The methods are described here: Lee and Wooldridge (2025)
We provide methods similar to, and motivated by, Callaway and Sant'Anna (2021, Journal of Econometrics). Initially, the main differences is we use a "lags only" approach, which averages all pre-treatment periods to obtain a reference outcome. CS (2021) use the period just before the intervention for any treated cohort. Both estimators employ conditional parallel trends assumptions. We then extend our approach to show that removing unit-specific trends and unit-specific seasonality allows for more general selection into treatment. In particular, units trending differently in the absences of the intervention can have different cohort treatment probabilities. In our empirical example, the effects of Walmart sitings on county-level retail and wholesale employment, removing county-specific annual trends largely removes the large difference in pre-trends. We apply the doubly robust estimator IPWRA after the county-specific detrending. If we do not remove the trends, none of the popular estimators solves the violation of conditional PT.
The estimation choices are regression adjustment (ra), ipw, and ipwra. Our paper discusses how one could use matching or machine learning with high-dimensional covariates, and we might try to incorporate that in the future.
lwdid defaults to large-sample inference using cluster-robust inference. But to be reliable, this assumes a minimal number of treated units per cohort. In applications with, say, N = 50 states and many years, staggered rollouts sometimes have only a small number of new treated units in some years. Often, some cohorts g have Ng = 1. In such cases, one can use the "small" option with lwdid to collapse the problem to either one or a sequence of cross-sectional difference-in-means regressions -- with or without detrending -- and perform exact inference under normality. lwdid also accommodates randomization inference.
The format of lwdid is very similar to csdid and jwdid (@FernandoRios). Almost all credit for the command goes to Soo Jeong Lee.
Comments are welcome!
We provide methods similar to, and motivated by, Callaway and Sant'Anna (2021, Journal of Econometrics). Initially, the main differences is we use a "lags only" approach, which averages all pre-treatment periods to obtain a reference outcome. CS (2021) use the period just before the intervention for any treated cohort. Both estimators employ conditional parallel trends assumptions. We then extend our approach to show that removing unit-specific trends and unit-specific seasonality allows for more general selection into treatment. In particular, units trending differently in the absences of the intervention can have different cohort treatment probabilities. In our empirical example, the effects of Walmart sitings on county-level retail and wholesale employment, removing county-specific annual trends largely removes the large difference in pre-trends. We apply the doubly robust estimator IPWRA after the county-specific detrending. If we do not remove the trends, none of the popular estimators solves the violation of conditional PT.
The estimation choices are regression adjustment (ra), ipw, and ipwra. Our paper discusses how one could use matching or machine learning with high-dimensional covariates, and we might try to incorporate that in the future.
lwdid defaults to large-sample inference using cluster-robust inference. But to be reliable, this assumes a minimal number of treated units per cohort. In applications with, say, N = 50 states and many years, staggered rollouts sometimes have only a small number of new treated units in some years. Often, some cohorts g have Ng = 1. In such cases, one can use the "small" option with lwdid to collapse the problem to either one or a sequence of cross-sectional difference-in-means regressions -- with or without detrending -- and perform exact inference under normality. lwdid also accommodates randomization inference.
The format of lwdid is very similar to csdid and jwdid (@FernandoRios). Almost all credit for the command goes to Soo Jeong Lee.
Comments are welcome!
