Many regressions and instrumental variable (IV) specifications can be understood as leveraging the “design” of observed shocks for credibly estimating causal effects or structural parameters. This three-day workshop will build up this design-based toolkit and illustrate some of its advantages over alternative identification strategies. Questions we will seek to answer include:
- "What controls do I need to include to avoid omitted variables bias?"
- "Do I need to worry about ’negative weighting’ of heterogeneous effects?"
- "How should I be clustering my standard errors?"
- "What’s the payoff to considering nonlinear/’structural’ analyses?"
The course will include two programming exercises, where different techniques will be illustrated in real-world applications.
This is a three-day (9 hour) intensive workshop, with 6-7 hours of lectures and two 30-minute coding demonstrations. The remaining time will be given to breaks. The coding demonstrations will feature me going through a real-world application, which will be handed out in advance if you’d like to attempt it on your own or in small groups beforehand.
- Lecture 1: Selection-on-Observables
- Lecture 2: Design vs. Outcome Models
- Lecture 3: Design-Based IV
- Lecture 4: Negative Weights
- Lecture 5: Clustering
- Lecture 6: Recentering
- Lecture 7: Nonlinear Models
Here are selected readings that accompany the course.