Surrogate-Assisted Targeted Learning for Delayed Outcomes under Administrative Censoring

This paper proposes a surrogate-assisted targeted minimum loss estimator that achieves asymptotic linearity and double robustness for causal inference with delayed outcomes under administrative censoring by leveraging a surrogate-bridge representation to avoid unstable inverse-probability weighting and eliminate second-order bias without directly estimating the conditional surrogate law.

The Gist

Here is an explanation of the paper using simple language, creative analogies, and metaphors.

The Big Problem: The "Rushed Graduation"

Imagine a school district trying to measure how well a new teaching method works. They want to know if students pass their final exams (the Primary Outcome).

However, there are two big problems:

1. The Delay: The final exams happen at the very end of the year.
2. The Rush: The school district is closing early for renovations (this is Administrative Censoring).

Because of the early closure, many students who started the new teaching method late in the year never get to take the final exam. The data is missing for them.

But, there is good news: The school has mid-term quizzes (the Surrogate) that happen much earlier. Everyone took these quizzes, regardless of when they started or when the school closed.

The Dilemma:

• If you only look at the students who finished the year and took the final exam (Complete Case Analysis), you might get a wrong answer. Maybe the late starters were different, or maybe the school closing early skewed the results.
• If you try to mathematically "fix" the missing data by giving huge weight to the few students who did finish (Inverse Probability Weighting), the math becomes unstable. It's like trying to balance a seesaw with a feather on one side and a boulder on the other; the slightest wobble sends the whole thing flying.

The Solution: The "Surrogate Bridge"

The authors, led by Lin Li, propose a new way to solve this called Surrogate-Assisted Targeted Learning.

Think of the missing final exam scores as a river you need to cross.

• Old Method: Try to build a bridge using only the few people who made it to the other side. This is shaky and dangerous.
• New Method (The Bridge): Build a bridge using the mid-term quizzes. Since everyone took the quizzes, you have a solid foundation. You use the relationship between the quizzes and the final exams (for the students who did finish) to "bridge" the gap and predict what the missing final exam scores would have been.

How the New Method Works (The Two-Step Dance)

The authors created a specific statistical tool called SA-TMLE. It works in two stages, like a dance:

1. Step 1: The Guess (Super Learner):
   The computer uses a "Super Learner" (a team of different AI models working together) to make a smart guess about how the mid-term quizzes relate to the final exams. It looks at the students who finished and learns the pattern.

2. Step 2: The Correction (The Targeted Flip):
   This is the magic part. Standard methods often make a small, hidden error because the math gets too complicated (a "nested" problem). The authors realized that if you just guess, you might miss the mark slightly.
   So, they added a second step to "nudge" the guess. They adjust the prediction just enough to make sure the math balances perfectly, without needing to know the exact probability of the school closing early. It's like a chef tasting a soup and adding a pinch of salt to fix the flavor, rather than trying to calculate the exact chemistry of the salt.

Why This is a Big Deal

The paper proves three main things:

• It's Stable: Even when the school closes very early and almost no one finishes the year, this method stays calm. It doesn't blow up like the old methods do.
• It's Double Robust: This is a fancy way of saying, "We have a safety net." The method will give you the right answer if either your model for the quizzes is perfect OR your model for the school closing is perfect. You don't need both to be perfect, just one.
• It Handles Groups: In these studies, students are in clusters (like classrooms or schools). The math accounts for the fact that students in the same classroom influence each other, which older methods often mess up.

The Real-World Test

The authors tested this on a real-world scenario: A public health trial in Washington State about treating Chlamydia.

• The Situation: Some clinics started the program late. Because the study ended on a fixed date, those late clinics didn't have enough time to see the long-term results (12-month test positivity).
• The Result: The new method (SA-TMLE) gave a clear, stable answer with a tight confidence interval. The old methods either gave a very wide, shaky answer (too much uncertainty) or a narrow but wrong answer (because they ignored the missing data).

The Takeaway

When you have a study where the main result is delayed and some data is missing because the study ended early, don't panic and don't just throw away the missing data.

Instead, use the early "surrogate" data you do have to build a bridge to the missing results. The new method described in this paper provides a sturdy, mathematically proven bridge that keeps your conclusions reliable, even when the data is messy.

Technical Summary

Here is a detailed technical summary of the paper "Surrogate-Assisted Targeted Learning for Delayed Outcomes under Administrative Censoring" by Lin Li.

1. Problem Statement

The paper addresses a specific semiparametric estimation problem arising in modern longitudinal studies, particularly Stepped-Wedge Cluster Randomized Trials (SW-CRTs), where:

• Delayed Primary Outcomes: The main outcome of interest ($Y$) is observed only after a substantial delay.
• Administrative Censoring: The study ends before all units (clusters) have had sufficient time to realize the delayed outcome. Consequently, $Y$ is missing for "late-crossing" clusters.
• Early Surrogates: A short-term surrogate outcome ($S$) is observed for all units early in the follow-up.

The Core Challenge:
Standard estimators fail in this "near-boundary" regime where the probability of observing the primary outcome ($g_\Delta$) approaches zero for late-crossing clusters:

• Inverse Probability Weighting (IPW): Becomes unstable due to extreme weights ($1/g_\Delta \to \infty$), leading to massive variance inflation.
• Complete-Case Analysis: Discards late-crossing clusters entirely, introducing bias if the missingness is not completely random, and loses efficiency.
• Parametric Mixed Models: Rely on strict model specifications (e.g., linear time trends) that are often misspecified in complex real-world settings.

2. Methodology: Surrogate-Assisted TMLE (SA-TMLE)

The authors propose a Surrogate-Assisted Targeted Minimum Loss Estimator (SA-TMLE) that avoids inverse observation weights in the target parameter itself.

A. Identification Strategy: The Surrogate Bridge

The paper establishes a nested bridge representation for the Average Treatment Effect (ATE), $\Psi(P_0) = E[Y(1) - Y(0)]$.

• Assumption: Surrogate-Mediated Missing at Random (MAR). Conditional on the surrogate $S$, the censoring indicator $\Delta$ is independent of the outcome $Y$. (i.e., $Y \perp \Delta \mid S, A, W, t$).
• The Formula: Instead of weighting by $1/g_\Delta$, the ATE is identified by integrating the observed-outcome regression over the conditional surrogate distribution:
  $$\Psi(P_0) = E_{W,t} \left[ E_{S|A=1,W,t}[E[Y|S, A=1, W, t, \Delta=1]] - E_{S|A=0,W,t}[E[Y|S, A=0, W, t, \Delta=1]] \right]$$
  This formulation replaces the unstable inverse weight with a support positivity condition on the regression of $Y$ on $S$, which is much easier to satisfy.

B. Semiparametric Theory & Structural Features

The paper derives the Efficient Influence Curve (EIC) and identifies two critical structural features:

1. Vanishing Censoring Component: Under the surrogate-mediated MAR assumption, the censoring mechanism ($g_\Delta$) contributes no separate tangent-space component to the efficient influence function. Estimating $g_\Delta$ does not improve efficiency bounds, and its misspecification does not break the estimator if the outcome model is correct.
2. Cluster-Level Aggregation: Because data is clustered, the EIC for a cluster is the sum (not the average) of individual EICs. Valid inference requires cluster-robust variance estimation.

C. The Two-Stage Targeting Construction

A standard one-step Debiased Machine Learning (DML) estimator fails here because the nested bridge functional generates a second-order cross-product remainder ($R_{SY}$) involving the product of errors in the outcome regression ($\hat{Q}_Y$) and the conditional surrogate density ($\hat{f}_S$).

• The Obstacle: Standard cross-fitting eliminates first-order terms but cannot eliminate $R_{SY}$ without requiring the difficult-to-estimate density $f_S$ to converge at a fast rate ($o_P(J^{-1/4})$).
• The Solution: The SA-TMLE uses a two-stage targeting step:
  1. Stage 1: Obtain initial nuisance estimates (outcome regression, surrogate distribution, propensities) using Super Learner.
  2. Stage 2 (Nested Fluctuation): Perform a targeted update on the integrated outcome model ($\bar{Q}_{int}$) using a clever covariate. This step enforces the empirical mean of the cluster-level EIC to be zero.
  • Result: This step mathematically absorbs the $R_{SY}$ remainder into the efficient score without requiring direct estimation of the conditional surrogate density $f_S$.

3. Key Contributions

1. Identification: Introduced a surrogate-bridge G-computation formula that identifies causal effects under administrative censoring without placing inverse observation weights in the target functional.
2. Theoretical Characterization: Proved that under surrogate-mediated MAR, the censoring mechanism contributes no tangent-space component to the EIC, and established the necessity of cluster-level summation for influence curves.
3. Estimator Construction: Developed a two-stage SA-TMLE that achieves asymptotic linearity and double robustness without requiring the estimation of the conditional surrogate density $f_S$, overcoming a limitation of standard DML for nested functionals.
4. Finite-Sample Theory: Provided Berry-Esseen bounds and variance decomposition showing that while the estimator is robust, finite-sample coverage depends on the magnitude of the second-order remainder relative to the sample size.

4. Simulation Results

The authors conducted extensive Monte Carlo simulations (1,000 replicates) across three scenario blocks:

• Block I (Cluster Counts): Evaluated performance across varying numbers of clusters ($J=10$ to $100$).
  • Result: SA-TMLE maintained near-zero bias (<0.004) and stable variance. In contrast, GLMM showed persistent bias due to misspecified time trends, and IPCW showed massive variance inflation and bias as $J$ increased.
• Block II (Double Robustness): Tested scenarios with misspecified nuisance models.
  • Result: SA-TMLE remained unbiased and achieved nominal coverage (0.92) when either the outcome model or the propensity model was correctly specified. It failed only when both were misspecified, confirming double robustness.
• Block III (Increasing Censoring): Varied administrative censoring severity (8% to 43%).
  • Result: As censoring increased, IPCW bias exploded (up to +0.32) and coverage collapsed to near 0%. SA-TMLE maintained near-zero bias, though coverage dipped slightly (to ~0.77) due to finite-sample remainder variance, but remained far superior to competitors.

5. Real-World Application

The method was applied to a design-calibrated analysis of the Washington State EPT Trial (Partner Therapy for Chlamydia).

• Context: A stepped-wedge trial with 23 clusters where late-crossing clusters had 86% administrative censoring for the 12-month outcome.
• Findings:
  • SA-TMLE: Produced a narrow confidence interval (width 0.034) covering the known oracle truth.
  • IPCW: Produced a confidence interval twice as wide (0.068) due to variance inflation from near-zero weights.
  • GLMM: Produced the narrowest interval (0.026) but relied on strict parametric assumptions.
  • Conclusion: SA-TMLE offered the best trade-off between robustness to model misspecification and precision in the presence of heavy censoring.

6. Significance

This paper provides a rigorous solution to a pervasive problem in public health and clinical trials: estimating delayed outcomes when administrative censoring creates "near-boundary" positivity violations.

• Practical Impact: It offers a viable alternative to unstable IPW and fragile parametric models for stepped-wedge trials and similar longitudinal designs.
• Theoretical Advance: It resolves a specific technical hurdle in semiparametric theory (the nested cross-product remainder) by demonstrating that a targeted learning approach can bypass the need to estimate complex conditional densities.
• Software: The authors provide an open-source R package (swcrtSurrTMLE) implementing the estimator, making the method accessible for immediate application.

In summary, the paper demonstrates that by leveraging early surrogate data through a nested bridge representation and a two-stage targeting procedure, researchers can obtain stable, doubly robust causal estimates even when primary outcomes are heavily censored by study design.