Dynamic Inverse Probability Weighting estimation

[Tom Ford]

Dear All,
I have a panel dataset, and I would like to do a Dynamic Inverse Probability Weighting to run my estimation.
In other words, I would like that the probability of being observed to vary across units and over time.

My concern is that in my main dataset I have a number of observations where the probit model "predicts success/faliure perfectly". Should I impute the I impute the probability of being observed to be ==1 in these cases, or should I keep them as missing values?

Take the following example

use http://www.stata-press.com/data/r17/nlswork, clear
(National Longitudinal Survey of Young Women, 14-24 years old in 1968)

* Generate group-time variable, used for estimation
egen occ_year = group(occ_code year)
(121 missing values generated)

* baseline model to estimate: impact of union on wages controlling for individual FEs, and occupation-time FEs
reghdfe ln_w union, a(idcode occ_year) cluster(occ_code)
eststo base

I calulate my IPW like so:

probit union ///
i.occ_year c.age c.tenure, ///
vce(cluster occ_code)

predict p_remain, pr
* generate weights
gen w=.
replace w=1/p_remain if union==1
replace w=1/(1-p_remain) if union==0
summarize w

* estimate IPW model
reghdfe ln_w union [pweight=w], a(idcode occ_year) cluster(occ_code)
eststo ipw

Note that my probit model drops a number of observations because they predict success or faliure perfectly.

note: 93.occ_year != 0 predicts failure perfectly;
93.occ_year omitted and 21 obs not used.

note: 95.occ_year != 0 predicts failure perfectly;
95.occ_year omitted and 28 obs not used.

.....

note: 145.occ_year != 0 predicts success perfectly;
145.occ_year omitted and 2 obs not used.

. esttab base ipw, se

--------------------------------------------
(1) (2)
ln_wage ln_wage
--------------------------------------------
union 0.0965*** 0.0788**
(0.0185) (0.0172)

_cons 1.737*** 1.773***
(0.00440) (0.00867)
--------------------------------------------
N 18495 18011
--------------------------------------------
Standard errors in parentheses
* p<0.05, ** p<0.01, *** p<0.001

Because I do not have weights for these observations my estimation sample of the IPW regression drops of over 400 observations.

The question I have is whether this is correct, or whether I should impute weights for those observations where the probit model predicts success or faliure perfectly. In this cases should my weight be 1?


Any broader suggestion on whether I am doing this correctly would be extremely helpful

Best