Unmeasured Confounding and More Recent Developments/Challenges in Causal Discovery
summary
Describes how constraint‑based causal discovery extends to unmeasured confounding by searching PAG/MAG space with the FCI algorithm; shows the latent‑projection rules and the “possible DSEP” trick that make conditional‑independence testing feasible, demonstrates with linear‑Gaussian simulations (n=10,000) using Fisher‑Z tests and the pcaug R package, compares failures of PC/GES, and highlights when Y‑shaped patterns identify unmeasured confounding versus pure ancestry.
AI-written summary · about
chapters (8)
- 0:00Introduction to latent variables
- 1:56Causal processes and dependence
- 5:07Distinguishing causal structures
- 11:51Graphical models for latent data
- 21:40Limitations of general ADMGs
- 28:07Maximal ancestral graphs (MAGs)
- 43:22FCI algorithm overview
- 52:10Practical demonstration
related talks
31mIntroduction to Causal Inference using pgmpy
24mA Friendly Introduction To Causality
29m🧠🧹 Causality - Mental Hygiene for Data Science
28mLearning Causal Representations From Unknown Interventions
32mNew Approaches To Learning Nonparametric (Latent) Causal Graphical Models
57mA New Perspective on High-Dimensional Causal Inference
31mA Gentle Introduction To Causal Inference
27mUnmeasured Spatial Confounding