How to apply csdid when panel identifier switch treatment status

[Rui Dong]

Hi, I ran into a weird situation in csdid:
I have listed firm - year panel data, however, my treatment group is accounting firms.
Listed firms sometimes switch accounting firms, therefore, my panel identifier's treament status change.
I put listed firm in ivar(), and put treatment time of accounting firms in gvar(), is this the correct way?
I also want to add accounting firm fixed effects in the regression, however, when I add i.accounting firm, all coefficients become omitted. Is this normal?

[FernandoRios]

Hi Rui,
1. This wont work. the Ivar() has to identify a panel. In your case, seems that you will not be identifying the panel.
However, your GVAR can identify based on the Firm panel, with the caveat that once a unit is treated, remains treated. There is no switching status with CSDID
2. You could add fixed effects in the regression, but only if there is balance and you use method (reg) for estimation.
Other methods do not perform well because of the estimation of too many parameters. (Incidental parameter problem)
In terms of balance, you NEED each subgroup in the fixed effect list to affect ALL treated and control groups. Otherwise you violate overlapping assumption. This reflects into no effects being estimated.
Fernando

[Rui Dong]

Hi Rui,
1. This wont work. the Ivar() has to identify a panel. In your case, seems that you will not be identifying the panel.
However, your GVAR can identify based on the Firm panel, with the caveat that once a unit is treated, remains treated. There is no switching status with CSDID
2. You could add fixed effects in the regression, but only if there is balance and you use method (reg) for estimation.
Other methods do not perform well because of the estimation of too many parameters. (Incidental parameter problem)
In terms of balance, you NEED each subgroup in the fixed effect list to affect ALL treated and control groups. Otherwise you violate overlapping assumption. This reflects into no effects being estimated.
Fernando

Thanks so much for your reply, professor. I am still a little bit confused, I think this is because I did a poor job explaining my question:

My main dataset is a listed firm-year panel, and these listed firms are audited by accounting firms annually.
The treaments happen to accounting firms instead of listed firms, for example, Apple and Tesla were audited by Deloitte in 2018, then Deloitte got treated in 2019, apple decided to fire Deloitte and hire KPMG instead in 2021. In the data, apple appear to be in the treated group in 2019 and 2020 (audited by Deloitte which is treated) but in the control group in 2021 (audited by KPMG which is not treated).

I tried to drop ivar, use the year which the listed firm's current accounting firm get treated as gvar, and add listed firm fixed effects in the regression, the coefficients are omitted. Maybe I should drop the listed firms that switch accounting firms to avoid this situation? Please let me know if you have any suggestions.

[FernandoRios]

Ok, Seems to me CSDID is not the way go for your setup. Perhaps you need to look into did_mutliplegt_dyn, which allows you to have a time varying treatment.
I have not worked on the command, nor gone over the paper details, so you may want to read the original paper to start

[Rui Dong]

Ok, Seems to me CSDID is not the way go for your setup. Perhaps you need to look into did_mutliplegt_dyn, which allows you to have a time varying treatment.
I have not worked on the command, nor gone over the paper details, so you may want to read the original paper to start

Thanks so much, professor, I will look into it.