{"id":"790da428-69ef-4156-ac7f-96bde5e67a08","arxiv_id":"2509.01969","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"New sufficient conditions and a reweighted formula identify mediation and path-specific effects under sample selection, applied to liver transplant listing disparities.","lead":"This paper derives a mediation formula that adjusts for sample selection bias using external data on the full referred population, enabling estimates of direct and indirect causal effects on liver transplant listing. It applies the method to clinical data and finds that insurance status affects listing through psychosocial review and through other pathways.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Case-study DAG has Disease-related as an additional X→Y mediator, so Theorem 2's single-mediator condition is not met; Eq. 6 is used outside its stated scope.","rationale":"The reader's weakest assumption concerned unmeasured post-exposure recanting witnesses; that is real and acknowledged in Section 5.3. My concern is more internal to the paper's own formal apparatus: the application's DAG contains Disease-related as a second child-of-X/parent-of-Y mediator, so Theorem 2's single-mediator condition is violated, and Theorem 3 would require the unmeasured Disease-related variables in the adjustment formula. The paper asserts that these variables need not be measured, but supplies no identification theorem for a single mediator of interest in the presence of parallel mediators under selection. This does not invalidate the general theoretical contribution, but it means the applied NIE/NDE estimates in Table 3 are not covered by the stated theorems unless an additional corollary is proved. The recommended disposition remains conditional acceptance, with the added condition of supplying that corollary or a counterexample. Hence I do not change the reader's overall verdict, but I identify a more precise gap.","tokens_in":21200,"tokens_out":22746,"duration_ms":253370,"concrete_test":"Derive E[Y(x, M1(x'))] from the DAG in Fig. 5 (X=SEP, M1=Psychosocial Review, M2=Disease-related, Y=Listed, Z=Z^T=Baseline SDOH, S depends only on Z) using the do-calculus/extended-graph steps of Appendix A.3, without conditioning on or marginalizing over M2. If the derivation succeeds, add it as a corollary and state any extra assumption (e.g., no M1-M2 edges); if it fails, construct a small parametric DGP satisfying the GAC conditions and the no-recanting-witness assumption in which Eq. (6) with M=M1 disagrees with the true E[Y(x, M1(x'))].","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2 is stated for 'a single mediator M = Ch(X) ∩ Pa(Y)' and Eq. 6 marginalizes only over M. In the case study (Section 5.2, Fig. 5), SEP (=X) has at least two children that are parents of Listed (=Y): Psychosocial Review and the set of Disease-related variables (the text explicitly says SEP affects disease etiology/severity/comorbidities, which affect listing). Therefore Ch(X)∩Pa(Y) is not the singleton M. Theorem 3 covers multiple mediators, but its formula (7) includes the product over all mediators on proper causal paths, including Disease-related, which is not measured. The paper asserts that these variables need not be measured because the direct effect amalgamates their pathways, but no theorem in the paper establishes that Eq. (6) identifies E[Y(x, Psychosocial(x'))] when there is an additional parallel mediator M2 affected by X and affecting Y, under sample selection. Standard no-selection mediation theory supports this in special cases, but the GAC/selection extension is not given. This is an omitted proof for the applied claim, not merely an untestable assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends Pearl's mediation formula to settings with sample selection bias. The main theoretical result, Theorem 2, states that under the Generalized Adjustment Criterion (GAC) for the total effect and an additional condition blocking backdoor paths between the mediator and outcome, the selected mediation formula (Eq. 6) identifies E[Y(x, M(x'))], and hence natural direct and indirect effects, when M is a single mediator equal to Ch(X)∩Pa(Y). Theorem 3 gives an analogous multiple-mediator path-specific formula. The authors support the theory with a simulation study based on the single-mediator DAG of Fig. 1(b) and apply the method to estimate the mediated and direct effects of insurance status (as a proxy for SEP) on liver-transplant listing through psychosocial review, comparing naive estimates with selection-adjusted estimates.","tokens_in":21492,"tokens_out":10730,"duration_ms":129557,"significance":"If the results are correct, the paper makes a useful contribution by combining two established strands—mediation identification and GAC-based selection-bias correction with external data—into a practically applicable procedure. The appendix contains detailed do-calculus proofs and the simulation demonstrates bias correction in one DGP. The case study addresses a clinically and socially important question: whether socioeconomic disparities in transplant listing act through the psychosocial evaluation. The paper clearly states assumptions and limitations, including the possible recanting-witness problem. However, the proofs are not machine-checked and contain notational looseness, and there is a substantial gap between the single-mediator theorem and the DAG used in the case study.","major_comments":[{"comment":"Theorem 2 is stated only for a single mediator M = Ch(X)∩Pa(Y). In the case-study DAG, SEP has two children that are parents of Listed: Psychosocial Review and Disease-related. Thus Ch(X)∩Pa(Y) is not a singleton. Eq. (6) marginalizes only over Psychosocial Review; Theorem 3 would require a product over all mediators, including Disease-related, which is not measured. The §5.2 statement that disease-related variables need not be measured because the direct effect amalgamates their pathways is not a consequence of any theorem in the paper. No result proves that Eq. (6) identifies E[Y(x, Psychosocial(x'))] under GAC/selection when an additional parallel mediator affected by X and affecting Y is present. Standard no-selection mediation theory supports this estimand in special cases, but the selection extension is missing.","section":"§5.2 / Fig. 5 / Theorem 2 / Eq. (6)"},{"comment":"The simulation uses exactly the single-mediator DGP of Fig. 1(b). It therefore cannot detect the mismatch above. If the applied estimand is to be defended, add a simulation with a parallel mediator M2 (affected by X, affecting Y, unmeasured) and show that the proposed focal-mediator estimator remains unbiased under GAC/selection, or give an analytic argument showing where M2 drops out. Without this, the empirical support for the central applied claim is absent.","section":"§4 simulation"}],"minor_comments":[{"comment":"Typographical issues: 'a a brief simulation study' (§4), 'We being by describing' (§A.1), 'selection has minimal on the effect estimates' (§5.3), 'that that socioeconomic position' (§6), and 'Compete algorithms' (§6).","section":"Throughout"},{"comment":"The notation Gpbd_{(X,M),Y} used in condition 2 is not defined; Definition 2 defines only Gpbd_{X,Y}. Please define the generalization.","section":"Theorem 2"},{"comment":"The inverse-probability weights w_i = P(S=1)/P(S=1|z_i^T) are valid only if selection depends on Z^T alone, i.e., S ⊥ (Z \\ Z^T) | Z^T. This is encoded in Fig. 5 but should be stated as an explicit assumption, especially since the case study relies on it.","section":"Eq. (8)"},{"comment":"The modification of CMAverse to incorporate selection weights is not described. A brief algorithmic description, or a statement of code/data availability, would substantially improve reproducibility.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The main barrier is the gap between Theorem 2's single-mediator scope and the case-study DAG. If the authors can supply a theorem/proof for a focal mediator in the presence of parallel mediators under GAC/selection, the applied analysis would be supported; otherwise the case study should be reframed as an illustration of the single-mediator result, with the DAG amended accordingly. The theoretical core appears plausible and worth publication after that issue is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The theoretical core of this paper is stronger than the application. Combining Correa et al.'s generalized adjustment criterion with the extended-graph approach to mediation is a legitimate new result, and the selected mediation formula (Eq. 6) plus the path-specific version (Eq. 7) are derived in enough detail that I believe they are correct for the stated conditions. The simulation is clean and matches the theory for the single-mediator DGP. That part deserves a serious referee.\n\nWhere it falls short is the case study. The stress-test note is right: Theorem 2 is stated for a single mediator M = Ch(X) ∩ Pa(Y), and in Figure 5 SEP has two children that are parents of Listed—Psychosocial Review and Disease-related. The paper's claim that disease-related variables need not be measured because the direct effect amalgamates their pathways is an assertion, not a theorem. Theorem 3 covers multiple mediators but requires measuring them all, which the case study does not. So the applied identification claim is unsupported. I'm not saying the claim is false—in standard no-selection mediation theory, the NIE through one mediator can sometimes be identified without measuring parallel mediators, especially under no-interaction assumptions. But that identification is not automatic, and the selection-bias version needs its own proof. The paper does not give it. This is a load-bearing gap, not a minor caveat.\n\nThe other assumptions are what they are—no recanting witnesses, selection only via measured SDOH—and the paper honestly flags them. But the only sensitivity check is a DAG with transportation and health literacy; there is no formal sensitivity analysis, and no data or code. Minor: the appendix has some typographical looseness in the proofs, though nothing that made me doubt the single-mediator theorem.\n\nWho gets value from this: methodologists working on selection bias in mediation, and applied researchers who have a true single-mediator setting. For the transplant application, I would not trust the estimated NIE/NDE until the multi-mediator justification is supplied. A serious referee should ask for exactly that before publication.\n\nMy recommendation: engage with the theory, but treat the case study as conditional on a missing proof. Send it to review, ideally with a request for the omitted multi-mediator identification argument and a sensitivity analysis.","headline":"A useful theoretical extension of mediation formulas to selection bias, but the liver transplant application runs outside the theorem's stated single-mediator condition.","tokens_in":21966,"tokens_out":3727,"would_cite":true,"duration_ms":43855,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that natural direct and indirect effects can be identified from selected samples when the Generalized Adjustment Criterion holds and all mediator-outcome backdoor paths are blocked by measured covariates and the selection i","keywords":["causal inference","mediation analysis","selection bias","path-specific effects","liver transplantation","social determinants of health","generalized adjustment criterion","inverse probability weighting"],"falsifier":"Measure transportation access and health literacy in the cohort of 497 referred patients and re-estimate the selection-adjusted NIE and NDE under the expanded DAG in Figure 6; a material shift in the estimates would falsify the no-recanting-witness condition. Alternatively, generate data with a hidden post-exposure common cause of the mediator and outcome and check whether Eq. (6) recovers the true nested counterfactual; if it does not, the sufficiency conditions are incomplete.","tokens_in":21134,"feed_emoji":"🩺","tokens_out":7031,"duration_ms":68848,"temperature":0.7,"texified_at":"2026-08-05T20:21:59.750964+00:00","pith_summary":"This paper tries to show that direct and indirect causal effects can still be estimated when the analysis sample is not a random slice of the target population, as long as the selection process is represented in a causal graph and baseline data on the full population are available. The authors prove a 'selected mediation formula' that reweights the usual mediation formula with inverse-probability-of-selection weights, and they give graphical conditions under which it identifies natural direct and indirect effects and, more generally, path-specific effects. They apply the formula to liver transplant listing, asking whether a patient's socioeconomic position, measured by private insurance, affects the listing decision directly or through the outcome of a psychosocial review. Their estimates show both pathways operate, and correcting for selection into the evaluated sample leaves those conclusions essentially unchanged. A sympathetic reader would care because dropout before a decision is common in transplant evaluation and in many other clinical processes, so off-the-shelf mediation estimates may be biased.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":7740,"prompt_tokens":798,"completion_tokens":6942,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":798,"completion_tokens_details":{"reasoning_tokens":6204}},"feed_headline":"Two paths drive liver-listing disparity after bias correction","feed_subtitle":"A new formula weights out dropout from transplant evaluation; insurance still works directly and through psychosocial review.","key_machinery":"The key machinery is the selected mediation formula (Eq. 6), built from two components. The first is the extended graph: each proper causal path from $X$ to $Y$ receives a new deterministic child $X^e$ of $X$, turning nested counterfactuals like $Y(x, M(x'))$ into ordinary do-interventions on the extended nodes. The second is the Generalized Adjustment Criterion, which supplies a selection-robust adjustment formula using conditional distributions given $S=1$ and an external marginal distribution for a baseline subset $Z^T$. Theorem 1 transfers validity of the criterion from the original graph to the extended graph, and the additional condition that all backdoor paths between $M$ and $Y$ are blocked by $Z$ and $S$","core_discovery":"The central claim is Theorem 2, with its multi-mediator generalization in Theorem 3. Under the Generalized Adjustment Criterion for the total effect of exposure $X$ on outcome $Y$ in a graph augmented with a selection node $S$, and provided every backdoor path between the mediator $M$ and $Y$ is blocked by $Z$ and $S$, the nested counterfactual $E[Y(x, M(x'))]$ equals the sum over $z$ and $m$ of $E[Y \\mid X=x, m, z, S=1] p(m \\mid X=x', z, S=1) p(z \\setminus z^T \\mid z^T, S=1) p(z^T)$. This identifies natural direct and indirect effects using only the selected sample plus external baseline information on $Z^T$. The same style of condition extends to edge-consistent path-specific effects with multiple mediators, producing a fusion of","pith_inferences":["If the no-recanting-witness assumption holds, the same identification strategy transfers to other clinical or administrative processes where a gatekeeping assessment mediates a disparity and dropout precedes the decision; the required data are only baseline variables for the full referred cohort.","The small gap between the naive and adjusted estimates in this cohort suggests that selection on the measured social determinants is weak relative to the effects; in populations with stronger selection, the formula would matter more, so replication in higher-bias cohorts is a natural next test.","A sensitivity analysis treating transportation and health literacy as latent common causes of psychosocial review and listing could bound how much of the reported NIE and NDE is artifact; the paper does not provide such bounds."],"forward_implications":["When the two conditions in Theorem 2 hold, researchers can estimate natural direct and indirect effects from selected data plus external baseline information, without observing outcomes for dropouts.","With multiple mediators, edge-consistent path-specific effects are identified by Eq. (7), so finer decompositions than a single indirect path are available under the same style of assumptions.","In the liver cohort, private insurance affects listing both through psychosocial review (NIE-RR 1.10) and through other pathways (NDE-RR 1.26), and both intervals exclude the null even after selection adjustment.","A practical corollary is that interventions on the psychosocial-review pathway alone would move only the mediated component; reducing listing disparities also requires acting on the direct pathway."],"supporting_citations":[{"why":"Supplies the Generalized Adjustment Criterion and the selection-bias adjustment formula with external baseline data that the selected mediation formula extends.","marker":"[Correa et al., 2018]"},{"why":"Defines the mediation formula and natural direct/indirect effects that Eq. (6) generalizes to selection settings.","marker":"[Pearl, 2001]"},{"why":"Provides the extended-graph construction used to express nested counterfactuals as do-interventions.","marker":"[Malinsky et al., 2019]"},{"why":"Establishes edge-consistent path-specific effects and recanting witnesses, which the mediator-outcome backdoor condition is designed to avoid.","marker":"[Avin et al., 2005]"},{"why":"Supplies the edge g-formula that the multi-mediator path-specific result fuses with selection adjustment.","marker":"[Shpitser and Tchetgen Tchetgen, 2016]"},{"why":"Provides the clinical cohort and evidence linking social determinants to evaluation completion and listing.","marker":"[Strauss et al., 2022]"},{"why":"Offers the counterfactual imputation estimator that the case study adapts with selection weights.","marker":"[Vansteelandt et al., 2012]"},{"why":"Software implementation modified to incorporate selection weights for the applied estimates.","marker":"[Shi et al., 2021]"}],"fun_headline_variants":["Reweighted mediation reveals direct and indirect effects on liver listing","Selection correction uncovers two pathways to liver-listing disparity","New formula weights out dropout: insurance and psychosocial review drive listing gap","Addressing sample bias: socioeconomic effects on liver transplant listing persist","Liver listing disparity: direct and indirect mediation after selection adjustment"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"There are no unmeasured post-exposure variables, such as transportation access or health literacy, that affect both the psychosocial review and the listing decision; if such variables exist, the estimated direct and indirect effects are biased.","fun_headline_variants_meta":{"raw":{"variants":["Reweighted mediation reveals direct and indirect effects on liver listing","Selection correction uncovers two pathways to liver-listing disparity","New formula weights out dropout: insurance and psychosocial review drive listing gap","Addressing sample bias: socioeconomic effects on liver transplant listing persist","Liver listing disparity: direct and indirect mediation after selection adjustment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001135,"raw_usage":{"total_tokens":4563,"prompt_tokens":771,"completion_tokens":3792,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":3707}},"tokens_in":515,"tokens_out":3792,"duration_ms":29216,"temperature":1.0,"reasoning_tokens":3707,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:01:56.202321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure transportation access and health literacy in the cohort of 497 referred patients and re-estimate the selection-adjusted NIE and NDE under the expanded DAG in Figure 6; a material shift in the estimates would falsify the no-recanting-witness condition. Alternatively, generate data with a hidden post-exposure common cause of the mediator and outcome and check whether Eq. (6) recovers the true nested counterfactual; if it does not, the sufficiency conditions are incomplete.","supporting_citations":[],"review_version":1}