`csdid`

A Stata package for difference-in-differences with multiple periods.

Welcome to the csdid Stata page! This is your one-stop shop for Difference-in-Differences (DiD) estimates under staggered (and block) designs. We cover a large family of settings, including balanced panel, unbalanced panel, and repeated cross-section data. All our procedures work with or without covariates. For inference, we enable analytical and bootstrap-based cluster-robust procedures. We also support simultaneous/uniform inference procedures suitable for making inference about the entire event-study path and addressing multiple-testing concerns.

The engine of csdid is Callaway and Sant’Anna (2021). All our procedures embrace treatment effect heterogeneity across covariate subgroups, groups/cohorts, and time periods. We also discuss how we can aggregate group-time average treatment effects, ATT(g,t), into more aggregated summary measures such as event studies (and other functionals).

How to install

To install the latest version of csdid

net install csdid, from("https://raw.githubusercontent.com/pedrohcgs/csdid-stata/main") replace

You can also install it using the ssc command (this version is not yet these)

ssc install csdid, replace

What it is built for

Our csdid DiD package is tailored to researchers who are comfortable with Stata and want an easy-to-use, fast, and reliable implementation of the staggered DiD methods in Callaway and Sant’Anna (2021). All the procedures we implement first estimate the building block ATT(g,t)’s and then aggregate them using well-understood and explicitly defined weights. We built the package to make the methods accessible to a broader audience, without requiring them to worry about technical implementation details.

One important aspect of csdid is that all our causal target parameters are well-defined with explicit weights as discussed in Callaway and Sant’Anna (2021), and that these causal parameters do not vary with the type of sampling process you have (e.g., the causal target parameter does not depend on whether the data is a balanced or unbalanced panel).

We caveat that one should be careful when comparing csdid estimates with DiD estimates from other packages, as they do not always target the same causal parameters. We discuss this in more detail here.

Speed

This version of csdid is substantially faster than versions 1.XX (including version 1.81 and the csdid2 variant). It can sometimes be 300x faster than Version 1.82, for example. We have documented these speed gains here. Our understanding is that, so far, we are the fastest implementation of staggered DiD in Stata. That is, we have a strong package, with a lot of options and strong statistical guarantees, that is also fast!

Also in R and Python

The same estimators, from the same team, are available as did for R and csdid for Python (pip install csdid).

The three implementations agree to machine precision once the same options are set. We hope this facilitates conversations between researchers who have different software preferences.

If you find any discrepancy, please raise an issue, and we will address it. But also make sure you are using the same options!

How to use

csdid y x1 x2, ivar(id) time(year) gvar(gvar)   // every ATT(g,t), doubly robust
estat event                                     // the event study, uniform bands
estat group                                     // one effect per group/cohort

The syntax is as simple as described above. Covariates go right after the outcome, and the default estimation method is a doubly robust DiD estimator; you set method(reg) or method(ipw) options to use regression-adjusted or (normalized/Hajek-based) IPW DiD estimators. The default comparison group is not-yet-treated units (if you use the nevertreated option, the comparison group becomes the never-treated units). The default inference procedure is based on a multiplier bootstrap procedure paired with simultaneous bands, so we address head-on the issues of multiple hypothesis tests; if you want to have pointwise analytical standard errors, use the analytical option. By default, all standard errors are clustered at the id level, though you can also use alternative clustering options, e.g., cluster(state) to cluster at a state level.

Guides

We have prepared some user guides to help you use and understand our commands. Beyond actually reading Callaway and Sant’Anna (2021), you can find all our guides here. A few that stand out include:


Getting started	A discussion about the three choices you need to make and how to use them in an example
Why not two-way fixed effects	What a TWFE coefficient actually averages once treatment timing is staggered, and why that is rarely what you are after
How csdid compares	We compare csdid with other packages, in different settings. This can clarify some differences that sometimes get ignored
Upgrading from Version 1.82	What changes in 2.0, which defaults moved, and what each old option is called now
Comparison groups	Whether to compare against not-yet-treated or never-treated units, and what to do when there are no never-treated units

Help Files

From inside Stata, you can get help for all our commands: help csdid, help csdid_postestimation, help csdid_estat, help csdid_stats, help csdid_plot.

Citing csdid

We ask that you cite both the method and the software when you use our csdid command: for the method, Callaway and Sant’Anna (2021); for the software, cite the version you ran, which csdid version reports for you.

@misc{CRS2026_csdidStata,
  author = {Callaway, Brantly and Rios-Avila, Fernando and Sant'Anna, Pedro H. C.},
  title  = {csdid: Difference-in-Differences with Multiple Time Periods in Stata},
  note   = {Stata module, version 2.0.0},
  year   = {2026},
  url    = {https://github.com/pedrohcgs/csdid-stata}
}