Causal Effect Identification from Multiple Incomplete Data Sources

Explore causal effect identification from multiple incomplete data sources in this 36-minute lecture by Dr Santtu Tikka from the University of Jyväskylä, Finland. Delve into the challenges of determining interventional probability distributions without parametric assumptions. Learn about a novel search algorithm utilizing do-calculus rules to address advanced data-generating mechanisms and various observational and experimental source distributions. Discover how this approach extends to causal inference under context-specific independence relations. Examine topics such as the data-fusion problem, identifiability issues in causal inference, missing data challenges, and context-specific independence. Gain insights into labeled directed acyclic graphs, CSI-separation, and the complexity of decision problems in causal effect identification. Explore practical examples, including case-control design and derivations, to solidify understanding of these complex concepts.

Intro
Starting point
The data-fusion problem
Identifiability problems in causal inference
The general identifiability problem
Motivation for a search-based approach
Search over the rules of do-calculus
Example on applying do-search
Missing data in causal inference
Example: case-control design.
Identifiability problems reassessed (with missing data)
Context-specific Independence
Alternative Representations for CSI
Labeled Directed Acyclic Graphs
Example on Context-specific DAGS
CSI-separation Example
Causal Effect Identification in LDAGS
Interventions in LDAGS
Complexity of the Decision Problem
Search over the rules of CSI-calculus
Search Example
Derivation of the Example
A Curious Example
Some Properties of the Search
Open Problems and Possible Future Work
References I