Adaptive Survival Estimator in Survival Analysis

• Adaptive Survival Estimator (ASE) is a semiparametric survival estimator that adapts data‐adaptive nuisance estimation under censoring using influence functions for robust inference.
• It integrates cross-fitting, doubly robust methods, and drift correction to achieve n½-rate asymptotic normality for treatment-specific survival curves and restricted mean survival time functionals.
• Empirical studies demonstrate ASE’s reduced bias, improved efficiency, and near-nominal coverage compared to conventional estimators under complex censoring and non-proportional hazards.

Adaptive Survival Estimator (ASE) refers explicitly to the framework introduced for adaptive experimentation with censored survival outcomes in discrete time and, in related usage, to adaptive, doubly robust survival estimators that combine influence-function-based estimation, cross-fitting, and data-adaptive nuisance learning under right censoring. Across these formulations, ASE is associated with treatment-specific survival curves, average survival effect curves, and restricted mean survival time functionals; with nuisance components such as treatment propensity, censoring, and event hazards; and with asymptotic inference based on efficient influence functions (EIFs), Gaussianization of drift terms, or martingale central limit theorems (Wang et al., 18 May 2026, Westling et al., 2021, Díaz, 2017).

1. Scope and principal formulations

The term has three closely related formulations in the cited literature. In "Adaptive Experimentation for Censored Survival Outcomes" (Wang et al., 18 May 2026), the Adaptive Survival Estimator is a sequential, censoring-aware framework for adaptive experiments that learns an efficiency-optimal allocation policy and estimates the average survival effect curve. In "Inference for treatment-specific survival curves using machine learning" (Westling et al., 2021), the paper does not explicitly label the estimator “ASE,” but it clearly qualifies as an adaptive survival estimator because it is doubly robust and cross-fitted, permits flexible, data-adaptive estimation of nuisance functions without Donsker restrictions, targets the entire treatment-specific survival function in continuous, discrete, or mixed time via an ensemble learner, and provides pointwise and uniform inference with monotonicity-corrected estimates. In "Statistical Inference for Data-adaptive Doubly Robust Estimators with Survival Outcomes" (Díaz, 2017), the estimator is described as an ASE for survival analysis, focusing on double robustness, $n^{1/2}$-rate asymptotic normality with data-adaptive nuisance learners, Gaussianization of the drift, and cross-fitting.

Formulation	Core target	Distinctive feature
Cross-fitted doubly robust estimator	$\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$	Continuous, discrete, or mixed time; monotonicity correction
Drift-corrected TMLE ASE	$\theta_0 = P(T_1 > \tau)$	Gaussianizing the drift and cross-fitting
Adaptive experimentation ASE	$\tau_t = E[S_t(X,1) - S_t(X,0)]$	Closed-form efficiency-optimal allocation policy

A plausible implication is that ASE is better understood as a class of semiparametric survival estimators than as a single algorithmic object. The shared architecture is adaptive nuisance estimation under censoring together with orthogonal or influence-function-based correction.

2. Targets, observed data, and identification

In the treatment-specific survival setting, the ideal causal target is

$\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$

where $W$ denotes baseline covariates, $A \in \{0,1\}$ is a binary exposure, and $T(a)$ is the potential event time under intervention $A=a$. The observed data are $O=(W,A,Y,\Delta)$, where $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$0 and $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$1. Under assumptions (A1)–(A5), including conditional exchangeability for treatment and censoring, conditional independence of event and censoring times given $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$2, and positivity of treatment and censoring, the target is identified by

$\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$3

with

$\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$4

The product-integral formulation accommodates continuous, discrete, or mixed time (Westling et al., 2021).

In the drift-corrected TMLE formulation, the observed data are $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$5 with discrete event time $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$6, censoring time $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$7, and $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$8. The primary estimand is the treatment-specific survival at a fixed time $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$9 under $\theta_0 = P(T_1 > \tau)$0 and censoring eliminated:

$\theta_0 = P(T_1 > \tau)$1

The methodology also applies symmetrically to $\theta_0 = P(T_1 > \tau)$2, to pointwise survival targets across a grid of times, and by linear functionals such as restricted mean survival time (Díaz, 2017).

In the adaptive experimentation formulation, the observed data are $\theta_0 = P(T_1 > \tau)$3, where $\theta_0 = P(T_1 > \tau)$4 is baseline covariates, $\theta_0 = P(T_1 > \tau)$5, $\theta_0 = P(T_1 > \tau)$6 is the event time, and $\theta_0 = P(T_1 > \tau)$7 is the censoring time. The target is the average survival effect curve

$\theta_0 = P(T_1 > \tau)$8

Scalar summaries can be formed by integrating in $\theta_0 = P(T_1 > \tau)$9, but ASE targets the full curve and the trace-optimal design minimizes the sum of component-wise variances across $\tau_t = E[S_t(X,1) - S_t(X,0)]$0 (Wang et al., 18 May 2026).

3. Efficient influence functions and estimator construction

A central object in all ASE formulations is the efficient influence function. For the treatment-specific survival curve at exposure level $\tau_t = E[S_t(X,1) - S_t(X,0)]$1, the EIF in the nonparametric model is

$\tau_t = E[S_t(X,1) - S_t(X,0)]$2

where

$\tau_t = E[S_t(X,1) - S_t(X,0)]$3

Plugging cross-fitted nuisance estimates into this expression yields the cross-fitted one-step estimator

$\tau_t = E[S_t(X,1) - S_t(X,0)]$4

with nuisance fits reused across all $\tau_t = E[S_t(X,1) - S_t(X,0)]$5 and sample splitting used to relax empirical process conditions (Westling et al., 2021).

For the fixed-time survival estimand in the treated arm, the EIF is

$\tau_t = E[S_t(X,1) - S_t(X,0)]$6

The standard doubly robust estimator has drift term

$\tau_t = E[S_t(X,1) - S_t(X,0)]$7

and the distinctive construction in this line of work is to “Gaussianize” this drift by showing

$\tau_t = E[S_t(X,1) - S_t(X,0)]$8

followed by targeted logistic tilting with offsets for $\tau_t = E[S_t(X,1) - S_t(X,0)]$9, $\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$0, and $\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$1. The resulting estimator is

$\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$2

where the updated nuisances solve the empirical EIF equation and the Gaussianizing equation $\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$3 (Díaz, 2017).

In adaptive experimentation, the non-centered EIF for $\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$4 under a fixed policy $\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$5 is

$\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$6

where

$\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$7

With sequential cross-fitting, the round-$\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$8 pseudo-outcome is

$\theta_{0,F}(t,a) := P_{0,F}(T(a) > t),$9

and ASE averages these pseudo-outcomes over rounds:

$W$0

This estimator targets the EIF with the current adaptive $W$1 (Wang et al., 18 May 2026).

4. Adaptivity, nuisance learning, and robustness structure

Adaptivity in ASE has two distinct meanings. One is data-adaptive nuisance estimation. The other is adaptation of the estimator or the allocation rule to the estimated data-generating mechanism. In the treatment-specific survival curve formulation, any binary regression can be used for the propensity, with SuperLearner combinations recommended, and the paper proposes a novel ensemble learner for the conditional survival functions $W$2 and $W$3. The losses

$W$4

and

$W$5

have population minimizers equal to $W$6 and $W$7 under (A1)–(A5). The iterative SuperLearner alternates minimization over convex combinations of candidate estimators and terminates when sup-norm changes are below threshold (Westling et al., 2021).

In the drift-corrected TMLE formulation, cross-fitting is paired with univariate regression smoothing of residual terms and targeted logistic tilting. The estimator is consistent if either the treatment-censoring mechanism $W$8 is consistently estimated or the outcome mechanism $W$9 is consistently estimated. The paper further states that the ASE achieves $A \in \{0,1\}$0-rate asymptotic normality if at least one nuisance converges at $A \in \{0,1\}$1, even if the other is inconsistent. This is the role of Gaussianizing the drift plus cross-fitting (Díaz, 2017).

In the treatment-specific survival curve formulation, the robustness structure is stronger. The estimator is consistent if either $A \in \{0,1\}$2 is consistent or both $A \in \{0,1\}$3 and $A \in \{0,1\}$4 are consistent, and condition (B3) allows “piecemeal” correctness across time $A \in \{0,1\}$5, yielding an infinite-dimensional multiple robustness akin to $A \in \{0,1\}$6-robustness in longitudinal discrete-time G-computation. This extends double robustness from a finite set of nuisance components to correctness that can vary across time (Westling et al., 2021).

In adaptive experimentation, adaptivity is expressed through the treatment allocation policy. The A-optimal allocation is

$A \in \{0,1\}$7

with

$A \in \{0,1\}$8

The policy generalizes classical Neyman allocation to survival settings by prioritizing patient strata where both event and censoring dynamics induce high uncertainty. In the no-ties setting, ASE is doubly robust: consistent if either event hazards or censoring hazards are estimated consistently (Wang et al., 18 May 2026).

5. Inference, asymptotics, and shape constraints

The treatment-specific survival curve estimator has pointwise and uniform asymptotic theory. Under conditions (B1)–(B3), $A \in \{0,1\}$9, and with (B4) there is uniform consistency:

$T(a)$0

If, additionally, $T(a)$1, $T(a)$2, $T(a)$3, and the product-rate conditions (B5)–(B6) hold, then

$T(a)$4

and the process

$T(a)$5

converges weakly in $T(a)$6 to a tight, mean-zero Gaussian process (Westling et al., 2021).

Finite-sample shape violations are handled by an explicit four-step monotonicity correction: compute $T(a)$7 on the observed time grid $T(a)$8; truncate to $T(a)$9; project onto monotone non-increasing functions via isotonic regression to obtain $A=a$0; and extend to all $A=a$1 by right-continuous stepwise interpolation. Inference is reported using $A=a$2. The paper gives a cross-fitted influence-function variance estimator, Wald confidence intervals, a recommended logit-scale interval, fixed-width uniform bands, variable-width logit-scale bands on $A=a$3, delta-method inference for contrasts, a consistent and asymptotically linear plug-in estimator for RMST, and a global test statistic for equality of survival curves over $A=a$4 (Westling et al., 2021).

For drift-corrected TMLE, the main theorem states

$A=a$5

with

$A=a$6

If all nuisances are consistent, then $A=a$7 and the estimator is efficient. Variance is estimated by the empirical variance of the estimated influence function, and Wald-type confidence intervals and hypothesis tests follow directly (Díaz, 2017).

For adaptive experimentation, the asymptotic result is sequential:

$A=a$8

where

$A=a$9

The proof uses a martingale difference decomposition, hazard estimation rates of order $O=(W,A,Y,\Delta)$0, bounded inverse probability of censoring weights under uniform overlap, and sequential cross-fitting to bypass empirical process constraints. If $O=(W,A,Y,\Delta)$1, ASE attains the A-optimal semiparametric efficiency bound component-wise (Wang et al., 18 May 2026).

6. Empirical behavior and substantive applications

In the continuous-time simulation study for treatment-specific survival curves, the proposed cross-fitted doubly robust estimator was compared with marginalized Cox proportional hazards and survtmle discretized into 12 intervals. The setup used nonlinear treatment assignment, exponential censoring with covariate dependence, non-proportional hazards under treatment, censoring about $O=(W,A,Y,\Delta)$2, observed event rate about $O=(W,A,Y,\Delta)$3 in control, and sample sizes from $O=(W,A,Y,\Delta)$4 to $O=(W,A,Y,\Delta)$5. The proposed estimator had bias approximately $O=(W,A,Y,\Delta)$6 across $O=(W,A,Y,\Delta)$7 and parameters, had the smallest mean squared error for treatment survival and risk ratio at all $O=(W,A,Y,\Delta)$8, was best for control survival for $O=(W,A,Y,\Delta)$9 and comparable at $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$00, achieved near-nominal pointwise coverage, and had uniform bands that were slightly anti-conservative at small $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$01 but otherwise near nominal. Marginalized Cox showed persistent bias and severe undercoverage under non-proportional hazards, while survtmle showed finite-sample bias that decreased with $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$02 but retained discretization challenges in continuous-time data (Westling et al., 2021).

The same paper applied the method to elective neck dissection for parotid carcinoma in a cohort of $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$03 patients with clinically node-negative, high-grade parotid cancer from NCDB 2004–2013. Unadjusted stratified Kaplan–Meier estimates at five years were $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$04 for END and $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$05 for no END, with logrank $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$06. After covariate adjustment using the proposed estimator, $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$07 and $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$08, survival difference and risk ratio curves suggested possible short-term benefit over $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$09–$\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$10 years but uniform bands included the null throughout, the global test over $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$11 yielded $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$12, and five-year RMST was $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$13 years under END versus $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$14 years under no END, with difference $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$15 years. The conclusion was no statistically significant effect on overall survival through five years after covariate adjustment (Westling et al., 2021).

In the drift-corrected TMLE simulations, with $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$16, discrete time, sample sizes from $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$17 to $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$18, and scenarios in which all nuisances were consistent, only $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$19 was consistent, or only $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$20 were consistent, ASE showed best performance when all nuisances were consistent and behaved similarly asymptotically to conventional doubly robust TMLE. When only one nuisance block was consistent, ASE had markedly smaller bias and better confidence-interval coverage than the conventional doubly robust estimator, and empirical standard error estimates from ASE’s influence function were accurate across scenarios (Díaz, 2017).

The clinical trial application in the same paper used the N9831 phase III randomized trial with $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$21, maximum follow-up of $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$22 years, and $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$23 baseline covariates. For the difference in treatment-specific survival at $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$24 years, ASE yielded $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$25 with standard error $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$26, conventional doubly robust TMLE yielded $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$27 with standard error $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$28, and Kaplan–Meier yielded $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$29 with standard error $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$30, suggesting Kaplan–Meier bias under informative censoring (Díaz, 2017).

For adaptive experimentation, synthetic experiments reported that ASE nearly matched oracle efficiency with relative MSE approximately $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$31 at $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$32, outperforming non-adaptive ASE-NA at approximately $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$33, Plug-in-NA at approximately $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$34, and A2IPW-Naïve at approximately $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$35–$\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$36. In semi-synthetic Twins data, ASE had relative MSE approximately $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$37 oracle at $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$38, versus approximately $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$39 for ASE-NA and Plug-in-NA, and approximately $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$40–$\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$41 for A2IPW-Naïve. ASE and ASE-MS maintained nominal coverage, whereas plug-in and censoring-agnostic baselines deteriorated with $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$42 due to bias (Wang et al., 18 May 2026).

7. Relation to adjacent methods, misconceptions, and limitations

ASE is closely related to outcome regression, inverse probability weighting, augmented inverse probability weighting, and TMLE, but the cited work distinguishes it from each of these in specific ways. In the treatment-specific survival curve setting, pure IPW relies on correct models for both treatment and censoring and can suffer substantial variance inflation, while marginalized Cox can be biased when proportional hazards fails, and existing doubly robust approaches often assume specific parametric models or discrete-time implementations that discretize time. The cross-fitted estimator instead allows continuous, discrete, or mixed time without discretization bias, uses cross-fitting to permit flexible machine learning for all nuisances, achieves multiple robustness across time and uniform-in-time asymptotics, and provides pointwise and uniform inference with monotone correction (Westling et al., 2021).

In the drift-corrected TMLE setting, the main distinction from standard doubly robust estimators is that asymptotic normality can fail when one nuisance is inconsistent, even though consistency is preserved. The estimator addresses this by Gaussianizing the drift term and by using cross-fitting to avoid entropy conditions. When all nuisances are consistent, the estimator is efficient; when only one nuisance block is consistent, the drift-corrected construction is designed to preserve $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$43-rate inference under the stated $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$44 conditions (Díaz, 2017).

In adaptive experimentation, a plausible misconception is that randomization alone makes censoring secondary. The framework explicitly derives a censoring-aware efficiency-optimal allocation because IPC weights inflate uncertainty when censoring is heavy, and even equal censoring hazards across arms do not imply equal censoring survival because $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$45 depends on $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$46 and $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$47 jointly. Another plausible misconception is that adaptivity removes overlap requirements; in fact the method imposes treatment overlap, survival overlap, censoring overlap, and uniform overlap up to horizon $\theta_0(t,a) := E_0[ S_0(t \mid a,W) ]$48, and uses clipping or truncation schedules to keep assignment probabilities interior and control augmented score magnitudes (Wang et al., 18 May 2026).

The limitations are correspondingly structural rather than merely computational. The adaptive experimentation ASE is developed in discrete time, with continuous-time settings requiring discretization or further development. All three formulations rely on conditional independence assumptions for treatment and censoring, and all emphasize positivity or overlap. Heavy censoring, near-degenerate hazards, and extreme inverse weights challenge stability; the recommended responses include truncation, bounded fold counts, monitoring overlap and fitted censoring survival, and restricting analysis horizons when needed. This suggests that ASE should be viewed not as a relaxation of survival-identification assumptions, but as a way to use data-adaptive learning while retaining semiparametric inferential guarantees under those assumptions (Wang et al., 18 May 2026, Westling et al., 2021, Díaz, 2017).