csdid: ATT omitted

[Harriet WANG]

Hi all,

I am using csdid to estimate treatment effects in a staggered DID, and I am running into problems once I add control variables. My data are an unbalanced firm-year panel.


Obs Mean Std. Dev. Min Max
Independent Variables
Treated 968 0.1384 0.3455 0.0000 1.0000
Dependent Variables
Green Patent 532 1.1838 1.4638 0.0000 7.1017
Control Variables
Firm Age 536 2.9363 0.2992 1.9459 3.8067
Firm Size 541 22.8617 1.7382 19.3825 30.5711
Leverage 541 -0.8723 0.5661 -3.0372 0.1113
Liquidity 632 -0.6472 0.4451 -2.7989 -0.0285
Shareholder Rate 508 3.5292 0.4721 1.9330 4.4129
(Treat=0: 834; Treat=1: 134)

My baseline specification is: csdid ptotal size liq large age lev , ivar(firmid) time(year) gvar(firstyear) method(dripw) vce(cluster firmid)

In some specifications, the estimation results become almost empty, or the ATT estimates disappear for many (g,t) cells.

When I run estat simple, I do not get a single ATT estimate.

I wondered whether the issue could be due to the limited number of observations, but unfortunately, the dataset cannot be expanded for the core variables because of data availability constraints.

Is there any way to solve this problem?

Many thanks in advance for any guidance or suggestions.


Best,

Harrite

[Carlo Lazzaro]

Harriet:
welcome to this forum.
As per FAQ, please also post what Stata gave you back when using the community-contributed module (as another FAQ recommends us to name all the modules elaborated by Stata users and approved by KitBaum) -csdid-. Thank you.

[FernandoRios]

The two likely cases are
Your Cohort time specific sample is too small.
or there is no combination of that cohort and time.
CSDID still produces the combination (as seen in the table) just because I coudlnt find a way to correct that at the time
F

[Gael Wilfried]

Hi Sir

I have a question regarding the CSDID estimator (and other Diff-in-Diff estimator with staggered treatment adoption). Is it possible to include lags of the dependent variable as covariates in order to account for dynamics?

In addition, my model includes some covariates. Can I use one-period lags of these covariates to reduce potential reverse causality?

Thanks in advance.

This my model :

global covars1 Lag_ln_itaxrevenueworld Lag_GDPpcConst Lag_AgriVadded Lag_UrbanPop

csdid ln_itaxrevenueworld $covars1, ivar(id) time(year) gvar(first_dr_year) method(dripw) rseed(1) agg(event) notyet saverif(_rif_)

[FernandoRios]

CAn you add lags. Yes
Should you? I cannot say either way. It will all depend on whether or not there is enough variation to estimate the ATTs

[Gael Wilfried]

Thank you for your feedback!

[Gael Wilfried]

Hi Sir,

I have another question. In my final results for World, the table does not show the number of observations. I restricted the window to +/- 5 years before and after the treatment. How can I include the number of observations in the table?

I checked the scalar e(N), but it shows the total number of observations for the baseline, before applying the window restriction. How can I get the number of observations within the restricted window (+/- 5 years around the treatment)?

Here is my code:


global covars1 Lag_ln_itaxrevenueworld Lag_GDPpcConst Lag_AgriVadded Lag_UrbanPop


*================================================= ====
* NOT-YET-TREATED
*================================================= ====

csdid ln_itaxrevenueworld $covars1, ivar(id) time(year) gvar(first_dr_year) method(dripw) rseed(1) agg(event) notyet

estat event, window(-5 5) estore(csdid_model1a)
estat pretrend, window(-5 5)
estat cevent, window(-5 5)

event_plot csdid_model1a, default_look graph_opt(xtitle("Periods since the event") ytitle("Average effect") ///
title("csdid") xlabel(-5(1)5)) stub_lag(Tp#) stub_lead(Tm#) together


estat event, window(-10 10) estore(csdid_model1b)
estat pretrend, window(-10 10)
estat cevent, window(-10 10)

event_plot csdid_model1b, default_look graph_opt(xtitle("Periods since the event") ytitle("Average effect") ///
title("csdid") xlabel(-10(1)10)) stub_lag(Tp#) stub_lead(Tm#) together




*================================================= ====
* NEVER-TREATED (default control group)
*================================================= ====

csdid ln_itaxrevenueworld $covars1, ivar(id) time(year) gvar(first_dr_year) method(dripw) rseed(1) agg(event)

estat event, window(-5 5) estore(csdid_model2a)
estat pretrend, window(-5 5)
estat cevent, window(-5 5)

event_plot csdid_model2a, default_look graph_opt(xtitle("Periods since the event") ytitle("Average effect") ///
title("csdid") xlabel(-5(1)5)) stub_lag(Tp#) stub_lead(Tm#) together

estat event, window(-10 10) estore(csdid_model2b)
estat pretrend, window(-10 10)
estat cevent, window(-10 10)

event_plot csdid_model2b, default_look graph_opt(xtitle("Periods since the event") ytitle("Average effect") ///
title("csdid") xlabel(-10(1)10)) stub_lag(Tp#) stub_lead(Tm#) together

*================================================= ====
* Import results
*================================================= ====

esttab csdid_model1a csdid_model2a using "csdid_results.rtf", ///
se(3) b(3) star label ///
title("CSDID RESULTS") ///
mtitles("Not-yet treated" "Never-treated") ///
keep(Tm* Tp* Pre_avg Post_avg) ///
replace


esttab csdid_model1b csdid_model2b using "csdid_results.rtf", ///
se(3) b(3) star label ///
title("CSDID RESULTS") ///
mtitles("Not-yet treated" "Never-treated") ///
keep(Tm* Tp* Pre_avg Post_avg) ///
replace

[FernandoRios]

unfortunately that is not easy to obtain.
Because of How CSDID uses nottreated and not-yet treated, its hard to keep tract of sample sizes, specially when using restricted ouputs.

[Gael Wilfried]

Okay, Thank you!