Doubly robust inference via calibration

Technical Summary: Doubly Robust Inference via Calibration

Problem Statement
The paper addresses a fundamental mismatch in the theory of debiased machine learning (DML) for estimating linear functionals of regression functions, such as the average treatment effect (ATE). While standard DML estimators (e.g., one-step estimators, TMLE) possess "double robustness" for consistency—meaning they remain consistent if either the outcome regression or the Riesz representer (e.g., propensity score) is estimated consistently—they typically require both nuisance functions to be estimated consistently and at a sufficiently fast rate (specifically, the product of their convergence rates must be $o_p(n^{-1/2})$) to achieve asymptotic normality.

This requirement creates a gap in statistical inference: if one nuisance estimator is inconsistent or converges too slowly (e.g., due to model misspecification, over-regularization, or intrinsic complexity), the standard product-rate condition fails. Consequently, the estimator loses asymptotic linearity, rendering standard confidence intervals and $p$-values invalid. Existing approaches to achieve "doubly robust asymptotic linearity" (DRAL)—where valid inference holds if only one nuisance is estimated well—often rely on parametric assumptions, high-dimensional sparsity, or complex iterative procedures (like TMLE) that are computationally demanding and difficult to generalize to arbitrary linear functionals.

Methodology: Calibrated Debiased Machine Learning (DML)
The authors propose Calibrated DML, a framework that achieves DRAL by enforcing empirical orthogonality between nuisance estimation errors through a post-hoc calibration step.

1. Core Insight: The authors establish a link between calibration (typically used in prediction to align fitted values with targets) and the debiasing of the cross-product remainder term $\langle \alpha_n - \alpha_0, \mu_n - \mu_0 \rangle$ in the asymptotic expansion of DML estimators.
2. The Calibration Step: Given initial cross-fitted nuisance estimators $\mu_n$ (outcome regression) and $\alpha_n$ (Riesz representer), the method applies a one-dimensional isotonic regression adjustment:
   • $\mu^*_n = f_n \circ \mu_n$, where $f_n$ minimizes the squared error loss between $Y$ and $f(\mu_n)$.
   • $\alpha^*_n = g_n \circ \alpha_n$, where $g_n$ minimizes the Riesz loss (or appropriate loss for the functional) between the target and $g(\alpha_n)$.
3. Enforcing Orthogonality: This calibration step enforces specific empirical orthogonality conditions (Equations 4 and 5 in the paper). Specifically, the residuals of the calibrated estimators become orthogonal to the projected errors of the other nuisance estimator.
4. The Estimator: The final estimator is a standard one-step debiased machine learning estimator constructed using these calibrated nuisances:
   $$\tau^*_n = \psi_n(\mu^*_n) + \frac{1}{n} \sum_{i=1}^n \alpha^*_n(A_i, W_i) \{ Y_i - \mu^*_n(A_i, W_i) \}$$
5. Inference: The authors propose a bootstrap-assisted procedure to construct confidence intervals. This method avoids the need to estimate additional nuisance functions (specifically the projection errors $r_0$ and $s_0$) required by previous Wald-type approaches, making the inference fully automated and computationally efficient.

Key Contributions

• DRAL via Calibration: The paper demonstrates that calibrating nuisance estimators via isotonic regression allows for doubly robust asymptotic linearity. Valid inference is possible if either the outcome regression or the Riesz representer is estimated consistently at a rate of $o_p(n^{-1/4})$, while the other can be inconsistent or converge arbitrarily slowly.
• Learner-Agnostic Framework: Unlike sparsity-based methods or specific TMLE constructions, this approach works with arbitrary black-box machine learning algorithms for the initial nuisance estimation. It does not require deriving new debiasing algorithms for each specific linear functional.
• Non-Iterative and Tuning-Free: The proposed "isocalibrated DML" is a non-iterative, two-step procedure. The calibration step uses isotonic regression, which is tuning-free and can be implemented with standard software.
• Theoretical Refinement: The authors refine the understanding of DRAL by showing it holds not only when one nuisance is inconsistent but also when both are consistent but one converges too slowly for the standard product-rate condition. They establish that the method achieves nonparametric efficiency when both nuisances are consistent and at least one converges faster than $n^{-1/4}$.
• Bootstrap Inference: They introduce a computationally efficient bootstrap procedure that yields doubly robust confidence intervals without requiring the analyst to know which nuisance estimator is consistent or to estimate complex auxiliary nuisance functions.

Results

• Theoretical Guarantees: Under a "partial orthogonality" condition (which holds if the estimation errors are not perfectly aligned in a specific way), the calibrated estimator is shown to be asymptotically linear with a closed-form influence function. The paper proves that the calibration step preserves the convergence rates of the initial learners (up to a negligible $O_p(n^{-1/3})$ term).
• Simulation Experiments:
  • Experiment 1: In settings with intentionally misspecified nuisance models, calibrated DML exhibited significantly lower bias and better confidence interval coverage compared to standard AIPW and DR-TMLE.
  • Experiment 2: In high-dimensional settings where the propensity score is difficult to estimate (rough, high-dimensional) but the outcome regression is smooth, calibrated DML maintained nominal coverage and lower RMSE compared to AIPW, which suffered from undercoverage.
  • Experiment 3: On semi-synthetic benchmark datasets (ACIC, Lalonde, Twins), calibrated DML consistently reduced bias and improved coverage relative to standard AIPW, particularly in scenarios where AIPW coverage was poor (e.g., dropping to 0.21 in some ACIC settings, while calibrated DML achieved 0.71–0.90).

Significance and Claims
The paper claims to provide a unified, learner-agnostic solution to the open problem of achieving doubly robust inference for generic linear functionals of the outcome regression using black-box machine learning. By showing that a simple isotonic calibration step enforces the necessary orthogonality conditions, the authors argue that valid inference can be achieved under weaker conditions than previously thought possible (relaxing the product-rate requirement to a one-sided rate requirement).

The authors position their work as a practical advancement over existing DRAL methods (like TMLE-based approaches) by offering a non-iterative, computationally efficient implementation that requires only a few lines of code to integrate into existing DML pipelines. They emphasize that while the method enables inference even when one nuisance is inconsistent, it remains regular and efficient when both are consistent, provided one converges sufficiently fast. The paper concludes that calibrated nuisance estimation is a general strategy for reducing bias from inconsistent nuisance estimation and enabling valid doubly robust inference.