COURSE 16L2100% FREE
Verified 2026-08-14

Pearl DAGs vs potential outcomes

Two major teaching traditions for causal inference: 1. **Potential outcomes / Rubin causal model** — each unit has potential outcomes under different treatments; experiments and careful observational designs estimate effects like ATEs. C...

What it is

Two major teaching traditions for causal inference:

  1. Potential outcomes / Rubin causal model — each unit has potential outcomes under different treatments; experiments and careful observational designs estimate effects like ATEs. Centered in Hernán & Robins *What If*, Stanford MS&E 226’s causality block, and much of STATS 361’s design language.
  1. Structural causal models / Pearl DAGs — causal graphs + interventions (\(do\)) + identification via graphical criteria and do-calculus. Centered in Pearl’s *Causality* and the UAI “probabilistic calculus of actions.”

MASTER’s Topic 5 asks learners to study both, not pick a tribal side.

HIGH PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

Bilingual map: left Pearl/DAG language (do-operator, identification, backdoor); right Potential Outcomes (Y(1), Y(0), ATE, ignorability). Arrows pairing equivalent ideas. Title: 'Two dialects, one goal'.

Educational Focus: Stops tribal confusion between DAG and PO communities.

Why it matters

AI+ML engineers meet DAGs in tooling (DoWhy) and potential outcomes in experiment culture (A/B, uplift). Speaking only one dialect causes needless confusion in cross-functional teams.

How it works (plain)

Potential outcomes: “What would Y have been if this person got treatment vs not?” Randomization makes those comparisons fair on average.

Pearl/DAGs: “Here is how the system is wired; what happens if we force X?” Graphs show which adjustments are needed.

Many modern workflows use both: graphs to argue assumptions, potential-outcomes estimands to name the target.

Everyday example

Doctor language: “Would this patient recover *if treated*?” (potential outcomes). Engineer language: “If we force the thermostat setting, which sensors and confounders matter?” (intervention on a graph).

Try it

Write the same question—“Does AI tutoring raise grades?”—once in potential-outcomes words and once as a small DAG. List one assumption each framing makes obvious.

Myths

⚠️ Myth: One school makes the other obsolete.
✓ Reality: Both are widely taught (Stanford MS&E 226 vs STATS 361 including graphical models; *What If* vs Pearl).
⚠️ Myth: DAGs don’t need assumptions.
✓ Reality: Missing arrows and wrong arrows are assumptions—just drawn.

Sources