Heterogeneous DID: jwdid "never" vs. "not-yet-treated" implementation

[Frances Pogacar]

I am using repeated cross-sectional data with a binary treatment and staggered treatment adoption. When I run jwdid and csdid both with the "not-yet-treated" control group, the two methods produce very different point estimates. However, when I run jwdid and csdid with the "never-treated" control group, I get the same point estimates between the two methods. Why might this be the case?

The following lines of code produce the same estimates:

(1) jwdid count, tvar(period) gvar(firsttreat) cluster(city) never
(2) csdid count, time(period) gvar(firsttreat) never cluster(city)

The following lines of code produce very different estimates:

(1) jwdid count, tvar(period) gvar(firsttreat) cluster(city) 
(2) csdid count, time(period) gvar(firsttreat) notyet cluster(city)

Here are some sample outputs after aggregating:

Method	ATT, simple	ATT, group 12	ATT, group 14	ATT, group 17
jwdid/csdid "never treated"	-.1564271	-.4647559	-.043473	-.0912151
jwdid "not-yet-treated"	.0584691	.063138	.0021105	.1439579
csdid "not-yet-treated"	-.1565513	-.4688338	-.0416988	-.0912151

[FernandoRios]

Short version
different models

longer versions
csdid with not yet tested makes no assumptions regarding pretrest
ent
jwdid with not yet assumes pta holds

also jwdid uses a full outcome prediction model whereas Csdid for crossection does not
you may need to go over the math to understand the specifics

[Frances Pogacar]

Thank you for your response, FernandoRios. Does jwdid with "never" also assume that pta holds, or is it more similar to csdid in that it makes no assumptions regarding pretrend? I ask because the documentation for jwdid mentions that when the never option is specified, "for each group/cohort, the period g-1 is from the specification" - does this mean that comparisons for each group/period are made only with respect to the period prior to that group's treatment (rather than the entire pre-treatment period and assuming pta holds)?

[FernandoRios]

JWDID with never is identical to CSDID reg. They both use only G-1 as baseline.

[Frances Pogacar]

OK great, thank you for clarifying and confirming! Is it possible for jwdid with not-yet to use G-1 as baseline (although this may not be implemented in the package), or are there issues that arise with this?

Finally, my understanding of jwdid is that repeated cross-sections is assumed when ivar() is not declared, and hence I was expecting that jwdid without ivar() would be identical to hdidregress twfe, which should also be for repeated cross-sectional data. However the two models only produce the same estimates when ivar() is declared in jwdid. Any reason why this is the case?

jwdid count, tvar(period) gvar(firsttreat_1000_02) cluster(city_cat)
jwdid count, ivar(int_uid) tvar(period) gvar(firsttreat_1000_02) cluster(city_cat)
hdidregress twfe (count) (treat), group(int_uid) time(period) vce(cluster city_cat) controlgroup(notyet)

Note: for this comparison, I used the not-yet as the control group, as my understanding is that the models under the 'never' option are not comparable since jwdid uses the G-1 baseline and hdidregress twfe does not.

Model	ATT, simple	ATT, group 12	ATT, group 14	ATT, group 17
jwdid without ivar()	.0584691	.063138	.0021105	.1439579
jwdid with ivar()	.0572045	.0638813	.0027032	.1381655
hdidregress twfe	.0572045	.0638813	.0027032	.1381655

[FernandoRios]

Hi Frances
1) Is it possible only using G-1 data?, Not sure, I think it may be possible. You would need to
a) built the model (mathematically) that reflects that particular specification
b) exclude G-1 and G-2 dummies and interactions from the data
This, however its not possible with JWDID as it is (I do have a feature that may mimic that, but would have to check

2) hdidregress ana jwdid do things slightly different, even tho they are based on the same model. To understand if its correct or not, and which one is better, I recommend redo the regression by hand.
I recalled speaking with Stata ppl that we were getting different results in some of the exercises i was running, but did not catch all of the differences.

What you report may just be one of those.

Which one is correct....its a different question

[Frances Pogacar]

Thank you, Fernando - your responses have been really helpful!