edid

The {edid} package implements the efficient difference-in-differences and event study estimators developed by Chen, Sant’Anna, and Xie (2025). I initially developed the code for a personal project, so the functionality is limited. I describe the limitations below. It also has several dependencies, though almost all are {tidyverse} packages. I imagine a more robust implementation will be released, but as far as I am aware, this is currently the only publicly available software for this class of estimators. I am happy to expand the functionality so please feel free to submit requests!

Installation

Install the package from this GitHub page by running the code below in R.

# install.packages("pak")
pak::pak("david-loeb/edid")

Overview

The package contains a series of functions designed to be run sequentially, beginning with data setup and ending with a data frame of results. The function edid() executes the entire process by running each of these functions under the hood. This is the most straightforward way to use the package. edid() requires the following set of arguments:

• Dataset
• Names of four variables in the dataset
  • y: The outcome variable
  • treat_time: A variable that specifies the time period in which each unit adopted the treatment; never treated units get a 0 on this variable
  • id: A unique identifier for each unit
  • time: The time variable
• num_t_pre: The number of pre-treatment periods to use for ATT estimation

The code below demonstrates how to use edid().

library(edid)

edid_results <- edid(
  edid_ex_data,
  y = "outcome",
  treat_time = "treat_adopt_time",
  id = "id",
  time = "time",
  num_t_pre = 3
)

The result will be a data frame with the efficient $ATT(g,t)$, event study, and calendar time aggregation results. You can also request a more complete set of model components, including the individual $ATT(g,t)$ and weights used to compute the efficient $ATT(g,t)$ by setting get_full_mod = TRUE in the edid() function. Additional options can be specified in edid() such as anticipation periods; see the documentation for details.

Plot results

The EDiD results can be plotted with the plot_edid() function. It uses {ggplot2} (and {patchwork} if plotting individual $ATT(g,t)$ ) under the hood. The following code demonstrates how to use plot_edid() and an example of the resulting plot. (Wouldn’t we all love to have results like this?)

# Event study plot
plot_edid(edid_results, type = "es")

The plots can be customized in a number of ways, including the confidence intervals to plot, color scheme, titles, point and line sizes, text sizes, font, and whether to display the legend. You can also save the plot result and customize it further with your own {ggplot2} pipeline. The plot below shows the full set of $ATT(g,t)$ results and all confidence intervals with some appearance customizations. (I increase the confidence interval size in the example to provide a more demonstrative plot.)

# Increase the size of CIs
edid_res_larger_CI <- edid_results |> 
  dplyr::mutate(
    ci_up_95_boot = ci_up_95_boot + se_boot * 100,
    ci_up_90_boot = ci_up_90_boot + se_boot * 50,
    ci_up_95_analytic = ci_up_95_analytic + se_analytic * 80,
    ci_up_90_analytic = ci_up_90_analytic + se_analytic * 30,
    ci_low_95_boot = ci_low_95_boot - se_boot * 100,
    ci_low_90_boot = ci_low_90_boot - se_boot * 50,
    ci_low_95_analytic = ci_low_95_analytic - se_analytic * 80,
    ci_low_90_analytic = ci_low_90_analytic - se_analytic * 30
  )

plot_edid(
  edid_res_larger_CI, 
  type = "attgt", 
  ci = "all", 
  point_size = 2.5,
  error_bar_vline_width = 1,
  text_size_base = 10
)

See the plot_edid() documentation for more details on the various plotting options.

Full set of package functions

A list of all the package functions with a brief description of each is below. See the package documentation for a detailed explanation of each function. The functions are run in the order listed to execute the modeling process.

• prep_data(): Takes a panel dataset and converts it to the proper format for the modeling functions.
• get_ytilde(): Returns the $\tilde{Y}_{g',t_{pre}}^{att(g,t)}$ values for each unit.
• get_influence_func(): Returns the influence function values for each unit.
• estimate_model(): Estimates the efficient $ATT(g,t)$, standard errors, and confidence intervals.
• get_results_attgt(): Extracts the efficient $ATT(g,t)$ and related results into a data frame.
• get_results_agg(): Aggregates the results into event study and/or calendar time effects and returns the results in a data frame.
• edid(): Runs all of the above functions and returns any or all of the results of the three previous functions.
• plot_edid(): Plots the results with {ggplot2}.

Limitations

{edid}’s biggest limitation is that it does not handle covariates. It also requires that all treatment groups use the same number of pre-treatment periods to estimate treatment effects, and it requires balanced panels. I am definitely open to expanding the functionality so don’t hesitate to reach out with requests / issues / feedback of any kind.

Acknowledgements

First I would like to thank the authors Drs. Xiaohong Chen, Pedro H. C. Sant’Anna, and Haitian Xie for their fantastic work developing this estimator. It has already proved useful in my own work. I highly recommend reading their paper. I also want to thank Dr. Brantly Callaway (and Dr. Pedro Sant’Anna again) for the multiplier bootstrap procedure developed in their renowned 2021 paper “Difference-in-Differences With Multiple Time Periods.” I used not only the procedure but much of the code directly from their implementation in the {did} package.