Regularity, Phase Transitions, and Uniform Inference for Proximal Counterfactual Quantile Processes

Technical Summary: Regularity, Phase Transitions, and Uniform Inference for Proximal Counterfactual Quantile Processes

Problem Statement
This paper addresses the challenge of estimating counterfactual distribution functions, quantile treatment effects (QTE), and lower-tail risk measures (such as CVaR) in observational studies where unmeasured confounding is present. While standard distributional causal inference relies on the assumption of no unmeasured confounding, proximal causal inference utilizes negative-control proxies (treatment-inducing $Z$ and outcome-inducing $W$) to identify causal effects.

Existing literature has established pointwise identification for average treatment effects (ATE) and specific quantiles using proximal bridge equations. However, a direct pointwise application of these methods to the entire counterfactual distribution function $\{F_a(y) : y \in \mathcal{Y}\}$ is theoretically incomplete. It fails to characterize the regularity boundary of the underlying inverse problem, does not yield the efficient covariance kernel for the CDF process, cannot produce density-free simultaneous quantile bands, and does not quantify how weak proxies degrade inference from root-$n$ regularity to slower rates. The core problem is to determine when the counterfactual distribution process can be estimated regularly, efficiently, and uniformly under proximal identification.

Methodology
The authors recast proximal counterfactual distribution estimation as a primal-dual inverse problem.

1. Primal-Dual Formulation:

   • Primal Bridge: For a treatment arm $a$ and threshold $y$, the outcome bridge $h_{a,y}$ solves $T_a h_{a,y} = g_{a,y}$, where $T_a$ is the conditional expectation operator and $g_{a,y}$ is the conditional expectation of the indicator $1\{Y \le y\}$.
   • Dual Bridge: The counterfactual CDF is a linear functional $\ell(h) = E\{h(W, X)\}$. The dual bridge $q_a$ is the Riesz representer of this functional, solving the adjoint equation $T_a^* q_a = 1$.
   • Identification: The CDF is identified via the primal representation $F_a(y) = E\{h_{a,y}(W, X)\}$ or the dual representation $F_a(y) = E\{1\{A=a\}q_a(Z, X)1\{Y \le y\}\}$.

2. Regularity and Phase Transitions:

   • The paper establishes that the existence of a square-integrable dual bridge $q_a$, combined with a minimal residual-moment condition, constitutes the exact regularity boundary.
   • Using a compact singular-system characterization (analogous to Picard conditions for Fredholm equations), the authors derive a phase transition. Root-$n$ regular estimation is possible if and only if the series $\sum_j \ell_{a,j}^2 / s_{a,j}^2$ converges, where $s_{a,j}$ are singular values of the operator and $\ell_{a,j}$ are coefficients of the target functional.
   • If this series diverges (indicating weak proxy relevance), finite-dimensional efficiency bounds diverge, and Gaussian inverse benchmarks exhibit slower minimax rates.

3. Process Theory and Estimation:

   • Efficient Influence Function (EIF): The paper derives the canonical gradient for the CDF process:
     $$\phi_{a,y}(O) = h_{a,y}(W, X) - F_a(y) + 1\{A=a\}q_a(Z, X)\{1\{Y \le y\} - h_{a,y}(W, X)\}$$
   • Uniform Inference: A cross-fitted one-step estimator is constructed to achieve uniform doubly robust expansions. The authors prove weak convergence of the counterfactual CDF process to a tight Gaussian limit with an explicit covariance kernel.
   • Density-Free Quantile Bands: Instead of estimating the counterfactual density (which is ill-posed), the method constructs simultaneous CDF bands and inverts them to obtain conservative, density-free simultaneous quantile bands.
   • Lower-Tail Risk: For Conditional Value at Risk (CVaR), a "shortfall bridge" representation is used. By differentiating the value of an optimized shortfall criterion (via Danskin's envelope theorem), the resulting influence function avoids the need for counterfactual density estimation.
   • Implementation: Estimators utilize closed-form linear algebra for finite-rank systems, convex Tikhonov regularization for ill-conditioned systems, and isotonic projection to enforce shape constraints.

Key Contributions

1. Primal-Dual Identification: Formalizes counterfactual CDF identification as a pair of primal and dual inverse problems.
2. Regularity Dichotomy: Proves that in a threshold-saturated observed-data model, pathwise differentiability of the CDF functional is equivalent to the existence of a square-integrable dual bridge and a finite residual moment.
3. Spectral Phase Transition: Characterizes the "weak-proxy" phenomenon, showing that root-$n$ inference fails exactly when the singular values of the proxy operator decay too rapidly relative to the target functional's coefficients.
4. Efficient Process Theory: Derives the efficient covariance kernel for the entire CDF process and establishes uniform weak convergence.
5. Density-Free Inference: Provides a method for constructing simultaneous quantile bands and lower-tail CVaR confidence intervals without estimating the counterfactual density, thereby avoiding an additional source of ill-posedness.
6. Finite-Rank Estimators: Develops closed-form and convex finite-rank estimators with explicit rate conditions for weak proxies.

Results

• Simulations:
  • Calibration: In finite-rank settings, the proposed Proximal Doubly Robust (PDR) estimator corrects bias and achieves near-nominal coverage, whereas naive observed-confounding methods remain biased.
  • Phase Transition: As proxy relevance decreases, the dual bridge norm and influence function standard deviation inflate, leading to wider confidence bands and eventual loss of informativeness, consistent with the theoretical phase transition.
  • Process Inference: The density-free simultaneous QTE bands maintain valid coverage even when density estimation is difficult (e.g., multimodal outcomes), whereas density-based delta bands undercover.
  • Machine Learning Integration: Learned feature maps (splines, trees, neural networks) improve approximation of bridge functions but preserve the final convex ridge structure; the proximal correction remains necessary even with flexible ML adjustment for observed covariates.
• Real Data (RHC Study): Applied to the Right Heart Catheterization data, the proximal PDR estimator substantially attenuated the estimated distributional shifts (QTE and CVaR) compared to naive ML-IPW. The resulting density-free bands contained zero, suggesting that the apparent large effects in standard analyses may be artifacts of unmeasured confounding.

Significance and Claims
The paper claims to provide the first comprehensive semiparametric theory for proximal counterfactual distribution processes. It does not claim novelty for pointwise identification (which follows from existing ATE theory applied to indicators) but argues that the pointwise approach is insufficient for distributional inference.

The significance lies in:

• Identifying the Boundary: Explicitly defining when proximal distributional inference is regular versus non-regular, moving beyond "identification in principle" to "estimability in practice."
• Spectral Diagnostics: Providing a concrete criterion (the Picard condition on singular values) to diagnose whether weak proxies will destroy root-$n$ inference.
• Robust Process Inference: Offering a practical, density-free route to simultaneous quantile inference that avoids the instability of estimating counterfactual densities, which is particularly valuable in ill-posed proximal settings.
• Unified Framework: Unifying the treatment of CDFs, QTEs, and CVaR under a single primal-dual bridge operator framework, showing that the regularity conditions for all these objects are governed by the same adjoint-range properties.

The authors remain modest, noting that the regularity dichotomy is proved for a saturated model (implying non-regularity in smaller submodels is model-specific) and that the minimax phase transition is established for a Gaussian singular-coordinate benchmark. They emphasize that the validity of the results ultimately depends on the scientific plausibility of the proxy assumptions, which cannot be verified by data alone.