arxiv: 2604.22532 · v1 · submitted 2026-04-24 · 💰 econ.EM

Recognition: unknown

Causal Identification under Interference: The Role of Treatment Assignment Independence

Julius Owusu, Monika Avila M\'arquez

Authors on Pith no claims yet

Pith reviewed 2026-05-08 09:08 UTC · model grok-4.3

classification 💰 econ.EM

keywords causal inferenceinterferencetreatment assignmentaverage direct effectssensitivity analysisobservational studiesinstrumental variablesdifference-in-differences

0 comments

The pith

Under restrictions on treatment assignment dependence, standard ITR-based formulas identify average direct effects even with arbitrary interference.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper examines what conventional causal identification strategies actually recover when units can interfere with one another. It shows that if treatment assignments satisfy certain independence restrictions across units, the usual formulas from selection-on-observables, instrumental variables, regression discontinuity, and difference-in-differences continue to identify well-defined average direct effects. These results require no knowledge of the interference network or exposure mappings. The authors also supply a sensitivity framework to gauge how much inference changes if the assignment-independence restrictions are violated.

Core claim

Under restrictions on dependence between treatment assignments across units, conventional ITR-based identification formulas identify well-defined causal objects: types of average direct effects (ADEs). These results hold for selection-on-observables, instrumental variables, regression discontinuity designs, and difference-in-differences without requiring knowledge of the interference structure or specification of exposure mappings.

What carries the argument

Treatment assignment independence under arbitrary interference, which preserves the causal content of standard ITR-based formulas as average direct effects.

If this is right

Standard estimates from common designs can be reinterpreted as average direct effects rather than total effects when interference is present but assignment independence holds.
Researchers need not model or estimate the full interference structure to obtain interpretable causal quantities.
Sensitivity analysis quantifies how much statistical conclusions change under varying degrees of violation of assignment independence.
Existing empirical studies in settings with potential spillovers can retain their validity for direct-effect questions under the independence restrictions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Many published estimates in networked economic settings could be re-read as direct-effect results without new data collection.
The framework opens the door to designing assignment mechanisms that deliberately satisfy the independence restrictions to improve robustness.
Extensions could test the assignment-independence restrictions directly using observed treatment data.

Load-bearing premise

Treatment assignments across units satisfy restricted forms of independence even though outcomes may interfere arbitrarily.

What would settle it

Data showing that treatment assignments are dependent in ways that violate the stated restrictions, together with estimates that deviate from the true average direct effects computed under full knowledge of the interference pattern.

Figures

Figures reproduced from arXiv: 2604.22532 by Julius Owusu, Monika Avila M\'arquez.

**Figure 1.** Figure 1: Sensitivity to CAI violation using Calibrated Data view at source ↗

read the original abstract

Empirical researchers routinely invoke the no-interference or \textit{individualistic treatment response} (ITR) assumption to identify causal effects in observational studies, despite concerns that interference across units may arise in many economic settings. This paper studies the causal content of standard ITR-based identification formulas when arbitrary interference is present. We show that, under restrictions on dependence between treatment assignments across units, conventional ITR-based identification formulas -- including those underlying selection-on-observables, instrumental variables, regression discontinuity designs, and difference-in-differences -- identify well-defined causal objects: types of \textit{average direct effects} (ADEs). These results do not require knowledge of the interference structure or specification of exposure mappings. We also propose a sensitivity analysis framework that quantifies the robustness of statistical inference to violations of treatment-assignment independence under arbitrary interference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows that treatment assignment independence lets standard ITR formulas recover average direct effects under arbitrary interference without exposure mappings, but the RD and DiD cases need extra conditions on auxiliary variables.

read the letter

The main takeaway is that under restrictions keeping treatment assignments independent across units, the usual selection-on-observables, IV, RD, and DiD formulas identify well-defined average direct effects even with interference present. This holds without specifying how units affect each other. The result is new in its clean separation of assignment dependence from the interference structure itself, and it gives a direct way to reinterpret a lot of existing empirical work as measuring direct effects rather than total effects. The sensitivity framework they sketch for violations of the independence restriction is a practical addition that applied people could actually use. The paper does well by staying close to the identification formulas researchers already run and showing what those formulas still deliver. The soft spot is the stress-test concern on RD and DiD. If the running variable or the group/time indicators are themselves shaped by interference, the local continuity or parallel-trends assumptions can break even when raw treatment indicators satisfy the paper's dependence restriction. The abstract does not flag auxiliary conditions on those variables, so the claim for those two designs looks narrower than stated. Without the full proofs it is hard to see exactly how they close that gap. This paper is for applied microeconomists and policy evaluators who face possible spillovers but want to keep using familiar estimators. It is worth sending to peer review because the core identification argument is grounded enough to deserve referee scrutiny, even if the RD and DiD sections need tightening.

Referee Report

3 major / 2 minor

Summary. The paper claims that under restrictions on dependence between treatment assignments across units (termed treatment assignment independence), standard ITR-based identification formulas for selection-on-observables, IV, RD, and DiD identify well-defined types of average direct effects (ADEs) even under arbitrary interference, without requiring knowledge of the interference structure or exposure mappings. It further proposes a sensitivity analysis framework to quantify robustness of inference to violations of these assignment independence restrictions.

Significance. If the results hold, the paper makes a meaningful contribution to the interference literature in econometrics by showing how conventional empirical strategies retain causal content (as ADEs) under weaker conditions than full no-interference. The avoidance of explicit exposure mappings is a practical strength, and the sensitivity framework provides a concrete tool for applied researchers. This could influence how observational studies with potential spillovers are interpreted and stress-tested.

major comments (3)

[§4] §4 (RD application): The identification result for regression discontinuity states that the standard RD formula recovers an ADE under treatment assignment independence. However, the running variable X_i is a deterministic function of the assignment rule; if interference allows neighbors' treatments or assignments to affect the distribution of X_i or the potential outcomes Y_i(t, t_{-i}) in a manner that violates local continuity or randomization, the result does not follow from assignment independence alone. The manuscript does not appear to impose or verify auxiliary conditions on the running variable or its potential dependence on T_{-i}.
[§4] §4 (DiD application): The DiD identification argument similarly relies on the group/time indicators satisfying the independence restriction. Yet if interference induces correlation between these indicators across units (e.g., via shared exposure affecting group formation), the parallel trends assumption underlying the DiD formula may fail even when raw treatment indicators T_i are independent. No explicit discussion or auxiliary assumption addresses this channel.
[§3.2] §3.2 (general identification): The central claim that ITR formulas identify ADEs 'without knowledge of the interference structure' is load-bearing. While the argument holds directly for selection-on-observables and IV under the stated conditional independence of T_i and T_{-i}, the extension to RD and DiD requires that the auxiliary variables (running variable, group indicators) are themselves unaffected by interference in ways that preserve the identifying assumptions; this is not formally stated or proven.

minor comments (2)

[§2] Notation for the different types of ADEs (e.g., marginal vs. conditional) should be introduced earlier and used consistently in the sensitivity section to improve readability.
[§5] The sensitivity analysis framework would benefit from an explicit mapping between the sensitivity parameters and the degree of violation of treatment assignment independence (e.g., a correlation bound).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments, which help clarify the scope of our results for design-based methods. We address each major comment below. Where the comments identify gaps in explicit assumptions or discussion, we will revise the manuscript to incorporate the necessary clarifications and auxiliary conditions without altering the core claims.

read point-by-point responses

Referee: §4 (RD application): The identification result for regression discontinuity states that the standard RD formula recovers an ADE under treatment assignment independence. However, the running variable X_i is a deterministic function of the assignment rule; if interference allows neighbors' treatments or assignments to affect the distribution of X_i or the potential outcomes Y_i(t, t_{-i}) in a manner that violates local continuity or randomization, the result does not follow from assignment independence alone. The manuscript does not appear to impose or verify auxiliary conditions on the running variable or its potential dependence on T_{-i}.

Authors: We agree that treatment assignment independence alone does not automatically preserve the continuity or local randomization assumptions if interference can affect the distribution of the running variable X_i. In standard RD applications, X_i is typically a pre-determined unit-specific characteristic (e.g., a score or distance measure observed before treatment), but the manuscript does not explicitly rule out channels where T_{-i} influences X_i or the relevant potential outcomes. We will revise §4 to add an auxiliary assumption that X_i is independent of T_{-i} conditional on the covariates (X_i ⊥ T_{-i} | covariates), ensuring the standard RD formula identifies the ADE at the cutoff under TAI. This makes the conditions explicit while preserving the result that no exposure mapping is required. revision: yes
Referee: §4 (DiD application): The DiD identification argument similarly relies on the group/time indicators satisfying the independence restriction. Yet if interference induces correlation between these indicators across units (e.g., via shared exposure affecting group formation), the parallel trends assumption underlying the DiD formula may fail even when raw treatment indicators T_i are independent. No explicit discussion or auxiliary assumption addresses this channel.

Authors: This observation is correct: TAI on the raw treatment indicators T_i does not automatically extend to the group and time indicators if interference can induce dependence through shared exposures or group formation. The manuscript applies the DiD formula under the maintained assumption that these indicators satisfy the independence restriction, but does not discuss potential violations via interference. We will revise §4 to include an explicit auxiliary assumption that the group/time indicators are independent of T_{-i} given covariates, or note that the sensitivity analysis framework can be applied to assess robustness if this fails. This addresses the channel without requiring knowledge of the full interference structure. revision: yes
Referee: §3.2 (general identification): The central claim that ITR formulas identify ADEs 'without knowledge of the interference structure' is load-bearing. While the argument holds directly for selection-on-observables and IV under the stated conditional independence of T_i and T_{-i}, the extension to RD and DiD requires that the auxiliary variables (running variable, group indicators) are themselves unaffected by interference in ways that preserve the identifying assumptions; this is not formally stated or proven.

Authors: We acknowledge that the direct proofs in §3.2 for selection-on-observables and IV follow immediately from conditional TAI without specifying the interference function, but the RD and DiD applications in §4 rely on the design-specific identifying assumptions (continuity, parallel trends) remaining valid for the auxiliary variables. The manuscript presents these as straightforward extensions but does not formally state the required independence of auxiliaries from T_{-i}. We will revise §3.2 to include a general proposition on auxiliary variables in design-based estimators and add a brief proof sketch for the RD and DiD cases. This clarifies that 'without knowledge of the interference structure' means we avoid exposure mappings but still require TAI plus independence of auxiliaries; the revision strengthens the formalization without changing the main results. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain.

full rationale

The paper establishes that, given restrictions on dependence between treatment assignments, standard ITR identification formulas recover well-defined ADEs without requiring knowledge of the interference structure. This follows from applying the paper's stated independence condition to the usual conditional independence or continuity assumptions underlying selection-on-observables, IV, RD, and DiD; the resulting objects are defined directly from the potential-outcome framework and the new restriction rather than by re-labeling fitted quantities or importing load-bearing results from the authors' prior work. No self-definitional steps, fitted-input predictions, or ansatz smuggling via self-citation appear in the derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard causal inference primitives (potential outcomes, consistency) plus the domain assumption of treatment assignment independence; no free parameters or invented entities are introduced in the abstract.

axioms (2)

standard math Standard causal inference assumptions including consistency, positivity, and the individualistic treatment response (ITR) framework.
Invoked to define the identification formulas and potential outcomes under interference.
domain assumption Restrictions on dependence between treatment assignments across units.
The key condition enabling the result that ITR formulas identify ADEs; stated but not derived from more primitive principles.

pith-pipeline@v0.9.0 · 5436 in / 1283 out tokens · 64677 ms · 2026-05-08T09:08:46.422424+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

6 extracted references

[1]

Estimate propensity scores ˆei = Λ(W ⊤ i ˆβ) from a logistic regression ofD i onW i
[2]

, K}by discretizing{ˆe i}N i=1 intoKquantile bins

Construct strataS i ∈ {1, . . . , K}by discretizing{ˆe i}N i=1 intoKquantile bins
[3]

For each stratums, setI s ={i:S i =s},n s =|I s|, and ˆes =n −1 s P i∈Is Di
[4]

Compute within-stratum ranksq i =r s(Yi),fori∈ I s, wherer s(·) assigns ranks to{Y j :j∈ I s}(using average ranks for ties)
[5]

Compute the observed test statistic:T obs =PK s=1 P i:Si=s qiDi
[6]

7.forj= 1,

For eachs, precompute h T (l) s (ms) i ms=0,...,ns l=1,...,L , the within-stratum test-statistic contribution by uniformly selectingm s treated units replacement. 7.forj= 1, . . . , Jdo Set the variableξequal toξ j ifξ= 1then fors= 1, . . . , Kdo Setπ s(ms) = ns ms ˆems s (1−ˆes)ns−ms form s = 0, . . . , ns i. SimulateB base draws{T (b)}Bbase b=1 as follo...

1985