In June 1998 a 4.8-kilometre bridge over the Jamuna river connected 26 million isolated Bangladeshis to Dhaka and cut freight costs in half. This tutorial rebuilds the difference-in-differences evaluation of that bridge from the ground up in Python, using the Padma hinterland — a symmetric region left isolated by a river whose own bridge was not started until 2015 — as the comparison group. It teaches the 2x2 logic, parallel trends, two-way fixed effects, event studies and honest sensitivity analysis on satellite nighttime lights, then runs the same machinery over census employment shares, rice yields and a public-goods placebo. The two doubly robust estimators of the original paper are rebuilt by hand in NumPy and pushed through both diff-diff and pyfixest. All 122 published coefficients are audited side by side with the replication, and the defects found inside the shipped Stata package are documented in full.
Reproducing Scott Cunningham's LaLonde test in Python — covariates rescue a difference-in-differences ATT only when they enter the control group's counterfactual trend, recovering the $1,794 experimental benchmark from a naive $3,621.
Extend synthetic difference-in-differences to staggered adoption, where units adopt treatment at different times, and apply it in Stata to parliamentary gender quotas across 119 countries — deriving the per-cohort estimator, its aggregation into the overall ATT, the modern sdid_event event study, and bootstrap, jackknife, and placebo inference.
Introduce and derive synthetic difference-in-differences, then apply it to California's Proposition 99 — comparing SDID with the original difference-in-differences and synthetic control (synth2), and how to run placebo inference with a single treated unit.
When the 'treatment' is a point in space, distance becomes the running variable. We walk through the parametric ring DiD and a data-driven nonparametric alternative, first on a simulated world with a known answer, then on Linden and Rockoff's home-prices study, and reconcile a parametric −5.78 % with a nonparametric −20.6 %.
A case study on the Affordable Care Act's Medicaid expansion --- working through 2x2 cell-means, TWFE, covariate-adjusted DRDID, 2xT and Callaway-Sant'Anna staggered event studies, and HonestDiD sensitivity --- to show how population weighting changes the target parameter when the units are regions of very different sizes.