Closed taiwoakinyemi closed 2 years ago
I added an issue about this to the csdid
github, it is here: friosavila/csdid_drdid#29.
I think this is not currently available on Stata, but let’s see what they say.
Brant
Thank you Brantly So, for the Stata version, I gave it a different name. When you use csdid, you need to use the option "long". It will reproduce what DID does for the universal base period.
👍 Thank you!
Brant
Hi Fernando,
Thanks for the update. I was expecting the coefficients for the universal base period to be 0 but it is not.
Another issue I have is the command code, when I used the below command lines I get different results, which is the more appropriate?
csdid wealthscore hheadage under5 i.state, time(syear) gvar(pyear) method(drimp) long saverif(rif) replace
estat all
estat event, window(-5 5)
csdid_plot
However, when I used the below command and when I did not specify option(long) the results are similar.
csdid wealthscore hheadage under5 i.state, time(syear) gvar(pyear) method(drimp) option(long) saverif(rif) replace
estat all
estat event, window(-5 5)
csdid_plot
Hi,
Yes, when you use the universal base period, the estimate in the pre-treatment period (i.e., e=-1) should be equal to 0. I'm guessing this is some issue with just getting the Stata syntax correct, but I'm not sure what the answer is. I saw you posted this on the csdid
issues page too. I think they'll be able to help you with this one.
Brant
So, it is really a programming design in how I approached this. Usually, you have the following:
Time = 1 2 3 4 5 6
Treat = 0 0 0 1 1 1
The Universal base in DID compares Everything with Time 3:
ATT(t-3) = T1-T3
ATT(t-2) = T2-T3
ATT(t-1) = T3-T3
ATT(t+0) = T4-T3
ATT(t+1) = T5-T3
ATT(t+2) = T6-T3
I, however, did it following a slightly different nomenclature:
ATT(t-2) = T3-T1
ATT(t-1) = T3-T2
ATT(t+0) = T4-T3
ATT(t+1) = T5-T3
ATT(t+2) = T6-T3
Basically Omitting the base period
Hi Fernando,
Yes, I totally agree with this --- it's not unique what to do in pre-treatment periods. In R, the default is "varying" and it is a different variation from either of the ones you mention. It reports
ATT(t-2) = T2-T1
ATT(t-1) = T3-T2
and the same in post-treatment periods.
In my view, the "universal" base period one is kind of weird. I think this is mainly just a legacy from event study regressions where this is more-or-less just what you can easily get the regression to report. Importantly, the pre-treatment estimates are not actually any kind of ATT parameter. The thing that this works well for is just visualizing the trend in the data; and, actually, I only think think that case works well if you have unit-specific linear trends (which...I am not sure where these come from either...).
I think the "varying" base period is the one that people should use in most applications (this is why it's the default). In particular, you can see things like anticipation effects in that one much more clearly.
I have post on my website about all this here
Thanks, Brant and Fernando.
HI, I wonder if there is a way to include the universal base period in stata csdid.