CSDID question

[Enrique Oriundo]

Hi all,

I want to apply the Callaway and Sant'Anna estimator to evaluate the impact of microfinancial program (subsidy to the supply side) on the credit amount borrowed by commercial banks in Uganda (information comes from records of those banks, not from surveys). This program is implemented by stages, so the csdid is the method I want to apply. However, I am not quite sure about what the best way is. I have the following structure:

csdid credit_amount [iw=bank_size], ivar(bank_id) time(year) gvar(cohort) method(dripw)

where bank_size is the weight that I want to apply since the size of banks are heterogeneous in baseline, cohort =1 for the bank that enters into the program in year T, 0 for the never treated. So, my questions are as follows:
- Am I weighting correctly the bank size? If I am understanding well, the 'iw' option works as pweights, so based on my framework, is that correct?
- I do not understand the method option. I've been looking in internet/chatgpt about what the best option to choose is, since I do not fully understand the help file that says:

drimp Sant’Anna and Zhao (2020) Improved doubly robust DiD estimator based on inverse probability of tilting and weighted least squares.

dripw (default) Sant’Anna and Zhao (2020) doubly robust DiD estimator based on stabilized inverse probability weighting and ordinary least squares.

reg Outcome regression DiD estimator based on ordinary least squares. When no covariates are specified, this method is used as default, because it provides the same point estimates and standard errors.

stdipw Inverse probability weighting DiD estimator with stabilized weights

ipw Abadie (2005) inverse probability weighting DiD estimator. May not perform well with unbalanced panel data

rc1 In combination with the methods {cmd drimp} and {cmd dripw}, this option request the doubly robust but not locally efficient repeated crossection estimators. Not available when using panel data.

I know that I should decide between drimp or dripw, but I am not sure. FernandoRios do you have any comment/suggestion about it?

Thanks in advance!

[FernandoRios]

DRIPW is more stable that DRIMP. But both are Doubly RObust.
I would go DRIPW

[Enrique Oriundo]

Hi FernandoRios

Thanks for your reply. I still have a question about the estimator. I want to account the bigger banks to have a greater impact on the ATT, so is my specification doing that? I know that in other contexts, the iw is preferred, but in the CSDID context, the iw works internally as pw, so may I use it based on my goal?

Thanks again!

[Jeff Wooldridge]

Enrique: the Callaway-Sant'Anna approach allows heterogeneous effects, but you can't see those effects unless you do the estimation "by hand." It's a bit complicated, so instead I'd recommend using jwdid -- also written by Fernando -- which implements the regression-based estimators I proposed. The output includes the so-called moderating effects of any of the control variables, which I assume in your case includes a bank size measure. To make it a "leads and lags" (event study) analysis, include the "never" option.