Which model is suitable to see ATE on zero inflated panel data? zip, zinb or hurdle (dhreg, xtdhreg) or xtpoisson?

[Lars Pete]

zip, zinb or hurdle (dhreg, xtdhreg)? or nbreg? or should I just add a constant to zero values and go with xtpoisson?


treament var could be binary (did = treated # post17)

. fre treat17

treat17
-----------------------------------------------------------
              |      Freq.    Percent      Valid       Cum.
--------------+--------------------------------------------
Valid   0     |       7633      99.45      99.45      99.45
        1     |         42       0.55       0.55     100.00
        Total |       7675     100.00     100.00          
-----------------------------------------------------------

or could be interacted with a stadardized var as treat17#zscore1 (DDD)

. sum zscore1

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
     zscore1 |      7,675    .0131534    .6424957  -.3452138   10.56285

. xtset countyfips year

Panel variable: countyfips (unbalanced)
 Time variable: year, 2010 to 2019
         Delta: 1 unit

The depvar is:

. . sum mdstot

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
      mdstot |     24,518    275.0965    1084.201          0      22991

Graph.gph

or if depvar is taken as proportion of county population.

. sum mdstotprop

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
  mdstotprop |      7,675    .1285506    .1763522          0   3.982446

[Jeff Wooldridge]

I would simply use xtpoisson with the fe option and control for log(population). You'll get a treatment effect measured as a percent when you multiple the coefficient by 100.

xtreg mdstot  lpopulation treat17 i.year, fe vce(robust)