{"id":"fc9f87a2-55e2-494f-9a55-ffe237cbc1dc","arxiv_id":"2505.03526","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In two-period difference-in-differences, under linear faithfulness, parallel trends is incompatible with pre-treatment outcomes causing treatment or with the two outcomes having distinct minimally sufficient confounding sets.","lead":"This statistics paper gives practical rules for deciding when difference-in-differences can be trusted, based on a causal diagram of the problem. It shows that certain diagram features, such as the pre-treatment outcome influencing treatment choice, contradict the key parallel-trends assumption unless coincidental cancellations occur.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1 is false: parallel trends does not imply mean independence from parents of A, so Condition 1's rejection criterion is invalid even when linear faithfulness holds.","rationale":"The reader's concern was that linear faithfulness is fragile; the actual load-bearing problem is more severe. Even granting Assumption 1 in full, the paper's Condition 1 is contradicted by a simple NPSEM. The counterexample is not a generic-position argument: coefficients are chosen so the graph has Y0->A and U confounding, parallel trends holds exactly because D and L are jointly normal with zero covariance, and all d-connected pairs retain nonzero conditional covariance. This isolates Lemma 1 as the false premise: parallel trends only constrains the first moment of D given A, not given pa(A). Since Lemma 2 and Condition 2 use the same lemma, the main rejection machinery is unsupported. The abstract's claim (i) is therefore not merely overstated; it is incorrect in the very setting the paper defines. Because a central result is demonstrably false, acceptance in any form is inappropriate until the conditions are substantially weakened or Lemma 1 is replaced by a correct sufficient condition. Verdict changes from CONDITIONAL to REJECT.","tokens_in":17826,"tokens_out":38008,"duration_ms":376191,"concrete_test":"Run the counterexample analytically or by simulation (n=1e5): estimate E[Y1^0-Y0^0 | A=1] - E[Y1^0-Y0^0 | A=0]; it is zero, despite Y0->A and unmeasured confounding. Bin on Y0 and estimate Cov(A,Y1^0|Y0); it is nonzero, confirming linear faithfulness. Independently, re-derive Lemma 1 from Definition 1 and identify the step where E(D|A)=E(D) is replaced by E(D|pa(A))=E(D); no such step is valid.","verdict_should_be":"REJECT","load_bearing_attack":"Condition 1 (Section 4.2) inherits its force from Lemma 1: parallel trends, E(Y1^0-Y0^0|A)=E(Y1^0-Y0^0), is claimed to imply E(Y1^0-Y0^0|pa(A))=E(Y1^0-Y0^0). This implication is false; mean independence from A does not propagate to A's parents. Explicit NPSEM satisfying the paper's setting: U~N(0,1), εY0~N(0,2), eA,e1~N(0,1) independent; Y0=U+εY0; A=1{2U+εY0+eA>0}; Y1^0=2U+e1. Let D=Y1^0-Y0^0=U-εY0+e1 and L=2U+εY0+eA. Then Cov(D,L)=2Var(U)-Var(εY0)=0 and (D,L) are jointly normal, so D⊥⊥A; parallel trends holds. The graph has Y0->A and the open confounded path A<-U->Y1^0, precisely the configuration Condition 1 declares incompatible with parallel trends under linear faithfulness. Yet linear faithfulness is satisfied: d-connected vertex pairs have nonzero conditional covariance, e.g. Cov(A,Y1^0|Y0) is proportional to Cov(L,Y1^0|Y0)=4/3>0. Hence Condition 1 is false as stated; the error is equation (8), which treats E(D|pa(A)) as constant.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper aims to provide graphical guidance for deciding whether the parallel trends assumption underlying difference-in-differences is plausible. Under a linear faithfulness assumption and a nonparametric structural equation model with unmeasured common causes, the authors claim that parallel trends implies three conditions: (1) no direct effect of the pre-treatment outcome Y0 on treatment A in the presence of unmeasured confounding; (2) the pre- and post-treatment outcomes have common minimally sufficient adjustment sets; and (3) no direct effect of Y0 on the untreated potential outcome Y1^0. They further claim that, absent these features, parallel trends is equivalent to an 'additive homogeneous confounding' condition with respect to a common sufficient set. The paper applies this framework to Medicaid expansion and insurance coverage. The writing is clear and the authors are transparent about the heuristic status of Condition 3, but the central lemmas used to derive Conditions 1 and 2 are false as stated.","tokens_in":18074,"tokens_out":12070,"duration_ms":115501,"significance":"If the central results were correct, the paper would provide a useful bridge between causal diagrams and the scale-dependent parallel trends assumption, allowing applied researchers to use substantive graphical knowledge to assess a key DID assumption. The paper also contributes a useful explicit statement of linear faithfulness and an application to a realistic policy question, and it includes reproducible R code for illustrative minimal sufficient set calculations. However, the main rejection criteria rest on incorrect implications; the counterexample below satisfies the paper's own assumptions and shows that Condition 1 and the additive homogeneous confounding necessity are invalid. The contribution as stated therefore does not stand, although the broader goal of connecting DID assumptions to graphical structure remains valuable.","major_comments":[{"comment":"Lemma 1 is false as stated. Let U, εY0, eA, e1 be independent mean-zero normal variables with variances 1, 2, 1, 1, and define Y0 = U + εY0, A = 1{2U + εY0 + eA > 0}, Y1^0 = 2U + e1. Put D = Y1^0 - Y0^0 and L = 2U + εY0 + eA. Then (D, L) is jointly normal with Cov(D, L) = 2Var(U) - Var(εY0) = 0, so D⊥⊥L; since A is a function of L and the independent eA, D⊥⊥A, so parallel trends holds. Yet pa(A) includes U and Y0, and E(D | U, Y0) = 2U - Y0 is not constant, so the conclusion E(D | pa(A)) = E(D) of Lemma 1 fails. This model has the Y0→A arrow and the open confounded path A←U→Y1^0 that Condition 1 declares incompatible with parallel trends, and it satisfies linear faithfulness: the d-connected vertices in this graph have nonzero conditional covariances, for example Cov(A, Y1^0 | Y0) > 0. Equation (8) is exactly the invalid step: parallel trends gives mean independence of D from A, not from the parents of A. Consequently Condition 1 is not a valid necessary condition.","section":"§3.7, Lemma 1; §4.2, Eq. (8)"},{"comment":"Lemma 2 and the claimed necessity of additive homogeneous confounding are also false. In the same model, M = {U, Y0} is a common sufficient set for Y0 and Y1^0: conditional on M, A depends only on eA, which is independent of Y0 and of Y1^0, so E(Yt^0 | A, M) = E(Yt^0 | M) for t = 0, 1. Parallel trends holds, but E(D | M) = E(D | U, Y0) = 2U - Y0, which is not constant, contradicting Eq. (6). The proof in Appendix A is invalid: the equality E{π(M)E(D|M)} = 0 is obtained only for the actual propensity score π(M) = E[A | M], but the proof then replaces π(M) by indicator functions of {E(D|M) ≥ 0} and {E(D|M) ≤ 0}, as if parallel trends held for every propensity score. That inference is not licensed. Since Lemma 3 and Condition 2 are derived from Lemma 2, the common-minimally-sufficient-set criterion is unsupported.","section":"§4.1, Lemma 2; Appendix A; Remark 1"},{"comment":"The treatment of Condition 3 is explicitly conditional and does not support the summary claim in §4.5 that 'no arrow from Y0 to Y1' is one of the conditions implied by parallel trends under the paper's assumptions. The proof in §4.4 only shows that, in an additively separable model, parallel trends cannot hold in both G0 and G1 without violating linear faithfulness; it does not establish that parallel trends is impossible in G1, nor does it quantify the 'strongly questioned' claim for general nonparametric models. The extension to the nonparametric setting is a conjecture. The authors are candid about this limitation, but the abstract and Section 4.5 present Condition 3 as part of the operative checklist, which exceeds what is proven.","section":"§4.4, Condition 3; §4.5 summary"}],"minor_comments":[{"comment":"The citation to Ghanem et al. is inconsistent: Section 3.7 cites Lemma F.3 as Ghanem et al. (2024), while the reference list and Section 2 identify the paper as Ghanem et al. (2022).","section":"References and Section 3.7"},{"comment":"The status of the edges among U1, U2, and U3 is described only informally; the text says U1, U2, and U3 can impact U4 but leaves their mutual relationships otherwise unspecified. A clearer statement of which edges are definitely present versus unknown would help the reader interpret the partially directed SWIG.","section":"§3.5, Figure 1"},{"comment":"The reliance on dagitty output for the minimal sufficient sets of Figure 4, with the comment that showing the result analytically is complex, leaves the reader without a verifiable argument for a claim that is used in the main text. A proof or a more detailed derivation would strengthen the paper.","section":"Appendix D"},{"comment":"Several minor language issues remain: 'canonical' is misspelled in the caption of Figure 1, Appendix B contains the phrase 'a colliders', and Assumption 2 is phrased in a way that is close to tautological ('are either not all positive or not all negative').","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The core of the manuscript—Condition 1 and the necessity of additive homogeneous confounding—is contradicted by a simple NPSEM that satisfies the paper's own linear faithfulness assumption. The proof of Lemma 2 confuses the actual propensity score with an arbitrary one. These are not presentation issues; the graphical rejection program would need to be reformulated. I found no evidence of any problematic citation or novelty-disclosure practice beyond the internal reference inconsistency noted in the minor comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, read this one if you want a clean example of why faithfulness arguments are delicate. The paper aims to give graphical conditions for when parallel trends cannot hold in DID. Good question, readable writing, and the connection to additive homogeneous confounding (Remark 1) is a genuinely useful reformulation. But the main machinery is broken. The stress-test is right: Lemma 1, restated from Ghanem et al., is false as used here. Parallel trends only gives zero covariance between the change in untreated outcomes and treatment; it does not give mean independence from the parents of treatment. The counterexample with U, Y0, A, and Y1^0 works: parallel trends holds, the graph has Y0->A and an open A<-U->Y1^0 path, and linear faithfulness is satisfied. So Condition 1 is simply wrong. Lemma 2 and Condition 2 inherit the problem, since their proofs lean on the same flawed step. The abstract overstates the result, and the body's qualification about needing unmeasured confounding does not rescue it. What is salvageable: the necessary and sufficient condition in terms of additive homogeneous confounding on a common sufficient set is correct and worth keeping, and the discussion of why pre-trends are neither necessary nor sufficient is sound. But the novel graphical rejection criteria are unsupported. This is not a minor fix; it is the load-bearing wall. If I were refereeing, I would reject in current form and ask the authors to rederive what actually follows from the covariance condition. It may be that weaker or different conditions survive, but the current paper as written should not be accepted. Would I cite it? No, not with the central lemma false. Worth a serious referee? Yes, because the topic matters and the authors are clearly capable; this deserves a careful review, not a desk rejection. Bring it to a reading group as a cautionary tale about conflating uncorrelatedness with mean independence.","headline":"The paper's main graphical rejection criteria rest on a false lemma; the supplied counterexample is valid, so Condition 1 is invalid and the central contribution collapses.","tokens_in":18637,"tokens_out":3863,"would_cite":false,"duration_ms":36096,"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":"A causal diagram can tell you when parallel trends is untenable.","keywords":["difference-in-differences","parallel trends","causal diagrams","DAG","unmeasured confounding","linear faithfulness","minimally sufficient sets","additive homogeneous confounding"],"falsifier":"Simulate a two-period linear structural equation model that exactly follows the graph with disjoint minimally sufficient sets (a U3 affecting A and Y0, a U4 affecting A and Y1, no Y0-to-A or Y0-to-Y1 edge), and choose coefficients so that E($Y1^{0}$-$Y0^{0}$|A)=E($Y1^{0}$-$Y0^{0}$) exactly. Then check every d-connected pair in the graph for nonzero covariance; a coefficient vector that satisfies both parallel trends and nonzero covariances would refute the claim that Condition 2 plus linear faithfulness rejects parallel trends.","tokens_in":17573,"feed_emoji":"📉","tokens_out":7768,"duration_ms":68813,"temperature":0.7,"pith_summary":"Difference-in-differences (DID) delivers causal estimates only when parallel trends holds, and researchers have had little guidance for judging that assumption ahead of time. This paper shows that a causal diagram can provide such guidance once one adds linear faithfulness: every pair of variables the graph connects has nonzero covariance. Under that assumption, parallel trends is incompatible with pre-treatment outcomes affecting treatment (together with unmeasured confounding of treatment and the post-treatment outcome), and with pre- and post-treatment outcomes having distinct minimally sufficient adjustment sets. The paper also argues that pre-treatment outcomes affecting post-treatment outcomes is a serious warning sign, though this is proven only in restricted semiparametric models. If all three warning signs are absent, the remaining content of parallel trends is exactly additive homogeneous confounding: a common confounder set must have constant additive association with the untreated potential outcome across the two periods.","feed_headline":"Three graph patterns signal parallel trends will fail","feed_subtitle":"Pre-treatment outcomes driving treatment, or time-specific confounders, mark the DID assumption as untenable.","key_machinery":"The machine that does the work is linear faithfulness (Assumption 1): whenever the graph does not d-separate (block) two variables, their conditional covariance is nonzero. This lets the authors convert 'the graph says this association should be present' into 'parallel trends would force this association to be absent.' The supporting machinery consists of directed single-world intervention graphs (SWIGs) for reading counterfactual independencies, minimally sufficient adjustment sets for the effect of treatment on each outcome, and Lemma 2, which states that parallel trends plus a common sufficient set M implies additive homogeneous confounding, E($Y1^{0}$-$Y0^{0}$|M)=E($Y1^{0}$-$Y0^{0}$).","core_discovery":"The central claim is that parallel trends, although scale-dependent, can be assessed with a scale-independent graph if linear faithfulness holds. In that setting, adopting parallel trends forces conditional mean equalities that the graph contradicts whenever (i) the pre-treatment outcome Y0 directly affects treatment A while unmeasured confounding connects A to the post-treatment outcome $Y1^{0}$, or (ii) the minimally sufficient adjustment sets for Y0 and $Y1^{0}$ differ, as when separate unmeasured confounders affect treatment with only one outcome. The paper further argues, without a full proof in the general nonparametric model, that (iii) an arrow from Y0 to $Y1^{0}$ should be regarded as suspect, because in partially linear and additively separable models parallel trends can hold with and without that arrow only through exact cancellation. When none of these features appears, the maximal graph compatible with parallel trends consists of a common confounder for both outcomes and a separate source of correlation between Y0 and Y1, and the assumption reduces to additive homogeneous confounding.","pith_inferences":["The checklist suggests a sensitivity analysis: for a graph that violates Conditions 1 or 2, one could quantify how large a linear-faithfulness violation would have to be for parallel trends to survive, turning the rejection into a graded warning.","The paper's logic implies that empirically observed parallel pre-trends cannot rescue a graph with disjoint sufficient sets; if the graph is right, the pre-trends must be a coincidence, which is a testable prediction when many similar policy evaluations are analyzed together.","Condition 3's conjecture could be probed by constructing a fully nonparametric model in which h(Y0) enters nonlinearly and checking whether parallel trends forces h to be uncorrelated with the common confounder set, a strict condition the paper demonstrates only in separable models."],"forward_implications":["Researchers can reject parallel trends before estimation when their causal diagram shows pre-treatment outcomes influencing treatment while unmeasured confounding between treatment and the post-treatment outcome remains.","A graph whose minimally sufficient adjustment sets differ between the pre- and post-treatment outcomes is incompatible with parallel trends under linear faithfulness; such graphs should steer analysts toward other designs.","An arrow from the pre-treatment to the post-treatment outcome should be treated as a warning flag, since in reasonable semiparametric models it makes parallel trends depend on exact coincidence.","Even a graph with none of the three features does not verify parallel trends; it only narrows the required justification to additive homogeneous confounding.","These results extend earlier warnings, which were confined to linear structural equation models, to nonparametric structural models and graphs."],"supporting_citations":[{"why":"Supplies Lemma F.3, the core result that parallel trends over treatment implies parallel trends over the parents of treatment; Lemma 2 and Conditions 1-2 build directly on it.","marker":"Ghanem et al. (2022)"},{"why":"Prior linear structural-equation analysis showing similar coefficient-cancellation conditions for parallel trends; the paper generalizes these to nonparametric models.","marker":"Kim and Steiner (2021)"},{"why":"Establishes that parallel trends is scale-dependent and equivalent to equality of confounding across periods, motivating the whole graphical challenge.","marker":"Lechner et al. (2011)"},{"why":"Provides SWIGs, which the paper uses to read counterfactual independencies and to define sufficient sets in the graphs.","marker":"Richardson and Robins (2013)"},{"why":"Source of the linear faithfulness assumption used to turn graphical connection into nonzero covariance.","marker":"Spirtes et al. (2001)"},{"why":"Properties of sufficient adjustment sets used in the proof that a common sufficient set always exists and in Lemma 3.","marker":"Shpitser et al. (2012)"},{"why":"Earlier nonparametric structural example that already isolates the Y0-to-A condition for parallel trends failure.","marker":"Weber et al. (2015)"},{"why":"Parametric DID analysis of time-varying confounding that motivates the interpretation of Condition 2.","marker":"Zeldow and Hatfield (2021)"}],"fun_headline_variants":["Causal diagrams reveal when parallel trends fail","Three graph signs to reject parallel trends","How to test DID's key assumption with graphs","Graph conditions for parallel trends in DID","Spotting invalid parallel trends in diff-in-diff"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is linear faithfulness: any two variables connected by the graph must have nonzero covariance, so exact cancellation of associations is ruled out; if such cancellations occur naturally, a graph can contain all three warning features while parallel trends still holds.","fun_headline_variants_meta":{"raw":{"variants":["Causal diagrams reveal when parallel trends fail","Three graph signs to reject parallel trends","How to test DID's key assumption with graphs","Graph conditions for parallel trends in DID","Spotting invalid parallel trends in diff-in-diff"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1406,"prompt_tokens":1063,"completion_tokens":343,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":276}},"tokens_in":679,"tokens_out":343,"duration_ms":3588,"temperature":1.0,"reasoning_tokens":276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:48:54.573860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a two-period linear structural equation model that exactly follows the graph with disjoint minimally sufficient sets (a U3 affecting A and Y0, a U4 affecting A and Y1, no Y0-to-A or Y0-to-Y1 edge), and choose coefficients so that E($Y1^{0}$-$Y0^{0}$|A)=E($Y1^{0}$-$Y0^{0}$) exactly. Then check every d-connected pair in the graph for nonzero covariance; a coefficient vector that satisfies both parallel trends and nonzero covariances would refute the claim that Condition 2 plus linear faithfulness rejects parallel trends.","supporting_citations":[],"review_version":1}