Pith. sign in

REVIEW 5 major objections 6 minor 26 references

Targeted Deep Architectures: A TMLE-Based Framework for Robust Causal Inference in Neural Networks

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Targeted Deep Architectures claims that neural-network causal estimates can be made asymptotically valid by freezing most weights and updating a small targeting subset along a gradient obtained from influence-function projections…

desk verdict TDA is a promising projection-based targeting idea for neural networks, but the efficiency claims outrun the proof; treat it as a heuristic until the convergence and coverage conditions are fixed. read the letter →

arxiv 2507.12435 v1 pith:7EF6B7LC submitted 2025-07-16 cs.LG

Yi Li , David Mccoy , Nolan Gunter , Kaitlyn Lee , Alejandro Schuler , Mark van der Laan This is my paper
classification cs.LG MSC 62G0562G2062N01
keywords targeteddeeparchitecturesmaximumlikelihoodestimationefficientinfluencefunctionsemiparametricefficiencydoublerobustnesscausalinferenceneuralnetworkssurvivalanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that causal parameters estimated by deep neural networks do not need external debiasing steps or new training losses: the TMLE fluctuation can live inside the network's own weights. It proposes to freeze most parameters and update a small 'targeting' subset along a gradient obtained by projecting the efficient influence function onto the network's loss gradients. The authors argue this removes first-order bias, yields root-n-consistent plug-in estimates with valid confidence intervals, and extends to functional targets like survival curves through a universal update that targets all time points simultaneously. If correct, this would give practitioners a straightforward way to get rigorous causal inference from any modern deep architecture using only first-order gradients.

What carries the argument

The central object is the 'targeting gradient': for a free weight subset $\theta_{\mathrm{targ}}$, compute score vectors $S(O_i) = \nabla_{\theta_{\mathrm{targ}}} \ell(\theta, O_i)$, then solve the penalized projection $\alpha^* = \arg\min_\alpha \sum_i [D(O_i) - \alpha^\top S(O_i)]^2 + \lambda\|\alpha\|^2$, giving the projected influence function $D_{\mathrm{proj}} = \alpha^{*\top} S$. This projected influence function is the local efficient influence function in the parametric submodel induced by unfreezing $\theta_{\mathrm{targ}}$. For several targets, the individual $\alpha^*_k$ are combined with weights $w_k = d_k / \sqrt{\sum_j d_j^2}$, where $d_k$ is the empirical mean of the $k$-th projected influence function, producing a single universal update direction.

What would settle it

Run TDA on the IHDP benchmark with a last-layer targeting subset and record the L2 projection residual norm after convergence; if the residual is not o_p(1), or if confidence-interval coverage stays near the naive plug-in level when the oracle efficient influence function lies well inside the gradient span, the first-order debiasing claim fails.

Watch

Extended reading notes

Core claim

The paper's central discovery claim is that a plug-in neural estimator can be turned into an asymptotically linear estimate of a causal parameter by embedding the TMLE fluctuation directly in weight space. The procedure freezes most of the network, selects a small free subset of weights, projects an influence function onto the loss gradients of that subset via penalized regression, and iteratively steps the weights along the resulting targeting gradient until the empirical mean of the projected influence function is near zero. The authors claim the final plug-in estimate removes first-order bias, inherits classical TMLE properties including double robustness and semiparametric efficiency, and that multi-dimensional targets can be handled by merging separate targeting gradients into one universal update.

Load-bearing premise

The load-bearing premise is that the span of the gradients of the loss with respect to the unfrozen weights contains a close L2 approximation to the oracle efficient influence function; the paper says this can be checked by the projection residual but never reports that residual.

Editorial extensions

If this is right

  • A neural network can be debiased for pathwise differentiable causal parameters without changing its architecture or adding a separate post-hoc model, by updating only a small subset of weights.
  • For multi-dimensional targets like entire survival curves, one universal gradient update can simultaneously drive the empirical means of several influence functions toward zero, preserving coherence across time points.
  • Under the stated regularity conditions, the TDA estimator is root-n-consistent, asymptotically normal, doubly robust, and attains the semiparametric efficiency bound.
  • On the IHDP benchmark for average treatment effects, TDA improves confidence-interval coverage to roughly 92 percent from 74 percent for the naive plug-in, and on simulated survival data it reduces time-averaged MSE by about 47 percent relative to the initial neural network.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If TDA is right, the projection residual from condition T3 should become a routine diagnostic: reporting it would tell practitioners when last-layer targeting is too small to support the claimed efficiency.
  • The universal update's weighting scheme resembles a multi-objective steepest-descent step in the projected influence-function space, suggesting TDA could be viewed as a general template for debiasing any gradient-trained model, not only neural networks.
  • The survival experiments show that the coverage gains come at the cost of wider confidence intervals, so in practice the method trades precision for validity rather than uniformly improving both.
  • Because TDA uses only first-order gradients, it is naturally testable on parameters where Hessian-based debiasing is infeasible, such as quantile treatment effects or sequentially defined longitudinal targets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes Targeted Deep Architectures (TDA), a framework that debiases plug-in causal estimators from neural networks by freezing most weights and iteratively updating a small targeting subset along a gradient obtained by projecting an influence function onto the span of the network's loss gradients. The authors claim that this procedure removes first-order bias, yields asymptotically valid confidence intervals, inherits classical TMLE double robustness and semiparametric efficiency, and extends naturally to vector-valued targets such as survival curves. The empirical sections report bias reduction and improved coverage relative to plug-in, targeted regularization, A-IPTW, post-hoc TMLE, and Kaplan-Meier baselines on the IHDP ATE benchmark and on simulated survival data with informative censoring.

Significance. If established, the framework would be a useful complement to post-hoc TMLE and Hessian-based autoDML: it is architecture-agnostic, requires only first-order gradients, and provides a single universal update for multi-dimensional targets. The manuscript ships reproducible code and a detailed experimental appendix, which are clear strengths. However, the main theoretical guarantee is conditional on assumptions that are stated but not verified, and the iterative algorithm contains a stopping-rule error that is load-bearing for the central claim. The paper is a promising proposal, but the efficiency and inference claims are not yet demonstrated.

major comments (5)
  1. [Algorithm 1, step 7] The stopping tolerance is dimensionally inconsistent with the claimed rate. The loop stops when |Pn[D*_proj]| <= eta_n with eta_n = sd(D*_proj) sqrt(n) log n. Since sd(D*_proj) is typically O_p(1), eta_n grows like sqrt(n) log n, so the condition is vacuous for large n and cannot imply Pn[D*_proj] = o_p(n^{-1/2}). The algorithm as written may terminate at an iterate whose first-order bias is not negligible, so Theorem 3.1 does not cover the actual procedure. Please correct the tolerance to a quantity that is o_p(n^{-1/2}) and prove that the resulting stopping rule is reached.
  2. [Section 3.2, condition T2] Condition T2 asserts that the TDA estimator satisfies the required first-order expansion 'by construction', but no proof is supplied that the sign-based update in Section 2.2, even with line search, converges to a solution of Pn[D*_proj] = o_p(n^{-1/2}). The update uses only the sign of the empirical mean and a line search on the loss, neither of which is shown to control the magnitude of the targeted score equation. A fixed-point or contraction argument, or an explicit convergence rate for gamma_t, is needed before Theorems 3.1 and 3.2 can be applied to the actual algorithm.
  3. [Section 3.2, condition T3] T3 requires that the span of the targeting gradients contains a good L2(P0) approximation to the oracle efficient influence function. The manuscript says this 'can be checked by the residual norm of your projection step', but no residual norms are reported for the ATE or survival experiments, and no argument is given that the final-layer gradients of the specific architectures satisfy this condition. The claimed asymptotic variance and coverage rates depend directly on this condition; please report projection residuals (e.g., ||D* - D_proj|| on the empirical distribution) before and after targeting, or prove the condition for the architectures considered.
  4. [Section 4.1 and Appendix A.1] The experimental protocol does not use sample splitting for the targeting step. Appendix A.1 states 'Targeting set (100%): Full data harnessed for the targeting procedure', meaning the same observations used to train the initial network are reused in the TDA updates. Condition T2 explicitly invokes sample splitting, and without it the empirical-process term (Pn - P0)(D_{n,P0} - D_{0,P0}) is not automatically negligible. Please either implement proper sample splitting or cross-fitting in the experiments or provide a theoretical justification for reusing the full sample.
  5. [Section 2.3] The multi-parameter universal update is not justified. With d_k = Pn[alpha_k^T S_theta] and w_k = d_k / (sum_j d_j^2)^{1/2}, the claim that one step in direction alpha* = sum_k w_k alpha_k 'reduces the empirical mean of each D_k' requires a calculation: moving theta changes S_theta and hence changes each d_k in a coupled way, and negative d_k enter with different signs. No monotonicity or contraction result is provided, so the simultaneous targeting property is not established. Please add a lemma stating the required conditions or revise the claim.
minor comments (6)
  1. [Section 2, paragraph after example] The phrase 'unentered influence function' is unclear; 'uncentered' appears to be intended.
  2. [Section 4.2, Remark and Metrics] There are typos: 'neural netwrok' should be 'neural network', and 'couterpart' should be 'counterpart'.
  3. [Table 1] The relationship between 'Variance' and 'CI Width' is unclear; if variance is the variance of the point estimate across replications, the displayed widths of 0.250 are not consistent with n=747 and the reported variances. Please clarify whether 'variance' refers to the influence-function variance or the estimator variance.
  4. [Section 3.1] The projection Pi_n defining the working parameter is described as 'often a Kullback-Leibler or L2 projection', but the actual projection used in the algorithm and experiments is never defined formally; please specify it for the neural submodel.
  5. [Algorithm 1] The penalty is written as lambda ||alpha||_{1/2}, while the text and experiments use lambda ||alpha||_2 or lambda ||alpha||_1; please standardize the notation.
  6. [Section 2.2, reference [7]] The tolerance formula is attributed to reference [7], but as written it cannot have the intended meaning of a mean-scale tolerance; please re-check the citation and the formula.

Circularity Check

2 steps flagged · score 6.0 of 10

TDA's central efficiency guarantee is asserted 'by construction' via the same first-order expansion it must prove (T2), while the load-bearing gradient-coverage condition (T3) is left unverified; empirical benchmarks are independent.

  1. self definitional [Section 3.2, Condition (T2)]
    "First-order expansion (ADML Condition B1): the TDA estimator ˆΨn satisfies ˆΨn = Ψn(P0) + (Pn − P0)Dn,P0 + op(n−1/2) with Dn,P0 the efficient influence function of Ψn in Mn at P0. By construction TDA as a TMLE for Ψn(P0) in Mn satisfies this, provided sample-splitting is used."

    This is exactly the asymptotic debiasing property the abstract claims as a result ('remove first-order bias and produce asymptotically valid confidence intervals'). Instead of being derived from Algorithm 1, the expansion is asserted 'by construction' because TDA is labeled a TMLE. The algorithm's stopping rule accepts |Pn[D*_proj]| ≤ sd(D*_proj)√n log n or a max-iteration/no-improvement plateau; neither condition is shown to imply Pn(D*_proj) = op(n^{-1/2}), so the first-order expansion is not established. Theorem 3.1's conclusion is therefore the hypothesis T2 restated, not a consequence of the update rule.

  2. self citation load bearing [Section 3 opening and Section 3.2, Condition (T3)]
    "Our consistency and efficiency results are a direct corollary of the Adaptive Debiased Machine Learning (ADML) framework of van der Laan et al. [12] ... (T3) Gradient-coverage ... amounts to requiring that the span of ∇θtargℓ contains a good L2(P0) approximation to the oracle EIF D0,P0; in neural nets it can be checked by the residual norm of your projection step."

    The cited ADML framework is prior work by the same research group (including present co-author Mark van der Laan), and its applicability to TDA rests on T2, which the paper only asserts 'by construction' (see previous step). The remaining key premise, T3, is the substantive content of the efficiency claim—that the neural gradient span can approximate the oracle EIF—yet the paper never reports the residual norm it says would verify the condition. Without that verification, the theorem is a conditional statement whose main conditions coincide with the desired conclusion; the abstract's unconditional efficiency and valid-CI claims are carried by this self-citation plus an unverified premise rather than by a proof specific to TDA.

full rationale

The paper has two genuinely independent parts: the IHDP and survival experiments benchmark against external baselines with public code, so the empirical claims are not circular. The theoretical narrative, however, contains a load-bearing circular step. Condition T2 states the first-order expansion that is the target result and justifies it only by calling TDA a TMLE ('by construction'); no argument shows that Algorithm 1's sign-update plus line search, or the stated stopping tolerance, drives Pn(D*_proj) to op(n^{-1/2}). The stated tolerance ηn = sd(D*_proj)√n log n actually grows with n and cannot deliver that rate. The advertised efficiency and valid-confidence-interval conclusions therefore reduce to an assumption rather than a derivation. The ADML citation [12] is from the same group (van der Laan is a co-author here), and its key condition T3—gradient span covering the oracle EIF—is only 'checkable' by a residual norm that is never reported. Hence the central theoretical guarantee inherits its content from an unverified self-citation plus a definitional assertion. I score 6 because the empirical evaluation and algorithmic construction have independent content, but the core 'removes first-order bias / asymptotically valid CIs' claim is partially forced by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its free parameters are the regularization penalty, the choice of which weights to unfreeze, and the number of survival time points. The main unstated burden is the assumption that neural gradients span the oracle EIF, which is load-bearing for the efficiency claims.

free parameters (4)
  • Projection penalty λ (ATE) = 0.01
    Penalty in the regularized projection (G^T G + λI)Δw = G^T r; chosen by hand as prespecified, not tuned or justified.
  • Projection penalty λ (survival) = 1e-5
    L1 penalty in the score projection for survival targeting; prespecified and not justified.
  • Targeting subset size and depth = last layer vs full outcome head
    Which weights are unfrozen is a modeling choice; in ATE they compare last-layer and two-layer outcome heads, in survival only the final layer is targeted.
  • Number of targeted time points = 50
    Survival curve is evaluated at 50 equally spaced time points; this choice affects the multi-parameter weighting and confidence intervals.
assumptions (5)
  • standard math Pathwise differentiability of Ψ and existence of an influence function
    Section 2 opening: 'such parameters admit an efficient influence function D*_P'; for survival they fall back to IPCW-type influence functions.
  • ad hoc to paper ADML conditions T1-T3 and C1-C2 hold for the neural submodel
    Section 3.2 states these as high-level conditions; T3 requires the span of ∇θtargℓ to approximate the oracle EIF, but no verification or projection residual is reported.
  • domain assumption Sample splitting is used for the first-order expansion
    Section 3.2 condition T2 says the first-order expansion holds 'provided sample-splitting is used'. The experiments use the full sample for TDA updates, so the condition is not met in the empirical implementation.
  • domain assumption Nuisance estimators are consistent, or at least one is, for double robustness
    Classical TMLE double robustness requires consistent estimation of at least one nuisance function (outcome or propensity/censoring); this is not demonstrated in the experiments.
  • ad hoc to paper The projected D*_proj equals the EIF of the working submodel
    Section 2.2 defines D*_para = α^T ∇ℓ via least squares projection; this is an EIF for the submodel only if the score span equals the submodel tangent space, which is assumed but not shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Targeted Deep Architectures: A TMLE-Based Framework for Robust Causal Inference in Neural Networks." pith.science (2026). https://pith.science/paper/7EF6B7LC

@misc{pith2026250712435,
  author       = {Pith},
  title        = {Pith review of: Targeted Deep Architectures: A TMLE-Based Framework for Robust Causal Inference in Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7EF6B7LC}},
  note         = {Machine review of arXiv:2507.12435}
}
read the original abstract

Modern deep neural networks are powerful predictive tools yet often lack valid inference for causal parameters, such as treatment effects or entire survival curves. While frameworks like Double Machine Learning (DML) and Targeted Maximum Likelihood Estimation (TMLE) can debias machine-learning fits, existing neural implementations either rely on "targeted losses" that do not guarantee solving the efficient influence function equation or computationally expensive post-hoc "fluctuations" for multi-parameter settings. We propose Targeted Deep Architectures (TDA), a new framework that embeds TMLE directly into the network's parameter space with no restrictions on the backbone architecture. Specifically, TDA partitions model parameters - freezing all but a small "targeting" subset - and iteratively updates them along a targeting gradient, derived from projecting the influence functions onto the span of the gradients of the loss with respect to weights. This procedure yields plug-in estimates that remove first-order bias and produce asymptotically valid confidence intervals. Crucially, TDA easily extends to multi-dimensional causal estimands (e.g., entire survival curves) by merging separate targeting gradients into a single universal targeting update. Theoretically, TDA inherits classical TMLE properties, including double robustness and semiparametric efficiency. Empirically, on the benchmark IHDP dataset (average treatment effects) and simulated survival data with informative censoring, TDA reduces bias and improves coverage relative to both standard neural-network estimators and prior post-hoc approaches. In doing so, TDA establishes a direct, scalable pathway toward rigorous causal inference within modern deep architectures for complex multi-parameter targets.

Figures

Figures reproduced from arXiv: 2507.12435 by the authors.

Figure 1
Figure 1. Traditional Post-Processing vs. Our Approach. Left: conventional practice appends a small parametric “fluctuation” model outside a frozen neural net, which is cumbersome in multi-parameter settings. Right: TDA embeds the targeting step into a local final-layer parametric submodel, allowing universal updates for all parameters of interest. way, TDA provides a direct route to achieve valid statistical inference (i.e.,… view at source ↗
Figure 2
Figure 2. Median survival curves with 95% percentile bands. Solid curves show the median survival estimate across replications for the initial model (blue), the targeted model (green), and the Kaplan–Meier estimator (cyan). Shaded ribbons depict the corresponding 2.5th–97.5th percentiles. The black line is the ground-truth survival function used in the simulation [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. How often does targeting help? Percentage of replications in which the targeted estimator has a smaller absolute error than the initial model (blue) and the Kaplan–Meier estimator (teal) at each time point. The horizontal dashed line indicates parity (50%); values above the line mean the targeted estimator is more accurate in the majority of runs. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 22 canonical work pages

  1. [1]

    Double/debiased machine learning for treatment and structural parameters, 2018

    Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. Double/debiased machine learning for treatment and structural parameters, 2018

    2018

  2. [2]

    RieszNet and ForestRiesz: Automatic Debiased Machine Learning with Neural Nets and Random Forests

    Victor Chernozhukov, Whitney K. Newey, Victor Quintas-Martinez, and Vasilis Syrgkanis. RieszNet and ForestRiesz: Automatic Debiased Machine Learning with Neural Nets and Random Forests, June 2022. arXiv:2110.03031 [cs, econ, stat]

    work page Pith review arXiv 2022

  3. [3]

    Sieve plateau variance estimators: a new approach to confidence interval estimation for dependent data

    Molly M Davies and Mark J van der Laan. Sieve plateau variance estimators: a new approach to confidence interval estimation for dependent data. 2014

    work page 2014

  4. [4]

    Bayesian nonparametric modeling for causal inference

    Jennifer L Hill. Bayesian nonparametric modeling for causal inference. Journal of Computa- tional and Graphical Statistics, 20(1):217–240, 2011

    work page 2011

  5. [5]

    Springer, 2003

    Mark J Laan and James M Robins.Unified methods for censored longitudinal data and causality. Springer, 2003

    work page 2003

  6. [6]

    Regularized Targeted Maximum Likelihood Estimation in Highly Adaptive Lasso Implied Working Models

    Yi Li, Sky Qiu, Zeyi Wang, and Mark van der Laan. Regularized targeted maximum likelihood estimation in highly adaptive lasso implied working models. arXiv preprint arXiv:2506.17214, 2025

    work page Pith review arXiv 2025

  7. [7]

    One-step targeted maximum likelihood estima- tion for targeting cause-specific absolute risks and survival curves

    Helene CW Rytgaard and Mark J van der Laan. One-step targeted maximum likelihood estima- tion for targeting cause-specific absolute risks and survival curves. Biometrika, 111(1):129–145, 2024

    work page 2024

  8. [8]

    Estimating individual treatment effect: generalization bounds and algorithms

    Uri Shalit, Fredrik D Johansson, and David Sontag. Estimating individual treatment effect: generalization bounds and algorithms. In International conference on machine learning, pages 3076–3085. PMLR, 2017

    work page 2017

Show all 26 references
  1. [9]

    Adapting neural networks for the estimation of treatment effects

    Claudia Shi, David Blei, and Victor Veitch. Adapting neural networks for the estimation of treatment effects. Advances in neural information processing systems, 32, 2019

    2019

  2. [10]

    Longitudinal targeted minimum loss-based estimation with temporal-difference heterogeneous transformer

    Toru Shirakawa, Yi Li, Yulun Wu, Sky Qiu, Yuxuan Li, Mingduo Zhao, Hiroyasu Iso, and Mark Van Der Laan. Longitudinal targeted minimum loss-based estimation with temporal-difference heterogeneous transformer. arXiv preprint arXiv:2404.04399, 2024

    2024 arXiv

  3. [11]

    Automatic debi- ased machine learning for smooth functionals of nonparametric m-estimands

    Lars van der Laan, Aurelien Bibaut, Nathan Kallus, and Alex Luedtke. Automatic debi- ased machine learning for smooth functionals of nonparametric m-estimands. arXiv preprint arXiv:2501.11868, 2025. 11

    2025

  4. [12]

    Adaptive debiased ma- chine learning using data-driven model selection techniques

    Lars van der Laan, Marco Carone, Alex Luedtke, and Mark van der Laan. Adaptive debiased ma- chine learning using data-driven model selection techniques. arXiv preprint arXiv:2307.12544, 2023

    2023

  5. [13]

    One-step targeted minimum loss-based estimation based on universal least favorable one-dimensional submodels

    Mark van der Laan and Susan Gruber. One-step targeted minimum loss-based estimation based on universal least favorable one-dimensional submodels. The international journal of biostatistics, 12(1):351–378, 2016

    2016

  6. [14]

    van der Laan and Sherri Rose

    Mark J. van der Laan and Sherri Rose. Targeted learning in data science: causal inference for complex longitudinal studies. Springer International Publishing, 2018

    2018

  7. [15]

    Response Surface B

    Mark J Van der Laan, Sherri Rose, et al. Targeted learning: causal inference for observational and experimental data, volume 4. Springer, 2011. 12 A Appendix: Experimental Details This appendix provides additional information on dataset characteristics, implementation details,...

    2011

  8. [16]

    Hybrid projection system: We compute δi residuals (following the EIF) for each sample, and build a design matrix of MSE gradients with respect to targeted parameters. 13

  9. [17]

    Regularized least squares : We solve (GT G + λI)∆w = GT r with λ = 0 .01 for robust projection

  10. [18]

    Parameter updates: We update parameters iteratively with line search

  11. [19]

    Comparison Methods

    Adaptive early stopping: When mean residual is sufficiently small or convergence plateaus. Comparison Methods. • Initial (Plug-in) uses ˆψAT E= 1 n Pn i=1(ˆµ1(Xi) − ˆµ0(Xi)) • TMLE adds a single ε-fluctuation step to the outcome predictions using the clever covariate H(X) = A ...

  12. [20]

    Score matrix computation: We construct a score matrix representing model gradients with respect to the final layer parameters, capturing the relationship between model parameters and the EIF

  13. [21]

    Censoring adjustment: We fit a separate neural network to model the conditional censoring process G(t|X), and use these estimates to weight observations in the targeting procedure, addressing the informative censoring

  14. [22]

    Universal targeting: For m time points of interest t1, . . . , tm, we: • Compute influence functions for each time point Dtj (O) • Project each onto the score space: αj = arg minα P i(Dtj (Oi) − αT Sθ(Oi))2 + λ|α|1 • Compute empirical means: dj = 1 n P i Dtj (Oi) • Form a weig...

  15. [23]

    ε-parametric submodel

    Iterative updates: Continue updating until convergence criteria are met or maximum iterations reached. This approach allows us to simultaneously target the entire survival curve, maintaining coherence and smoothness across time points. Confidence Interval Construction. For the...

  16. [24]

    For each candidate submodel Mk, run TDA to obtain bΨk plus an approximate standard error bsek

  17. [25]

    Form a confidence interval bΨk − 1.96 bsek, bΨk + 1.96 bsek

  18. [26]

    ε-fluctuations

    Compare the lower or upper interval bound (depending on whether bΨk is trending upward or downward). Once the bound plateaus or expands again, stop. Intuitively, if a smaller Mk already drives Pn[D∗] near zero, further expansions waste degrees of freedom and may inflate varian...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.