[FernandoRios]

Different methods all together
and difference can arise on how perfect prediction is handled
too many to type them right now
One thing you can do is check propensity scores for the controls

[Awais Farid Khan]

FernandoRios Thanks for your reply. I will check that...I am on an inner debate which to prefer efficiency or precision based on my data.....because the outcomes for all methods are same without covariates but differ when covariates are added.....just seraching for plausable justification for the use of methods....i will be obliged if you can point some of the pitfalls as signs of imperfect predictions.

My second question was about CSDID and CSDID2...I am prefering CSDID2 as it allows "pretrend" and "honestdid" options....

This is with CSDID

Difference-in-difference with Multiple Time Periods

Number of obs = 353
Outcome model : weighted mean
Treatment model: stabilized inverse probability

Coefficient Std. err. z P>z [95% conf. interval]

ATT 4.917568 1.705818 2.88 0.004 1.574226 8.26091

Control: Not yet Treated


this is with CSDID2

Difference-in-difference with Multiple Time Periods
Outcome model : weighted mean
Treatment model: stabilized inverse probability
------------------------------------------------------------------------------
| Robust
| Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
SimpleATT | 5.107128 1.79526 2.84 0.004 1.588483 8.625772
------------------------------------------------------------------------------

Is this difference within acceptable range that we can expect....your comments will be most appreciated.

[FernandoRios]

It really depends on collinearity
csdid depends.on stata approach to detect and drop collineae variables
cadid2 uses a different method.
That suggest your specificat9may need addressing

[Awais Farid Khan]

@FernandoRios ....it there a way to use pretrends and honesdid commands after CSDID?...

after csdid-pretrends
Coefficient vector must be at least # pre + # post

after csdid - honestdid
Coefficient vector must be at least # pre + # post

[Awais Farid Khan]

I resolved the issue how to use csdid , pretrends and honestdid together...thanks

[Michael Cooper]

Hello FernandoRios ,

I am working on a project updating our traditional OLS TWFE model results to use csdid2. In our context, we have a model where it makes sense to exclude time periods while a unit is being treated. The reason for this is so that we can consistently interpret Lead1 as the quarter before treatment started, and Lag1 as the quarter after treatment ended, regardless of individual variations in time spent being treated. Using csdid, this method did not appear to work, as either the model returned only "x"'s when the periods were excluded, or gives event study values with a different interpretation when using an "if" command. My way around this was creating new time variables relative to treatment after excluding the periods so that there are always observations relative to the treatment time.

My two questions are:
1. Does excluding particular time periods in the panel sound problematic for addressing the TWFE bias and producing event studies using csdid?
2. How might time fixed effects be impacted in this estimation? For instance, I was adding a control for the original time fixed effects back into the model after adjusting for the excluded periods since I still wanted the model to be adjusted for real time shocks:

csdid2 outcome i.calendar_time , ivar(ID) time(csdid_time) gvar(csdid_treat) long2 method(reg)

Here is an example of how the data would look after this adjustment, with leads shown as an example:

Original
ID	Treat Start	Treat End	Time	Lead1	Lead2	Lead3
1	2001q3	2002q1	2001q1	0	0	0
1	2001q3	2002q1	2001q2	0	0	1
1	2001q3	2002q1	2001q3	0	1	0
1	2001q3	2002q1	2001q4	1	0	0
1	2001q3	2002q1	2002q1	0	0	0
1	2001q3	2002q1	2002q2	0	0	0
1	2001q3	2002q1	2002q3	0	0	0
2	2002q1	2002q1	2001q1	0	0	0
2	2002q1	2002q1	2001q2	0	0	1
2	2002q1	2002q1	2001q3	0	1	0
2	2002q1	2002q1	2001q4	1	0	0
2	2002q1	2002q1	2002q1	0	0	0
2	2002q1	2002q1	2002q2	0	0	0
2	2002q1	2002q1	2002q3	0	0	0
Treatment Periods Excluded
ID	Treat Start	Treat End	Time	csdid Time	csdid treat	Lead1	Lead2	Lead3
1	2001q3	2002q1	2001q1	1	3	0	1	0
1	2001q3	2002q1	2001q2	2	3	1	0	0
1	2001q3	2002q1	2002q2	3	3	0	0	0
1	2001q3	2002q1	2002q3	4	3	0	0	0
2	2002q1	2002q1	2001q1	1	5	0	0	0
2	2002q1	2002q1	2001q2	2	5	0	0	1
2	2002q1	2002q1	2001q3	3	5	0	1	0
2	2002q1	2002q1	2001q4	4	5	1	0	0
2	2002q1	2002q1	2002q2	5	5	0	0	0
2	2002q1	2002q1	2002q3	6	5	0	0	0

Thank you for your work and for any help with these questions!

[FernandoRios]

Well
csdid works under the assumption you have complete data
or at least regular spaced data
Onve it ids the unit of.time (evry year or evey couple.of years or every 5 years) thenit ids every thing else

In your casse because you are mixing dates, it may be trying to find.periods that do not exist giving you the Xs

[Michael Cooper]

Well
csdid works under the assumption you have complete data
or at least regular spaced data
Onve it ids the unit of.time (evry year or evey couple.of years or every 5 years) thenit ids every thing else

In your casse because you are mixing dates, it may be trying to find.periods that do not exist giving you the Xs

Thank you for the response! That is correct, there will be periods that won't exist and the data will have to be non-regularly spaced to exclude those quarters. Do you think csdid can still be used appropriately in this circumstance, given my workaround of creating a new, consistent time variable and adding the original calendar time fixed effects as a regression control?

[FernandoRios]

No, i dont think it will.
Creating the new time variable works, but adding calendar fixed effects will most likely do nothing (or at least have unforseen consequences)

[Michael Cooper]

Okay, that is what I expected. With our models that included a comparison group, this workaround seemed fine and consistent with other methods (the "unforeseen consequences I'm assuming were just balanced across both groups), but when we applied it to a staggered-only identification with no never-treated and no comparison group, we started to get strange results that were very sensitive to small changes.

One last question: when using long differences, csdid/csdid2 naturally uses t-1 as the default baseline period. I did not see a way to change this to use a different reference period. Is this because the method requires comparing to the period immediately before treatment for long differences, or is it just an uncommon request that hasn't been added to the program?

[FernandoRios]

THe second.
with csdid it was a bit more complicated to add an "anticipation" option.
with csdid2, there is one...but i do not remember what it is...(and currently have no access to github from work computer ...non academic work)
I ll try to post tonight the option
F

[Michael Cooper]

Got it, thanks! I don't see an anticipation option, but it is fairly straightforward to adjust the gvar manually to build in an anticipation effect.

We were actually implementing a more niche case: instead of modeling a treatment effect with anticipation, we only adjust the baseline period from -1 to a different period (purely for visualization purposes on a version of our event study plot). So preserving the "not-yet-treated" status of the other periods so that they can still be used as not-yet-treated comparisons.

[FernandoRios]

the option on csdid2 its antici(#).
This takes only positive numbers, and it "shifts" the treatment by # periods backwards. The plots adjust by also starting like that.
HTH

[David Sanders-Ellis]

Hi FernandoRios

I'm using csdid and have hit a bit of a conceptual wall I'm hoping you can help me get through.

I'm evaluating an education policy using individual annualised unbalanced panel data to see if it reduced school absences for children in care. The policy was adopted at a local area level (my condition) in different years, and I have data for 11 years, with 7 different annual adoption cohorts and a minimum of two pre-policy periods (with the last group never treated).

The issue I'm facing is that children tend to have more school absences when they are older. As such in the staggered event study, the earlier and later event times will have a naturally different sample composition e.g. in time -5 will only have 6 to 11 year olds, while in time +5 it will be only 11- 16 year olds. I'm using age as one of my covariates, but I don't know whether long or long2 was a better option for my main study. Am I correct that in this case, long2 is the better choice so long as parallel trends hold, as it uses the -1 universal base, which is closest to the 1st post-period?

Is the pretrend test sufficient to evidence this decision, or should I also look at the event study plot and check for CIs crossing the 0?

In my second anlysis, I'll be using the not yet treated option to leverage the data as much as possible. Does this change whether long or long2 is the better option?

Thanks in advance!

[FernandoRios]

Hi David
1. long and long2 will make no difference for what you have in mind (although long2 should be THE option for you to follow)
2. If it is a panel data, your Post and Pre groups are always the same. So each ATTGT is valid. However, when comparing the information across time you will have the issue that your data is suffering attrition, thus you cant really compare ATTGts across time.
3. Perhaps one way to avoid the problem is to use a group that is stable across time. And see if the results are consistent.
4. The point above really matters more for the treated group. Although for the non-treated ones, Pedro's Papers suggests that is a good option, you also have to keep in mind that this advice is under the assumption that "control groups" are similar across time.

HTH
F