COURSE 16L2100% FREE
Verified 2026-08-14

Machine learning and causal inference

**Prediction** answers “what is likely?” **Causal inference** answers “what if we change something?” Stanford MS&E 226 teaches those as separate pillars (with inference as a third). Stanford GSB **MGTECON 634 — Machine Learning and Causa...

What it is

Prediction answers “what is likely?” Causal inference answers “what if we change something?” Stanford MS&E 226 teaches those as separate pillars (with inference as a third). Stanford GSB MGTECON 634 — Machine Learning and Causal Inference focuses on when and how ML methods can estimate counterfactual policy effects (ATE and personalized policies) in experiments and observational studies.

MEDIUM PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

Venn: ML predictive excellence vs causal identification needs. Overlap: flexible nuisance models (DML etc.) with warning 'ML ≠ automatic causality'.

Educational Focus: Clarifies when ML helps causal work and when it doesn't.

Why it matters

Teams often train a high-AUC model and then treat feature weights as levers. EconML’s premise: use ML *inside* causal estimators to flexibly model heterogeneity and nuisances—without pretending prediction loss equals interventional effect.

How it works (plain)

Use ML to fit complicated relationships (propensity scores, outcome surfaces, CATE models). Keep a causal target and assumptions in charge. Prefer experimental assignment when you can; otherwise use credible observational strategies (IV, longitudinal designs, etc.—listed in MGTECON 634’s description). Validate with sensitivity and policy-evaluation holdouts.

Everyday example

A travel site can’t force membership, but can randomize a easier signup nudge—then estimate the effect of *membership* (not just the nudge) with an IV-style learner (EconML DRIV case narrative).

Try it

Take a predictive model you like. Write one predictive question and one interventional question it does *not* answer. What data would identify the interventional one?

Myths

⚠️ Myth: Deeper nets identify deeper causes.
✓ Reality: Capacity helps approximation; identification is about design/assumptions (DoWhy separation).
⚠️ Myth: Personalized treatment effects are always ethical to deploy.
✓ Reality: CATE estimates can enable manipulation—pair with policy and fairness review.

Sources