{"id":"78400173-d616-488a-b695-849b605aff3e","arxiv_id":"2606.22165","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"New cumulative natural direct and indirect effects are defined to decompose local causal effects into direct and indirect components for continuous and ordinal treatments while preserving skew-symmetry and additivity.","lead":"The paper introduces cumulative natural direct and indirect effects (CNDE and CNIE) for causal mediation analysis with continuous treatments by decomposing local causal effects. These address interpretability issues in standard NDE and NIE by ensuring skew-symmetry and additivity.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Differentiability of Y_x w.r.t. continuous x is presupposed but not implied by standard assumptions","rationale":"The reader correctly flags that identification rests on standard assumptions, but the central construction for continuous treatments introduces an additional, unstated regularity requirement (existence of ∂_x Y_x) that is logically prior to those assumptions. This is the single load-bearing gap for the continuous-treatment claim; the ordinal extension sidesteps it. The verdict should therefore be moved from UNVERDICTED to CONDITIONAL pending explicit statement and verification of the differentiability condition.","tokens_in":1712,"tokens_out":374,"duration_ms":29359,"concrete_test":"In the linear-interaction example of §4, replace the outcome model with a piecewise-linear function that has a kink at a fixed x0 while preserving the same conditional expectations and the four standard assumptions; recompute the integrated CNDE and CNIE and verify whether direct + indirect still equals the total effect and whether skew-symmetry continues to hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"CNDE/CNIE are constructed by decomposing the local total effect E[∂_x Y_x] into local direct and indirect components, then integrating to obtain cumulative effects. This step requires that the potential-outcome map x ↦ Y_x be differentiable (at least almost everywhere) so that the partial derivative exists and the local decomposition is well-defined. The paper states that identification holds under the usual quartet of consistency, positivity, and no-unmeasured-confounding assumptions, yet none of those four conditions entails differentiability or even continuity of Y_x. Consequently the claimed skew-symmetry and additivity properties are only guaranteed inside the subclass of data-generating processes that already satisfy an extra regularity condition not listed among the “standard causal assumptions.”","agreement_with_reader":"disagree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces cumulative natural direct and indirect effects (CNDE and CNIE) for continuous treatments in causal mediation analysis. These are constructed by first decomposing the local total effect E[∂_x Y_x] into local direct and indirect components and then integrating to obtain cumulative versions. The authors claim that CNDE and CNIE provide a decomposition of the total effect that satisfies skew-symmetry and additivity (unlike standard NDE/NIE), establish decomposition and identification results under standard causal assumptions, extend the framework to ordinal treatments via discrete analogues, and illustrate the measures in linear mediation models with interaction and on a real dataset.","tokens_in":1876,"tokens_out":668,"duration_ms":16971,"significance":"If the decomposition, identification, and algebraic properties hold, the proposed measures would supply a coherent alternative for mediation analysis with continuous or ordinal treatments that avoids certain paradoxical interpretations arising from failures of skew-symmetry and additivity. The explicit construction from local effects, the extension to ordered discrete treatments, and the comparison with existing linear-model results constitute a targeted methodological contribution in causal inference.","major_comments":[{"comment":"§3 (definition of CNDE/CNIE): the construction decomposes and integrates E[∂_x Y_x], which presupposes that the potential-outcome map x ↦ Y_x is differentiable (at least almost everywhere). The identification results later invoke only the standard quartet of consistency, positivity, and no unmeasured confounding for the three relations; none of these conditions entails differentiability or even continuity of Y_x. Consequently the claimed skew-symmetry and additivity hold only conditionally on an extra regularity assumption not listed among the “standard causal assumptions.”","section":"§3"},{"comment":"§4 (identification and decomposition theorems): the paper states that decomposition and identification results are established, yet the provided text does not display the explicit integral expressions for CNDE and CNIE nor the step-by-step verification that integration preserves skew-symmetry and additivity for arbitrary (sufficiently regular) response surfaces. Without these derivations the central claim that the cumulative effects “yield a decomposition … that preserves both skew-symmetry and additivity” cannot be fully assessed.","section":"§4"},{"comment":"§5 (linear-model illustration): the comparison with NDE/NIE is performed only inside canonical linear models with interaction. It remains unclear whether the reported numerical differences between CNDE/CNIE and NDE/NIE persist, or reverse, under non-linear data-generating processes that still satisfy the differentiability requirement.","section":"§5"}],"minor_comments":[{"comment":"Notation for the local total effect E[∂_x Y_x] is introduced without an explicit statement of the measure with respect to which the derivative is taken or the domain on which it is assumed to exist.","section":"§3"},{"comment":"The abstract refers to “canonical linear mediation models with interaction” but does not specify the precise interaction term; the main text should state the model equation explicitly before presenting numerical results.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive comments. We address each major comment below and indicate the corresponding revisions.","responses":[{"response":"We agree that the definition of the local effects and their integration requires differentiability of the potential-outcome map (at least almost everywhere). The identification results rely on the standard assumptions, but the construction of CNDE and CNIE additionally presupposes this regularity condition. We will explicitly add differentiability (a.e.) to the list of assumptions in the revised manuscript and clarify that skew-symmetry and additivity hold under the standard causal assumptions together with this regularity condition.","revision_made":"yes","referee_comment":"[§3] §3 (definition of CNDE/CNIE): the construction decomposes and integrates E[∂_x Y_x], which presupposes that the potential-outcome map x ↦ Y_x is differentiable (at least almost everywhere). The identification results later invoke only the standard quartet of consistency, positivity, and no unmeasured confounding for the three relations; none of these conditions entails differentiability or even continuity of Y_x. Consequently the claimed skew-symmetry and additivity hold only conditionally on an extra regularity assumption not listed among the “standard causal assumptions.”"},{"response":"We will insert the explicit integral expressions for CNDE and CNIE in §4. We will also add a step-by-step verification showing that, for sufficiently regular response surfaces, the integration operation preserves skew-symmetry and additivity. These additions will make the central claim fully verifiable from the text.","revision_made":"yes","referee_comment":"[§4] §4 (identification and decomposition theorems): the paper states that decomposition and identification results are established, yet the provided text does not display the explicit integral expressions for CNDE and CNIE nor the step-by-step verification that integration preserves skew-symmetry and additivity for arbitrary (sufficiently regular) response surfaces. Without these derivations the central claim that the cumulative effects “yield a decomposition … that preserves both skew-symmetry and additivity” cannot be fully assessed."},{"response":"The linear-model section is intended as an illustration in a setting where NDE/NIE are commonly studied and where their failure of skew-symmetry and additivity is transparent. The algebraic properties of CNDE/CNIE are established generally under the stated assumptions and do not rely on linearity. We will add a clarifying sentence noting that the magnitude and direction of numerical differences are model-dependent and that the linear case is chosen to highlight the interpretative advantage of the new measures. A full exploration of non-linear DGPs is left for future work, as it would require specifying particular non-linear surfaces while preserving differentiability.","revision_made":"partial","referee_comment":"[§5] §5 (linear-model illustration): the comparison with NDE/NIE is performed only inside canonical linear models with interaction. It remains unclear whether the reported numerical differences between CNDE/CNIE and NDE/NIE persist, or reverse, under non-linear data-generating processes that still satisfy the differentiability requirement."}],"tokens_in":1522,"tokens_out":660,"duration_ms":19010,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"Kawakami and Tian introduce cumulative natural direct and indirect effects for continuous treatments. They start from the local total effect given by the expected derivative of the potential outcome, split that into local direct and indirect pieces, and integrate to get the cumulative versions. The result is a decomposition of the total effect that keeps both skew-symmetry and additivity, which the usual NDE and NIE do not.\n\nThe construction is genuinely new. It is not a relabeling of existing quantities but comes from working directly with the derivative and then accumulating. The ordinal extension follows the same logic and preserves the properties over ordered levels. In the linear mediation models they examine, the new measures behave as claimed and avoid the paradoxical interpretations that can arise with NDE and NIE.\n\nThe main limitation is the regularity condition. Defining the local effects requires the map from treatment to potential outcome to be differentiable, at least almost everywhere. The paper states that identification holds under the standard quartet of consistency, positivity, and no unmeasured confounding, yet none of those conditions implies differentiability or even continuity. The stress-test note is correct on this point. The results therefore apply cleanly inside the subclass of data-generating processes that already satisfy the extra regularity, such as the linear cases shown, but the paper should flag that condition explicitly rather than treating it as covered by the usual assumptions.\n\nThe work is aimed at researchers who do causal mediation with continuous or ordinal treatments and who care about algebraic properties of the effect measures. A reader already working in that niche would find the definitions and the linear illustrations useful. It deserves peer review because the problem it targets is real and the proposed fix is distinct from prior definitions, even though the assumption gap needs tightening.","headline":"The paper defines CNDE and CNIE by decomposing local effects E[∂x Y_x] to fix skew-symmetry and additivity for continuous treatments, but the differentiability step sits outside the listed assumptions.","tokens_in":2360,"tokens_out":433,"would_cite":false,"duration_ms":26052,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Cumulative natural direct and indirect effects decompose total effects while preserving skew-symmetry and additivity for continuous treatments.","keywords":["causal mediation analysis","natural direct effects","natural indirect effects","continuous treatments","skew-symmetry","additivity","effect decomposition"],"falsifier":"A numerical example or dataset in which the sum of the estimated CNDE and CNIE differs from the total effect, or in which reversing the treatment direction fails to produce the negative of the original decomposition, would falsify the claimed preservation of additivity and skew-symmetry.","tokens_in":2593,"feed_emoji":"📊","tokens_out":650,"duration_ms":18273,"temperature":0.7,"pith_summary":"Standard natural direct and indirect effects can produce paradoxical interpretations in mediation analysis because they fail to satisfy skew-symmetry and additivity. The paper defines cumulative natural direct and indirect effects by decomposing the local causal effect given by the expected derivative of the outcome with respect to a continuous treatment. These new measures integrate the local direct and indirect components to recover a decomposition of the total effect that obeys both required properties. Discrete analogues are defined for ordinal treatments that inherit the same structural guarantees. The construction relies on the usual causal identification assumptions and is illustrated on linear models with interaction.","feed_headline":"Cumulative effects restore additivity in mediation decompositions","feed_subtitle":"New CNDE and CNIE measures integrate local direct and indirect effects so the total effect splits without symmetry violations.","key_machinery":"Cumulative natural direct and indirect effects, obtained by integrating local direct and indirect effects derived from the derivative of the outcome with respect to the treatment.","core_discovery":"We introduce the cumulative natural direct effect (CNDE) and cumulative natural indirect effect (CNIE) for continuous treatments by decomposing the local causal effect E[∂_x Y_x] into local direct and indirect effects. These measures satisfy a decomposition of the total effect that is skew-symmetric and additive. We extend the framework to ordinal treatments with discrete analogues that preserve these properties. Identification results are established under standard causal assumptions.","pith_inferences":["The construction could be applied to other continuous-treatment effect decompositions where symmetry properties are required.","In settings with treatment-mediator interactions, the cumulative measures may produce more stable policy interpretations than conventional natural effects.","Extensions to time-varying treatments might follow by replacing the derivative with an appropriate infinitesimal operator."],"forward_implications":["The total effect decomposes additively into CNDE and CNIE without sign or ordering paradoxes.","The decomposition remains valid when treatment is reversed, satisfying skew-symmetry.","Discrete cumulative effects defined over ordered treatment levels inherit the same decomposition properties.","Standard identification formulas apply directly to the new measures under the usual no-confounding conditions."],"fun_headline_variants":["CNDE and CNIE yield additive total effect decomposition","Cumulative natural effects preserve skew symmetry in mediation","Local effects decompose with additivity via CNDE CNIE","CNDE CNIE extend additive mediation to ordinal treatments"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Identification of the cumulative effects from observed data requires the standard assumptions of consistency, positivity, and no unmeasured confounding for the treatment-mediator, treatment-outcome, and mediator-outcome relations.","fun_headline_variants_meta":{"raw":{"variants":["CNDE and CNIE yield additive total effect decomposition","Cumulative natural effects preserve skew symmetry in mediation","Local effects decompose with additivity via CNDE CNIE","CNDE CNIE extend additive mediation to ordinal treatments"]},"model":"grok-4.3","cost_usd":0.005614,"raw_usage":{"total_tokens":2604,"prompt_tokens":663,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":56140500,"prompt_tokens_details":{"text_tokens":663,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1881,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":663,"tokens_out":60,"duration_ms":13253,"temperature":1.0,"reasoning_tokens":1881,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T11:25:17.520303+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A numerical example or dataset in which the sum of the estimated CNDE and CNIE differs from the total effect, or in which reversing the treatment direction fails to produce the negative of the original decomposition, would falsify the claimed preservation of additivity and skew-symmetry.","supporting_citations":[],"review_version":1}