fastdid - fast Difference-in-Differences

fastdid is a lightning-fast implementation of Callaway and Sant'Anna's (2021) staggered Difference-in-Differences (DiD) estimators. DiD setup for millions of units used to take hours to run. With fastdid, it takes seconds.

To learn more about the staggered Difference-in-differences estimators implemented, visit Callaway and Sant'Anna's website.

Installation

You can install fastdid from GitHub.

# install.packages("devtools")
devtools::install_github("TsaiLintung/fastdid")

Usage

fastdid is the main function provided by fastdid. When using fastdid, you need to provide the dataset (data), specify the names of the relevant columns (-var), and the type of target (aggregated) parameters (result_type such as "group_time", "time", "dynamic", or "simple".) Here is a simple call.

#loading the package
library(fastdid)

#generate simulated data
simdt <- sim_did(1e+03, 10, cov = "cont", second_cov = TRUE, second_outcome = TRUE)
dt <- simdt$dt

#calling fastdid
result <- fastdid(data = dt, #the dataset
                  timevar = "time", cohortvar = "G", unitvar = "unit", outcomevar = "y", #name of the columns
                  result_type = "group_time") #the result type

You can control for covariates by providing the name of the data columns, and choose the method among doubly-robust ("dr"), inverse probability weight ("ipw"), and outcome regression ("or").

result <- fastdid(data = dt, 
                  timevar = "time", cohortvar = "G", unitvar = "unit", outcomevar = "y",
                  result_type = "group_time",
                  control_option = "dr", #choose the control method
                  covaraitesvar = c("x", "x2")) #add covariates

While the default is to coerce the data into a balanced panel, you can allow for unbalanced panel. Note that currently only "ipw" is available when allow_unbalance_panel = TRUE.

result <- fastdid(data = dt, 
                  timevar = "time", cohortvar = "G", unitvar = "unit", outcomevar = "y",
                  result_type = "group_time",
                  allow_unbalance_panel = TRUE, #allow for unbalanced panel
                  covaraitesvar = c("x", "x2"))

Clustered standard error can be obtained from multiplier bootstrap.

result <- fastdid(data = dt,
                  timevar = "time", cohortvar = "G", unitvar = "unit", outcomevar = "y",
                  result_type = "group_time",
                  clustervar = "x", boot = TRUE) #add clustering by using bootstrap

Estimation for multiple outcomes can be done in one call by providing a vector of outcome column names (saves a lot of time when controlling for covariates since logit estimates can be recycled across outcomes).

#calling fastdid
result <- fastdid(data = dt, #the dataset
                  timevar = "time", cohortvar = "G", unitvar = "unit", outcomevar = c("y", "y2"), #name of the columns
                  result_type = "group_time") #the result type

Performance

fastdid is magnitudes faster than did, and 15x faster than the fastest alternative DiDforBigData for large dataset.

Here is a comparison of run time for fastdid, did, and DiDforBigData (dfbd for short) using a panel of 10 periods and varying samples sizes.

Unfortunately, the Author's computer fails to run did at 1 million sample. For a rough idea, DiDforBigData is about 100x faster than did in Bradley Setzler's benchmark. Other staggered DiD implementations are even slower than did.

fastdid also uses less memory.

For the benchmark, a baseline group-time ATT is estimated with no covariates control, no bootstrap, and no explicit parallelization. Computing time is measured by microbenchmark and peak RAM by peakRAM.

fastdid and did

As the name suggests, fastdid's goal is to be fast did. Besides performance, here are some comparisons between the two packages.

Estimates

fastdid's estimators is identical to did's. As the performance gains mostly come from efficient data manipulation, the key estimation implementations are analogous. For example, 2x2 DiD (estimate_did.R and DRDID::std_ipw_did_panel), influence function from weights (aggregate_gt.R/get_weight_influence, compute.aggte.R/wif), and multiplier bootstrap (get_se.R and mboot.R).

Therefore, the estimates are practically identical. For point estimates, the difference is negligible (smaller than 1e-12), and is most likely the result of floating-point error. For standard errors, the estimates can be slightly different sometimes, but the difference never exceeds 1% of did's standard error estimates.

Interface

fastdid should feel very similar to att_gt. But there are a few differences:

Control group option:

fastdid	did	control group used
both	notyettreated	never-treated + not-yet-but-eventually-treated
never	nevertreated	never-treated
notyet		not-yet-but-eventually-treated

Aggregated parameters: fastdid aggregates in the same function.

fastdid	did
group_time	no aggregation
dynamic	dynamic
time	calendar
group	group
simple	simple

Other

1. fastdid only offers inverse probability weights estimators for controlling for covariates when allowing for unbalanced panels.
2. fastdid use universal base periods as default.
3. fastdid only reports the pointwise confidence intervals, instead of the simultaneously valid confidence intervals (check section 4.1 of Callaway and Sant'Anna's (2021) for more detail.)

Roadmap

fastdid is still in active development. Many features are planned to be added:

• Multiple outcomes ✅
• Min/max event time and balanced composition ✅
• DR and OR estimators ✅
• Allowing for unbalanced panels ✅ (well, not fully because DR and OR still need to be added)
• Anticipation ✅
• Varying base periods ✅
• User-provided aggregation scheme
• User-provided control formula
• simultaneously valid confidence bands
• Further optimization

Source version

Since fastdid is not on CRAN yet, it needs to be converted to R scripts to be used in some restricted environments. This can be done with development/build_source.R. After changing the working directory, the script will produce development/fastdid_VERNAME.R, which can be sourced to mimic the functionalities of the package.

Update

0.9.3 (2024/5/7)

• add anticipation and varying base period option
• add min and max control cohort difference
• add time-varying control (reference)
• add filtervar

0.9.3.1 (2024/5/24): fix the bug with univar == clustervar (TODO: address problems with name-changing and collision)

0.9.2 (2023/12/20)

• add support to doubly robust and outcome regression estimators
• add support to unbalanced panels (simple and ipw only)
• add support to balanced composition option in dynamics aggregation
• fixed argument checking that was not working properly
• set the default to copying the entire dataset to avoid unexpected modification of the original data (thanks @grantmcdermott for the suggestion.)

0.9.1 (2023/10/20)

• now supprts estimation for multiple outcomes in one go!
• data validation: no longer check missing values for columns not used.

Acknowledgments

fastdid is created by Lin-Tung Tsai, Maxwell Kellogg, and Kuan-Ju Tseng.