{"id":"effc0f27-d0ca-4dec-a18e-0f9fa333f14a","arxiv_id":"2608.06654","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PD-Robust is a triply robust pseudo-outcome estimator for subgroup-specific effects on recovery trajectories among patients who would survive regardless of exposure, and it identifies males under 85 as losing up to 23 days at home after a hospital-acquired condition.","lead":"Researchers built a statistical method that estimates how a hospital complication changes the number of days older dementia patients spend at home after a hip fracture, while accounting for patients who die. Applied to Medicare data, it finds that men under 85 with a complication may lose up to 23 days at home over six months, a difference larger than the 8-day threshold considered clinically meaningful.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sensitivity analysis for principal ignorability is scalar and not stratified by the effect-modifying subgroup; the 23-day estimate for males under 85 may not be robust to heterogeneous violations.","rationale":"The central methodological contribution is the PD-Robust estimator and Theorem 2.2. I checked the algebraic structure of the pseudo-outcomes in Section 2.4: under monotonicity and Assumptions 1-3, phi1 and phi0 have expectations e11 * m11_1 and e11 * m11_0 respectively, provided the relevant principal-ignorability equality holds; the estimating equation then targets E[g(e11(tau-f))]=0, which is the projection estimand. The cancellation patterns for correct pairs among PS/PPS/CM in Section 2.4 are plausible, and the simulation results support the triple-robustness claim. Thus I do not see an internal inconsistency in the method. The weakest link is the applied causal claim: the 23-day effect in males under 85 is identified only under Assumption 4, and the sensitivity analysis in Section 2.5/Figure 4B gives only a homogeneous, scalar perturbation. Since the headline subgroup is defined by the same covariates that the working model interacts, a violation of principal ignorability that is larger for that subgroup is exactly the scenario that could overturn the clinical conclusion; the current figure does not address it. I therefore agree with the reader's CONDITIONAL verdict and recommend no change.","tokens_in":19972,"tokens_out":16889,"duration_ms":138347,"concrete_test":"Re-run the Section 2.5 sensitivity analysis stratified by the effect-modifying covariate profile: set epsilon0 = c for the male-under-85 subgroup, with c varying from 1 down to 0.5 (and also 1.2 to check the direction), while holding epsilon0=1 for all other profiles, and recompute the aggregated six-month DAH effect for male-under-85 with bootstrap 95% confidence intervals. If the point estimate or lower confidence bound crosses the 8-day threshold for any c >= 0.7, the clinical significance claim depends on an unverified homogeneous-violation assumption. This is implementable with the provided PD_Robust R package by passing subgroup-specific epsilon0 weights.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The applied claim that males under 85 lose up to 23 DAH, exceeding the 8-day threshold, is the load-bearing conclusion for clinical significance. This estimate is identified only under Assumption 4 (principal ignorability), which is untestable. The sensitivity analysis in Section 2.5 and Figure 4B varies a single scalar epsilon0(t,t*,X) and assumes epsilon0<=1, but treats violation as homogeneous: it does not report subgroup-specific sensitivity curves for the very subgroup (male, under 85) that drives the headline. Because the working model includes sex-by-age interaction, the projection parameter for that subgroup can behave differently from the population-averaged effect; a violation of principal ignorability concentrated in that subgroup (e.g., HAC selectively removes younger males with poor unexposed recovery, making their epsilon0 substantially below 1) could shrink the estimated six-month effect below 8 days even though the overall sensitivity pattern looks stable. The current figure thus does not establish that the clinical significance conclusion is robust to the most relevant assumption violation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes PD-Robust, a frequentist projection-based method for estimating heterogeneous treatment effects on a longitudinal outcome that is truncated by death, within a principal stratum of patients who would survive regardless of exposure. The target parameter beta_{t0} is defined as the projection of tau(t, t*, Xtilde) = E[Y^1(t) - Y^0(t) | Xtilde, U(t*)=(1,1)] onto a parametric working model f(t, t*, Xtilde; beta_t). Under Assumptions 1-4, the authors construct pseudo-outcomes from propensity score, principal score, and conditional outcome models and estimate beta by solving Equation (4); Theorem 2.2 claims triple robustness and asymptotic normality. The case study applies the method to Medicare claims for older adults with ADRD after hip fracture, estimating HAC effects on six-month days-at-home and reporting that males under 85 lose up to 23 DAH, exceeding the 8-day clinically meaningful threshold. Simulations under various combinations of correctly and incorrectly specified nuisance models show low bias and near-nominal coverage, and the paper provides an R package and practical guidance.","tokens_in":20254,"tokens_out":6807,"duration_ms":65427,"significance":"If Lemma 2.1 and Theorem 2.2 hold, the paper makes a substantive methodological contribution: a triply robust, finite-dimensional projection estimator for principal-stratum heterogeneous effects, together with model diagnostics, sensitivity tools, and principal-stratum profile characterization. The application addresses an important clinical question, and the concrete subgroup finding is falsifiable and clinically relevant. The paper also ships an R package and detailed implementation guidance, which is a practical strength. The main weakness is that the most prominent applied claim—males under 85 lose up to 23 DAH—relies on an untestable principal ignorability assumption, and the sensitivity analysis presented for this assumption is not stratified by the subgroup that drives the headline result.","major_comments":[{"comment":"The sensitivity analysis for Assumption 4 (principal ignorability) varies a single scalar epsilon0(t, t*, X) over (0,1] and does not report subgroup-specific sensitivity curves for the males-under-85 subgroup that drives the abstract's 'up to 23 fewer DAH' claim. Because Model 3 includes a sex-by-age interaction, a violation of principal ignorability concentrated in younger males (e.g., HAC selectively removes younger males with poor unexposed recovery) could change the subgroup projection parameter without affecting the overall sensitivity pattern shown in Figure 4B. Please report sensitivity results separately for the males-under-85 subgroup, or use X-dependent epsilon0 values that represent violations concentrated in that subgroup, and state whether the estimated effect remains above the 8-day threshold under those violations. If the current implementation already permits X-dependent epsilon0, the figure and text should describe how those covariate-specific values were chosen.","section":"Section 2.5 and Figure 4B"},{"comment":"Lemma 2.1 and Theorem 2.2 are the load-bearing theoretical results of the paper, but their proofs and all regularity conditions are deferred to the Supplementary Material, which is not included in the manuscript text. The main text should state the extra conditions needed for asymptotic normality or clearly identify the supplement statement and confirm that the supplement is part of the submission. As written, the triple robustness and asymptotic normality claims cannot be fully verified from the manuscript alone.","section":"Theorem 2.2 and Lemma 2.1"},{"comment":"The simulation evaluation focuses almost entirely on the first element of beta_t, while the case-study conclusion relies on interaction coefficients that distinguish males under 85 from other subgroups. To support the applied claim, please add simulation results for the interaction parameters (or at least for the subgroup-specific linear combination corresponding to males under 85), especially under partial model misspecification, so that the finite-sample behavior of the parameters driving the headline is assessed directly.","section":"Section 4.2 and Table 2"}],"minor_comments":[{"comment":"Please clarify whether 'up to 23 fewer DAH' is a six-month cumulative difference or a month-specific estimate for the male-under-85 subgroup, and identify the model and figure from which this number is taken.","section":"Abstract and Section 3"},{"comment":"The definition of \\hat\\phi_{0,i} uses \\hat\\psi_{S1,i} before \\hat\\psi_{S1,i} is defined; please reorder the displayed definitions to avoid circular reading.","section":"Section 2.4, Eq. (4)"},{"comment":"The 'standardized t-statistic' diagnostic for the PPS model is informal; please state a threshold or reference distribution for judging when the t-statistics are small enough to indicate adequate PPS fit.","section":"Figure 3B"},{"comment":"There are typographical errors: 'devision' should be 'deviation' in Table 2, and 'confidence internals' should be 'confidence intervals' in Figure 5.","section":"Table 2 and Figure 5 captions"},{"comment":"The exposure ignorability sensitivity analysis is described only by reference to Section S1.6.2 of the Supplement; please provide at least a brief account of its identifying assumptions and interpretation in the main text.","section":"Section 2.5"}],"recommendation":"major_revision","confidential_remarks":"The methodological idea appears sound and the simulation evidence is generally supportive, but the applied conclusion of clinical significance is not yet backed by a subgroup-specific sensitivity analysis for the untestable principal ignorability assumption. I did not have access to the supplementary material; since Lemma 2.1 and Theorem 2.2 are central to the manuscript, the editor should ensure that the supplement, including proofs and regularity conditions, is made available to reviewers in any revision. The paper fits the journal's scope well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper. It does something new: a triply robust estimator for effect modification on a trajectory truncated by death, using principal stratification and a projection-based estimand. The pseudo-outcome construction is clean, and the simulation evidence supports the triple-robustness claim. The case study on Medicare claims is well executed and finds a clinically meaningful 23-day reduction in days at home for males under 85, which is the kind of applied payoff that makes the method worth attention.\n\nThe paper earns credit for the R package and code, for clear model diagnostics, and for being upfront about limitations in the discussion.\n\nSoft spots: the main theorems and sensitivity derivations live in the supplementary material, which I could not inspect; that is okay for a preprint but means the strongest formal claims are unverified here. More concerning for the applied claim: the sensitivity analysis for principal ignorability varies a scalar epsilon_0(t,t*,X) and treats the violation as homogeneous. The headline result is driven by males under 85, and a violation concentrated in that subgroup would not be caught by the overall sensitivity plot. The stress-test note makes that point well. So the 23-day estimate should be read as conditional on a fairly strong, untestable assumption. That is not a flaw in the method; it is a limitation of the application. The authors explicitly call for extending to partial identification, so they are aware.\n\nOverall: solid methods paper, honest about assumptions, with a credible simulation study. The applied claim is interesting but more fragile than the main text suggests. Worth a serious referee; I would encourage referees to ask for subgroup-specific sensitivity curves and access to the supplementary proofs.","headline":"A genuinely new triply robust estimator for effect modification under truncation by death; the case study's headline subgroup is more assumption-dependent than the sensitivity analysis acknowledges.","tokens_in":20700,"tokens_out":2778,"would_cite":true,"duration_ms":24020,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that its pseudo-data estimator PD-Robust yields consistent, asymptotically normal subgroup-effect estimates under truncation by death when any two of three nuisance models are correctly specified, and that the method…","keywords":["heterogeneity of treatment effects","principal stratification","truncation by death","triple robustness","days at home","Medicare claims","effect modification","pseudo-outcome estimation"],"falsifier":"A re-analysis that lets the principal-ignorability ratio $\\epsilon_0(t,t^*,X)$ vary with covariates rather than stay a fixed scalar, and shows the 23-day estimate dropping below the 8-day threshold under mild violations, would falsify the clinical conclusion; more directly, a dataset with both potential survival outcomes observed, such as a fully followed randomized trial, could verify whether the principal ignorability equalities hold at all.","tokens_in":19738,"feed_emoji":"🏥","tokens_out":11934,"duration_ms":95766,"temperature":0.7,"pith_summary":"The paper develops PD-Robust, a frequentist semiparametric estimator for effect modification when the outcome trajectory is truncated by death, and uses it to ask who is hurt most when hospitalized dementia patients acquire a hospital-acquired condition (HAC) after hip fracture. The central methodological claim is triple robustness: solving the PD-Robust estimating equation yields a consistent estimator of the projection parameter $\\beta_{t0}$ if any two of the propensity score, principal score, and conditional outcome mean models are correctly specified, and the estimator is asymptotically normal. Applying the method to Medicare claims, the paper reports that males under 85 who would survive regardless of exposure lose up to 23 days at home over six months post-discharge due to HAC, exceeding the 8-day threshold considered clinically meaningful. The paper also provides model diagnostics, sensitivity analyses for the untestable principal ignorability assumption, and tools to characterize the patient profile of the unobserved always-survivor stratum. A reader would care because survivor-only analyses are biased and treating death as zero days at home conflates death with institutional care, whereas this approach offers an identifiable causal estimate for a clinically interpretable subpopulation.","feed_headline":"23 days at home lost to hospital-acquired conditions in men under 85","feed_subtitle":"A death-aware statistical method isolates the true cost of hospital harm for dementia patients after hip fracture.","key_machinery":"The load-bearing machinery is the PD-Robust pseudo-data construction. For each subject at each time $t$, the procedure builds a pseudo-outcome $\\hat\\phi_{1,i} - \\hat\\phi_{0,i}$ and a pseudo-scalar $\\hat\\psi_{S1,i}$ from the fitted propensity score, principal score, and conditional mean models, then solves the estimating equation $\\sum_i g(t,t^*,\\tilde X_i;\\beta_t)\\{\\hat\\phi_{1,i} - \\hat\\phi_{0,i} - \\hat\\psi_{S1,i}\\, f(t,t^*,\\tilde X_i;\\beta_t)\\} = 0$. The defining identities are $E(\\phi_1 - \\phi_0) = E[\\{Y^{a=1}(t)-Y^{a=0}(t)\\}I\\{U(t^*)=(1,1)\\}]$ and $E(\\psi_{S1}) = P(U(t^*)=(1,1))$, which hold whenever at least one pair of the three nuisance models is correct; because the product of two estimation errors enters the influence function, any two correct models suffice, producing triple robustness. The estimand itself is the projection parameter $\\beta_{t0} = \\arg\\min_{\\beta_t} E[(\\tau^{(1,1)}(t,t^*,\\tilde X) - f(t,t^*,\\tilde X;\\beta_t))^2 \\mid U(t^*)=(1,1)]$, which maps the possibly high-dimensional conditional effect onto an interpretable structural working model such as the linear form $\\eta^T(\\tilde X)\\beta_t$.","core_discovery":"The paper's central discovery is that the causal effect of an exposure on a longitudinal outcome among the always-survivor principal stratum, the patients who would be alive at a fixed time $t^*$ regardless of exposure, can be recovered under principal ignorability through the projection-based estimand $\\beta_{t0}$ defined in Equation (3), and that this parameter admits a triply robust estimator. Theorem 2.2 states that the estimator obtained by solving Equation (4) converges in probability to $\\beta_{t0}$ if any two of the three nuisance working models, for the propensity score $\\pi(X)$, the principal score $e^{(1,1)}(t^*,X)$, and the conditional outcome mean $\\mu_{as}(t,t^*,X)$, are correctly specified, and that the estimator is asymptotically normal. On the applied side, the paper claims this identifies a clinically significant heterogeneity: within the always-survivor stratum, males under 85 years experience up to 23 fewer days at home in the six months after discharge when HAC occurs, compared with the 8-day minimal clinically meaningful difference, while the estimated effects for females are near zero.","pith_inferences":["Because the sensitivity analysis perturbs only a scalar $\\epsilon_0(t,t^*,X)$, the 23-day estimate has not been stress-tested against violations of principal ignorability that differ by age or sex; letting the ratio vary with covariates is a direct next stress test.","The near-agreement with the adapted DR-learner may be specific to this cohort's relatively low death rate and large always-survivor stratum; in populations where survival differs sharply by exposure, the two approaches would likely diverge, which is a testable prediction.","The same pseudo-outcome template could be carried over to composite endpoints such as days alive and at home, or to count outcomes with a log-link working model, generalizing the projection estimand beyond the linear working models used here."],"forward_implications":["Estimating subgroup effects in studies with heavy post-treatment mortality no longer forces the choice between survivor-only analyses, which are biased, and coding death as zero, which distorts institutional-care utilization; PD-Robust targets the always-survivor stratum directly.","Consistency survives misspecification of any one of the three nuisance models: a wrong propensity score, principal score, or outcome regression alone does not break the estimate, so applied conclusions rest on a weaker modeling requirement than standard doubly robust procedures.","The reported 23-day reduction for males under 85 is a causal claim about patients who would survive to month 6 regardless of HAC, not an observed association, and it exceeds the 8-day threshold regarded as clinically meaningful.","Clinicians and payers can use the profiling tools to identify which observed patients most resemble the always-survivor stratum, making the causal estimate actionable rather than confined to a latent subpopulation.","Because the method ships with an R package, the same pipeline can be applied to other claims-based trajectory questions without re-deriving the influence function."],"supporting_citations":[{"why":"Defines principal stratification, the framework that gives the always-survivor stratum its causal meaning.","marker":"Frangakis and Rubin (2002)"},{"why":"Argues that death is an absorbing state and that outcomes after death are undefined, motivating principal-stratum estimation instead of survivor-only or zero-imputed outcomes.","marker":"Rubin (2006)"},{"why":"Introduces the principal score and principal ignorability identification that the PPS nuisance function relies on.","marker":"Ding and Lu (2017)"},{"why":"Establishes multiply robust estimation and the covariate-balancing diagnostic under principal ignorability that PD-Robust builds on.","marker":"Jiang, Yang and Ding (2022)"},{"why":"Extends principal stratification to continuous post-treatment variables and semiparametric estimation, grounding the projection-based estimand.","marker":"Lu, Jiang and Ding (2025)"},{"why":"Supplies the pseudo-outcome and debiased machine learning template behind the PD-Robust estimating equation.","marker":"Semenova and Chernozhukov (2021)"},{"why":"The existing Bayesian machine learning approach to heterogeneous survivor causal effects that this frequentist method positions itself against.","marker":"Chen et al. (2024)"},{"why":"Sets the 8-day clinically meaningful threshold for days at home used to judge the 23-day finding.","marker":"Auriemma et al. (2023)"}],"fun_headline_variants":["Men under 85 with dementia lose 23 days at home to hospital harm","Death-aware method quantifies 23-day hospital harm for certain men","Hospital-acquired conditions cost men under 85 23 days at home","New tool shows hospital harm costs men under 85 with ADRD 23 days"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"What everything rests on is principal ignorability, that once covariates are accounted for, patients who die under a given exposure have the same average outcome as patients who survive under that same exposure, an assumption no data can verify and one the paper checks only against a single scalar violation pattern.","fun_headline_variants_meta":{"raw":{"variants":["Men under 85 with dementia lose 23 days at home to hospital harm","Death-aware method quantifies 23-day hospital harm for certain men","Hospital-acquired conditions cost men under 85 23 days at home","New tool shows hospital harm costs men under 85 with ADRD 23 days"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000354,"raw_usage":{"total_tokens":1977,"prompt_tokens":1047,"completion_tokens":930,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":849}},"tokens_in":663,"tokens_out":930,"duration_ms":8308,"temperature":1.0,"reasoning_tokens":849,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:03:12.291183+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A re-analysis that lets the principal-ignorability ratio $\\epsilon_0(t,t^*,X)$ vary with covariates rather than stay a fixed scalar, and shows the 23-day estimate dropping below the 8-day threshold under mild violations, would falsify the clinical conclusion; more directly, a dataset with both potential survival outcomes observed, such as a fully followed randomized trial, could verify whether the principal ignorability equalities hold at all.","supporting_citations":[],"review_version":1}