[Paper Review] Estimation based on nearest neighbor matching: from density ratio to average treatment effect

Zhexiao Lin, Peng Ding+1 more|arXiv (Cornell University)|Dec 27, 2021

Statistical Methods and Inference4 citations

TL;DR

This paper reinterprets nearest neighbor (NN) matching as a consistent estimator of the density ratio when the number of neighbors M diverges with sample size, enabling semiparametric efficiency in average treatment effect (ATE) estimation. It establishes that NN matching achieves the minimax optimal rate for Lipschitz density estimation and provides a doubly robust, efficient ATE estimator, positioning it as a precursor to double machine learning methods.

ABSTRACT

Nearest neighbor (NN) matching as a tool to align data sampled from different groups is both conceptually natural and practically well-used. In a landmark paper, Abadie and Imbens (2006) provided the first large-sample analysis of NN matching under, however, a crucial assumption that the number of NNs, $M$, is fixed. This manuscript reveals something new out of their study and shows that, once allowing $M$ to diverge with the sample size, an intrinsic statistic in their analysis actually constitutes a consistent estimator of the density ratio. Furthermore, through selecting a suitable $M$, this statistic can attain the minimax lower bound of estimation over a Lipschitz density function class. Consequently, with a diverging $M$, the NN matching provably yields a doubly robust estimator of the average treatment effect and is semiparametrically efficient if the density functions are sufficiently smooth and the outcome model is appropriately specified. It can thus be viewed as a precursor of double machine learning estimators.

Motivation & Objective

• To re-express nearest neighbor matching in the context of density ratio estimation.
• To resolve the long-standing inefficiency of fixed-M NN matching in ATE estimation.
• To establish that diverging M enables consistent and rate-optimal density ratio estimation.
• To demonstrate that NN matching yields a doubly robust, semiparametrically efficient ATE estimator.
• To position NN matching as a foundational method for double machine learning estimators.

Proposed method

• Reinterprets the intrinsic statistic $K_M(x)$ from Abadie and Imbens (2006) as a density ratio estimator in a two-sample setting.
• Uses $k$-d trees to achieve sub-quadratic, nearly linear time complexity for NN computation.
• Establishes that the NN-based density ratio estimator is one-step, computationally efficient, and rate-optimal under Lipschitz smoothness.
• Derives asymptotic variance bounds and shows convergence in probability to the true density ratio when $M \to \infty$.
• Connects the density ratio estimator to bias-corrected ATE estimation via double robustness principles.
• Demonstrates that with appropriate $M$, the resulting ATE estimator achieves the semiparametric efficiency bound.

Experimental results

Research questions

• RQ1 Can nearest neighbor matching with diverging $M$ consistently estimate the density ratio between two samples?
• RQ2 Does the NN-based density ratio estimator achieve the minimax optimal rate over Lipschitz density functions?
• RQ3 How does the choice of $M$ affect the efficiency and bias of ATE estimation in NN matching?
• RQ4 Can NN matching be reinterpreted as a double-robust, semiparametrically efficient estimator of ATE?
• RQ5 Is there a theoretical link between NN matching and double machine learning estimators?

Key findings

• The statistic $K_M(x)$ from Abadie and Imbens (2006) consistently estimates the density ratio when $M \to \infty$, even though it was previously used only for ATE estimation.
• The NN-based density ratio estimator is information-theoretically optimal, achieving the minimax lower bound over Lipschitz density functions.
• The estimator is computationally efficient, with nearly linear time complexity via $k$-d trees, and avoids estimating individual densities.
• With a diverging $M$, the ATE estimator becomes semiparametrically efficient under smooth density and correctly specified outcome models.
• The method achieves double robustness: it is consistent if either the propensity score or the outcome model is correctly specified.
• The analysis shows that the fixed-$M$ assumption in Abadie and Imbens (2006) leads to inefficiency, which is resolved by allowing $M \to \infty$.