CSDID (Bootstrap C.I. vs. Asymptotic normal C.I.)

[Marlene Fiesole]

Hi,

I recently started using the CSDID command with the agg(event) option, namely event aggregation.
I would like to have the usual event plot with the resulting coefficients and the Wild Boostrap Confidence intervals.
I know that an easy way would be to use csdid_plot,but since the command has not many options to personalize the graph (e.g. decide how many leads and lags to include) I decided to opt for the event_plot command. My main issue is that event_plot doesn't graph the wild bootstrap confidence intervals but the default asymptotic normal ones.

Two main questions:
1) Is there an easy way to obtain an event plot with wild boostrap Confidence intervals and decide how many lags/leads to show?
2) If Callaway Sant'Anna 2021 suggest the usage of wild boostrap, when and why should we think to use asymptotic normal confidence intervals that are default in CSDID?

Thanks in advanced for your help!

[FernandoRios]

1. It is better to use csdid to create the attgts and save the rifs then create the event plots with the option window
csdid_stat event, window(#1 #2)
then use csdid plot
2. this are two different things All together
If you are concerned about multiple testing adjustment
then wild bootstrap is better.
other wise asymptotic CI is what other papers do already
f

[Marlene Fiesole]

Hi Fernando,

thanks a lot, this really helps!

I have another question related to csdid_plot.

Opposed to other commands (e.g. event_plot) csdid_plot seems to estimate also the coefficient and confidence intervals for the period before treatment, usually referred to as "reference period" in t = g-1, where g is the time where treatment switched on.

Could it be that actually it is omitting the reference period directly and plotting only the other coefficients? This latter option would make more sense to me.

How should I interpret this?

Once again thanks a lot for your help and time
Marlene

[FernandoRios]

It’s not cadid_plot
you need to estimate the model with long2 option

[Marlene Fiesole]

Thanks a lot!

[Melody Burns]

It’s not cadid_plot
you need to estimate the model with long2 option

If long2 option is not used, how do we interpret the results in the csdid_plot?

[FernandoRios]

the -pre-treatment- are SHORT differences. If even one is Significant, there is a violation of PTA.
For Post-treatment, interpretation is as usual

[Melody Burns]

Thank you!

[paulvonhippel]

I am finding that csdid is using the wild bootstrap, even though the help file says it produces analytic standard errors by default. How do I get analytic standard errors and prevent it from running the bootstrap? Here is my code, thanks!

csdid obese ib(5).ethnicity_code, time(spring) gvar(groupyear) agg(simple) cluster(school_id) dripw notyet

If it helps, I should say that children are clustered within schools.

[paulvonhippel]

I am also getting a warning: "Units always treated found. These will be ignored." Which seems odd to me since I thought I had excluded always-treated units from the data. Can csdid show me which units it's talking about? Thanks!

[paulvonhippel]

Finally, what do the dots mean in the output? After running csdid, I see dots like this:
......XXX......
In other commands, each dot refers to a single bootstrap iteration, and I've been assuming that's what they mean here. If they mean something else that would be good to know.

[FernandoRios]

CSDID is a command that runs a large number of DRDID models. If a particular iteration was successful, you will see a "dot". "X's" on the other hand represent failed cases.
To obtain Wbootstrap standard errors you would need to add "wboot" option.

Regarding always treated. It will depend on your tvar and gvar variables. If you can Tabulate them, I can show them (But I will add that as a message for future iteration!)

[paulvonhippel]

Thank you, Fernando. Another question occurred to me after I first posted. Does this model include school fixed effects?

csdid obese ib(5).ethnicity_code, time(spring) gvar(groupyear) agg(simple) cluster(school_id) dripw notyet

Schools are identified by school_id, which appears in the cluster statement but not elsewhere.
I tried to add school fixed effects with the ivar(school_id) option, but was rejected because there is more than one observation (child) per school.

The corresponding TWFE model would include school fixed effects, so I'm surprised not to see them explicitly in CSDID.

[FernandoRios]

No, CS does not automatically add fixed effects. Instead, it implicitly uses Cohort Fixed effects (gvar) when using repeated crossection.
When using Panel data, the method uses First Differences, instead of fixed effects, but for 2x2 its the same.

[paulvonhippel]

Is there a danger of bias from omitting the school fixed effects? There are different numbers of children in each school.
I suppose I could add a dummy for each school. Any reason not to?