Valid post-selection inference for penalized G-estimation
Pith reviewed 2026-05-23 04:50 UTC · model grok-4.3
The pith
Penalized G-estimation in structural nested mean models admits valid post-selection inference via two extended methods.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
We extend two different methods to develop valid inference for penalized G-estimation that investigates effect modification of proximal treatment effects within the structural nested mean model framework. We show the asymptotic validity of the proposed methods.
What carries the argument
Penalized G-estimation inside the structural nested mean model framework, which identifies effect modifiers of proximal treatment effects while incorporating data-driven selection.
If this is right
- Asymptotically correct confidence intervals and p-values are available after data-driven selection of effect modifiers.
- Type I error inflation is reduced relative to naive sandwich variance estimators.
- The procedures apply directly to high-dimensional covariate settings in causal models for treatment effect heterogeneity.
- Finite-sample behavior is characterized through extensive simulation comparisons with existing approaches.
Where Pith is reading between the lines
- The framework supports more reliable discovery of treatment effect modifiers when the set of candidates cannot be fixed in advance.
- Application to repeated-session medical data suggests the methods can handle clustered or longitudinal outcomes under the same selection-and-inference pipeline.
- The asymptotic results open the door to checking whether similar extensions hold for other penalized causal estimators outside the structural nested mean model class.
Load-bearing premise
The sampling distribution approximations that underlie the two post-selection methods stay accurate once the penalization and G-estimation steps have been performed.
What would settle it
Large-sample simulations or the hemodiafiltration dataset in which the empirical coverage of the proposed intervals falls substantially below the nominal level.
read the original abstract
Understanding treatment effect heterogeneity is important for decision making in medical and clinical practices, or handling various engineering and marketing challenges. When dealing with high-dimensional covariates or when the effect modifiers are not predefined and need to be discovered, data-adaptive selection approaches become essential. However, with data-driven model selection, the quantification of statistical uncertainty is complicated by post-selection inference due to difficulties in approximating the sampling distribution of the target estimator. Data-driven model selection tends to favor models with strong effect modifiers with an associated cost of inflated type I errors. Although several frameworks and methods for valid statistical inference have been proposed for ordinary least squares regression following data-driven model selection, fewer options exist for valid inference for effect modifier discovery in causal modeling contexts. In this article, we extend two different methods to develop valid inference for penalized G-estimation that investigates effect modification of proximal treatment effects within the structural nested mean model framework. We show the asymptotic validity of the proposed methods. Using extensive simulation studies, we evaluate and compare the finite sample performance of the proposed methods and the naive inference based on a sandwich variance estimator. Our work is motivated by the study of hemodiafiltration for treating patients with end-stage renal disease at the Centre Hospitalier de l'Universit\'e de Montr\'eal. We apply these methods to draw inference about the effect heterogeneity of dialysis facility on the repeated session-specific hemodiafiltration outcomes.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper extends two post-selection inference methods to penalized G-estimation within the structural nested mean model (SNMM) framework to enable valid inference after data-driven selection of effect modifiers for proximal treatment effects. It establishes the asymptotic validity of the proposed extensions and evaluates their finite-sample performance via simulations against naive sandwich variance estimators. The methods are motivated by and applied to a hemodiafiltration study examining dialysis facility effect heterogeneity on repeated outcomes.
Significance. If the asymptotic validity holds under the paper's conditions, the work addresses a notable gap by providing valid post-selection inference tools for high-dimensional causal effect modification discovery in SNMMs. This is relevant for clinical applications where effect heterogeneity must be discovered rather than pre-specified. The simulation comparisons and real-data application add practical value; the extension of existing post-selection frameworks to this causal setting is a clear strength.
major comments (2)
- [§3] §3, the statement of asymptotic validity for the extended methods: the claim that the sampling distribution approximations carry over after penalization and G-estimation requires an explicit lemma or expansion showing that the dependence between the data-driven selection and the SNMM estimating equations does not alter the limiting distribution; without this, the validity extension rests on an unverified assumption.
- [Table 2] Table 2 (simulation results): coverage rates are reported near nominal levels, but the design does not include scenarios with strong selection pressure or misspecified penalty tuning; this weakens support for the robustness of the asymptotic approximation in finite samples under the conditions most relevant to the central claim.
minor comments (4)
- [Abstract] Abstract: the phrase 'proximal treatment effects' is used without a brief parenthetical definition or reference, which may reduce accessibility for readers outside the SNMM literature.
- [§1] §1: the introduction should include explicit citations to the two specific post-selection inference methods being extended, rather than only describing them generically.
- [§2.2] Notation in §2.2: several symbols (e.g., the penalty parameter and the selection indicator) are introduced without a consolidated table of definitions, making cross-referencing cumbersome.
- [Figure 1] Figure 1 caption: the description of the panels does not specify the sample size or penalty value used, reducing interpretability of the displayed results.
Simulated Author's Rebuttal
We thank the referee for the positive evaluation and constructive comments. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation of the asymptotic results and simulation evidence.
read point-by-point responses
-
Referee: [§3] §3, the statement of asymptotic validity for the extended methods: the claim that the sampling distribution approximations carry over after penalization and G-estimation requires an explicit lemma or expansion showing that the dependence between the data-driven selection and the SNMM estimating equations does not alter the limiting distribution; without this, the validity extension rests on an unverified assumption.
Authors: We agree that an explicit lemma would make the argument more transparent. In the revised manuscript we will insert a new lemma (placed in §3 or the appendix) that derives the limiting distribution of the post-selection penalized G-estimator by showing that the selection event is asymptotically independent of the SNMM estimating equations under the stated regularity conditions, thereby confirming that the post-selection approximation carries over without additional bias terms. revision: yes
-
Referee: [Table 2] Table 2 (simulation results): coverage rates are reported near nominal levels, but the design does not include scenarios with strong selection pressure or misspecified penalty tuning; this weakens support for the robustness of the asymptotic approximation in finite samples under the conditions most relevant to the central claim.
Authors: We appreciate the suggestion. Although the existing design already varies dimension, sample size, and penalty strength, we acknowledge that scenarios with very strong selection pressure and deliberately misspecified tuning parameters would provide stronger finite-sample evidence. In the revision we will augment the simulation study with these additional settings and update Table 2 and the accompanying discussion accordingly. revision: yes
Circularity Check
No significant circularity detected
full rationale
The paper frames its contribution as an extension of two existing post-selection inference methods to the setting of penalized G-estimation inside the structural nested mean model, accompanied by new asymptotic validity proofs and simulation studies. No derivation step is shown to reduce by construction to a fitted parameter, a self-referential definition, or a load-bearing self-citation chain; the central claims rest on independent asymptotic arguments rather than tautological renaming or re-use of the same data quantities. The work is therefore self-contained against external benchmarks.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Structural nested mean model assumptions for proximal treatment effects hold
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.