Interactive demo · Probability and Statistical Inference · Chapter 17

Markov equivalence: which arrows can data orient?

Different DAGs can impose exactly the same conditional independences: X → Y → Z, X ← Y ← Z and X ← Y → Z all say X ⊥ Z | Y and nothing else. Theorem 17.13: two DAGs are Markov equivalent if and only if they have the same skeleton and the same unshielded colliders (a → c ← b with a, b not adjacent, drawn in gold). Edit the DAG with the edge buttons; panel (b) lists every DAG of its equivalence class and the essential graph (arrows shared by the whole class; the other links can be turned either way). Data are then simulated from your DAG with random tables, and every DAG on the same variables is fitted by maximum likelihood. Panel (c) shows one score per class, because all members of a class reach the same maximised likelihood with the same number of parameters: no amount of data separates them.

Variables

Edges (click to cycle: none, →, ←)

Data simulated from this DAG

Score

Try these

Tables: P(v = 1 | parents) = 0.05 + 0.9u for each parent configuration, u uniform. Scores in the book's convention (Section 13.6), larger is better: AIC = ℓ − k and BIC = ℓ − (k/2) log n, where ℓ is the maximised log-likelihood (each table fitted by its empirical conditional frequencies) and k the number of free parameters. An arrow that would close a directed cycle is skipped.

In the lecture notes