{"id":"1824009e-4b85-4dba-8fd1-981163b199f3","arxiv_id":"2508.14770","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Dropping people with non-existent outcomes can bias causal effect estimates; the paper uses principal stratification to estimate the effect on outcome existence and the effect among those who would have the outcome regardless.","lead":"This paper studies what happens when researchers drop people who do not have an outcome at all, such as dropping the unemployed when studying wages. It shows this common shortcut can hide real causal effects, and offers a statistical method to separate the effect on having an outcome from the effect on its value.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Identification of the always-survivor estimand is unverified from the corrupted text; the regression-simulation framework likely needs monotonicity or strong parametric assumptions not stated in the abstract.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption being unmeasured confounding and monotonicity. My stress-test converges on a related but more specific concern: the always-survivor estimand is latent, and the proposed 'regression and simulation' framework must contain a non-testable identifying assumption beyond ignorability. This concern is exactly why the reader's conditions (point identification or sensitivity analysis, and non-reduction to prior estimators) are appropriate. I do not move the verdict because the full text is unreadable; the conditional verdict correctly requires the authors to demonstrate these properties. I would not strengthen to UNVERDICTED because the abstract's conceptual claim is plausible and consistent with the principal-stratification literature, and the reader's low confidence already reflects the corruption. The proposed concrete test would settle whether the identification concern actually lands once a clean text is available.","tokens_in":11483,"tokens_out":5299,"duration_ms":70468,"concrete_test":"Obtain a clean copy of the manuscript and re-derive the identification argument in the selection-on-observables section. Check whether it explicitly assumes monotonicity of existence in treatment or uses a principal-score/structural model. Then run a Monte Carlo experiment with a known DGP under (i) no unmeasured confounding but non-monotonic existence (e.g., 15% defiers) and (ii) an unmeasured confounder affecting both treatment and existence. Apply the proposed estimator in each scenario and compare to the true always-survivor average effect. If the estimator is biased in (i), the method relies on monotonicity; if biased in (ii), the method does not actually adjust for unmeasured confounding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central methodological claim is that the regression-simulation framework identifies the average effect among the latent always-survivor stratum under selection-on-observables. This claim is insecure. First, the always-survivor stratum is latent: even with no unmeasured confounding, the observed data distribution does not identify its average potential outcomes unless one assumes monotonicity of existence in treatment or an equivalent structural restriction. The abstract does not state such a restriction. In the parenthood application, monotonicity is doubtful: parenthood may cause some to leave employment and others to enter or intensify work, creating both 'defiers' and 'compliers' in existence. If the paper assumes monotonicity, it must defend it; if not, the estimand is only partially identified. Second, a 'regression and simulation' procedure must encode that identifying assumption explicitly, e.g., via a principal-score model or a structural model for S(1), S(0) given X. If it simply imputes missing outcomes from a regression fitted on observed survivors, it targets the effect among observed survivors, not among always-survivors, and is biased under selection. Because the supplied full text is corrupted, these derivations cannot be checked. The abstract's 'adjust for measured confounders' does not resolve the latent-stratum problem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses a common practice in social stratification research: dropping cases for which the outcome is non-existent (e.g., wages for non-employed persons). The authors argue that when a treatment affects both whether the outcome exists and its value, this practice can bias estimated causal effects, and they show that effects of both beneficial and harmful treatments can be underestimated. They propose two principal-stratification estimands: (1) the average effect on outcome existence and (2) the average effect on the outcome among the latent subgroup whose outcome would exist under either treatment condition (the always-survivor stratum). The paper extends this to selection-on-observables settings via a regression-and-simulation framework, and illustrates the approach with an application to the effects of parenthood on labor market outcomes. The conceptual point is plausible and consistent with the truncation-by-death literature, but the supplied full text is heavily corrupted and unreadable, so the identification arguments, simulation details, and empirical results cannot be verified.","tokens_in":11651,"tokens_out":3023,"duration_ms":37175,"significance":"If the methodological claims could be verified, the paper would make a useful contribution to applied inequality research by separating existence effects from value effects and by cautioning against silent conditioning on outcome presence. The proposed approach draws on established principal-stratification ideas, which is a strength. However, the manuscript as submitted is not in readable form: the full text is mojibake, and the only clear statement is the abstract. No proofs, derivations, simulation code, or reproducible results are available to check. The central conceptual claim is not new in the causal inference literature, but the paper's contribution would be in translating it for applied researchers and providing a practical estimator. That contribution cannot be assessed from the current submission.","major_comments":[{"comment":"The submitted full text is unreadable: it consists of corrupted character sequences rather than coherent prose, equations, or tables. No identification proof, simulation setup, or empirical analysis can be checked. This is a load-bearing issue because the paper's central claim rests on the regression-simulation framework correctly identifying the always-survivor effect. The authors must provide a complete, readable manuscript before any substantive review can occur.","section":"Full text (entire submitted manuscript)"},{"comment":"The abstract states that the framework 'adjust[s] for measured confounders' and enables principal stratification estimates, but it does not state the identifying restriction needed to recover the always-survivor stratum. Even under ignorability conditional on covariates, the distribution of potential outcomes among the always-survivor group is not identified from observed data without an additional structural assumption such as monotonicity of existence in treatment. The provided text does not show whether such an assumption is made and defended. If the framework simply imputes missing outcomes from a regression fitted on observed survivors, it targets the effect among observed survivors, not among always-survivors. The authors must state and justify the identifying assumptions explicitly.","section":"Abstract (regression-simulation framework)"},{"comment":"The parenthood application is mentioned in the abstract but the corresponding analysis is not readable. Given the identification concern above, the application is especially vulnerable: monotonicity of employment existence with respect to parenthood is doubtful, since parenthood may lead some people to leave employment and others to enter or intensify work. Without a defense of monotonicity or a sensitivity analysis relaxing it, the empirical estimates are not interpretable as always-survivor effects. The authors should clarify how the application handles this.","section":"Empirical example (parenthood and labor market)"}],"minor_comments":[{"comment":"The document contains an unrelated arXiv identifier ('arXiv:2508.14772v2 [cond-mat.str-el]') and other extraneous artifacts. These should be removed in a clean submission.","section":"Full text"},{"comment":"The phrase 'non-existent outcomes' is intuitive but the paper should more explicitly link it to the established 'truncation-by-death' literature in the abstract, which would help situate the contribution.","section":"Abstract"},{"comment":"Once a readable manuscript is provided, the authors should include a formal definitions section with notation for potential outcomes, existence indicators, and principal strata, since these are central to the proposed estimands.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central idea is plausible and potentially useful, but the submitted text is entirely unreadable. I am recommending major_revision rather than reject because the scientific idea appears defensible in principle. However, if the authors cannot provide a complete, readable manuscript—including derivations and simulation details—the paper should not proceed; in that case rejection would be appropriate. I would also encourage the editor to consider desk-rejecting corrupted submissions as a policy matter."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take. The abstract of 2508.14770 makes a correct and useful point: when a treatment affects both whether an outcome exists and its value, dropping the non-existent cases can bias estimated effects in either direction. The truncation-by-death literature already says this, but the paper frames it specifically for inequality and stratification research, where conditioning on employment or other post-treatment states is endemic. That framing is a real service.\n\nThe proposed estimands are standard principal stratification: an effect on existence and an effect on the outcome among the latent always-survivor group. The apparent novelty is a regression-and-simulation implementation for selection-on-observables settings, plus a parenthood illustration. That is a plausible applied contribution, and if the simulation works as advertised, it gives applied people a tool they didn't have in a convenient form.\n\nNow the soft spots. The supplied full text is corrupted mojibake, so I can only evaluate the abstract. That means I cannot verify the identification arguments, the simulation, or the empirical results. The abstract says 'adjust for measured confounders,' but that alone does not identify the always-survivor mean. The always-survivor stratum is latent; without a monotonicity assumption or a principal-score model, the observed distribution does not pin down its average potential outcome. The abstract does not state such a restriction. In the parenthood application, monotonicity is doubtful—parenthood can push some out of employment and others into it. If the full paper assumes monotonicity, it needs to be defended; if it uses a principal-score approach, the paper needs to show how it differs from existing estimators. This is the key stress-test concern, and the reader's conditional verdict is appropriate.\n\nWhat is good here: the paper identifies a common practice hazard, offers estimands that separate existence from value, and does so in a way applied researchers might actually use. The conceptual core is sound, and the parenthood example is a natural test case. This is worth referee time if the full manuscript is readable and the identification section confronts these issues.\n\nMy recommendation: don't desk reject. Ask the authors for a clean version, then send it to a referee who knows principal stratification. If the always-survivor estimand is identified only under an unstated assumption, that is a fixable revision. The paper is serious.\n\nI'd bring it to reading group once the text is fixed, and I wouldn't cite it before I can see how the regression-simulation works.","headline":"The conceptual point is solid and applied researchers will care, but the supplied full text is unreadable and the always-survivor estimand needs an explicit identifying restriction; worth sending to peer review after a clean manuscript is provided.","tokens_in":12196,"tokens_out":2989,"would_cite":false,"duration_ms":34131,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62P25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that routine practice of dropping cases whose outcome doesn't exist—like wages for the unemployed—can understate causal effects, and replaces it with two principal-stratification estimands that separate the effect on whethe","keywords":["non-existent outcomes","principal stratification","causal inference","complete-case analysis","selection on observables","treatment effects","labor market inequality","parenthood effects"],"falsifier":"Generate simulated data with known potential outcomes in which a treatment changes both whether the outcome exists and its value among always-reporters, then compare the standard complete-case estimate with the paper's two estimands. The paper predicts the complete-case estimate is biased away from the true always-reporter effect; data or experiments where the two agree across a wide range of such settings would refute the central claim.","tokens_in":11276,"feed_emoji":"📉","tokens_out":5635,"duration_ms":72717,"temperature":0.7,"pith_summary":"This paper targets a problem hidden in plain sight in inequality research: many outcomes, such as wages, exist only for part of the population, and researchers routinely drop the people without them. It shows that when a treatment changes both who gets the outcome and what its value is, that dropping can make the true causal effect look smaller than it is—for both beneficial and harmful treatments. The proposed fix is to report two causal effects instead of one: the average effect on whether the outcome exists, and the average effect on its value among the latent subgroup for whom the outcome exists in either treatment condition. A regression-and-simulation implementation extends this approach to the selection-on-observables designs common in observational social science, and the paper illustrates it with parenthood and labor-market outcomes. If the argument holds, researchers can stop silently conditioning on outcome presence and instead separate existence effects from value effects.","feed_headline":"Dropping non-existent outcomes can hide true effects","feed_subtitle":"A causal framework splits the effect on whether an outcome exists from its effect among those who always have it.","key_machinery":"Principal stratification with an 'existence' event: each unit has a potential existence indicator under treatment and under control, placing it in a latent principal stratum (always-reporter, never-reporter, reporter only under treatment, reporter only under control). The two target estimands are the average causal effect on the existence indicator and the average causal effect on the value among the always-reporter stratum—the subgroup whose outcome would exist in either condition. Identification comes from selection-on-observables: regressions for existence and value, conditional on measured confounders, are used to simulate the full potential-outcome distribution and average over the impl","core_discovery":"The central claim is that conditioning on outcome presence does not estimate a causal effect on any well-defined population; it mixes people whose outcome exists under both treatment states with people whose outcome exists only under one state. If treatment shifts existence, this mixture changes, so the complete-case comparison can understate effects of both beneficial and harmful treatments. The paper proposes two principal-stratification estimands—the average effect on outcome existence and the average effect on outcome value among the latent always-reporter subgroup—and provides a regression-simulation implementation for selection-on-observables designs, illustrated on parenthood and labo","pith_inferences":["The same two-estimand logic applies to any outcome defined only conditionally—test scores among test-takers, reports among victims, earnings among taxpayers—so the framework is a general template rather than a labor-market fix.","The paper's selection-on-observables implementation could be paired with a formal sensitivity analysis that asks how much unmeasured confounding or non-monotonicity would be needed to erase an estimated always-reporter effect; the estimands give that analysis a well-defined target.","With randomized treatment assignment, the always-reporter value effect could be estimated nonparametrically, which suggests the observational framework is one practical special case of a broader identification strategy.","Published null or small effects on conditional outcomes may, in some settings, be artifacts of conditioning on outcome presence; re-analyzing existing datasets with the two-estimand decomposition would test that possibility directly."],"forward_implications":["Complete-case estimates should no longer be interpreted as the effect of the treatment on outcome values; the paper's decomposition supplies the quantities that answer that question.","If a treatment increases outcome existence, the standard estimate will tend to understate the effect among always-reporters; if it decreases existence, understatement can occur in the harmful direction as well.","In observational settings with measured confounders, the regression-simulation framework yields point estimates for both existence and always-reporter value effects.","In the parenthood and labor-market example, the effect on employment and the effect on wages among always-employed people become separable, so the usual 'motherhood penalty' estimates are interpretable rather than mixed with selection.","The proposed first estimand makes the effect on whether an outcome exists an explicit, reportable causal quantity rather than a nuisance to be conditioned away."],"supporting_citations":[],"fun_headline_variants":["Non-existent outcomes bias causal estimates: a fix","Causal effects on existence vs. values: a split","Why dropping non-existent outcomes can hide causal effects","When outcomes don't exist, causal effects get hidden"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The estimates are only valid if, after measured confounders are controlled, treatment is as good as randomly assigned for both outcome existence and outcome value, and the existence effect is monotone; neither condition can be verified from observed data.","fun_headline_variants_meta":{"raw":{"variants":["Non-existent outcomes bias causal estimates: a fix","Causal effects on existence vs. values: a split","Why dropping non-existent outcomes can hide causal effects","When outcomes don't exist, causal effects get hidden"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001088,"raw_usage":{"total_tokens":4338,"prompt_tokens":653,"completion_tokens":3685,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":397,"completion_tokens_details":{"reasoning_tokens":3623}},"tokens_in":397,"tokens_out":3685,"duration_ms":30688,"temperature":1.0,"reasoning_tokens":3623,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:17:31.239657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate simulated data with known potential outcomes in which a treatment changes both whether the outcome exists and its value among always-reporters, then compare the standard complete-case estimate with the paper's two estimands. The paper predicts the complete-case estimate is biased away from the true always-reporter effect; data or experiments where the two agree across a wide range of such settings would refute the central claim.","supporting_citations":[],"review_version":1}