Heterogeneous Treatment Effects

When average treatment effects hide more than they reveal.

Overview

Average treatment effects can mask important variation. A job training program might help some workers dramatically while doing nothing for others. This notebook demonstrates how to estimate treatment effect heterogeneity using modern machine learning methods.

Methods

Causal Forests

• Non-parametric estimation of conditional average treatment effects (CATEs)
• Automatic discovery of effect heterogeneity
• Valid confidence intervals via honest splitting

Double Machine Learning

• Semi-parametric approach combining ML prediction with causal inference
• Robust to model misspecification
• Efficient estimation of average and heterogeneous effects

When to Use This

Good fit:

• You have experimental or quasi-experimental data
• Effects likely vary by observable characteristics
• You need to target interventions efficiently

Not ideal:

• Small sample sizes (needs 1,000+ observations)
• Treatment assignment depends on unobservables

Key Assumptions

1. Unconfoundedness: Treatment assignment is as-good-as-random conditional on observables
2. Overlap: All subgroups have some treated and control units
3. SUTVA: No spillovers between units