9 · Research designs II
Wednesday, October 21, 2026
On Zoom. This meeting is held on Zoom, at the usual time. Join from the Zoom page on Canvas, which asks for your NetID.
What you should be able to do after this meeting. Draw your project’s design as a graph. Name the single assumption carrying the estimate — the one you would have to defend to a sceptic.
Handouts
Where to read more
Remix ch. 8 on panel data and fixed effects, ch. 9 on difference-in-differences, ch. 10 if your rollout is staggered, and ch. 11 on synthetic control. The Effect ch. 16 to 18 cover fixed effects, event studies and difference-in-differences. Mostly Harmless ch. 5.
Session A · 2:00–3:15
Designs that use change over time, each drawn before it is described. Fixed effects first, opened by women on corporate boards, where the association with firm value changed sign once each firm was compared with itself: what the fixed effects remove, the confounders that change over time and survive them, measurement error that they amplify, and a common shortcut that is not a fixed effect. Then difference-in-differences and event studies as one design built in four steps. Two groups and two periods, opened by John Snow’s comparison of two London water companies before and after one of them moved its intake in 1852: cholera deaths fell by 57 per 10,000 residents in the districts it partly supplied, relative to districts it did not reach. A simulated coffee chain that launches a mobile-ordering app shows what each subtraction removes and what parallel trends assumes, and two studies show it failing: trainees whose earnings dipped just before job training, and Philadelphia’s beverage tax pushing sales into the stores outside the city. More periods turn the comparison into an event study, whose pre-trend tests can pass when parallel trends fails. Adoption on different dates changes which units serve as controls, and two-way fixed effects can then report the wrong sign. Treatment that comes in doses needs a stronger assumption to compare one dose with another. Then few treated units: standard errors with one treated state, and synthetic control for California’s 1988 tobacco program. Last, the finance event study, opened by stock prices on the day of earnings announcements: the market model, abnormal returns, and other news in the window. The session closes with all six designs of meetings 8 and 9 in one table.
Session B · 3:30–4:30
Every session B opens with a ten-minute team check-in: what you did since last week, what is stuck, and who is doing what next. Written down, and handed in with that session’s work.
Seven published studies that cover the designs of meetings 8 and 9, each posed as the question and the setting its authors faced: working from home at a Chinese travel agency, family heirs as chief executives, close shareholder votes on governance, calorie labels at Starbucks, bank deregulation and inequality, the cost of terrorism in the Basque Country, and the sudden deaths of company directors. For each, the room proposes a design and names the assumption it needs before the slide shows what the authors did. Then teams are drawn at random to put up their own project’s design, draw the graph, state the assumption carrying the estimate, and defend it while the room tries to break it. Every team presents at some point, without notice. The design that survives goes in your identification section, and the attack that landed goes in your limitations paragraph.
Your team’s output is submitted before you leave.
Due
Project description — up to three pages, on Canvas by Friday, October 23