Question about control variables when using JWDID

[Joe Dui]

I'm using the JWDID command to perform staggered DID estimation. I use not-yet treated and never treated samples as control group. I added some control variables, and got the results. But I don't know if my model is written correctly. It seems not right, because there are many results of interactions when I use JWDID. For example, cohort#year#c.__tr__#c._x_urban ； year#c.urban .

Thank you all.
The model I write：

[FernandoRios]

I would suggest checking the papers cited in the help file
In my paper that explains jwdid, we present with lots of details the correspondence of the different models

[Joe Dui]

I would suggest checking the papers cited in the help file
In my paper that explains jwdid, we present with lots of details the correspondence of the different models

Thank you Fernando. I will read these papers again.

[Jeff Wooldridge]

The only problem with your current equation is that it is too general. The extended TWFE equation I proposed has X interacted with just the cohort dummies, D(i,g), just the time dummies, and the cohort x time dummies in the treated periods. I'm attaching slides from a recent short course. See slide 34 for the equation in expectation form. If you use fixed effects, the time constant variables get dropped. slides_3_did_cemfi.pdf

[Joe Dui]

The only problem with your current equation is that it is too general. The extended TWFE equation I proposed has X interacted with just the cohort dummies, D(i,g), just the time dummies, and the cohort x time dummies in the treated periods. I'm attaching slides from a recent short course. See slide 34 for the equation in expectation form. If you use fixed effects, the time constant variables get dropped. [ATTACH]n1763968[/ATTACH]

Thank you Professor Wooldridge, your answer and the slides are really helpful.

[Joe Dui]

The only problem with your current equation is that it is too general. The extended TWFE equation I proposed has X interacted with just the cohort dummies, D(i,g), just the time dummies, and the cohort x time dummies in the treated periods. I'm attaching slides from a recent short course. See slide 34 for the equation in expectation form. If you use fixed effects, the time constant variables get dropped. [ATTACH]n1763968[/ATTACH]

I made some updates based on the slide. It seems better now, but I'm not sure if it is complete. I am posting it here for discussion. Thank you very much for taking the time to look at this.

[Jeff Wooldridge]

My one caution is that tau is commonly used to denote the treatment effects, and you're using on the time dummies. Also, I'd be consistent with the subscripts. X should have an i subscript everywhere.

[Joe Dui]

My one caution is that tau is commonly used to denote the treatment effects, and you're using on the time dummies. Also, I'd be consistent with the subscripts. X should have an i subscript everywhere.

Thank you so much for your guidance on my model Professor Wooldridge. Your insights have been invaluable to my research.

[Myat Thida Win]

FernandoRios I am studying conflict's impact on the rice mills. I want to use the continuous treatment variable which interacts conflict exposure with conflict intensity. I am wondering if jwdid supports continuous treatment. This is my code for the binary variable case:

jwdid ln_ttl_prod, ivar(id) tvar(time) gvar(cohort) cluster (ts)

where ln_ttl_prod is the log of production. I want to incorporate cohort_dummy x time_dummy x conflict_intensity. Please let me know. Thank you.

[FernandoRios]

yes it does. But only latest version.
Also intensity is allowed only under very strict assumptions (intensity is 0-1) and the impact within ATTGT is linear

You can of course make it even more flexible, manually interacting square and even cubic terms, but the model becomes more complex. and potentially harder to estiamte

[Myat Thida Win]

Thanks! I just updated jwdid in Stata, but still could not figure out how to apply for the continuous treatment variable. I have had several conversations with Dr. Wooldridge, and manually I have been calculating the following of the conflict effects:

xtreg ln_ttl_prod c.d7#(c.f7 c.f8 c.f9 c.f10 c.f11 c.f12 c.f13 c.f14 c.f15)#c.w ///
c.d8#(c.f8 c.f9 c.f10 c.f11 c.f12 c.f13 c.f14 c.f15)#c.w ///
c.d9#(c.f9 c.f10 c.f11 c.f12 c.f13 c.f14 c.f15)#c.w ///
c.d10#(c.f10 c.f11 c.f12 c.f13 c.f14 c.f15)#c.w ///
c.d13#(c.f13 c.f14 c.f15)#c.w ///
i.time,fe vce (cluster ts) ///

where d7-d13 are the cohort dummies, f7-f15 are the time dummies, and w is the time-varying total number of conflict incidents.

Could you please suggest how I should include w in jw command? Thank you.

[FernandoRios]

Sorry I should ahve added my latest notes and where to get latest version
https://friosavila.github.io/rnd_ppt...a#/title-slide
here is what I presented earlier this year, showing how to add continuous variables.
it would be pretty much what you already did, but with less of the typing more of the analysis.
F

[Myat Thida Win]

FernandoRios Thank you so much for sharing the ppt! It is super helpful. I think my case falls into case (2): treatment heterogeneity as the number of total conflicts exposed could be different even within each cohort group. I ran the following codes:

jwdid ln_ttl_prod, ivar(id) tvar(time) gvar(cohort) xattvar(w) cluster (ts)

where w is the time-varying total number of conflicts in each time period. I used both estat simple, ores(w), and estat simple to check ATE. The results are entirely different from what I calculated manually.

I also tried case (1): treatment intensity. I got the error message that I could only include either gvar(gvar) or trtvar(trtvar), and trtvar has to be between 0 and 1. The conflict is a total number of incidents. How should I adapt this to 0 and 1 in case my study falls under case #1?

Thank you. I really appreciate your help.

[Myat Thida Win]

FernandoRios thank you so much for your previous response and for sharing the material to add continuous variables.

I divided the total conflict in each round by the largest number of conflicts to get the conflict intensity ranging from 0 to 1. I added the intensity (in 0 to 1) using trtvar. I found the exact same results as manual calculation. Thanks so much.

I want to ask one more follow-up question. I would like to have an event-study plot. Does jwdid also support for the event study plot with continuous treatment variables? Please note that treatment is zero and intensity is also zero for pre-treatment rounds.

[FernandoRios]

It should but have not tested that