arxiv: 2605.03204 · v1 · submitted 2026-05-04 · 📊 stat.ME · econ.EM

Recognition: unknown

Estimating peer effects in noisy, low-rank networks via network smoothing

Alex Hayes , Keith Levin

Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:24 UTC · model grok-4.3

classification 📊 stat.ME econ.EM

keywords peer effectsnoisy networkslow-rank matrix estimationmeasurement errornetwork smoothingadjacency matrixsocial networksasymptotic equivalence

0 comments

The pith

Peer effects on noisy networks are asymptotically equivalent to those on the expected adjacency matrix when it is low-rank.

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

The paper shows that peer effect estimates computed from the expected adjacency matrix match those from the true unobserved network in large samples. This holds provided the expected matrix has low-rank structure, which converts noisy network data into a standard low-rank matrix estimation problem. Researchers can then plug in any suitable denoising technique matched to their sampling method, such as for egocentric samples or missing edges. The equivalence matters because direct use of noisy observed networks typically makes peer effect estimators inconsistent.

Core claim

Our key result shows that peer effects over a true unobserved network are asymptotically equivalent to peer effects over the expected adjacency matrix. This result reduces peer effect estimation in noisy networks to low-rank matrix estimation targeting the expected adjacency matrix. We develop our theory for weighted networks observed with additive noise, but simulations suggest the approach can be applied more generally when there is a low-rank estimation method suited to a particular noise structure. We demonstrate via simulations that our approach applies to egocentric samples, aggregated relational data, and networks with missing edges, each requiring a different low-rank estimation.

What carries the argument

Low-rank estimator of the expected adjacency matrix, which acts as the smoothed network on which peer effects are computed.

If this is right

Standard peer effect estimators become consistent when applied to the low-rank estimate of the expected adjacency matrix rather than the raw noisy network.
The theory covers weighted networks with additive noise and extends in practice to other noise types via appropriate low-rank methods.
The same peer effect model can be fit without modification once the network has been replaced by its low-rank estimate.
Different data collection designs (egocentric, aggregated, missing edges) each admit their own low-rank estimator while preserving the asymptotic equivalence.

Where Pith is reading between the lines

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

Social scientists could collect cheaper, noisier network data and still recover peer effects if the expected structure is low-rank.
The result points toward designing network surveys around low-rank recovery rather than minimizing measurement error directly.
Similar smoothing may apply to dynamic or multiplex networks if their expected slices remain low-rank.

Load-bearing premise

The expected adjacency matrix is low-rank and the observed noise permits a consistent low-rank estimator of it.

What would settle it

A simulation or dataset where the expected adjacency matrix is low-rank yet peer effect estimates from the low-rank estimate diverge from those on the true network at large sample sizes.

Figures

Figures reproduced from arXiv: 2605.03204 by Alex Hayes, Keith Levin.

**Figure 1.** Figure 1: Adjacency matrices, corresponding expectations, and estimated expectations view at source ↗

**Figure 2.** Figure 2: Monte Carlo estimates of mean squared error of view at source ↗

**Figure 3.** Figure 3: Monte Carlo estimates of mean squared error of view at source ↗

**Figure 4.** Figure 4: Monte Carlo estimates of mean squared error of view at source ↗

**Figure 5.** Figure 5: Monte Carlo estimates of mean squared error of view at source ↗

**Figure 6.** Figure 6: Point estimates and asymptotic 95% confidence intervals for view at source ↗

read the original abstract

Peer effect estimation requires precise network measurement, yet most empirical networks are noisy, rendering standard estimators inconsistent. To address measurement error in networks, we propose a method to estimate peer effects in networks whose expected adjacency matrix is low-rank. Our key result shows that peer effects over a true unobserved network are asymptotically equivalent to peer effects over the expected adjacency matrix. This result reduces peer effect estimation in noisy networks to low-rank matrix estimation targeting the expected adjacency matrix. We develop our theory for weighted networks observed with additive noise, but simulations suggest approach can be applied more generally when there is a low-rank estimation method suited to a particular noise structure. We demonstrate via simulations that our approach applies to egocentric samples, aggregated relational data, and networks with missing edges, each requiring a different low-rank estimation method.

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 reduces peer-effect estimation to low-rank recovery of the expected adjacency matrix via an asymptotic equivalence, but derives the result only for additive noise on weighted networks.

read the letter

The central move is clean: under a low-rank expected adjacency matrix, peer effects on the unobserved true network are asymptotically the same as peer effects on that expectation. This lets you replace the noisy observed network with a low-rank estimate before running the usual peer-effect regression. For the additive-noise weighted case the paper actually derives the equivalence, which is the part that feels new and useful. The simulations then show that swapping in a different low-rank estimator handles egocentric samples, aggregated relational data, and missing edges, and the numbers look reasonable in those settings. That practical flexibility is the main thing the work contributes beyond the theory for the additive case. The soft spot is exactly what the stress-test flags. The equivalence is proven only for additive noise; the other measurement-error regimes get simulation support but no corresponding derivation showing the equivalence still holds or that the resulting estimator stays consistent. If the low-rank assumption is even moderately off, or if the noise structure does not match one of the tested estimators, the method could degrade without warning. The paper does not explore sensitivity to these issues. This is for network econometricians or social-science researchers who already believe their adjacency matrix is approximately low-rank and who have one of the covered noise patterns. A reader facing exactly those conditions would get a usable smoothing step and some simulation evidence that it works. I would send it to peer review. The core reduction is worth referee attention even if the authors need to tighten the scope or add theory for the simulation-only cases.

Referee Report

2 major / 2 minor

Summary. The paper claims that peer effects computed on a true unobserved network are asymptotically equivalent to those computed on the expected adjacency matrix when the latter is low-rank. This equivalence reduces peer-effect estimation in noisy networks to a low-rank matrix recovery problem targeting the expectation. Theory is developed for weighted networks observed under additive noise; simulations are used to suggest that any suitable low-rank estimator can be substituted for other noise structures (egocentric sampling, aggregated relational data, missing edges).

Significance. If the asymptotic equivalence holds with explicit conditions and rates, the result supplies a clean reduction of a measurement-error problem to low-rank estimation, a structure that appears frequently in social and economic networks. The explicit theory for the additive-noise case and the demonstration that different low-rank estimators can be plugged in for different sampling schemes are concrete strengths that could make the method usable in applied work.

major comments (2)

[Abstract; simulation section (likely §5)] The central asymptotic equivalence (abstract and the key result referenced in the introduction) is derived only for weighted networks under additive noise. For the three additional settings highlighted in the abstract and simulation section (egocentric samples, aggregated relational data, missing edges), the manuscript supplies only simulation evidence that a matched low-rank estimator can be used; no corresponding derivation is given showing that the equivalence itself continues to hold or that the resulting peer-effect estimator remains consistent.
[Theory development section (likely §3–4)] The statement that the approach 'reduces peer effect estimation in noisy networks to low-rank matrix estimation' (abstract) is load-bearing for the paper's contribution, yet the manuscript does not state explicit conditions on the low-rank estimator (e.g., convergence rate, bias, or operator-norm error) that are required for the asymptotic equivalence to carry over from the additive-noise case.

minor comments (2)

Notation distinguishing the realized adjacency matrix A from its expectation E[A] should be introduced once and used consistently; current usage occasionally blurs the two in the simulation descriptions.
[Simulation section] The simulation tables would benefit from reporting the exact low-rank estimator employed for each noise structure (e.g., nuclear-norm, SVD truncation, or matrix completion method) alongside the peer-effect estimates.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the scope of our theoretical contributions. We address each major point below and will revise the manuscript to improve precision and transparency.

read point-by-point responses

Referee: [Abstract; simulation section (likely §5)] The central asymptotic equivalence (abstract and the key result referenced in the introduction) is derived only for weighted networks under additive noise. For the three additional settings highlighted in the abstract and simulation section (egocentric samples, aggregated relational data, missing edges), the manuscript supplies only simulation evidence that a matched low-rank estimator can be used; no corresponding derivation is given showing that the equivalence itself continues to hold or that the resulting peer-effect estimator remains consistent.

Authors: The referee is correct that the formal asymptotic equivalence between peer effects on the unobserved network and on its expectation is derived exclusively for the weighted additive-noise model. For egocentric sampling, aggregated relational data, and missing edges, the manuscript offers only simulation evidence that an appropriate low-rank estimator can be substituted, without claiming or proving that the same equivalence holds or that consistency is preserved. We will revise the abstract, introduction, and simulation section to explicitly distinguish the proven result from the simulation-based illustrations of broader applicability, and we will add a limitations paragraph noting that full theoretical extensions would require case-specific noise models and proofs. revision: partial
Referee: [Theory development section (likely §3–4)] The statement that the approach 'reduces peer effect estimation in noisy networks to low-rank matrix estimation' (abstract) is load-bearing for the paper's contribution, yet the manuscript does not state explicit conditions on the low-rank estimator (e.g., convergence rate, bias, or operator-norm error) that are required for the asymptotic equivalence to carry over from the additive-noise case.

Authors: We agree that the reduction claim requires explicit conditions on the low-rank estimator. The current proof implicitly relies on the estimator converging to the true expectation in operator norm at a rate sufficient to preserve the asymptotic equivalence of the peer-effect estimators. We will add a dedicated subsection in the theory development (likely §3 or §4) that states the precise requirements, including the necessary operator-norm convergence rate (e.g., ||Â − A||_op = o_p(1) or faster, depending on the moment conditions), bounded bias, and any additional assumptions needed for the equivalence to hold. This will make the load-bearing statement fully rigorous. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central asymptotic equivalence derived from model assumptions

full rationale

The paper's key result establishes an asymptotic equivalence between peer effects computed on the unobserved true network and those on the expected adjacency matrix, under the stated low-rank assumption and additive noise model for weighted networks. This equivalence is obtained via direct analysis of the probabilistic model rather than by fitting parameters to data or by self-referential definition. The reduction of estimation to low-rank matrix recovery follows as a consequence of the proven equivalence, not as a tautology. For egocentric sampling, aggregated relational data, and missing edges, the paper explicitly limits the formal theory to additive noise and uses simulations only to suggest applicability, without claiming or deriving a general equivalence for those cases. No load-bearing step reduces by construction to its own inputs, and no self-citation chain is invoked to justify the uniqueness or validity of the core theorem.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that the expected adjacency matrix is low-rank and that the noise admits a low-rank estimator; no free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption The expected adjacency matrix is low-rank
This is the key modeling assumption that enables reduction to low-rank matrix estimation.

pith-pipeline@v0.9.0 · 5424 in / 1105 out tokens · 38185 ms · 2026-05-08T17:24:21.075877+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

How to Generate Personal Net- works: Issues and T ools for a Sociological Perspective

DOI: 10.1111/j.2517-6161.1974.tb00999.x. Bhatia, Rajendra (1997). Matrix Analysis. Springer. Bidart, Claire and Johanne Charbonneau (Aug. 2011). “How to Generate Personal Net- works: Issues and T ools for a Sociological Perspective”. In: Field Methods 23.3, pp. 266–

work page doi:10.1111/j.2517-6161.1974.tb00999.x 1974
[2]

A Review of Software for Spatial Econometrics in R

DOI: 10.1177/1525822X11408513. Bivand, Roger, Giovanni Millo, and Gianfranco Piras (June 2021). “ A Review of Software for Spatial Econometrics in R”. In:Mathematics 9.11, p. 1276. DOI:10.3390/math9111276. Boucher, Vincent and Aristide Houndetoungan (Oct. 2025). “Estimating Peer Effects Us- ing Partial Network Data”. In: Review of Economics and Statistics...

work page doi:10.1177/1525822x11408513 2021
[3]

Regression Adjustments for Estimating the Global Treatment Effect in Experiments with Interference

DOI: 10.1016/j.jeconom.2022.10.005. Chin, Alex (Sept. 2019). “Regression Adjustments for Estimating the Global Treatment Effect in Experiments with Interference”. In: Journal of Causal Inference 7.2, p. 20180026. DOI: 10.1515/jci-2018-0026. Cho, Juhee, Donggyu Kim, and Karl Rohe (Apr. 2019). “Intelligent Initialization and Adap- tive Thresholding for Iter...

work page doi:10.1016/j.jeconom.2022.10.005 2022
[4]

Best Spatial T wo-Stage Least Squares Estimators for a Spatial Autoregres- sive Model with Autoregressive Disturbances

DOI: 10.1017/S0266466602182028. — (Jan. 2003). “Best Spatial T wo-Stage Least Squares Estimators for a Spatial Autoregres- sive Model with Autoregressive Disturbances”. In: Econometric Reviews 22.4, pp. 307–

work page doi:10.1017/s0266466602182028 2003
[5]

Asymptotic Distributions of Quasi-Maximum Likelihood Estimators for Spatial Autoregressive Models

DOI: 10.1081/ETC-120025891. — (Nov . 2004). “ Asymptotic Distributions of Quasi-Maximum Likelihood Estimators for Spatial Autoregressive Models”. In: Econometrica 72.6, pp. 1899–1925. DOI: 10.1111/j. 1468-0262.2004.00558.x. Lee, Lung-Fei, Xiaodong Liu, and Xu Lin (July 2010). “Specification and Estimation of So- cial Interaction Models with Network Struct...

work page doi:10.1081/etc-120025891 2004