Triple difference

[abshev]

I'm running a triple difference analysis with staggered treatments. It is straightforward to obtain point estimates for triple difference estimators; however, it is not clear to me how to use the did package for inference in this setting. I found some slides from Dr. Sant'anna that suggest making use of the influence functions and the multiplier bootstrap for this purpose, but the details aren't given. I was wondering if you could shed some light on how one might do this with the did package.

[raffaelemancuso]

Hello,

It is straightforward to obtain point estimates for triple difference estimators;

It is not so straightforward to me how to obtain point estimates for triple difference estimators

Would you mind explaining to me how you did it?

[0trm]

did has no DDD estimator, but triplediff (Ortiz-Villavicencio and Sant'Anna, on CRAN) does. It uses the same interface as att_gt and has multiplier-bootstrap inference built in:

library(triplediff)

res <- ddd(yname = "y", tname = "year", idname = "id",
           gname = "first_treat",   # first period treated; 0 or Inf if never
           pname = "eligible",      # 1 for units in the eligible partition
           xformla = ~ x1 + x2, data = df,
           control_group = "nevertreated", base_period = "varying", est_method = "dr",
           boot = TRUE, nboot = 999, cband = TRUE, cluster = "state")
summary(res)
es <- agg_ddd(res, type = "eventstudy")
summary(es)

It changes the point estimates too, not only the standard errors. Ortiz-Villavicencio and Sant'Anna (2025) show that the difference of two DiDs is generally invalid when identification needs covariates, and that with staggered adoption, pooling all not-yet-treated units as the comparison group can add bias. So subtracting two att_gt runs isn't a valid DDD once xformla has covariates. Checked on triplediff 0.2.4.