An Introduction to Potential Outcomes, DAGs and Single-World Intervention Graphs
summary
Defines potential outcomes as paired counterfactuals per unit, models binary treatment/outcome with four response types (helped, hurt, always/never recover), and shows random assignment yields ACE identified under randomization as difference in arm means. For observational data the lecture builds the feasible joint-distribution polytope, visualizes a simplex, and derives bounds that still straddle zero (example bounds −0.35 to +0.65), then points to using these polytopes for instrumental-variable bounds.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:13History of causation
- 3:37The potential outcomes model
- 8:05Defining the causal framework
- 11:53Categorizing patient types
- 15:04Fundamental problem of inference
- 20:24Average causal effects
- 23:36Geometric representation
- 39:49Observational study analysis
- 44:46Bounds and polytopes
- 57:25Conclusion and Q&A
related talks
24mA Friendly Introduction To Causality
30mCausal Estimands And Interventions
41mOn the Path to Causal Inference
31mA Gentle Introduction To Causal Inference
18mConstantine
34mInterpretability and Algorithmic Fairness
41mKeynote: Personalization, Bandits, and Causal inference
12mCausal Inference And Data-Driven Decision-Making