pith. sign in

arxiv: 2604.28047 · v1 · submitted 2026-04-30 · 📊 stat.ME

Data-Adaptive and Model-Robust Covariate Adjustment for Time-to-Event Outcomes in Stratified Randomized Trials

Raphael C. Kim , Brian Gilbert , Ramin Zabih , Michele Santacatterina , Ivan Diaz This is my paper

Pith reviewed 2026-05-07 07:07 UTC · model grok-4.3

classification 📊 stat.ME
keywords covariate adjustmenttime-to-event outcomesstratified randomizationtargeted minimum loss estimationsurvival analysisclinical trialsdata-adaptive methodsmodel robustness
0
0 comments X

The pith

A targeted minimum loss estimation method permits data-adaptive covariate selection for efficient inference on survival outcomes in stratified randomized trials.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a statistical procedure for incorporating baseline covariates into analyses of time-to-event outcomes in trials that randomize within strata. It relies on targeted minimum loss-based estimation to let the data choose which covariates to use rather than requiring researchers to name them in advance. The procedure is constructed to stay robust if the working models are wrong and still produce efficient estimates of quantities such as survival probabilities or restricted mean survival time. A sympathetic reader would care because most trials record many baseline variables whose value for adjustment is hard to guess beforehand, and conventional methods either sacrifice precision or lose validity when models fail. The approach explicitly respects the stratified randomization so that the resulting inference remains correct.

Core claim

The paper establishes that targeted minimum loss-based estimation, extended with data-adaptive covariate selection, yields model-robust and asymptotically efficient estimates of functionals of the survival curve while properly accounting for stratification in the randomization design. This holds even when the set of prognostic covariates is unknown a priori, as shown by the theoretical development and by simulations that compare precision against unadjusted and non-adaptive estimators.

What carries the argument

Targeted minimum loss-based estimation (TMLE) adapted for data-adaptive covariate selection and stratification weights.

If this is right

  • The method produces more precise estimates of treatment effects on time-to-event outcomes without pre-specifying covariates.
  • Double robustness is retained, so estimates remain consistent if either the outcome or censoring model is wrong.
  • Efficiency gains appear in simulations when the prognostic covariate set is not known in advance.
  • Inference remains valid for functionals of the survival curve such as restricted mean survival time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Trial analysts could reduce pre-commitment to specific covariates during design.
  • The same adaptive TMLE structure might extend to other outcome types or randomization schemes.
  • Routine software implementation would let trial teams apply the adjustment without custom coding.

Load-bearing premise

The data-adaptive covariate selection step must preserve the double-robustness and efficiency properties of TMLE under stratified randomization.

What would settle it

A simulation or re-analysis of trial data in which the proposed estimator loses its efficiency gain or yields confidence intervals with coverage below nominal levels when the outcome regression model is misspecified.

Figures

Figures reproduced from arXiv: 2604.28047 by Brian Gilbert, Ivan Diaz, Michele Santacatterina, Ramin Zabih, Raphael C. Kim.

Figure 1
Figure 1. Figure 1: Simulation results comparing our proposed estimators to unadjusted view at source ↗
Figure 2
Figure 2. Figure 2: Forest Plot of our analyses, with a median threshold change. view at source ↗
Figure 3
Figure 3. Figure 3: Forest Plot of our analyses, under mean threshold. view at source ↗
Figure 4
Figure 4. Figure 4: Forest Plot of our analyses, under 25th percentile cutoff. view at source ↗
read the original abstract

Time-to-event outcomes are commonly used as primary endpoints in randomized clinical trials. Despite this, relatively little work incorporates baseline covariate information while also accounting for stratified randomization, a common form of randomization. Moreover, leveraging efficiency gains using these approaches typically requires pre-specifying a subset of covariates that are most predictive of the outcome -- a challenging task in practice, as most trials collect dozens of potentially prognostic baseline variables. In this work, we build on existing literature to propose a data-adaptive and model-robust covariate adjustment method for time-to-event outcomes. Our approach, based on targeted minimum loss-based estimation, allows for data-adaptive covariate selection and model-robust efficient inference on functionals of the survival curve while accounting for stratification. Through extensive simulations and analysis, we showcase the simplicity and improved precision of our method when the covariate set is not known a priori.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a targeted minimum loss-based estimation (TMLE) procedure for covariate-adjusted inference on functionals of the survival curve (e.g., survival probabilities) in stratified randomized trials with time-to-event outcomes. The method incorporates data-adaptive selection of baseline covariates while aiming to retain double robustness and asymptotic efficiency, properly accounting for the stratification in the randomization. The authors argue that this avoids the need to pre-specify prognostic covariates and demonstrate, via simulations, gains in precision and simplicity relative to unadjusted or fixed-covariate analyses when the relevant covariate set is unknown a priori.

Significance. If the central claim holds, the work would provide a useful, practical extension of TMLE to a common trial design, enabling more efficient use of the many baseline variables typically collected without requiring subjective pre-specification or risking efficiency loss. The model-robustness property is especially relevant for regulatory settings. Credit is due for focusing on stratified randomization and for including simulation evidence of precision gains; however, the absence of publicly available code or detailed implementation of the selection step limits immediate assessment of reproducibility.

major comments (2)
  1. [§3–4 (Proposed estimator and asymptotics)] Section on the proposed TMLE estimator and its asymptotic theory (likely §3–4): The manuscript asserts that data-adaptive covariate selection preserves the double robustness and efficiency of TMLE under stratified randomization. Standard TMLE theory for survival outcomes requires that nuisance estimators (outcome regression, censoring hazard) satisfy Donsker conditions or are obtained via cross-fitting so that the remainder term is o_p(n^{-1/2}). It is not clear from the description whether the selection step (whatever algorithm is used) is performed on the full sample or via cross-fitting, nor how stratification (which couples the propensity scores within blocks) is handled in the selection. This is load-bearing for the claim of model-robust efficient inference; a concrete statement of the conditions under which the influence function remains orthogonal, or an explicit cross-fitting/ens
  2. [Simulation studies (tables/figures)] Simulation section and associated tables: The reported precision improvements are presented as evidence that the method works when the covariate set is unknown. However, the simulations do not appear to include stress tests with a large number of candidate covariates relative to stratum-specific sample size, nor do they isolate the effect of performing selection without cross-fitting. Such scenarios are precisely where the Donsker/cross-fit concern raised by the skeptic would manifest in finite samples; without these results, it is difficult to confirm that the efficiency gains are not an artifact of the particular simulation design.
minor comments (2)
  1. [§2–3] Notation for the stratified propensity score and the targeted fluctuation submodel could be clarified; it is not immediately obvious how the stratification weights enter the efficient influence function.
  2. [Introduction] The abstract and introduction would benefit from a brief comparison table or explicit statement of how the new procedure differs from existing TMLE implementations for survival data (e.g., those in the survtmle or tmle packages) in the presence of stratification.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments, which have helped us strengthen the presentation of our theoretical results and simulation evidence. We address each major comment below and have revised the manuscript accordingly to improve clarity and rigor.

read point-by-point responses

  1. Referee: Section on the proposed TMLE estimator and its asymptotic theory (likely §3–4): The manuscript asserts that data-adaptive covariate selection preserves the double robustness and efficiency of TMLE under stratified randomization. Standard TMLE theory for survival outcomes requires that nuisance estimators (outcome regression, censoring hazard) satisfy Donsker conditions or are obtained via cross-fitting so that the remainder term is o_p(n^{-1/2}). It is not clear from the description whether the selection step (whatever algorithm is used) is performed on the full sample or via cross-fitting, nor how stratification (which couples the propensity scores within blocks) is handled in the selection. This is load-bearing for the claim of model-robust efficient inference; a concrete statement of the conditions under which the influence function remains orthogonal, or an explicit cross-fitting/ens

    Authors: We agree that the original description of the data-adaptive selection procedure was insufficiently precise regarding cross-fitting and the handling of stratification. In the revised manuscript we have expanded Section 3 to state explicitly that covariate selection is performed inside a cross-fitting scheme: the sample is partitioned into K folds, selection of the candidate covariate set (via the algorithm described in Section 2.3) is carried out on the training folds, and the resulting nuisance estimators (outcome regression and censoring hazard) are fitted on the held-out folds. This construction ensures the remainder term is o_p(n^{-1/2}) and that the efficient influence function remains orthogonal. Stratification is incorporated by using the known, fixed stratum-specific randomization probabilities directly in the targeting step and by allowing the selection algorithm to operate either within strata or with stratum indicators included as mandatory covariates; both variants are shown to preserve double robustness. We have added a new Theorem 3.2 that states the precise regularity conditions (including the cross-fit Donsker requirement) under which asymptotic efficiency and model-robustness hold. revision: yes

  2. Referee: Simulation section and associated tables: The reported precision improvements are presented as evidence that the method works when the covariate set is unknown. However, the simulations do not appear to include stress tests with a large number of candidate covariates relative to stratum-specific sample size, nor do they isolate the effect of performing selection without cross-fitting. Such scenarios are precisely where the Donsker/cross-fit concern raised by the skeptic would manifest in finite samples; without these results, it is difficult to confirm that the efficiency gains are not an artifact of the particular simulation design.

    Authors: We acknowledge that the original simulation design did not include high-dimensional regimes or explicit isolation of the cross-fitting step. The revised manuscript now contains an expanded simulation study that adds two new scenarios: (i) p = 80 candidate covariates with stratum-specific sample sizes as small as 40, and (ii) direct head-to-head comparisons of the same selection algorithm run with and without cross-fitting. The new results (Tables S3–S5 and Figure S2 in the supplement) show that efficiency gains relative to the unadjusted estimator are retained under cross-fitting even in these stressed settings, while the non-cross-fitted version exhibits modest finite-sample bias and efficiency loss when stratum sizes are small. These additions directly address the Donsker/cross-fit concern and confirm that the reported gains are not artifacts of the original design. revision: yes

Circularity Check

0 steps flagged

No circularity: extension of TMLE theory to stratified survival outcomes with adaptive selection remains non-tautological

full rationale

The paper extends targeted minimum loss-based estimation (TMLE) to time-to-event functionals under stratified randomization, incorporating data-adaptive covariate selection while claiming preservation of double robustness and efficiency. The derivation chain relies on standard TMLE influence-function arguments and asymptotic theory for the efficient influence function, adjusted for stratification blocks and adaptive nuisance estimation. No equation or claim reduces the target parameter or reported efficiency gain to a fitted quantity by construction; the adaptive selection step is justified by cross-fitting or Donsker-class arguments rather than being defined in terms of the final estimator. No load-bearing self-citation chain or ansatz smuggling is present in the provided abstract and method description. The result is therefore self-contained against external TMLE benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the theoretical efficiency and robustness properties of TMLE carrying over to the stratified, data-adaptive setting; no new entities are postulated.

free parameters (1)
  • hyperparameters for data-adaptive covariate selection
    Any ensemble or cross-validation procedure used for adaptive selection typically requires choices of library size, tuning grids, or stopping rules.
axioms (2)
  • domain assumption TMLE nuisance estimators converge at rates sufficient to preserve asymptotic efficiency and robustness for survival functionals
    Standard regularity conditions for TMLE applied to the stratified randomization design.
  • domain assumption Stratified randomization is correctly incorporated into the estimating equations or influence functions
    The method claims to account for stratification; this must hold for unbiasedness.

pith-pipeline@v0.9.0 · 5461 in / 1490 out tokens · 86030 ms · 2026-05-07T07:07:24.239208+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    Outcome model.h a(Xi) =E[Y i |X i, Ai =a]

    work page

  2. [2]

    D.1 Methods D.1.1 AIPW Our AIPW approach is as follows

    Propensity score model.p A(Xi) =P[A i = 1|X i] As before, we will allow for the propensity score to depend on covariates using a Taylor expansion to accout for this. D.1 Methods D.1.1 AIPW Our AIPW approach is as follows. We present this generalized to cross-fitting. AIPW under Stratified Randomization

    work page

  3. [4]

    This can include data-adaptive variable selection methods

    Fork∈[K], (a) Train an outcome model ˆha,k fora∈ {0,1}using all folds but fold k, denotedV (−k). This can include data-adaptive variable selection methods. (b) Construct an estimate for the propensity score for strataw∈ W denoted by pa(Xi; ˆβk(w)) where the model is parametric and estiamted via Maximum Like- lihood Estimation (MLE)

    work page

  4. [5]

    (AIPW1)Consistency of outcome model estimates, to some possibly misspecified outcome regression function limit ˜h

    Construct an estimator of the ATE by ˆθAIP W = ˆθ1,AIP W − ˆθ0,AIP W for ˆθa,AIP W = 1 K KX k=1 1 |Vk| X i∈Vk I[Ai =a] pa(Xi; ˆβk(w)) (Yi − ˆha,k(Xi)) + ˆha,k(Xi) We make the followingAIPW Assumptions. (AIPW1)Consistency of outcome model estimates, to some possibly misspecified outcome regression function limit ˜h. We assume that our outcome model is cons...

    work page 2024

  5. [6]

    Split the data intoKfolds, denotedV 1 . . .V K

    work page

  6. [7]

    This can include data-adaptive variable selection methods

    Fork∈[K], (a) Train an outcome model ˆha,k fora∈ {0,1}using all folds but fold k, denotedV (−k). This can include data-adaptive variable selection methods. (b) Construct an estimate for the propensity score among each stratum variablew∈ Was pa(Xi; ˆβk(w))

    work page

  7. [8]

    Update initial predictions with TMLE estimator

    work page

  8. [9]

    We make the followingTMLE Assumptions

    Construct an estimator of the ATE via ˆθT M LE = ˆθ1,T M LE − ˆθ0,T M LE for ˆθa,T M LE = 1 K KX k=1 1 |Vk| X i∈Vk [ I[Ai =a] pa(Xi; ˆβk(w)) (Yi−ˆh(u) a,k(Xi; ˆϵ))+ˆh(u) a,k(Xi; ˆϵ)] We will modify the assumptions and handle the expansion a bit differently than our initial estimator. We make the followingTMLE Assumptions. (TMLE1)Consistency of outcome mod...

    work page 2025

  9. [10]

    Additionally, we have p |Vk|(βw − ˆβk(w)) = OP (1)

    ∂ ∂β ˆh(u) 1,k(Xi; ˆϵ, β)|β=βw] via Chebyshev’s. Additionally, we have p |Vk|(βw − ˆβk(w)) = OP (1). By elementary M-Estimation properties, the remainder term p |Vk| · ||β w − ˆβk(w)||2OP (1) =o P (1). Note that the first component can be shown to beo P (1) by(TMLE1). For the second component, consider as in Section A.4.1 Van Lancker et al. (2024) taking ...

    work page 2024