{"id":"7fb2ca71-5f50-471b-8acd-56e2c7c795d0","arxiv_id":"2512.21176","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Under unknown interference, the two-group two-period DiD estimand equals the total effect on the treated minus the spillover effect on the control, and identifies neither separately without additional assumptions.","lead":"This technical note shows that when units interfere with each other, the standard difference-in-differences (DiD) estimator does not measure a single treatment effect. Instead it measures the difference between the total effect on treated units and the spillover effect on control units, so separate effects need extra assumptions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 7 is false when tau1 = tau0 != 0; Assumption 11 needs strict inequality.","rationale":"The reader's strongest claim, Proposition 3, is mathematically correct and the paper's central decomposition of the DiD estimand into tau1 - tau0 under Assumption 7 is sound. The reader's verdict of CONDITIONAL is justified, but the specific reason is the false boundary case in Proposition 7, not the (unavoidable) untestability of Assumption 7. The reader's weakest_assumption field points to Assumption 7, which is a standard parallel-trends assumption, but the actual load-bearing flaw for the paper's acceptance is Proposition 7's division by tau1 without excluding tau1 = tau0. This is a concrete, fixable error that does not undermine the main decomposition but does invalidate an advertised practical result. Thus I partially agree with the reader: the paper needs a conditional revision, but the primary concern is the boundary failure of Proposition 7, not the untestability of Assumption 7.","tokens_in":12069,"tokens_out":13605,"duration_ms":121367,"concrete_test":"Evaluate Proposition 7 with tau1 = tau0 = 2 (e.g., construct potential outcomes satisfying Assumptions 1,2,7 with E[Y1(1,1,0)-Y1(0,0,0)|G=1] = 2 and E[Y1(0,0,1)-Y1(0,0,0)|G=0] = 2). Then DiD = 0, sgn(DiD) = 0, and sgn(tau1) = 1, falsifying the proposition. Re-prove the result under Assumption 11' (|tau1| > |tau0|) to confirm the correction.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central identification result (Proposition 3) is correct: under Assumptions 1, 2, and 7, DiD = tau1 - tau0. However, the paper's additional claim (Proposition 7) that under Assumption 11 (|tau1| >= |tau0|) sgn(DiD) = sgn(tau1) fails at the boundary tau1 = tau0 != 0. If tau1 = tau0 = c > 0, then DiD = 0, so sgn(DiD) = 0, but sgn(tau1) = 1. The proof divides by tau1 and asserts 1 - tau0/tau1 > 0, which requires strict inequality. This is an internal inconsistency in a result advertised as letting researchers sign the total effect on the treated. The fix is to strengthen Assumption 11 to |tau1| > |tau0| or handle equality separately. No issue arises for Proposition 3 itself, but the paper's broader claim about sign identification is unsound as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This technical note examines what the two-group, two-period difference-in-differences (DiD) estimand identifies when SUTVA's no-interference assumption is violated. The authors define two causal estimands that are well-defined under arbitrary interference: the total average treatment effect on the treated (TATT, τ1) and the average spillover effect on the control (ASC, τ0). Under a modified parallel-trends assumption (Assumption 7), they show that DiD identifies the difference τ1 − τ0, not either effect separately. They then explore identifying assumptions: constant no-treatment trends (Proposition 4), bounded trends (Proposition 5), sign restrictions on τ0 (Proposition 6), and a magnitude dominance condition (Proposition 7). The results are illustrated by revisiting Card and Krueger (1994).","tokens_in":12273,"tokens_out":11000,"duration_ms":97585,"significance":"The paper provides a clean and useful decomposition of the DiD estimand in a setting where interference is unrestricted. Proposition 3 is a simple algebraic identity, correctly proved, and it makes transparent that under interference a non-zero DiD is consistent with many combinations of treated-group and control-group effects. The partial-identification results in Propositions 4–6 are also correct and offer practitioners explicit assumptions under which policy-relevant parameters can be bounded or signed. The application to Card and Krueger nicely demonstrates how the original conclusion can change once spillovers are allowed. The main weakness is Proposition 7, which is false at a boundary and whose proof is therefore flawed; this needs correction before the paper can be accepted. Overall, the note is a worthwhile contribution to the DiD-under-interference literature, and the authors are honest about the untestable nature of Assumption 7 and the limits of their results.","major_comments":[{"comment":"Proposition 7 states that under Assumption 11, |τ1| ≥ |τ0|, we have sgn(τ1) = sgn(DiD). This is false when τ1 = τ0 ≠ 0: the assumption holds with equality, DiD = 0, so sgn(DiD) = 0 while sgn(τ1) ≠ 0. The proof divides by τ1 and asserts that Assumption 11 implies 1 − τ0/τ1 > 0; however, the non-strict inequality only gives 1 − τ0/τ1 ≥ 0, and equality occurs precisely when τ1 = τ0. Please either strengthen Assumption 11 to the strict inequality |τ1| > |τ0|, or explicitly exclude the case τ1 = τ0 (equivalently, DiD = 0) in the statement and proof. Also state the implicit assumption τ1 ≠ 0 in the division step. The application to Card and Krueger uses DiD = 2.75 ≠ 0, so the boundary case does not arise there, but the general theorem as written is incorrect.","section":"Section 3.2, Proposition 7 and Assumption 11"}],"minor_comments":[{"comment":"The sentence 'estimating the DiD estimand allows for testing whether the intervention had a different average effect on the treated and control groups' could be sharpened: DiD tests whether τ1 ≠ τ0, not whether the average effects are 'different' in any broader sense. This is clear from the context, but a precise statement would avoid ambiguity.","section":"Section 2.4, after Proposition 3"},{"comment":"The proof uses sgn(τ1(1 − τ0/τ1)) = sgn(τ1) and implicitly assumes τ1 ≠ 0. Please add a remark that the case τ1 = 0 is trivial (given |τ1| ≥ |τ0| this forces τ0 = 0) or handle it explicitly before the division.","section":"Section 3.2, proof of Proposition 7"},{"comment":"The notation Y_i1(1,1,0) and Y_i1(0,0,1) is used extensively. A brief note that the second and third arguments are vectors of ones and zeros of the appropriate dimensions (same-group and opposite-group units) would help readers unfamiliar with the partition.","section":"Section 2.1, notation"},{"comment":"The row label 'Difference' in Table 2 is slightly ambiguous because -2.89 is the difference between New Jersey and Pennsylvania in the pre-period, while 0.59 and -2.16 are within-state changes. Adding a note would improve readability.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid technical note with a correct main identification result (Proposition 3) and useful partial-identification results. The falsity of Proposition 7 at the boundary is a genuine technical error, but it is easily repaired by strengthening Assumption 11 to a strict inequality or by explicitly excluding the τ1 = τ0 case. I would be willing to consider a revised version; the revision is localized and should not require new substantive work. The paper's framing that SUTVA has received 'little attention' in DiD is somewhat overstated given the emerging literature, but the authors do cite recent work and position their contribution appropriately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this note: the main identification result is correct and worth having on the record, and Proposition 7, as stated, is false on a boundary case. That second point is a real but minor flaw.\n\nThe paper formalizes what the two-group, two-period DiD estimand identifies when SUTVA fails. Under a parallel trends assumption on Y(0,0,0), Proposition 3 shows DiD = tau1 - tau0, the difference between the total effect on the treated and the spillover effect on controls. The proof is straightforward and correct. This is a useful clarification. Practitioners often talk about spillovers informally; having the exact contrast made explicit — and having it shown that DiD alone cannot separate the two effects — is a genuine service. The authors are also careful to note the relation to Sävje et al. (2021) and Xu et al. (2025), so the novelty claim is modest but fair.\n\nThe subsequent propositions are mostly simple corollaries or sensitivity bounds. Propositions 4 and 5 are fine given their assumptions; Proposition 6 is a direct implication of the decomposition. These are not deep results, but they are organized well and applied sensibly in the Card–Krueger re-analysis.\n\nThe soft spot is Proposition 7. Under Assumption 11 (|tau1| >= |tau0|), the paper claims sgn(DiD) = sgn(tau1). That fails when tau1 = tau0 != 0: DiD = 0, so sgn(DiD)=0, while tau1 has a sign. The proof divides by tau1 and asserts 1 - tau0/tau1 > 0, which requires strict inequality. Strengthening Assumption 11 to |tau1| > |tau0| or handling equality separately fixes it. This is the only place I found where the math actually slips. The main decomposition is untouched.\n\nA larger caveat, present in any DiD-with-spillovers analysis, is that Assumption 7 refers to the never-observed potential outcome Y1(0,0,0); it is untestable. The authors do not pretend otherwise. That limits the practical force of the results, but it is not a flaw in the formal argument.\n\nBottom line: this is a competent technical note, not a major advance. The core contrast result is correct and citable, and the boundary bug in Proposition 7 is easy to patch. Send it to referees; the topic is relevant, the write-up is clear, and the reviewer time required is modest.","headline":"A short, correct formalization of what DiD identifies under unknown interference; the core decomposition is solid, but Proposition 7 has a boundary flaw that needs a strict-inequality fix.","tokens_in":12804,"tokens_out":2030,"would_cite":true,"duration_ms":19723,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In the presence of unknown interference, the difference-in-differences estimand identifies the difference between the total effect on the treated and the average spillover effect on the control group — not either effect alone.","keywords":["difference-in-differences","interference","spillover effects","SUTVA","parallel trends","causal identification","partial identification","minimum wage"],"falsifier":"Find a setting with known interference where a spillover-free comparison group is available. Estimate the time trend of the no-treatment outcome Y(0,0,0) in the treated and control groups from that comparison group; if the trends differ, then the DiD estimand equals τ1 − τ0 plus the trend gap, contradicting the Proposition 3 reading. Concretely, in a cross-border minimum wage study, use a non-bordering state as a no-spillover reference and compare the pre-treatment employment trends of the treated and control states.","tokens_in":1410,"feed_emoji":"📊","tokens_out":2043,"duration_ms":62127,"temperature":0.7,"pith_summary":"The paper asks what a standard difference-in-differences estimate captures when one unit's treatment can affect another unit's outcome, so the usual no-interference assumption fails. It shows that, under a modified parallel trends assumption, the DiD estimand equals the total average treatment effect on the treated minus the average spillover effect on the control group. Neither of the two effects is identified separately; the same DiD value is compatible with infinitely many pairs of effects. The paper then catalogues assumptions — bounded no-treatment trends, sign restrictions on the spillover effect, or a magnitude-dominance condition — under which the two components become partially identified or their signs recoverable. The point matters because many real DiD applications, such as cross-border minimum wage studies, plausibly involve spillovers.","feed_headline":"DiD with spillovers estimates a gap, not an effect","feed_subtitle":"When interference is possible, the classic DiD number only pins down τ1 minus τ0; extra assumptions are needed to sign either.","key_machinery":"The machinery is a decomposition of the post-treatment assignment vector into unit i's own treatment, the treatment vector of i's group, and the treatment vector of the opposite group, keeping interference unrestricted. This lets the paper define potential outcomes indexed by the vectors (1,1,0) and (0,0,1) and state Assumption 7, the parallel trends condition on the never-observed no-treatment outcome Y(0,0,0). That assumption substitutes the observed pre-period group gap for the unobserved post-period gap, turning the observed DiD into τ1 − τ0. The proof is a rearrangement that isolates Y_i1(0,0,0) as a common counterfactual subtracted from both groups.","core_discovery":"Under unknown interference, the paper defines two well-defined causal estimands: τ1 = E[Y_i1(1,1,0) − Y_i1(0,0,0) | G=1], the total effect of the intervention on treated units, and τ0 = E[Y_i1(0,0,1) − Y_i1(0,0,0) | G=0], the average spillover effect on control units. Proposition 3 shows that DiD = τ1 − τ0 under Assumptions 1, 2, and 7. The paper's central negative claim is that without further assumptions the DiD number alone is uninformative about the sign or magnitude of either effect: a zero DiD can mean no effects anywhere or two equal nonzero effects, and a positive DiD only orders the two effects. Positive proposals follow: bounded-trend assumptions give interval bounds on each effect","pith_inferences":["Not stated in the paper: the framework implies that researchers should pre-specify which side of the contrast their policy question targets — the treated effect, the spillover on controls, or the difference — because the same DiD number supports very different policy readings depending on the assumed sign of τ0.","A testable extension suggested by the paper: use a third, plausibly isolated region as a no-spillover reference to estimate the no-treatment trend gap between the DiD groups and correct the contrast, or use variation in exposure intensity to estimate τ0 directly.","If treatment effects are heterogeneous and interference operates through general equilibrium channels, the sign of τ0 is often ambiguous; the paper's Assumption 10 cannot resolve that ambiguity, pointing toward design-based strategies such as deliberately placing control units outside spillover range.","An operational consequence not drawn in the note: for each group, one can report the value of k at which the partial-identification interval for τ_g crosses zero, giving a simple 'robustness frontier' that shows how large a differential trend would be needed to overturn a policy conclusion."],"forward_implications":["In any DiD application where cross-group spillovers are plausible, the reported coefficient should be read as a differential effect, not as a causal effect on the treated group alone.","A DiD of zero cannot be cited as evidence of no effect; it only indicates that the total effect on the treated equals the spillover effect on the control.","If researchers can sign the spillover effect (Assumption 10), the DiD estimate becomes a one-sided bound on the treated effect.","If researchers can bound the no-treatment time trend within ±k for each group, both effects are interval-identified without needing parallel trends.","Reinterpreting published DiD results under interference changes conclusions: for example, the classic New Jersey–Pennsylvania minimum wage estimate of 2.75 FTE workers only establishes that New Jersey's total effect exceeded Pennsylvania's spillover by 2.75 workers."],"fun_headline_variants":["DiD with interference: one number, two unknowns","Spillover breaks DiD: it's a gap, not an effect","Under interference, DiD identifies a contrast, not a cause","Diff-in-diff can't split effects when spillover exists"],"cache_read_input_tokens":14208,"weakest_assumption_plain":"The load-bearing premise is Assumption 7: the no-treatment potential outcome Y_i1(0,0,0) would have followed the same expected time trend in treated and control groups — yet this quantity is never observed for any unit in the post-period, making the assumption untestable from the data at hand.","fun_headline_variants_meta":{"raw":{"variants":["DiD with interference: one number, two unknowns","Spillover breaks DiD: it's a gap, not an effect","Under interference, DiD identifies a contrast, not a cause","Diff-in-diff can't split effects when spillover exists"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000754,"raw_usage":{"total_tokens":3191,"prompt_tokens":742,"completion_tokens":2449,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":2377}},"tokens_in":486,"tokens_out":2449,"duration_ms":18678,"temperature":1.0,"reasoning_tokens":2377,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:10:36.465694+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a setting with known interference where a spillover-free comparison group is available. Estimate the time trend of the no-treatment outcome Y(0,0,0) in the treated and control groups from that comparison group; if the trends differ, then the DiD estimand equals τ1 − τ0 plus the trend gap, contradicting the Proposition 3 reading. Concretely, in a cross-border minimum wage study, use a non-bordering state as a no-spillover reference and compare the pre-treatment employment trends of the treated and control states.","supporting_citations":[],"review_version":1}