Skip to main content
Back

Harvard and Princeton Prove Double Machine Learning Is Leaving Accuracy on the Table

A new estimator beats Double Machine Learning's error rate for causal inference, eliminating first-order noise without requiring extra model assumptions

Paper, Full Access
  • DML is suboptimal: The paper proves Double Machine Learning's error rate is not optimal when the propensity score (E[T|X]) can't be consistently estimated.
  • New sharper estimator: A new estimator achieves rate n⁻¹/² + δ_μ^a + (δ_μ^s)², squaring the stochastic error term vs. DML's linear dependence.
  • No extra assumptions needed: First-order stochastic noise from the nuisance model can be eliminated without imposing Hölder smoothness or sparsity.
  • New tuning rule: Favor under-smoothing your nuisance ML model (δ_μ^a ≈ (δ_μ^s)²), not the classical balanced bias-variance tradeoff.
  • Broad applicability: Extends to average treatment effect estimation and a wide class of semiparametric linear functionals. Paper on arXiv.
  • No code released yet: The paper is theoretical with illustrative simulations; no open-source implementation is available.

Double Machine Learning (DML) has become the go-to framework for causal inference when you need to estimate a single number -- like an average treatment effect -- from data where the confounders are complex enough to require a neural net or random forest to model. The idea is elegant: use flexible ML models to handle the messy parts, then do clean statistical inference on the quantity you actually care about. But a new paper from researchers at Harvard and Princeton asks a sharper question: is DML actually doing this as well as it theoretically could?

The answer, it turns out, is no -- and the gap is not a minor one. The paper, "Optimally taming biases in black-box models for efficient semiparametric estimation," proves that DML's standard error rate is suboptimal in a common and practically important regime, and proposes a new estimator that provably closes the gap.

The hidden cost of black-box nuisance models

To understand the problem, you need to know what a nuisance function is. In a causal inference setup, you often care about one parameter -- say, the effect of a drug on an outcome -- but the data also contains confounders (age, health history, etc.) that you need to model away. That confounder model is the nuisance function: you don't care about it directly, but you need it to get your answer right.

DML is a semiparametric framework that estimates low-dimensional causal parameters while controlling high-dimensional nuisance functions, leveraging orthogonal moment equations and cross-fitting algorithms to reduce bias. The key insight behind DML is Neyman orthogonality -- a mathematical property that makes the final estimate insensitive to small errors in the nuisance model. It employs Neyman orthogonality to nullify first-order bias and cross-fitting to decouple nuisance estimation from target parameter estimation, thus reducing overfitting.

The problem is in how errors compound. DML is methodologically scalable and theoretically grounded, ensuring that nuisance estimation errors enter the final error only via multiplicative remainder terms. In practice, this means if your nuisance model has error rate