see #1093 to get a general solution
and #1221 missing exposure, related bug
all three NegativeBinomial versions have the same llnull in the test results RandHIE.
are these "regression" tests, only 'nb2' has a comment # Stata
Stata reports different llnull for nb1 and nb2, and it depends on exposure.
correction to Stata:
If there is no exposure or offset, then llnull is the same for nb1 and nb2
If there is an exposure, then llnull differs between nb1 and nb2
likely explanation: Stata has nb1 as dispersion(constant) and nb2 as dispersion(mean)
If there is no exposure or offset, then the estimated mean is constant int he case of constant-only exog.
fixed in #1610 for keywords in llnull,
note: without exposure this affected only NegativeBinomial geometric
exposure and offset for negbin will be added in PR #1612
with only one test case: nb2 with exposure and cluster robust standard standard errors
will automatically have correct llnull