Machine Learning
What is the main data structure used in causal modelling?
Graphs! More specifically, directed acyclic graphs.
Proposed by @rctatman
Card 1 of 259. Answer: Graphs! More specifically, directed acyclic graphs.
Graphs! More specifically, directed acyclic graphs.
Proposed by @rctatman
Card 1 of 259. Answer: Graphs! More specifically, directed acyclic graphs.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Graphs, specifically directed acyclic graphs (DAGs). Nodes are variables and an arrow X → Y means X directly causes Y; acyclic means no chain of arrows loops back, so nothing causes itself. Judea Pearl set out this approach in his book Causality (2000), and his 2011 Turing Award cited his calculus for causal reasoning.
A DAG tells you what to adjust for. Suppose Age → Exercise, Age → Disease and Exercise → Disease. The backdoor path Exercise ← Age → Disease mixes age's effect into the association, so you condition on age: P(disease | do(exercise)) = Σₐ P(disease | exercise, a) P(a), where do() means setting exercise by intervention.
Interviewers then ask about colliders: in A → C ← B, A and B are independent, but conditioning on C makes them dependent, so adjusting for everything can add bias. The graph comes mostly from domain knowledge, since observational data usually narrow it only to a set of equivalent DAGs. The Neyman-Rubin potential-outcomes framework needs no graph.