[Ridwan Sheikh]

T

[FernandoRios]

Ahhh, ok this finally explains the problem.
and sorry i didn't get this in your previous explanations.
So, when i ask you to look into the e(gtt) was to see how many observations were being used per ATTGT. It didn't occurred to me that you were working with state level panel data.
Based on the number of cohorts, it seems to me that many cohorts have only 1 or 2 treated states . That may not be enough for the analysis, specially if you further constrain the data to be fully balanced.
It is like trying to estimate variances based on 1 or 2 observations. They will not be reliable. (this is infact what may be happening in the background.

Let me know if this makes sense
Best
Fernando

[Ridwan Sheikh]

Thank you FernandoRios , this helps me understand it now. I think that is right.
Yes, i am working on a state-level panel data and i have just 1 or 2 states in treatment group in some cohorts (usually the cohorts from G₂₀₀₄ - G₂₀₁₅ have just 1 or 2 states) . As of now i am considering all the states that belong to each treatment cohort from G₁₉₉₄ - G₂₀₁₅ . But most of the states adopted the said legislation between 1994-2004 (only few states 1 or 2 after 2004). Therefore, if i only consider cohorts between G₁₉₉₄ - G₂₀₀₄ , i can have significant no. of states in each cohort then . I think this may circumvent the problem and improve the significance of group-time, event-study and calendar-time coefficients.
I hope my intuition to it seems right ?

Again, Thank you very much. You had been awesome.
I may be a slow learner (sorry)

Regards,
(Ridwan)

[Ridwan Sheikh]

Dear FernandoRios .
I have one last question about event-study plots.
I was running the following code to get the plot:[

csdid y if first_legislated !=0 , cluster(state) time(year) gvar(first_legislated) notyet method(dripw) saverif(A1)

use A1, clear
csdid_stats event, wboot estore(event)
esttab event, se
csdid_plot , style(rcap)

This generates the plot but the x-axis (periods to treatment) gives event-time with big gaps like (-20, -10, 0, 10, 20). How to generate the plot with consecutive event times without gaps .
I want the x-axis (periods to treatment) something like (-20, -19, -18, -17.....0, 1, 2, 3........17, 18, 19, 20)

[FernandoRios]

Hi Ridwa
sorry didn’t reply this
so, if the continuous periods are not showing up it suggests that there aren’t any to plot
otherwise you could simply grab the output after estat event and plot it using other commands including coefplot

[Emily Rata]

Hi, I am running a repeated cross-section on the British Household Survey Panel and want to use this method for the DD with different treatment timing. "Pidp" identifies the individuals, while Year the time variable. My treatment is whether the individual is in temporary or permanent contract (variable is "job"). The simplest form is the following:

csdid index2, ivar(pidp) time(Year) gvar(job) reg

However, I got the following error:

request may not be combined with by
r(190);


Any clue on why is this happening?

Thanks,
Emily.

[Emily Rata]

Hi, I am running a repeated cross-section on the British Household Survey Panel and want to use this method for the DD with different treatment timing. "Pidp" identifies the individuals, while Year the time variable. My treatment is whether the individual is in temporary or permanent contract (variable is "job"). The simplest form is the following:

csdid index2, ivar(pidp) time(Year) gvar(job) reg

However, I got the following error:

request may not be combined with by
r(190);


Any clue on why is this happening?

Thanks,
Emily.

[FernandoRios]

Hi there
a few pointers
1. instead of "reg" you should me typing method(reg)
2. Is job a binary variable? Remember that it should be thee year when that person received treatment, and 0 if it was never treated.

Fernando

[ZEYNEP YAVIC]

Hello @FernandoRios,

Thank you very much for developing the csdid command. I have a quick question regarding diff in diff with multiple treatment times. I am interested in the moderation effect of the treatment, not its direct effect on the DV. If I had a 2X2 traditional diff in diff structure with a single treatment time, I'd probably use a triple interaction to identify the moderation effect (timedummy*treatmentdummy*moderator). Does this command allow for an interaction term (similar to a triple interaction in a regular diff in diff regression), and if not, do you know any other method that can help me identify how treatment moderates the relationship between my independent and dependent variables with a diff in diff - multiple treatment times?

Thank you,
Zeynep

[FernandoRios]

Hi Zeybnep
unfortunately no, The way csdid operates, it doesnt allow for triple interactions.
For what you are thinking, perhaps the most appropriate approach would be to use something like the regression approach by Wooldridge(2021), or the two step approach of Gardner (2021)
There is also did_imputation
HTH
F

[ZEYNEP YAVIC]

Thank you very much, for your quick reply and helpful recommendations @FernandoRios!

I have a follow-up question if don't mind. I did try to use did_imputation with interaction terms but I was not able to figure out how to do it with did_imputation either. Do you know if it allows for interaction terms?

Thank you,
Zeynep

[FernandoRios]

no, i havent used did_imputation myself, except for some comparison examples.
The author, however, seems to be quite accessible. You could send him an email with questions too

[Leon Schmidt]

Thanks a lot for publishing this command! One clarifying question: If I understand the general approach correctly, the option "notyet" uses only the non-yet treated units as control group (i.e., it excludes the never treated units as control group). However, from reading the helpfile I get the impression that "notyet" invokes that both the never and not yet treated units are used as controls. Which of the two options is correct? And would it be OK to use both groups as comparison? And if one only wants the not yet treated as controls, should I drop all observations that never get treated before running the command? Thanks again!

[FernandoRios]

Hi Leon,
I think that is a matter of semantics. But will try to clarify in the future.
If you look into Callaway and Sant'Anna (2021), it suggests the group never treated is such that \[ G=\infty \[. This implicitly says that those not treated, may be treated at some point in the unforeseen future. Thus, for all practical purposes is also Not Yet treated!.

Thus, when you do not use notyet, it uses only those not treated in the span of the data.
If you use notyet, it also includes as controls those observations not yet treated. I don't think there is anything wrong about using either one. But you can justify it either way.
And, if you want to exclude those not treated (ever) you can simply type " if gvar!=0" thus excluding those observations.

HTH
Fernando

[Leon Schmidt]

Thank you very much for clarifying this and all the great work, Fernando!

Ah OK, that makes sense! I always thought that never-treated and notyet-treated are distinct groups, but it makes sense that notyet includes never-treated units (within the period of analysis). Since different parallel trends assumptions hold when using the comparison different groups, do you know which parallel trends assumption holds when specifying "if gvar != 0", i.e., only looking at switchers (including only notyet treated units and removing the never treated units)? My guess is that the "stronger" assumption in their paper holds, but I´m not sure.

All the best
Leon