- 1What is your identification strategy?In 1854 John Snow compared cholera deaths among London households supplied by two water companies, one drawing above the city’s sewage outflows and one below, and the comparison holds up because of how the households came to be served. What quasi-experimental designs are, why regression with control variables was hard to trust, and six designs that each supply a different counterfactual.6 min
- 2Comparing groups without baseline dataThe multiple-group post-test-only design compares a group that received an intervention with one that did not, at a single point in time, and everything rests on whether the groups were alike beforehand. Two villages with two malaria prevention strategies, then the Millennium Villages Project and what retrospective matching can and cannot establish.4 min
- 3Difference-in-differences and maternal health vouchers in BangladeshComparing the change over time in a treated group with the change in a comparison group, so the comparison group’s trend stands in for what would have happened anyway. Parallel trends and event-study estimates, worked through five rounds of DHS data on Bangladesh’s Maternal Health Voucher Scheme, which raised facility delivery with no evidence of lower stillbirth, neonatal or infant mortality.5 min
- 4Interrupted time series and the kidney transplant wait-time policyWhen a policy reaches everyone at once there is no untreated group, so a unit’s own past becomes the counterfactual. Changes in level and changes in slope, segmented regression, and the threats that decide whether the design holds — concurrent events, autocorrelation, seasonality — read against the 2023 policy that modified kidney transplant wait times for Black candidates in the United States.6 min
- 5Regression discontinuity and HPV vaccination in OntarioPrograms that assign treatment at a threshold leave people just above and just below the line nearly alike, which is what makes the comparison credible. Running variables and cutoffs, sharp against fuzzy designs, the density, placebo-cutoff and covariate-balance diagnostics, and a study of whether Ontario’s school-based HPV vaccination program changed sexual behavior.6 min
- 6Instrumental variables, and choosing a designAn instrument changes who gets treated without affecting the outcome any other way, and that exclusion restriction is the part that usually fails. Encouragement designs, compliers and never-takers, monotonicity and the local average treatment effect, a randomized campaign that helped households in Tangiers get piped water, and a closing guide to choosing among the six designs.9 min