Semiparametric Inference under Dual Positivity Boundaries:Nested Identification with Administrative Censoring and Confounded Treatment

1. Problem Statement

The paper addresses a complex missing-data structure common in longitudinal observational studies and comparative effectiveness research (CER). The problem arises when two distinct sources of incomplete observation occur simultaneously:

1. Administrative Censoring: A long-term outcome ($Y$) is censored for a substantial fraction of the cohort (e.g., patients dropping out before the study ends). The censoring probability ($g_\Delta$) can approach zero for certain subgroups.
2. Treatment Confounding: In observational settings, treatment assignment ($A$) depends on covariates ($W$), leading to a propensity score ($g_A$) that can approach 0 or 1 (positivity violations).

The Challenge:
Standard doubly robust (DR) estimators typically combine inverse probability weighting for both mechanisms ($g_\Delta^{-1} g_A^{-1}$). When both boundaries are active, this compounds instability. Furthermore, if a short-term intermediate variable ($S$) is available for most subjects (before censoring), standard DR methods fail to leverage the "nested" structure where $S$ acts as a surrogate, potentially missing an opportunity to avoid the unstable inverse censoring weights entirely.

2. Methodology: Nested Identification

The paper proposes a nested identification strategy that separates the two positivity boundaries into distinct structural roles.

• Identification Functional:
  The target parameter (Average Treatment Effect, $\Psi$) is identified via a nested functional that integrates the outcome regression over the conditional distribution of the intermediate variable $S$:
  $$\Psi(P_0) = E_W \left[ E_{S|A=1,W}[\bar{Q}_Y(S, 1, W)] - E_{S|A=0,W}[\bar{Q}_Y(S, 0, W)] \right]$$
  where $\bar{Q}_Y(S, a, W) = E[Y | S, A=a, W, \Delta=1]$.

• Structural Asymmetry:

  • Censoring Boundary ($g_\Delta$): Removed from the identification formula entirely. It does not appear in the target parameter but remains in the Efficient Influence Function (EIF) as a nuisance for estimation.
  • Treatment Boundary ($g_A$): Remains in the identification (via the targeting step) and appears in both the EIF and the identification.
  • Result: This creates a dual-boundary structure where the two boundaries play fundamentally different roles, unlike standard symmetric DR problems.

• Estimator Construction (SA-TMLE):
  The authors develop a Targeted Minimum Loss Estimator (TMLE) with a two-stage targeting construction:

  1. Initial Estimation: Estimate nuisance parameters ($\bar{Q}_Y, \bar{Q}_{int}, g_A, g_\Delta$) using Super Learner.
  2. Stage 1 Targeting: Fluctuate $\bar{Q}_Y$ using a clever covariate involving both $g_A$ and $g_\Delta$.
  3. Stage 2 Targeting: Fluctuate the intermediate regression $\bar{Q}_{int}$ using a clever covariate involving only $g_A$. This algebraically eliminates the cross-product remainder term ($R_{SY}$) associated with estimating the density of $S$, without needing to estimate $f_S$ directly.

3. Key Contributions & Theoretical Results

The paper establishes four major theoretical results regarding this dual-boundary structure:

A. Fluctuation Coupling (Proposition 3)

• Finding: When a targeted estimator solves the nested score equation, the convergence rate of the intermediate regression ($\bar{Q}_{int}$) depends on the quality of the outcome model ($\bar{Q}_Y$) but not on the quality of the treatment propensity model ($g_A$) at the leading order.
• Mechanism: The score residuals have a conditional mean of zero given the covariates. Therefore, any propensity reweighting (even if misspecified) preserves the Central Limit Theorem (CLT) rate.
• Implication: Propensity misspecification does not introduce bias in the point estimate but limits the convergence rate to the product-rate boundary.

B. Three-Component Robustness Geometry (Theorem 4, Corollary 5)

• Finding: The remainder term decomposes into two bilinear products:
  1. $R^{(1)}_2$: Outcome $\times$ Censoring ($\bar{Q}_Y \times g_\Delta$).
  2. $R^{(2)}_2$: Intermediate $\times$ Treatment ($\bar{Q}_{int} \times g_A$).
• Asymmetry: Unlike classical DR (where either outcome or propensity correct suffices) or multiply robust estimation (where any one of $K$ models suffices), this structure is asymmetric:
  • The treatment propensity ($g_A$) must be consistent (at any rate) for $\sqrt{n}$-consistency.
  • Correct outcome modeling alone is insufficient if $g_A$ is misspecified; it places the estimator at a "near-boundary" regime with persistent variance inflation.
  • The censoring propensity ($g_\Delta$) is interchangeable with the outcome model via double robustness.

C. Dual Near-Boundary Variance Decomposition (Proposition 6)

• Finding: The sampling variance includes two remainder-variance terms and a covariance term ($c_{12}$).
• Interaction: When both boundaries concentrate in the same covariate region (e.g., the most confounded patients are also the most censored), $c_{12} > 0$. This amplifies the total variance gap beyond the sum of individual contributions.
• Oracle Decomposition: The paper proposes a method to isolate which boundary dominates the variance by substituting the true propensity score into the variance estimator.

D. Uniform Jackknife Validity (Theorem 10)

• Finding: The standard sandwich variance estimator fails to be consistent in "near-boundary" regimes (where remainders are non-negligible).
• Solution: A leave-one-cluster-out jackknife is proven to be consistent for the total variance across all regimes (strong decay, single-boundary, and dual-boundary). It restores nominal coverage uniformly.

4. Simulation Results

The authors conducted extensive simulations (observational cohort with cluster structure) to validate the theory:

• Coverage: The jackknife achieved uniform coverage (0.94–0.97) across all 21 simulation cells, whereas the sandwich estimator failed in near-boundary regimes (undercoverage or overcoverage depending on finite-sample weight inflation).
• Fluctuation Coupling: Misspecifying the treatment propensity ($g_A$) had zero detectable effect on the point estimate bias, confirming Proposition 3.
• Variance Ratio ($\hat{\rho}$): The ratio of Jackknife variance to Sandwich variance tracked the severity of the dual boundaries.
  • Under correct specification, $\hat{\rho} < 1$ (Sandwich overcovers due to extreme weights).
  • Under outcome misspecification, $\hat{\rho} > 1$ (Sandwich undercovers), revealing the near-boundary regime.
• Comparison: The nested SA-TMLE achieved significantly lower variance (narrower intervals) than standard AIPW when administrative censoring was the dominant concern, as it avoids the product of inverse weights.

5. Significance and Implications

• Theoretical Advancement: This work characterizes a new class of semiparametric problems with asymmetric partial robustness, distinct from classical double robustness. It clarifies that in nested identification with confounding, the treatment propensity is a critical "hard" constraint.
• Practical Guidance:
  • Inference: Researchers should use the jackknife (or HC-corrected sandwich) rather than the standard sandwich estimator when dual boundaries are suspected, as the latter is unreliable in near-boundary regimes.
  • Modeling Priority: Estimation resources should be prioritized for the outcome model and the treatment propensity (which must be consistent). The censoring propensity is the lowest priority.
  • Strategy: Using nested identification to avoid inverse censoring weights is beneficial, but it shifts the positivity burden entirely to the treatment assignment, requiring careful handling of the treatment propensity.

Conclusion:
The paper provides a rigorous framework for inference in observational studies with administrative censoring and confounding. It demonstrates that while nested identification avoids the instability of inverse censoring weights, it introduces a complex dual-boundary variance structure that requires specific theoretical adjustments (fluctuation coupling analysis) and robust inference tools (jackknife) to ensure valid statistical conclusions.