Generic Machine Learning Inference on Heterogeneous Treatment Effects in Randomized Experiments, with an Application to Immunization in India
Victor Chernozhukov, Mert Demirer, Esther Duflo, Iván Fernández-Val
Summary
The paper develops a generic method to use any machine learning algorithm to draw valid statistical inference about features of heterogeneous treatment effects in randomized experiments. Rather than estimating the conditional average treatment effect function itself (which ML may estimate inconsistently), it targets summary parameters such as the best linear predictor of the effect on ML proxies, sorted average effects across groups, and average characteristics of the most/least affected units, using sample splitting and aggregation over many splits to obtain robust confidence intervals. It illustrates the approach with an immunization study in India.
Key findings
- Provides estimands (BLP, GATES, CLAN) that summarize heterogeneity and remain valid even when the underlying ML estimate of the effect function is biased or inconsistent.
- Uses repeated sample splitting with median-based aggregation of p-values and intervals to deliver uniformly valid inference across data splits.
- Demonstrates the framework empirically, including an application to an immunization intervention in India.
Subjects & keywords
Cite this paper
Victor Chernozhukov, Mert Demirer, Esther Duflo, & Iván Fernández-Val (2018). Generic Machine Learning Inference on Heterogeneous Treatment Effects in Randomized Experiments, with an Application to Immunization in India. NBER Working Paper Series (No. 24678). https://doi.org/10.3386/w24678
@misc{chernozhukov2018generic,
author = {Victor Chernozhukov and Mert Demirer and Esther Duflo and Iván Fernández-Val},
title = {Generic Machine Learning Inference on Heterogeneous Treatment Effects in Randomized Experiments, with an Application to Immunization in India},
journal = {NBER Working Paper Series (No. 24678)},
year = {2018},
doi = {10.3386/w24678},
url = {https://doi.org/10.3386/w24678}
}