Skip to content

Difference-in-differences in Python

Short answer. With staggered adoption (units treated in different years), do not stop at a two-way fixed-effects regression. In Python, StatsPAI runs the whole sequence in a few calls: sp.feols for the TWFE benchmark, sp.bacon_decomposition to see which comparisons drive it, sp.callaway_santanna plus sp.aggte for a heterogeneity-robust ATT and event study, and sp.honest_did for sensitivity to parallel-trends violations. On the castle-doctrine panel used below, these numbers match Stata (reghdfe, bacondecomp, csdid) and R (fixest, did); see the full replication.

When this applies

  • Estimand: the average treatment effect on the treated (ATT), overall, by cohort or by time since treatment.
  • Identifying assumption: parallel trends between treated cohorts and the comparison group (here, never-treated states), plus no anticipation.
  • Data: a panel (or repeated cross-sections) with a unit id, a time variable, the outcome, and the first treated period of each unit.
  • Not the right tool when treatment switches on and off (see sp.did_multiplegt_dyn), when there is no credible comparison group (see synthetic control), or when you have one treated unit.

Install

pip install statspai

1. The two-way fixed-effects benchmark

The data: 50 U.S. states, 2000–2010, with state "castle doctrine" laws adopted in different years (Cheng and Hoekstra's study, as redistributed with the Mixtape). The outcome is log homicides per 100,000.

import statspai as sp

df = sp.datasets.castle_doctrine(event_time=True)   # adds the cohort column `gvar`

twfe = sp.feols("l_homicide ~ post | sid + year", data=df, cluster="sid")
print(twfe.summary())
post 0.0694 0.0559 1.2424 0.2200 -0.0429 0.1817

StatsPAI warns while fitting: post switches on at five different dates, and 2.4% of the TWFE weight falls on comparisons that use already-treated states as controls. That is the problem the next two steps address.

2. Which comparisons drive the TWFE number?

bacon = sp.bacon_decomposition(df, y="l_homicide", treat="post", time="year", id="sid")
print(bacon["decomposition"].groupby("type")["weight"].sum().round(3))
Earlier vs Later Treated 0.077
Later vs Earlier Treated 0.024
Treated vs Untreated 0.899

3. Heterogeneity-robust ATT (Callaway–Sant'Anna)

cs = sp.callaway_santanna(df, y="l_homicide", g="gvar", t="year", i="sid",
                          control_group="nevertreated")
att = sp.aggte(cs, type="simple")
print(att.summary())
ATT: 0.110 ***
Std. Error: (0.039)
[95% CI]: [0.034, 0.186]

The robust ATT (0.110) is about 60% larger than TWFE (0.069). These are analytic standard errors with pointwise intervals. R's did defaults to a multiplier bootstrap with uniform bands; pass bstrap=True, cband=True to sp.aggte for those.

4. Event study

es = sp.aggte(cs, type="dynamic")
print(es.detail[["relative_time", "att", "se"]].round(3).to_string(index=False))
es.plot()
relative_time att se
...
0 0.097 0.040
1 0.112 0.049

Check before you report

  • Pre-trends. Read the leads in the event study. Here the earliest leads (−9, −7) rest on very few cohorts; test them jointly with sp.pretrends_test and report sp.honest_did bounds rather than a pass/fail verdict. See sensitivity analysis.
  • Cohort coding. The adoption year is only partly treated. Coding the cohort as effyear or effyear + 1 moves the ATT from 0.110 to 0.019, so report both or drop the adoption year (details in the replication guide).
  • Estimator choice. Sun–Abraham (sp.sun_abraham), Borusyak–Jaravel–Spiess imputation (sp.bjs) and Gardner's two-stage (sp.gardner_did) make different comparisons. The DiD chooser says which fits your data.

Stata and R equivalents

Step Stata R StatsPAI
TWFE reghdfe y post, absorb(id year) vce(cluster id) fixest::feols(y ~ post \| id + year, cluster = ~id) sp.feols("y ~ post \| id + year", data=df, cluster="id")
Decomposition bacondecomp bacondecomp::bacon() sp.bacon_decomposition()
Callaway–Sant'Anna csdid, estat simple did::att_gt() + aggte() sp.callaway_santanna() + sp.aggte()
Sensitivity honestdid HonestDiD sp.honest_did()

Check the evidence for any function with sp.parity_status("callaway_santanna").

Other Python options

differences and csdid implement Callaway–Sant'Anna, pyfixest offers Gardner's did2s, local projections and saturated event studies, and DoubleML has a DML-based staggered DiD. See the library comparison.


Tested with StatsPAI 1.43.1 on 2026-11-12. The code and numbers on this page are run by tests/test_docs_task_pages.py. Report a problem.