Causal Discovery Algorithms Intro

Causal Discovery Algorithms: An Intro

The one idea: So far in this track, you drew the DAG from domain knowledge and let it tell you what to control for. Causal discovery algorithms flip the workflow — they look at patterns of correlation and conditional independence in observational data and try to infer the DAG's structure. They're a genuinely useful second opinion, but they hand you a family of graphs that look identical statistically, not one confirmed truth — and no algorithm can tell you which member of that family is real.

⚠️ The trap: "I ran a discovery algorithm on my data and it output a DAG, so now I have the causal structure." That feels right and it's wrong. These algorithms recover a Markov equivalence class — a set of DAGs that all imply the same conditional independencies — not a single graph. Several of those DAGs disagree about which arrows point which way, and the algorithm has no way to break the tie. Only domain knowledge, timing, or an intervention can pick the real one out of the set.

Every DAG lesson so far started the same way: you, the analyst, draw the graph from what you know about the business — schooling causes wages, engagement causes both adoption and retention, price responds to demand. That's the honest workflow, and Cunningham's point from the reading-a-DAG lesson still holds — no volume of data alone reveals whether a variable sits in a fork or a collider. Causal discovery is the family of methods that tries anyway: given a table of observational data and no hand-drawn graph, output a plausible causal structure by testing which variables are…