COURSE 16L2100% FREE
Verified 2026-08-14

Causal graphs literacy

**Directed acyclic graphs (DAGs)** for causal inference: nodes are variables; arrows are assumed direct causal influence. UCLA’s workshop teaches DAG basics, elemental confounding structures, and the **backdoor criterion**. Pearl’s tradi...

What it is

Directed acyclic graphs (DAGs) for causal inference: nodes are variables; arrows are assumed direct causal influence. UCLA’s workshop teaches DAG basics, elemental confounding structures, and the backdoor criterion. Pearl’s tradition uses these graphs plus the \(do\)-operator; tools like dagitty (cited by UCLA) automate adjustment-set suggestions *given* a graph.

HIGH PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

DAG: Treatment T ← Confounder Z → Outcome Y, and T → Y. Fork paths marked; adjustment set {Z} highlighted. Title: 'Confounder (fork)'.

Educational Focus: Fork is the #1 causal graph pattern learners must own.
HIGH PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

DAG chain: T → M → Y with optional direct T → Y. Label 'mediator'. Note 'don't blindly condition on mediator if estimating total effect'.

Educational Focus: Mediator literacy prevents wrong adjustments.
HIGH PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

Collider DAG: T → C ← Y. Warning: 'conditioning on collider opens bias path'. Example bubble: hospital selection / Berkson's style toy.

Educational Focus: Collider bias is counterintuitive—needs its own panel.

Why it matters

Without graphs, teams smuggle causal claims into correlational dashboards. Graphs make assumptions visible so they can be debated—and so identification tools (DoWhy, etc.) have something to operate on.

How it works (plain)

Draw arrows you believe. Debate them. Find causal paths vs backdoor (spurious) paths. Block backdoors without blocking the causal path you care about. If you can’t, say the effect is not identified from the assumed graph.

Everyday example

“AI tutoring → grades” with arrows from prior skill into both tutoring use and grades—prior skill is a confounder fork.

Try it

Draw a 4-node graph for AI tutoring and grades including study time and prior skill. Mark one backdoor and one causal path.

Myths

⚠️ Myth: Drawing arrows makes them true.
✓ Reality: Graphs encode assumptions—they don’t certify them (UCLA + Pearl messaging).
⚠️ Myth: Acyclic means the real world has no feedback.
✓ Reality: DAGs are a modeling discipline; feedback often needs time-indexed variables or different frameworks.

Sources