Two-step negative binomial estimation with clustered bootstrap standard errors

[Kamyar Nasseh]

Hello, I am estimating a negative binomial count model with a generated regressor to control for endogeneity. I want to bootstrap my standard errors by 3-digit zipcode. Each row in my data table represents a unique individual However, after 500 replications, my standard errors are very big. I don't know why. I probably can do ivpoisson as this estimation technique does not require me to make any assumptions regarding the distribution of my dependent variable (Number of Dentists in a Practice). But. I would like to figure out what is going wrong in my code. My IVs are log_emp_per_firm_cty and log_firm_cty.

Here is my wrapper:

program drop twosri
program twosri, eclass
tempname b V
capture drop uhat

reg log_hhi_zip yr_2014 yr_2015 exper expersq male pop_density dds_density log_real_med_hhinc metro_1mil metro_250k_1mil metro_lt250k pediatric oth_specialty dent_hpsa_whole black_race hispanic_race asian_race other_race race_missing log_emp_per_firm_cty log_firm_cty midwest south west, cluster(geozip_3)

predict uhat, residual

nbreg tot_dds yr_2014 yr_2015 exper expersq male pop_density dds_density log_real_med_hhinc metro_1mil metro_250k_1mil metro_lt250k pediatric oth_specialty dent_hpsa_whole black_race hispanic_race asian_race other_race race_missing midwest south west log_hhi_zip uhat, cluster(geozip_3)





matrix `b' = e(b)

ereturn post `b'

end



bootstrap _b, reps(500) seed (10101) cluster(geozip_3) idcluster(newidv2) group(adaid) nowarn: twosri



Thanks,

Kamyar