{"id":"2e78fabe-32a2-48f1-b5df-ebf3286f6489","arxiv_id":"2411.15908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A randomized-lasso based selective inference method yields uniformly valid, bounded confidence intervals for time-varying causal effect moderation under non-Gaussian data.","lead":"This paper develops a two-step method for selective inference on time-varying causal effect moderation, using a randomized lasso to select moderators and then constructing a pivot that conditions on the selection event. The method produces bounded confidence intervals with valid coverage in simulations, even where the existing polyhedral approach produces infinite intervals or undercoverage.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pivot as implemented uses estimated H and K matrices, but Theorem 4.2 and its proof treat H and K as known population quantities; without a uniform perturbation argument, the coverage guarantee does not cover the actual procedure.","rationale":"The paper's central theoretical result, Theorem 4.2, is a conditional statement: if Assumptions 2, 3, and 4 hold, then the pivot constructed from the population matrices H and K yields uniformly asymptotically valid conditional intervals. The proof is technically detailed and, under those assumptions, appears to go through. However, the method that is actually implemented in the simulations and data application must compute H and K from data, because these are unknown population moments of the semiparametric model. The manuscript never states that H and K are assumed known, nor does it provide a theorem or even a remark about the effect of plugging in consistent estimators. This is not merely a technicality: the pivot's uniformity relies on the exact density of the change of variables in Theorem 4.1, and replacing H,K with estimates changes the conditioning event, the truncation interval, and the function F. Without a uniform bound on the estimation error, the central claim 'our method achieves valid coverage' is not established for the actual procedure. The reader's weakest_assumption identified Assumption 3 and the fixed-p/high-p mismatch, both of which are legitimate; the H,K issue is at least as load-bearing because it affects the validity of the theorem even in the intended fixed-p, large-n regime. I do not recommend changing the conditional verdict, because the mathematical core may be salvageable by adding a standard plug-in analysis, and the simulations suggest the approach works in practice. But the paper should be required to address this gap explicitly before full acceptance.","tokens_in":42675,"tokens_out":20855,"duration_ms":200732,"concrete_test":"In the Section 6 Gaussian data-generating process (where the true H and K are computable), run the proposed method twice on the same 500 Monte Carlo replications at n=120, p=50, low-signal setting: once using oracle H,K (computed from the known model) and once using sample-estimated H,K as a real user would. Compare empirical FCR and average interval length for the selected moderators. If the estimated-H,K version has FCR above 0.10 by more than Monte Carlo error (e.g., >0.12) while the oracle version attains 0.90 coverage, then the missing H,K theory is consequential. Additionally, derive the first-order term in the pivot's Edgeworth expansion due to plug-in H,K; if this term is not o(1) uniformly over distributions satisfying Assumptions 2-4, Theorem 4.2 does not justify the implemented method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 defines H = E[(1/n)Σ X_i^T ∇²ψ(X_i,E β_E^n; Y_i) X_i] and K = Cov((1/√n)Σ X_i^T ∇ψ(X_i,E β_E^n; Y_i)) as population moments. The master statistics bg^{E.j}_n, the matrices P_1^{E.j}, P_2^{E.j}, η^{E.j}, Λ, Q^{E.j}, the truncation interval [I_-^{E.j}, I_+^{E.j}] in Proposition 4.4, and the pivot F(x,y) in (11) all depend on H and K. In the WCLS and EMEE semiparametric settings, these population matrices are not known; a practitioner must estimate them from data. The paper discusses nuisance function estimation in Section 5.2 but is completely silent on estimating H and K. Theorem 4.2 states the coverage guarantee for the pivot computed with the population H and K, and the proofs in Theorem 4.1 and Theorem 11.2 treat eP, T_n, and the change-of-variables density as fixed. If Ĥ and K̂ are plugged in, the conditional distribution of the pivot is no longer the one analyzed. The absence of any uniform perturbation bound for Ĥ and K̂ means the main theorem does not actually apply to the method as implemented, even in the fixed-p regime. This is a concrete, load-bearing gap between the central claim and the proposed algorithm.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a two-step selective inference procedure for time-varying causal effect moderation in mobile-health settings. In the first step, a Gaussian-randomized lasso is applied to a weighted centered least-squares (WCLS) criterion to select a sparse set of effect moderators. In the second step, the authors construct a pivot for each selected coefficient by conditioning on a subset of the selection event, marginalizing over the randomization, and applying a probability integral transform. The main theoretical result, Theorem 4.2, claims that the resulting selective confidence intervals achieve uniform asymptotic conditional coverage over a class of data-generating distributions satisfying Assumptions 2–4. The paper also reports simulations under Gaussian, Laplace, and exponential errors showing that the method attains nominal false coverage rates with shorter bounded intervals than the polyhedral approach of Zhao et al. (2021), and it applies the method to data from the VALENTINE mHealth study.","tokens_in":42963,"tokens_out":6833,"duration_ms":68799,"significance":"If the main theorem is correct, the paper offers a genuinely useful extension of Gaussian-randomization selective inference from fixed-X Gaussian regression to semiparametric WCLS/EMEE-type estimators for causal excursion effects. The pivot construction is carefully motivated and the proofs, especially the Stein-based uniform asymptotic arguments in the appendix, are nontrivial. The empirical comparison is informative and suggests the method can improve on both data splitting and polyhedral selective inference in low-signal, heavy-tailed settings. The claim of bounded intervals is a practical advantage. However, the significance is tempered by a gap between the theoretical object for which coverage is proved (with population H and K) and the procedure as implemented (which must estimate these matrices), and by the fixed-p nature of the asymptotic guarantee relative to the high-dimensional motivation.","major_comments":[{"comment":"The main theorem does not cover the procedure as implemented because H and K are treated as known population matrices. Section 4.2 defines H and K as population expectations, and the matrices P_1, P_2, Λ, Q, η, the truncation interval in Proposition 4.4, and the pivot F in Eq. (11) all depend on H and K. Theorem 4.2 and the proofs in Theorem 11.2 and Lemma 11.7 treat these matrices as fixed. In the WCLS and EMEE settings of Section 5.1, H and K contain population expectations over the unknown data-generating distribution and must be estimated from data. Section 5.2 discusses nuisance function estimation only, not estimation of H and K. If plug-in estimates are used, the joint distribution of the master statistics and the change-of-variables density analyzed in Theorem 4.1 are no longer the ones analyzed, and no uniform perturbation bound is supplied. The coverage guarantee therefore applies to an oracle version of the method, not to the algorithm used in the simulations and data analysis.","section":"§4.2, §4.4, §5.2"},{"comment":"The asymptotic theory is stated in the fixed-p, growing-n regime, but the paper frames the problem as high-dimensional. Section 4.2 explicitly says the master statistics have an asymptotic normal distribution 'in the fixed p and growing n regime,' and Theorem 4.2's uniformity is over distributions F_n with p fixed. The abstract and Section 2.3 motivate the method by high-dimensional moderation analysis, and the simulations use n=120, p=50, which is not a regime covered by the theorem. Unless the authors either restrict their claims to fixed p or extend the theory to p=p_n growing with n under explicit conditions, the central 'high-dimensional' claim is not supported by the stated uniform asymptotic guarantee.","section":"§4.2, §6, abstract"},{"comment":"Assumption 3 is load-bearing but its verification is asserted rather than proved. The text states that since √n(\\tilde ω_n − ω_n)=o_p(1), the condition is 'automatically satisfied' when r_n does not grow with n. This does not follow: o_p(1) does not imply the existence of a Lebesgue density q_n for \\tilde ω_n, nor the uniform sup-norm closeness of the ratio G_n/F that Assumption 3 requires. The proof of Lemma 11.6 uses Assumption 3 to replace G_n by F, so if this assumption fails, the pivot's uniformity and hence the coverage guarantee break down. The authors should either prove that their leading examples satisfy Assumption 3 under explicit conditions or state it as a high-level primitive and verify it in the WCLS and EMEE settings.","section":"§4.5, Assumption 3"}],"minor_comments":[{"comment":"The text says the intervals aim to achieve an FCR of '0.1%' but the nominal level is 0.10 (10%); please correct the typo.","section":"§6.1"},{"comment":"The definition of T^{E·j} contains a notational typo: 'H,EV^{E·j}' should presumably be 'H_E V^{E·j}'. Please clarify.","section":"§4.4"},{"comment":"In Assumption 3, q_n is written as q_n(·; 0_p, Ω), but q_n is a general Lebesgue density and is not necessarily Gaussian; the second argument is unexplained and should be removed or redefined.","section":"§4.5, Assumption 3"},{"comment":"There is an incomplete sentence: 'This estimator is then used as a plug-in value in the penalized estimation to identify a more focused subset of important moderators. to the remaining 70% of the data, where we' — the second half appears to be a fragment and should be completed.","section":"§7.3"},{"comment":"The notation F_n is used both for a single data-generating distribution and for the collection of distributions, which is confusing; please use a different symbol for the collection, e.g., \\mathcal{F}_n.","section":"Theorem 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is built heavily on the authors' own prior work, which is understandable but means the key novelties should be stated relative to Panigrahi (2023) and Panigrahi et al. (2023a) more explicitly. The most serious issue is the H/K estimation gap: if the authors can either prove a uniform perturbation result for plug-in estimates or clearly restrict the claims to an oracle setting, the contribution would be solid. The fixed-p versus high-dimensional mismatch is also important for framing. I would be willing to look at a revision that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main take: this paper is a real extension of randomized selective inference to time-varying causal moderation, and the uniform asymptotic pivot result in Theorem 4.2 is new. But the theorem as stated covers a pivot computed with population H and K. Any implementation in the WCLS or EMEE setting has to estimate H and K, and the paper is silent on how, and offers no perturbation argument for those estimates. That is a concrete gap between the coverage guarantee and the algorithm being run.\n\nWhat the paper does well: it adapts the pivot machinery from Panigrahi–Fry–Taylor and Panigrahi's model-free carving to non-Gaussian, i.i.d. data, gives a uniform asymptotic conditional guarantee, and applies it to WCLS and EMEE estimators for causal excursion effects. The bounded-interval property is real and directly addresses a known failure mode of polyhedral conditioning. The proof structure—Theorems 11.1–11.3, Propositions 11.4–11.7—is involved but coherent under the stated assumptions. This is serious technical work, not repackaging.\n\nSoft spots, in decreasing importance. First, the H/K gap above. The master statistics, the matrices P_1, P_2, eta, Lambda, Q, the truncation interval, and the pivot all depend on H and K. In the semiparametric setting these must be estimated. The paper does not state a theorem for the estimated versions, and the proofs treat H and K as fixed population quantities. Without a uniform perturbation bound, Theorem 4.2 does not apply to the procedure as implemented. This is fixable but requires real work.\n\nSecond, the framing mismatch: the abstract says high-dimensional, but the theory is fixed-p, growing-n, and the simulations use n=120, p=50. That is not high-dimensional in the usual sense. The method may still be useful, but the claim should be tuned to the actual regime.\n\nThird, Assumption 3 is a density-level closeness condition on the perturbed randomization. The paper says it is automatically satisfied when r_n is bounded, but the assumption itself is stronger than the crude op(1) convergence of the randomization error, and no sufficient conditions are checked for the simulation data-generating processes. It deserves a clearer treatment.\n\nMinor issues: Table 1's caption says a 300-by-30 design while Section 6 says n=120, p=50; the code link is mentioned but no URL is given; simulation details are incomplete.\n\nVerdict: conditional accept. The paper is for selective-inference and mHealth methodologists. It deserves a serious referee. The central theorem is plausible but needs to be aligned with the estimated H and K that the method actually uses.","headline":"A serious, technically strong extension of randomized selective inference to time-varying causal moderation, but the main coverage theorem is stated for population H and K while the implemented procedure must estimate them—that gap needs fixing before the claimed guarantee covers the algorithm.","tokens_in":43525,"tokens_out":2958,"would_cite":true,"duration_ms":28649,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F25","62G20","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"After using a randomized lasso to choose moderators, conditioning on a slice of the selection event yields uniformly asymptotically valid confidence intervals for time-varying causal effect moderation, with bounded length in settings…","keywords":["effect moderation","selective inference","randomization","semiparametric inference","time-varying causal effects","randomized lasso","confidence intervals","micro-randomized trials"],"falsifier":"Take the low-signal, Laplace-error simulation setting of Section 6 ($n=120$, $T=30$, $p=50$) and repeat the 500-replicate coverage study, but draw the randomization $\\sqrt{n}\\omega_n$ from a heavy-tailed $t$ distribution with the same covariance instead of a Gaussian. If the empirical conditional coverage of the 90% intervals falls clearly below 0.9, then Assumption 3's uniform closeness of the randomization density fails in exactly the regime the simulations showcase, and Theorem 4.2's guarantee is not what produces the reported coverage.","tokens_in":42408,"feed_emoji":"📊","tokens_out":9042,"duration_ms":78156,"temperature":0.7,"pith_summary":"Choosing which features moderate a time-varying treatment effect creates a statistical bind: a high-dimensional analysis of all candidate moderators is hard to interpret and masks the true moderators, while separate marginal analyses produce many false positives from correlated features. This paper proposes a two-step solution for causal effect moderation. A lasso with added Gaussian noise selects a smaller working model, and inference then conditions on a carefully chosen part of the selection event through a truncated-normal pivot, using all of the data rather than discarding a split-half. The paper's main theorem gives uniformly asymptotic validity of the resulting confidence intervals across a broad class of non-Gaussian distributions, and simulations show nominal coverage with bounded, signal-adaptive intervals where polyhedral selective inference undercovers or returns infinitely long intervals and data splitting undercovers at low signal.","feed_headline":"Selective intervals stay valid at low signal strength","feed_subtitle":"A randomized-lasso pivot yields bounded, shorter intervals where existing conditional methods fail.","key_machinery":"The load-bearing object is the pivot $$$P^{{E\\cdot j}}$(\\hat $b^{{E\\cdot j}}$_n,\\hat $g^{{E\\cdot j}}$_n;\\$beta^{{E\\cdot j}}$_n)= \\frac{\\int_{-\\infty}^{\\sqrt{n}\\hat $b^{{E\\cdot j}}$_n}\\$\\varphi$(x;\\sqrt{n}\\$beta^{{E\\cdot j}}$_n,(\\$sigma^{{E\\cdot j}}$)^2)F(x,\\sqrt{n}\\hat $g^{{E\\cdot j}}$_n)\\,dx}{\\int_{-\\infty}^{\\infty}\\$\\varphi$(x;\\sqrt{n}\\$beta^{{E\\cdot j}}$_n,(\\$sigma^{{E\\cdot j}}$)^2)F(x,\\sqrt{n}\\hat $g^{{E\\cdot j}}$_n)\\,dx},$$ where $F$ integrates the Gaussian randomization density over the truncation interval $[I^{E\\cdot j}_-,I^{E\\cdot j}_+]$ determined by the conditioning event. The randomized lasso's K.K.T. stationarity conditions are what connect selection to the master statistics; the independent Gaussian noise makes the change of variables from the randomization to the selection variables exact in the Gaussian case, and Assumption 3 extends that Gaussian weight to general distributions. Inverting the pivot gives the selective confidence intervals.","core_discovery":"The central discovery is that the randomized-lasso selection event can be sliced into a one-dimensional truncation region plus a lower-dimensional conditioning statistic, and that slicing is enough for valid inference. Using the K.K.T. conditions of the randomized lasso, the paper rewrites selection in terms of master statistics; conditioning on $\\hat S_n=S$ and $\\hat V^{E\\cdot j}_n=V^{E\\cdot j}$ truncates the statistic $\\hat U^{E\\cdot j}_n$ to an interval $[I^{E\\cdot j}_-,I^{E\\cdot j}_+]$. After further conditioning on the nuisance statistic $\\hat g^{E\\cdot j}_n$, marginalizing the truncated Gaussian randomization density, and applying a probability integral transform, the pivot in (12) is exactly uniform in the Gaussian fixed-design case and asymptotically uniform in the semi-parametric case. Theorem 4.2 states that, under Assumptions 2, 3, and 4, the selective interval $(L^{E\\cdot j}_{n,\\alpha}, U^{E\\cdot j}_{n,\\alpha})$ satisfies $$\\lim_{n\\to\\infty}\\sup_{F_n\\in\\mathcal{F}_n} \\left|P\\left(\\$beta^{{E\\cdot j}}$_n\\in ($L^{{E\\cdot j}}$_{n,\\$\\alpha$}, $U^{{E\\cdot j}}$_{n,\\$\\alpha$})\\mid \\hat S_n=S,\\hat $V^{{E\\cdot j}}$_n=$V^{{E\\cdot j}}$\\right)-(1-\\$\\alpha$)\\right|=0.$$ Because the conditioning event is a strict subset of the full selection event $\\{\\hat E=E\\}$, the tower property of expectation transfers the conditional guarantee to the unconditional coverage and to false-coverage-rate control.","pith_inferences":["A natural next test, not performed in the paper, is to push the method into $p>n$ regimes: the theorem is proved for fixed $p$ with growing $n$, while the motivating language describes high-dimensional moderation, so coverage in that regime is an open question that simulations with $n=120$, $p=50$ do not settle.","The same conditioning construction suggests a tuning principle: the analyst can choose how much randomization variance $\\tau^2$ to inject, trading selection accuracy against interval length; the paper's $\\Omega=\\tau^2 I_p$ choice is one point on that frontier, and data-dependent choices could shorten intervals further.","If the uniform-closeness assumption on the perturbation density can be replaced by a coupling or Edgeworth-type correction, the pivot approach could extend to discrete or non-Gaussian randomization schemes without losing the bounded-interval property."],"forward_implications":["Practitioners can use the full dataset for both moderator selection and inference, so intervals reflect the strength of the observed signal instead of inheriting the fixed width of a split sample.","Coverage remains at the nominal level under non-Gaussian, autocorrelated errors and low signal strengths, the regime where the polyhedral method undercovers and commonly produces infinitely long intervals.","The uniform guarantee holds for parameter sequences growing as $r_n=o(n^{1/6})$, so the conditional coverage statement covers more than local alternatives fixed as $n$ grows.","The same pivot construction extends to other causal contrasts with a linear loss formulation, including relative-risk excursion effects for binary outcomes, via the loss-framework reformulation in the appendix."],"supporting_citations":[{"why":"Establishes the conditional selective-inference framework and the truncated-normal/polyhedral construction that this paper extends and later compares against.","marker":"Lee et al. (2016)"},{"why":"Prior selective inference for effect moderation via the lasso; supplies the comparison method whose intervals can be infinite and undercover at low signal.","marker":"Zhao et al. (2021)"},{"why":"Defines the causal excursion effect and the weighted centered least squares estimator used to estimate the selected moderation model.","marker":"Boruvka et al. (2018)"},{"why":"Provides the exact Gaussian-randomization selective-inference construction whose pivot and change-of-variables derivation this paper generalizes to asymptotic semi-parametric settings.","marker":"Panigrahi et al. (2023a)"},{"why":"Supplies the Stein-type bound used in the proof of the uniform asymptotic coverage result (Lemma 11.7).","marker":"Panigrahi (2023)"},{"why":"Gives the Neyman-orthogonal nuisance estimation used in the WCLS objective so that nuisance misspecification does not drive the inferential guarantees.","marker":"Shi and Dempsey (2023)"}],"fun_headline_variants":["Randomized lasso slicing yields valid selective intervals","Bounded intervals where existing conditional methods fail","New pivot for time-varying effect moderation inference","Sliced selection event gives uniform valid coverage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The coverage guarantee rests on Assumption 3, that the density of the perturbed randomization variable $\\sqrt{n}\\tilde{\\omega}_n$ is uniformly close to the Gaussian density inside the pivot's selection-probability integral; if that approximation fails, the pivot is no longer uniform and the claimed $(1-\\alpha)$ conditional coverage can break down, and the theorem is also stated for fixed $p$ with growing $n$, so the high-dimensional settings mentioned in the abstract are not covered by the proof.","fun_headline_variants_meta":{"raw":{"variants":["Randomized lasso slicing yields valid selective intervals","Bounded intervals where existing conditional methods fail","New pivot for time-varying effect moderation inference","Sliced selection event gives uniform valid coverage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1258,"prompt_tokens":1064,"completion_tokens":194,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":137}},"tokens_in":680,"tokens_out":194,"duration_ms":2608,"temperature":1.0,"reasoning_tokens":137,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:44:32.178049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the low-signal, Laplace-error simulation setting of Section 6 ($n=120$, $T=30$, $p=50$) and repeat the 500-replicate coverage study, but draw the randomization $\\sqrt{n}\\omega_n$ from a heavy-tailed $t$ distribution with the same covariance instead of a Gaussian. If the empirical conditional coverage of the 90% intervals falls clearly below 0.9, then Assumption 3's uniform closeness of the randomization density fails in exactly the regime the simulations showcase, and Theorem 4.2's guarantee is not what produces the reported coverage.","supporting_citations":[{"cited_title":"S., and Ertefaie, A","cited_arxiv_id":null,"evidence_quote":"Prior selective inference for effect moderation via the lasso; supplies the comparison method whose intervals can be infinite and undercover at low signal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the causal excursion effect and the weighted centered least squares estimator used to estimate the selected moderation model."},{"cited_title":"and Dempsey, W","cited_arxiv_id":null,"evidence_quote":"Gives the Neyman-orthogonal nuisance estimation used in the WCLS objective so that nuisance misspecification does not drive the inferential guarantees."}],"review_version":1}