Causal Reductions
Talkat Quantum Physics and Statistical Causal Models (Simons Institute workshop, 2022) · 2022-04-27..04-29·
summary
Proposes a causal reduction that collapses arbitrary latent confounders into a single W (matching X in the observational regime) and proves observational p(x,y) and interventional {p(y|do(x))} are not variationally independent, recovering Manski natural bounds for discrete variables. Uses MLE over combined observational + interventional data (discrete: bounds clip estimates; continuous: parameterizes kernels with normalizing flows), and reports empirical gains—roughly a twofold reduction in required interventional samples across ten simulated examples.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:08Motivation
- 3:49Causal Reductions: High-level Idea
- 5:39Causal reductions: Step 2
- 9:20Causal reductions: Result
- 11:14Causal reductions: Step 4
- 11:39Causal reductions: Step 1
- 12:20Generalization
- 15:03Maximum likelihood estimation (discrete case)
- 16:25Limit of infinite observational data
- 21:20Continuous Case: Reparameterization Trick
- 24:37Linear-Gaussian equality constraints
- 27:27Normalizing Flows: Simulation Results
- 30:19Contributions
- 30:51Conclusion & Discussion
- 35:45Causal reductions: Step 3
related talks
30mCausal Estimands And Interventions
31mA Gentle Introduction To Causal Inference
24mA Friendly Introduction To Causality
11mCausal Inference For Socio-Economic & Engineering Systems
01h 08mAn Introduction to Potential Outcomes, DAGs and Single-World Intervention Graphs
01h 02mUnmeasured Confounding and More Recent Developments/Challenges in Causal Discovery
28mLearning Causal Representations From Unknown Interventions
49mCausal Inference in Complex Systems: Network Interference, Strategic Agents, and Beyond