -.0138526   .0021754    -6.37   0.000    -.0181162    -.009589
invt_network_size |   .0094635   .0000565   167.59   0.000     .0093528    .0095742

[Jeff Wooldridge]

If you do not specify any option, the NegBin model is being assumed correct -- including the assumption on the variance -- and there is no need to clustering. The vce(robust) option means you're allowing the NegBin assumption to fail, but it does not account for clustering. The vce(cluster cbsacode) option accounts for correlation due to either cluster correlation or clustered assignment of the intervention. Unfortunately, what you're finding is not atypical when clustering.

However, it's not clear that you should cluster at the cbsacode level. The trust the data were not from a cluster sample at that level, correct? Then at what level is the internetdummy assigned? The patent level? If so, then I wouldn't cluster.