Event study in Stata about lost-stats.github.io

[grantmcdermott]

I think you're correct. It doesn't make sense to have time_to_treatment==0 represent two different states. (Schrödringer's zero?) Now, from a practical perspective, I don't think it matters since the treat and ib#shited_ttt variables will cancel things it. But I agree it's logically confusing.

Taking a higher-level look at this event study page, there are several other things we could change. We shouldn't be using a proprietary data format (dta) for one thing, and the R code as currently written won't work with the impending fixest 0.9.0 release. (It could also be considerably shortened.) I'll make some changes and then let someone else pitch in about the Stata code.

[grantmcdermott]

I've pushed a couple of updates. But really I think this page needs a closer look. I just ran the Stata implementation of Sun Abraham and got different results to the R version... I don't have time to troubleshoot now and it might just be a copy-paste error on my side. But I'm hoping @pdeffebach or @NickCH-K can double check.

Similarly, I tried the Callaway-Sant'Anna estimator and got different results once again. I spoke v. briefly to Pedro and he thinks it's likely due to covariates, which SA don't handle in their paper. Again, though, it could be user error on my part. Here's the code I was trying locally:

library(did)

## Change never treated to zero
dat[, year_treated := ifelse(treat==0, 0, `_nfd`)]

mod_cs = att_gt(yname = "asmrs",
gname = "year_treated",
idname = "stfips",
tname = "year",
xformla = ~ pcinc + asmrh + cases,
data = dat)

summary(mod_cs)

cs = aggte(mod_cs, type = "dynamic", na.rm = TRUE)
cs

ggdid(cs)

[NickCH-K]

I'll look at this when I get a chance

[NickCH-K]

Shifting the untreated treated-time to a big positive number seems fine to me; I do like the logic that the zeroes there let things drop out.

On the R side the sunab() function looks neat! It's a bit of a concern it doesn't match Stata, as the Stata package is by Sun herself.

[aalamer1]

Yes. You have to assign to very positive number that is not included in the wanted interval. Keeping this as it is will apply the regression for states that treat=0 while we just want to include states with treat=1. I tried it and got different results.

[NickCH-K]

@aalamer1 do you have a fixed version of the code?

[aalamer1]

@NickCH-K I am working on my own dataset but modified replace time_to_treat = 0 if missing(_nfd)
to be replace time_to_treat = 10000 if missing(_nfd).

Did you try in R code they posted? I tried to replicate it but generate errors of collinearity it is not working. I do not know how did they generate the graph and confirmed the results?

[Swarup8289]

Hi all: In the code:
reghdfe asmrs ib`true_neg1'.shifted_ttt pcinc asmrh cases, a(stfips year) vce(cluster stfips)

Shouldnt there be the interaction term of " treat## ibtrue_neg1'.shifted_ttt" instead of just "ibtrue_neg1'.shifted_ttt" ?

Thanks in advance!

[NickCH-K]

Yes, I believe you're right