csdid
A Stata package for difference-in-differences with staggered treatment adoption.
csdid estimates group-time average treatment effects, ATT(g,t), and the event studies and summaries built from them. It covers the designs applied researchers actually face: staggered adoption across many cohorts, block designs where every treated unit adopts at once, covariates through doubly robust estimation, sampling weights, unbalanced panels, repeated cross sections, anticipation windows, and simultaneous confidence bands over the whole event-study path. Every cell it estimates can be inspected, and every decision it makes along the way is reported. Based on Callaway and Sant’Anna (2021).
net install csdid, from("https://raw.githubusercontent.com/pedrohcgs/csdid-stata/main") replace
What it is built for
Staggered adoption is the hard case, and block adoption is the easy special case of it. Either way, the target is a population quantity, and it should not change because the sample arrived in a different shape. In our simulations, making survey waves unequal (and changing nothing else) drops the other commands’ coverage to 2–38%. csdid stays at 95%.
Doubly robust by default
The default estimator is consistent when either the outcome model or the treatment-probability model is correct. In our experiments, breaking the outcome model biases every one-model command by +0.30 to +0.49 while csdid dr reads −0.06 with nominal coverage; breaking the propensity score instead leaves dr on target again. Note that double robustness protects against one wrong model, and not against both being wrong at once.
We compared, so you do not have to guess
We ran csdid against five other Stata commands on one population with known truths. The differences that matter are the ones that are easy to miss: targets that move with the sampling, covariate assumptions that restrict more than their option names suggest, and standard errors that differ across commands even where the point estimates agree exactly. We would not compare estimates across estimators unless you know how each one behaves; the guide is our attempt to make that knowledge available.
Fast enough that care is free
A million-row event study with clustered standard errors finishes in 1.8 seconds with analytical standard errors (2.4 seconds at the bootstrap default), and every workload we measured runs 4.9× to 28× faster than the SSC version. Those are wall-clock times on our hardware. They say nothing about the accuracy of either version.
Transparency is the point
We built csdid so that its output can be taken apart. Every ATT(g,t) comes with its own standard error, the aggregation weights are available in closed form, the sample decisions are stored in e(), overlap failures are reported as refusals and not as numbers, and every simulation in the comparison guide can be recomputed from a seed with the scripts in the code appendix. We see the choice of an estimator as a conversation between empirical researchers and econometricians, and that conversation goes better when both sides can recompute the numbers in front of them. None of this replaces judgment about the identifying assumptions in your application.
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; two omitted-option defaults differ deliberately in Stata, and NEWS.md documents both.
In one screen
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 cohort
Covariates go right after the outcome. The comparison group is not-yet-treated units by default (nevertreated switches). Inference is a multiplier bootstrap with simultaneous bands (analytical gives pointwise analytical standard errors). Every example on this site runs from a clean Stata session, on a fixed, dated copy of the data.
Guides
Three places to start; the full set is on the guides page.
| Getting started | the estimand, the three choices you make, and a first estimate |
| How csdid compares | same data, five other estimators: targets, misspecification, precision, inference, speed |
| Code appendix | every script behind the comparison guide, click to expand |
Reference
From inside Stata: help csdid, help csdid_postestimation, help csdid_estat, help csdid_stats, help csdid_plot.