REVIEW 1 major objections 4 minor 34 references
Difference-in-Differences in the Presence of Unknown Interference
T0 review · 1 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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).
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 (1)
- [Section 3.2, Proposition 7 and Assumption 11] 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.
minor comments (4)
- [Section 2.4, after Proposition 3] 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 3.2, proof of Proposition 7] 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 2.1, notation] 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.
- [Table 2] 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.
Circularity Check
No circularity: the central decomposition is a direct algebraic identity under explicit assumptions, and partial-identification bounds follow by algebra rather than by construction.
full rationale
The paper's central result, Proposition 3, is a direct algebraic identity. The DiD estimand is rewritten as E[Y_i1(1,1,0)-Y_i0(0,0,0)|G_i=1] - E[Y_i1(0,0,1)-Y_i0(0,0,0)|G_i=0], and Assumption 7 is then used to replace the pre-period group gap in Y0(0,0,0) with the post-period gap in Y1(0,0,0). This yields DiD = tau1 - tau0. There is no fitting, no normalization, and no parameter that is defined in terms of the target. The target estimands TATT and ASC appear only as the algebraic outcome of the decomposition, not as inputs to Assumption 7. Propositions 4-6 are similarly direct consequences of the stated assumptions: Assumption 8 says the no-treatment potential outcome is constant over time on average, so tau_g equals the observed before-after change; Assumption 9 gives bounds by asserting the trend lies in [-k,k]; Assumption 10 translates a sign restriction on tau0 into one-sided bounds on tau1 via Proposition 3. None of these asserts the conclusion in different notation. Proposition 7 has a boundary defect (the proof divides by tau1 and requires |tau1|>|tau0|), but that is a correctness issue, not circularity: Assumption 11 does not contain sgn(DiD)=sgn(tau1). The only self-citation is Forastiere et al. (2021), which is cited as background for SUTNVA under partial interference and is not load-bearing for any derivation in this note. The Card-Krueger discussion is illustrative interpretation, not part of the identification chain. No self-definition, no fitted-input-called-prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled via citation. The derivation is self-contained once the potential-outcome notation and stated assumptions are accepted.
Assumptions & free parameters
free parameters (1)
- k =
user-specified (not estimated)
assumptions (12)
- domain assumption Full 2^N-per-period potential outcomes Y_it(W) indexed by the complete treatment assignment matrix (Section 2.1)
- domain assumption Superpopulation sampling: expectations are over a superpopulation and are conditional on group membership (Section 2.1)
- standard math Linearity of expectation and rearrangement of conditional means
- domain assumption Assumption 1: No treatment anticipation
- domain assumption Assumption 2: No hidden versions of treatment / consistency
- domain assumption Canonical benchmark assumptions 3-4: no-interference and parallel trends under no-interference
- domain assumption Partial-interference benchmark assumptions 5-6: partial interference and parallel trends under partial interference
- domain assumption Assumption 7: Parallel trends under unknown interference for Y(0,0,0)
- domain assumption Assumption 8: Constant average no-treatment outcome over time within each group
- domain assumption Assumption 9: No-treatment time trend bounded by k in each group
- domain assumption Assumption 10a/b: sign of the average spillover effect on controls
- domain assumption Assumption 11: magnitude dominance |tau1| >= |tau0|
Cite this review
Pith. "Pith review of Difference-in-Differences in the Presence of Unknown Interference." pith.science (2026). https://pith.science/paper/BG7ONPE5
@misc{pith2026251221176,
author = {Pith},
title = {Pith review of: Difference-in-Differences in the Presence of Unknown Interference},
year = {2026},
howpublished = {\url{https://pith.science/paper/BG7ONPE5}},
note = {Machine review of arXiv:2512.21176}
}
read the original abstract
The stable unit treatment value (SUTVA) is a crucial assumption in the Difference-in-Differences (DiD) research design. It rules out hidden versions of treatment and any sort of interference and spillover effects across units. Even if this is a strong assumption, it has not received much attention from DiD practitioners and, in many cases, it is not even explicitly stated as an assumption, especially the no-interference assumption. In this technical note, we investigate what the DiD estimand identifies in the presence of unknown interference. We show that the DiD estimand identifies a contrast of causal effects, but it is not informative on any of these causal effects separately, without invoking further assumptions. Then, we explore different sets of assumptions under which the DiD estimand becomes informative about specific causal effects. We illustrate these results by revisiting the seminal paper on minimum wages and employment by Card and Krueger (1994).
Reference graph
Works this paper leans on
-
[1]
and Imbens, G
Arkhangelsky, D. and Imbens, G. (2024). Causal models for longitudinal and panel data: a survey.The Econometrics Journal, 27(3):C1–C61
2024
-
[2]
Aronow, P. M. and Samii, C. (2017). Estimating average causal effects under general interference, with application to a social network experiment.The Annals of Applied Statistics, 11(4):1912–1947
2017
-
[3]
Braghieri, L., Levy, R., and Makarin, A. (2022). Social Media and Mental Health.American Economic Review, 112(11):3660–3693
2022
-
[4]
Butts, K. (2023). Difference-in-Differences Estimation with Spatial Spillovers. arXiv preprint arXiv:2105.03737
arXiv 2023
-
[5]
Caires, F. B. (2024). Internal Organization of Firms and Minimum Wage Spillovers. SSRN working paper 5136707
2024
-
[6]
and Sant’Anna, P
Callaway, B. and Sant’Anna, P. H. C. (2021). Difference-in-Differences with multiple time periods.Journal of Econometrics, 225(2):200–230
2021
-
[7]
and Krueger, A
Card, D. and Krueger, A. B. (1994). Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania.The American Economic Review, 84(4):772–793
1994
-
[8]
Cengiz, D., Dube, A., Lindner, A., and Zipperer, B. (2019). The Effect of Minimum Wages on Low-Wage Jobs*.The Quarterly Journal of Economics, 134(3):1405–1454
2019
Show all 34 references
-
[9]
C., and Xu, D
Chen, Q., Chen, Z., Liu, Z., Su´ arez Serrato, J. C., and Xu, D. Y. (2025). Regulating Conglomerates: Evidence from an Energy Conservation Program in China.American Economic Review, 115(2):408–447. de Chaisemartin, C. and D’Haultfœuille, X. (2020). Two-Way Fixed Effects Estima...
2025
-
[10]
and Saggio, R
Fenizia, A. and Saggio, R. (2024). Organized Crime and Economic Growth: Evidence from Municipalities Infiltrated by the Mafia.American Economic Review, 114(7):2171–2200
2024
-
[11]
M., , and Mealli, F
Forastiere, L., Airoldi, E. M., , and Mealli, F. (2021). Identification and Estimation of Treatment and Interference Effects in Observational Studies on Networks.Journal of the American Statistical Association, 116(534):901–918
2021
-
[12]
Ghanem, D., Sant’Anna, P. H. C., and W¨ uthrich, K. (2025). Selection and parallel trends. arXiv preprint arXiv:2203.09001
2025 arXiv
-
[13]
Goodman-Bacon, A. (2021). Difference-in-differences with variation in treatment timing.Journal of Econo- metrics, 225(2):254–277
2021
-
[14]
Grossi, G., Mariani, M., Mattei, A., Lattarulo, P., and Oner, O. (2023). Direct and spillover effects of a new tramway line on the commercial vitality of peripheral streets. A synthetic-control approach. arXiv preprint arXiv:2004.05027. 13
2023 arXiv
-
[15]
Grossman, J. B. (1983). The Impact of the Minimum Wage on Other Wages.Journal of Human Resources, 18(3):359–378
1983
-
[16]
Hettinger, G., Roberto, C., Lee, Y., and Mitra, N. (2024). Doubly robust estimation of policy-relevant causal effects under interference.Journal of the Royal Statistical Society Series C: Applied Statistics
2024
-
[17]
and Oeß, E.-M
Huber, M. and Oeß, E.-M. (2024). A joint test of unconfoundedness and common trends. arXiv preprint arXiv:2404.16961
2024 arXiv
-
[18]
Imbens, G. W. and Rubin, D. B. (2015).Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge University Press, Cambridge
2015
-
[19]
C., Plotnick, R., van Inwegen, E., Vigdor, J., and Wething, H
Jardim, E., Long, M. C., Plotnick, R., van Inwegen, E., Vigdor, J., and Wething, H. (2022). Minimum-Wage Increases and Low-Wage Employment: Evidence from Seattle.American Economic Journal: Economic Policy, 14(2):263–314
2022
-
[20]
Jha, P., Neumark, D., and Rodriguez-Lopez, A. (2024). What’s Across the Border? Re-Evaluating the Cross-Border Evidence on Minimum Wage Effects. NBER working paper 32901
2024
-
[21]
Kuehn, D. (2016). Spillover Bias in Cross-Border Minimum Wage Studies: Evidence from a Gravity Model. Journal of Labor Research, 37(4):441–459
2016
-
[22]
McKinnish, T. (2017). Cross-state differences in the minimum wage and out-of-state commuting by low-wage workers.Regional Science and Urban Economics, 64:137–147
2017
-
[23]
and Roth, J
Rambachan, A. and Roth, J. (2023). A More Credible Approach to Parallel Trends.Review of Economic Studies, 90(5):2555–2591
2023
-
[24]
and Sant’Anna, P
Roth, J. and Sant’Anna, P. H. C. (2023). When Is Parallel Trends Sensitive to Functional Form?Econo- metrica, 91(2):737–747
2023
-
[25]
Roth, J., Sant’Anna, P. H. C., Bilinski, A., and Poe, J. (2023). What’s trending in difference-in-differences? A synthesis of the recent econometrics literature.Journal of Econometrics, 235(2):2218–2244
2023
-
[26]
Rubin, D. B. (1980). Randomization Analysis of Experimental Data: The Fisher Randomization Test Comment.Journal of the American Statistical Association, 75(371):591–593
1980
-
[27]
Shirley, P. (2018). The response of commuting patterns to cross-border policy differentials: Evidence from the American Community Survey.Regional Science and Urban Economics, 73:1–16
2018
-
[28]
and Xiao, Z
Sun, K. and Xiao, Z. (2025). Difference-in-Differences Under Network Interference. arXiv preprint arXiv:2509.24259
2025
-
[29]
and Abraham, S
Sun, L. and Abraham, S. (2021). Estimating dynamic treatment effects in event studies with heterogeneous treatment effects.Journal of Econometrics, 225(2):175–199. S¨ avje, F., Aronow, P. M., and Hudgens, M. G. (2021). Average treatment effects in the presence of unknown inter...
2021
-
[30]
and Wong, A
Truffa, F. and Wong, A. (2025). Undergraduate Gender Diversity and the Direction of Scientific Research. American Economic Review, 115(7):2414–2448. 14
2025
-
[31]
VanderWeele, T. J. and Hern´ an, M. A. (2013). Causal Inference Under Multiple Versions of Treatment. Journal of causal inference, 1(1):1–20
2013
-
[32]
Wooldridge, J. M. (2021). Two-Way Fixed Effects, the Two-Way Mundlak Regression, and Difference-in- Differences Estimators. SSRN working paper 3906345
2021
-
[33]
Xu, R. (2025). Difference-in-Differences with Interference. arXiv preprint arXiv:2306.12003
2025 arXiv
-
[34]
Xu, Y., Zhao, A., and Ding, P. (2025). Factorial Difference-in-Differences. arXiv preprint arXiv:2407.11937. 15
2025
Reviewed August 3, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.