[Saurabh Trivedi]

Thanks FernandoRios for the response. Yes, I have installed the drdid and csdid both. Sorry I could not get your next question. I have attached the full data in the previous post. As mentioned in the code, I have used Date as the time series variable and SCAP_Date as the gvar. I am not even sure if these are used and setup correctly. Thanks again for your time.

[FernandoRios]

Ok
I haven’t seen the data yet but based on the screen shot
data is not correctly setup
Unless you are using daily data, try changing the scale of that variable to your preferred frequency
recreate the gvar
and then show me a two way table. Between time variable and group variable( time gvar)
thanks

[Saurabh Trivedi]

Thanks FernandoRios . I have attached my datafile here. Main time variable id is Date (quarterly frequency) and the firm level id is GVKey. The dummy variable SCAP takes the value of '1' if a State has implemented a certain law (treatment), else it takes a value of '0' if it never implemented such a law in the sample time-period. The variable SCAP_Date shows the dates on which the law was implemented by a particular State (different dates for different states), else it takes the value of '0' if the law was never implemented. I have another dummy variable POST_SCAP which takes the value of '1' only after a state has implemented the law else it remains 0. Sorry, I am not able to create gvar. Should it be POST_SCAP?

[FernandoRios]

Your data needs lots of work
1. your gvar is defined (based on your definition) by scap_date (the time when people were treated)
2. some of you "gvkey's" have more than one Scap_date data(never or an actual date) you need to clean that and define which one is correct, or drop them
3. Your data is currently in days. need to transform everything into year-quarters, to make the data managle.
HTH
A quick clean up may produce this:
F

[FernandoRios]

Alternatively, if you think your post_scap is correct, you can try using xthdidregress, (stata 18) which will make the necessary corrections for you (although you still need to modify your data variable

[Kal El]

Hi Sandra
ok so
Not sure why you cant run the model without covariates. I just reinstalled a version on my machine (using fra install) and I'm able to use it without controls.
Inregards to calendar. It is the Group /cohort specific Treatment effect.
Event is the average of the effects with respect to the relative year

Finally for group specific plots

ssc install frause, replace
fra install csdid2, replace

clear all
frause mpdta

csdid2 lemp, ivar(countyreal) tvar(year) gvar(first) long
estat event, rgroup(2004) plot
estat event, rgroup(2006) plot
estat event, rgroup(2007) plot

Hope this helps
Fernando

Hi Fernando,

I'm having some trouble getting csdid2 installed. I'm getting the following result:

. ssc install frause, replace
checking frause consistency and verifying not already installed...
all files already exist and are up to date.

. fra install csdid2, replace
command fra is unrecognized
r(199);

[FernandoRios]

Oops sorry for that.
Please follow the instructions here https://friosavila.github.io/chatgpt/fra_03_30_2023/
to install "fra"
then you can install csdid2.
F

[Sandra Macedo]

Hi FernandoRios, oops, me again. I hope you are doing well. I'm having problems with csdid and csdid2 again.

I've just updated to Stata 18, but for some reason, I cannot reproduce my results on csdid2 using fixed effects because csdid2 gives the same results with/without fixed effects. Not sure what is happening because I cant open the saved ster files with the estimates either.

csdid2 scores overage maeduc i.stateid  if treat==1, tvar(year) gvar(cohortuf) method(dripw) agg(simple) cluster(schoolid)
Producing Long Gaps by default
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 
.....
Difference-in-difference with Multiple Time Periods
Outcome model  : least squares
Treatment model: inverse probability
------------------------------------------------------------------------------
             |               Robust
             | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
   SimpleATT |   .0510869    .002529    20.20   0.000     .0461303    .0560436
------------------------------------------------------------------------------

. estat event
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
     Pre_avg |   .0205535   .0033877     6.07   0.000     .0139138    .0271932
    Post_avg |   .0500537   .0025308    19.78   0.000     .0450934    .0550139
         tm2 |   .0205535   .0033877     6.07   0.000     .0139138    .0271932
         tp0 |   .0097071   .0032017     3.03   0.002     .0034318    .0159823
         tp1 |   .0576214    .003083    18.69   0.000     .0515788     .063664
         tp2 |   .0496524   .0030823    16.11   0.000     .0436111    .0556937
         tp3 |   .0832337   .0030235    27.53   0.000     .0773078    .0891596
------------------------------------------------------------------------------

. 

csdid2 scores overage maeduc if treat==1, tvar(year) gvar(cohortuf) method(dripw) agg(simple) cluster(schoolid)
Producing Long Gaps by default
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 
.....
Difference-in-difference with Multiple Time Periods
Outcome model  : least squares
Treatment model: inverse probability
------------------------------------------------------------------------------
             |               Robust
             | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
   SimpleATT |   .0510869    .002529    20.20   0.000     .0461303    .0560436
------------------------------------------------------------------------------

. estat event
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
     Pre_avg |   .0205535   .0033877     6.07   0.000     .0139138    .0271932
    Post_avg |   .0500537   .0025308    19.78   0.000     .0450934    .0550139
         tm2 |   .0205535   .0033877     6.07   0.000     .0139138    .0271932
         tp0 |   .0097071   .0032017     3.03   0.002     .0034318    .0159823
         tp1 |   .0576214    .003083    18.69   0.000     .0515788     .063664
         tp2 |   .0496524   .0030823    16.11   0.000     .0436111    .0556937
         tp3 |   .0832337   .0030235    27.53   0.000     .0773078    .0891596
------------------------------------------------------------------------------

I can't figure out what's happening. I 've installed all packages, wondering if it is an update of csdid2 or in one of the packages.


Another problem: when I use csdid instead, I get the an error message

. csdid scores overage maeduc i.stateid if treat==1 , time(year) gvar(cohortuf ) method(dripw) agg(simple)
xxxxxestimates post: matrix has missing values
r(504);

I'm sure there must be something I'm missing. Thank you so much for helping.

[FernandoRios]

I think it’s doing what is supposed to
my best guess is that your treatment is at the state level
thus you can’t use state as fixed effect.
csdid is showing that by refusing to produce results
csdid2 is a bit smarter and tries to drop all Collinear variables before running a model
hth

Fernando

[Sandra Macedo]

I think it’s doing what is supposed to
my best guess is that your treatment is at the state level
thus you can’t use state as fixed effect.
csdid is showing that by refusing to produce results
csdid2 is a bit smarter and tries to drop all Collinear variables before running a model
hth

Fernando

oh FernandoRios, YES indeed. Why do we inadvertently (or conveniently) need to stick fixed effects in all specifications? LOL It was a stupid question, but I am glad I asked. Thanks a lot for clarifying. (BUT it did work before...)

[Anne Simon]

Dear Fernando,

someone already asked in March 2022 whether it is possible to interact the treatment variable in csdid with another variable, e.g. gender, and you replied that this is not possible. That's why I tried to use sample splits and compare treatment effects across different groups (e.g. male vs. female) by first estimating the treatment effect for each group separately and then testing, whether the difference in treatment effects between the two groups is statistically significant.

However, the code that I posted below does not seem to work - I keep getting an error message when using the suest command (and I found online that suest may not work with all user-written packages - but I may also be doing something wrong):

unable to generate scores for model Female
suest requires that predict allow the score option
r(322);

   
csdid     $outcome if female==1 [iw = wgt], time(tvar_new) gvar(gvar_new) notyet method(reg) agg(simple)
    estat simple
    qui: matrix col2=r(table)
    qui: matrix list col2
    eststo Female
   
    csdid     $outcome if female == 0 [iw = wgt], time(tvar_new) gvar(gvar_new) notyet method(reg) agg(simple)
    estat simple
    qui: matrix col3=r(table)
    qui: matrix list col3
    eststo Male

suest Female Male, vce(cl year_district_id)
test ...

Do you have any advice on how I can use the csdid results and compare them with a post-estimation command?

Thanks a lot in advance!

Kind regards,
Anne

[FernandoRios]

Hi Anne,
No easily answer unfortunately.
If you want to have the ability of further clustering results (as you seem to want to do ), you have to create separate RIFs for men and women, merge samples, and then try draw inference.
I have a pdf that perhaps may be useful
F

[Anne Simon]

Hi Fernando,

that's exactly what I needed, thanks a lot. Of course it's a bit more complicated but looks like a neat solution - will try it that way!

Best,
Anne

[Anil Rup]

Thank you so much for writing the wonderful csdid FernandoRios! I started using it recently and it is just great! This may be a very simple question but I am struggling to interpret various pre-trend tests that the program spit out. I have following results after estimation:
estat all
Test will be based on asymptotic VCoV
If you want aggregations based on WB, use option saverif() ad csdid_stats
Pretrend Test. H0 All Pre-treatment are equal to 0
chi2(45) = 81.7403
p-value = 0.0007

and then following:
*Producing a simple Plot for last results. Two different styles
. estat event, window(-9 9)
Test will be based on asymptotic VCoV
If you want aggregations based on WB, use option saverif() ad csdid_stats
ATT by Periods Before and After treatment
Event Study:Dynamic effects
------------------------------------------------------------------------------
| Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
Pre_avg | 1.377292 1.249099 1.10 0.270 -1.070898 3.825482
Post_avg | 12.49752 5.192085 2.41 0.016 2.32122 22.67382
Tm9 | -8.080732 7.265344 -1.11 0.266 -22.32054 6.15908
Tm8 | 5.598478 5.353457 1.05 0.296 -4.894104 16.09106
Tm7 | -2.086951 3.468995 -0.60 0.547 -8.886055 4.712154
Tm6 | 4.414266 3.246138 1.36 0.174 -1.948048 10.77658
Tm5 | -1.489702 2.190728 -0.68 0.497 -5.783451 2.804047
Tm4 | .2051074 2.022997 0.10 0.919 -3.759894 4.170109
Tm3 | 3.786438 2.131636 1.78 0.076 -.3914916 7.964367
Tm2 | 8.519847 2.240238 3.80 0.000 4.129062 12.91063
Tm1 | 1.528876 1.548208 0.99 0.323 -1.505555 4.563307
Tp0 | -1.345033 1.9418 -0.69 0.489 -5.150891 2.460825
Tp1 | 4.807234 2.515641 1.91 0.056 -.1233311 9.737799
Tp2 | 8.066187 3.037854 2.66 0.008 2.112102 14.02027
Tp3 | 14.20465 3.46684 4.10 0.000 7.409768 20.99953
Tp4 | 13.06161 4.38758 2.98 0.003 4.46211 21.66111
Tp5 | 18.62453 5.329493 3.49 0.000 8.178916 29.07014
Tp6 | 21.59786 6.454643 3.35 0.001 8.946991 34.24873
Tp7 | 17.31846 8.301969 2.09 0.037 1.046902 33.59002
Tp8 | 13.76577 11.6833 1.18 0.239 -9.133085 36.66462
Tp9 | 14.87394 20.70136 0.72 0.472 -25.69998 55.44786

My understanding is that the first test above rejects PTA and the second the 'Pre-avg' does not reject PTA. Then you have, apparently, for PTA for individual years in the form of Tmx and for some of those years, PTA is rejected. Based on the first pre-test, I cannot really interpret my results as causal, and then based on Pre_avg, I can. Then I have individual years giving me mixed pre-tests results. Can you help me? Thank you so much.

[FernandoRios]

Well, the pre-trend test is very restrictive.
the one you see in the event aggregation is not a pre-trend test perse. Just shows the average effect over all pre-trend periods.
Perhaps something you can do is simply test all pre-trend values using -test-
HTH
fernando