Pith. sign in

REVIEW 3 major objections 4 minor 76 references

Design-Based and Network Sampling-Based Uncertainties in Network Experiments

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Correlations among exposure variables contaminate OLS spillover estimates in network experiments, even when treatment is random.

desk verdict A useful design-based extension of contamination-bias logic to network experiments, though the headline quantitative claims rest on a linearity assumption that is substantive for the continuous exposure mappings in its own application. read the letter →

arxiv 2506.22989 v3 pith:O7JQTETX submitted 2025-06-28 econ.EM stat.ME

classification econ.EMstat.ME MSC 62D0562G2091D30
keywords networkexperimentsspillovereffectscontaminationbiasdesign-basedinferencesamplingheterogeneoustreatmentexposuremappingHACestimator
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Network experiments are usually analyzed by running OLS on summary measures of neighbors' treatments, such as the share of treated friends. This paper argues that when treatment effects differ across people, correlations between those summary measures contaminate every OLS coefficient: the coefficient on one exposure variable can be negative even when every individual's true effect from that exposure is positive. The paper further shows that sampling the network adds a second source of contamination, so the sample-level estimand can remain biased even when the population-level estimand is clean. If the argument holds, the common practice of reading OLS spillover coefficients as causal effects requires an extra no-correlation condition, and the paper proposes a conservative network HAC variance estimator for the cases in which the condition holds.

What carries the argument

The object carrying the argument is the elementwise residual $U_{n,i,(k)}$: the part of exposure variable $k$ left after projecting it on the other exposure variables, the same construction as in Frisch-Waugh-Lovell. It converts the multivariate OLS coefficient into a weighted average of individual effects, but because the residual is uncorrelated with the other exposures only in aggregate, the per-unit cross terms $E[U_{n,i,(k)} X'_{n,i,(-k)}]$ survive and become contamination weights. The second engine is the sampled exposure mapping $\tilde{T}_{n,i}$ computed from the sampled network, whose mismatch with $T_{n,i}$ enters the sample-level estimand; the paper pairs this with a network $\psi$-dependence condition and a HAC variance estimator to obtain limit theorems.

What would settle it

Re-estimate the empirical application using the overlap-free exposure mapping: the paper predicts the coefficient on weak connections will drop from about -0.15 to about -0.07; if it stays strongly negative, the contamination mechanism is not the explanation. Alternatively, in a simulation with true $\theta_{n,i,(3)} = 0$ and correlated exposure elements, OLS on the contaminated mapping should yield a nonzero third coefficient whose size tracks the cross-correlation.

Watch

Extended reading notes

Core claim

The central claim is that, under a linear potential-outcome model with heterogeneous effects, the OLS estimand decomposes into individual effects with weights that mix the exposure variables: equation (8) writes the $k$-th coefficient as a ratio of two sums, with the numerator including the contamination term $\sum_i E[U_{n,i,(k)} X'_{n,i,(-k)}] \theta_{n,i,(-k)}$ from other exposure components. Because these cross-terms need not vanish, residualizing on covariates does not remove them, and the coefficient on one spillover variable can be negative while all underlying effects $\theta_{n,i,(k)}$ are positive. When the network is sampled rather than fully observed, the observed and true exposure mappings differ, and even a clean population-level estimand can be contaminated at the sample level; the two estimands coincide in large samples only under correct specification and complete observation of relevant links. The paper also derives the asymptotic distribution of the OLS estimator and gives a conservative network HAC variance estimator, and it shows in simulations and in an application to weather-insurance diffusion that contamination can be large enough to flip the sign of a second-neighbor spillover estimate.

Load-bearing premise

The potential outcome is assumed exactly linear in the exposure mapping; for continuous exposures such as share of treated friends this is a real restriction, and if the true outcome model is nonlinear the claimed weighted-average and contamination decomposition no longer hold.

Editorial extensions

If this is right

  • A researcher who sees a positive OLS coefficient on 'share of treated friends' cannot conclude the spillover is positive unless the exposure variables are conditionally uncorrelated.
  • Sampling-based network data collection does not fix the problem: it can create contamination even when the population estimand is clean, so external validity fails.
  • The no-correlation condition of Corollary 2, plus positive covariance between observed and true exposure, restores a convex-combination interpretation with nonnegative weights.
  • When the exposure mapping is discrete and satisfies overlap, IPW estimators avoid contamination; for continuous exposures, regression is the fallback but needs the no-correlation check.
  • The proposed network HAC variance estimator with eigenvalue clipping gives conservative standard errors in large samples when no contamination is present.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A practical diagnostic suggested by the argument is to report the correlation matrix of exposure-mapping elements and re-estimate with an overlap-free exposure mapping; large coefficient changes flag contamination.
  • The same contamination logic should apply to any regression-based aggregator that uses exposure mappings as regressors, including instrumental-variables and debiased estimators, which the paper does not develop.
  • Because homogeneous treatment effects eliminate contamination, heterogeneity tests could serve as a low-cost screen before deciding whether to abandon OLS.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This paper studies the causal interpretation of OLS estimators for spillover effects in network experiments when treatment assignment is random and network links are sampled. The authors define population-level and sample-level causal estimands as solutions to moment conditions, and under Assumption 2 (potential outcomes exactly linear in a true exposure mapping) and Assumption 4 (linear conditional expectation of the exposure mapping given covariates), they decompose each OLS coefficient into a weighted average of own heterogeneous effects plus a contamination term from other exposure elements (Theorem 1 and Corollary 1). They show that network sampling can create sample-level contamination even when the population-level estimand is clean, provide consistency and asymptotic normality results under local network dependence (Theorems 2--5), propose a network HAC variance estimator with an eigenvalue modification (Theorems 6--7), and illustrate the issues with simulations on a real network and a re-analysis of Cai et al. (2015).

Significance. The decomposition of OLS estimands into an own-effect component and a contamination component is a useful extension of Goldsmith-Pinkham, Hull and Kolesár (2022) to network experiments with continuous exposure mappings, and the population-versus-sample estimand distinction is a genuine conceptual contribution. The asymptotic theory is built on credible network-dependence machinery from Kojevnikov, Marmer and Song (2021), and the paper includes detailed proofs, simulations, and a survey of applied practice. If Assumption 2 is maintained, the sign-reversal warning is well supported and practically important. The main limitation is that the headline claim is exact only under the linearity of the potential outcome model; for the continuous exposure mappings that motivate the application, this is a substantive functional-form restriction rather than a without-loss assumption. The empirical illustration is suggestive but does not by itself demonstrate contamination bias. These concerns are addressable by reframing the scope of the claims and adding robustness analysis.

major comments (3)
  1. [Section 2.3; Theorem 1; Corollary 1, Eq. (8)] The contamination decomposition in Eq. (8) is derived from Assumption 2, which requires Y*_{n,i}(t) = t'θ_{n,i} + ν_{n,i} to hold exactly for all t. The paper notes that this is without loss when the exposure mapping takes finitely many values, but the running example and the empirical application use the share of treated friends and second-neighborhood shares, which are continuous. For continuous exposure mappings, exact linearity is a genuine restriction: if the true outcome model is nonlinear, the OLS estimand is not the weighted average of the θ_{n,i} in Eq. (8), and the clean split into an own-effect term and a contamination term does not obtain. The headline warning that an OLS coefficient can be negative even when all own effects are positive is therefore not established for the continuous-exposure setting emphasized in the paper. I recommend either explicitly restricting the causal-interpretation claims to linear potential outcome models, or adding a formal approximation/robustness result and a simulation with a nonlinear DGP showing that the qualitative contamination phenomenon persists.
  2. [Section 6, Table 4] The empirical comparison between the 'with overlaps' and 'no overlaps' specifications is interpreted as evidence of contamination bias, but the two regressions use different exposure mappings and therefore estimate different causal estimands even under the paper's own framework. The observed difference in the weak coefficient could reflect the change in the estimand itself rather than contamination in the sense of Eq. (9). Without ground-truth values of θ_{n,i}, the data cannot identify which part of the difference is contamination. I suggest softening the claim that the application 'demonstrates' contamination bias and instead presenting the result as consistent with the theoretical decomposition under the maintained linearity and no-omitted-variable assumptions.
  3. [Section 4, Theorems 6--7] The paper repeatedly calls the proposed variance estimator 'conservative,' but Theorem 7 formally states only that the estimator converges to the true variance plus bias terms eB+_n or bB+_n. It would be helpful to state explicitly that these bias matrices are positive semidefinite under the maintained assumptions and to spell out the resulting coverage implication for Wald confidence sets. As written, a reader cannot immediately verify that the eigenvalue modification delivers the advertised conservativeness rather than merely a positive semidefinite estimator.
minor comments (4)
  1. [Equation (8)] The notation θ_{n,i,(-k)} is used before a formal definition; a brief definition as the subvector of θ_{n,i} excluding the k-th element would improve readability.
  2. [Assumption 8] The expression eΣ^{-(1+p/2)}_n is ambiguous because eΣ_n is a matrix; the intended matrix norm should be stated explicitly.
  3. [Section 6, footnote 11] The statement that the five-friend limit in Cai et al. (2015) is taken at face value sits somewhat uneasily with Section 2.5's censoring framework; the authors should either justify why censoring is negligible here or apply the censoring-adjustment approach from Example 4.
  4. [Table 3] The table reports average absolute deviations of the estimator from the estimands, but not Monte Carlo standard errors for the displayed means; adding them would help readers gauge the precision of reported simulation quantities.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the estimand decompositions are algebraic consequences of the stated moment conditions and linear potential-outcome model, and no fitted parameter is relabeled as a prediction.

full rationale

The paper's central decomposition results (Theorem 1 and Corollary 1) are derived directly from the definitions of the population and sample moment-condition estimands in equations (2)-(5), combined with Assumption 2 (linear potential outcomes) and Assumption 4 (linear propensity/conditional mean of the exposure mapping). These are algebraic identities conditional on the assumptions, not fitted quantities later presented as predictions. The simulation in Section 5 constructs the data from specified heterogeneous theta values and then computes OLS estimands to illustrate the decomposition; the contamination magnitudes are consequences of the model, not fitted parameters disguised as findings. The empirical application in Section 6 uses real data and compares exposure-mapping specifications; it does not use the theoretical decomposition to construct the outcome. There are no load-bearing self-citations: the authors cite prior work such as Abadie et al. (2020) and Goldsmith-Pinkham et al. (2022) for context and comparison, but the present claims rest on the paper's own moment conditions and proof. The linearity assumption in Assumption 2 is a substantive functional-form restriction for continuous exposure mappings, as the paper itself acknowledges by noting it is without loss only when |T_n| is finite; this is a limitation or a potential threat to external validity, not circularity. Overall, the derivation chain is self-contained and no step reduces by construction to its own inputs.

Assumptions & free parameters 1 free parameters · 7 assumptions · 0 invented entities

The central contamination result rests on the linear potential outcome model and the linear propensity score assumption. The inference results additionally rest on i.i.d. Bernoulli sampling, local dependence, and sparsity. No new physical entities are introduced. The only hand-chosen number is the HAC bandwidth K.

free parameters (1)
  • HAC bandwidth K = 2 in the empirical application (Section 6); user-specified in general
    The theory assumes K is known from the exposure mapping's neighborhood range (Assumption 5). In practice it is user-chosen; the application sets K = 2 because the exposure mapping includes second-neighborhood terms. If the true exposure mapping has longer range, K must be larger to avoid undercoverage.
assumptions (7)
  • domain assumption Assumption 2: potential outcomes are linear in the true exposure mapping, Y*_{n,i}(t) = t' theta_{n,i} + nu_{n,i}.
    Needed for Theorem 1 and Corollary 1 to express OLS estimands as weighted sums of theta_{n,i}. Without loss when the exposure mapping is finite, but substantive for continuous exposure mappings.
  • domain assumption Assumption 4: linear propensity score, E[T_{n,i} | R_n] = L_n Z_{n,i} and E[eT_{n,i} | R_n] = eL_n eZ_{n,i}.
    Removes omitted variable bias and is key to identifying the estimands as weighted averages of individual treatment effects. Can be satisfied by including conditional expectations of the exposure mapping as covariates.
  • domain assumption Assumption 1(i): unit sampling indicators R_{n,i} are i.i.d. Bernoulli(rho_n).
    Rules out cluster and multi-wave sampling; used in Lemmas 4 and 5 and throughout the asymptotic theory. The paper acknowledges this restriction in Remark 1.
  • domain assumption Assumption 7: no misspecification and no mismeasurement for sampled units, plus structural restrictions on covariates and direct treatment terms.
    Needed for consistency of OLS for the population-level estimand (Theorem 3) and for the unconditional variance results in Theorems 6 and 7. The paper notes it may fail for higher-order exposure mappings.
  • domain assumption Assumption 5: exposure mapping and misspecified exposure mapping depend only on treatment within a K-neighborhood.
    Imposes local dependence, which drives the psi-dependence structure, the asymptotic normality results, and the 2K HAC window.
  • domain assumption Assumptions 6, 8, 9, and 10: network sparsity and summability conditions on neighborhoods.
    Standard network-dependent asymptotics conditions; required for the CLT, variance estimator consistency, and the eigenvalue-modified HAC estimator.
  • standard math Assumption 3: uniform boundedness and full-rank variation of outcomes, exposure mappings, and covariates.
    Regularity conditions ensuring estimands and inverses exist and that moments are controlled; standard in design-based inference.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Design-Based and Network Sampling-Based Uncertainties in Network Experiments." pith.science (2026). https://pith.science/paper/O7JQTETX

@misc{pith2026250622989,
  author       = {Pith},
  title        = {Pith review of: Design-Based and Network Sampling-Based Uncertainties in Network Experiments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O7JQTETX}},
  note         = {Machine review of arXiv:2506.22989}
}
read the original abstract

Ordinary least squares (OLS) estimators are widely used in network experiments to estimate spillover effects. We study the causal interpretation of, and inference for the OLS estimator under both design-based uncertainty from random treatment assignment and sampling-based uncertainty in network links. We show that correlations among regressors that capture the exposure to neighbors' treatments can induce contamination bias, preventing OLS from aggregating heterogeneous spillover effects for a clear causal interpretation. We derive the OLS estimator's asymptotic distribution and propose a network-robust variance estimator. Simulations and an empirical application demonstrate that contamination bias can be substantial, leading to inflated spillover estimates.

Figures

Figures reproduced from arXiv: 2506.22989 by the authors.

Figure 1
Figure 1. Comparison of induced subgraph sampling (left) and star sampling (right). (a) Induced subgraph sampling (b) Star sampling Note: Blue nodes indicate sampled units, while light gray nodes denote non-sampled units. Solid black links are observable to the researcher; dashed gray links are unobserved. We denote the observed covariates by Zen,i, which may differ from Zn,i due to network sampling. For example, if Zn,i incl… view at source ↗
Figure 3
Figure 3. Networks with triangle links i i1 i2 (a) Without censored links i i1 i2 (b) With censored (dashed) links Example 11. Consider the setup in Example 10 but with censoring caused by naming up to four friends. As illustrated in Figure 3b, suppose that the sampled network link between i1 and i2 is not observed due to the censoring. Then, i2 is misclassified as a second neighborhood friend in the observed network while i2… view at source ↗
Figure 4
Figure 4. Flowchart for Valid Inference with Linear Regression Start Covariates satisfy Assumption 4? Regressors satisfy the no correlation condition in Corollary 2? Exposure mapping correctly specified and relevant network information observed (Assumption 7(i))? Network HAC estimator correctly applied? Valid inference: con￾fidence interval >95% asymptotically Include required covariates Modify the exposure mapping or flag re… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 74 canonical work pages

  1. [1]

    Abadie, A., Athey, S., Imbens, G. W. and Wooldridge, J. M. (2020) Sampling-based versus design-based uncertainty in regression analysis, Econometrica, 88, 265--296

  2. [2]

    Abadie, A., Athey, S., Imbens, G. W. and Wooldridge, J. M. (2023) When should you adjust standard errors for clustering?, The Quarterly Journal of Economics, 138, 1--35

  3. [3]

    Angrist, J. D. (1998) Estimating the labor market impact of voluntary military service using social security data on military applicants, Econometrica, 66, 249--288

  4. [4]

    Aronow, P. M. and Samii, C. (2017) Estimating average causal effects under general interference, with application to a social network experiment, Annals of Applied Statistics, 11, 1912--1947

  5. [5]

    G., Duflo, E

    Banerjee, A., Chandrasekhar, A. G., Duflo, E. and Jackson, M. O. (2013) The diffusion of microfinance, Science, 341, 1236498

  6. [6]

    and Mobarak, A

    Beaman, L., BenYishay, A., Magruder, J. and Mobarak, A. M. (2021) Can network theory-based targeting increase technology adoption?, American Economic Review, 111, 1918--1943

  7. [7]

    and Hull, P

    Borusyak, K. and Hull, P. (2023) Nonrandom exposure to exogenous shocks, Econometrica, 91, 2155--2185

  8. [8]

    and Hull, P

    Borusyak, K. and Hull, P. (2024) Negative weights are no concern in design-based specifications, AEA Papers and Proceedings, 114, 597--600

Show all 76 references
  1. [9]

    and Sadoulet, E

    Cai, J., de Janvry , A. and Sadoulet, E. (2015) Social networks and the decision to insure, American Economic Journal: Applied Economics, 7, 81--108

  2. [10]

    and Yang, D

    Carter, M., Laajaj, R. and Yang, D. (2021) Subsidies and the African green revolution: Direct effects and social network spillovers of randomized input subsidies in Mozambique , American Economic Journal: Applied Economics, 13, 206--29

  3. [11]

    and Lewis, R

    Chandrasekhar, A. and Lewis, R. (2016) Econometrics of sampled networks, Unpublished manuscript

  4. [12]

    Conley, T. G. and Udry, C. R. (2010) Learning about a new technology: Pineapple in Ghana , American Economic Review, 100, 35--69

  5. [13]

    and Jones, K

    Dizon, F., Gong, E. and Jones, K. (2020) The effect of promoting savings on informal risk sharing, Journal of Human Resources, 55, 963--998

  6. [14]

    Fernando, A. N. (2021) Seeking the treated: The impact of mobile extension on farmer information exchange in India , Journal of Development Economics, 153, 102713

  7. [15]

    and Ding, P

    Gao, M. and Ding, P. (2023) Causal inference in network experiments: regression-based analysis and design-based properties, arXiv preprint arXiv:2309.07476

  8. [16]

    and Koles \'a r, M

    Goldsmith-Pinkham, P., Hull, P. and Koles \'a r, M. (2022) Contamination bias in linear regressions, Tech. rep., National Bureau of Economic Research

  9. [17]

    (2022) Name your friends, but only five? the importance of censoring in peer effects estimates using social network data, Journal of Labor Economics, 40, 779--805

    Griffith, A. (2022) Name your friends, but only five? the importance of censoring in peer effects estimates using social network data, Journal of Labor Economics, 40, 779--805

  10. [18]

    and Yanagi, T

    Hoshino, T. and Yanagi, T. (2024) Causal inference with noncompliance and unknown interference, Journal of the American Statistical Association, 119, 2869--2880

  11. [19]

    I., Kovarik, J

    Hsieh, C.-S., Hsu, Y.-C., Ko, S. I., Kovarik, J. and Logan, T. D. (2024) Non-representative sampled networks: Estimation of network structural properties by weighting, Journal of Econometrics, 240, 105689

  12. [20]

    and Song, K

    Kojevnikov, D., Marmer, V. and Song, K. (2021) Limit theorems for network dependent random variables, Journal of Econometrics, 222, 882--908

  13. [21]

    Kolaczyk, E. D. and Cs \'a rdi, G. (2014) Statistical analysis of network data with R, vol. 65, Springer

  14. [22]

    Leung, M. P. (2020) Treatment and spillover effects under network interference, Review of Economics and Statistics, 102, 368--380

  15. [23]

    Leung, M. P. (2022) Causal inference under approximate neighborhood interference, Econometrica, 90, 267--293

  16. [24]

    and Tang, X

    Lewbel, A., Qu, X. and Tang, X. (2023) Ignoring measurement errors in social networks, The Econometrics Journal, p. utad028

  17. [25]

    and Kremer, M

    Miguel, E. and Kremer, M. (2004) Worms: Identifying impacts on education and health in the presence of treatment externalities, Econometrica, 72, 159--217

  18. [26]

    (2024) Policy targeting under network interference, Review of Economic Studies, p

    Viviano, D. (2024) Policy targeting under network interference, Review of Economic Studies, p. rdae041

  19. [27]

    and Wooldridge, J

    Xu, R. and Wooldridge, J. M. (2022) A design-based approach to spatial correlation, arXiv preprint arXiv:2211.14354

  20. [28]

    and Piraino, P

    Abel, M., Burger, R. and Piraino, P. (2020) The value of reference letters: Experimental evidence from South Africa , American Economic Journal: Applied Economics, 12, 40--71

  21. [29]

    Afridi, F., Dhillon, A., Li, S. X. and Sharma, S. (2020) Using social connections and financial incentives to solve coordination failure: A quasi-field experiment in India's manufacturing sector, Journal of Development Economics, 144

  22. [30]

    and Sutter, M

    Alan, S., Corekcioglu, G. and Sutter, M. (2022) Improving workplace climate in large corporations: A clustered randomized intervention, Quarterly Journal of Economics, 138, 151--203

  23. [31]

    G., Hanna, R

    Alatas, V., Banerjee, A., Chandrasekhar, A. G., Hanna, R. and Olken, B. A. (2016) Network structure and the aggregation of information: Theory and evidence from Indonesia , American Economic Review, 106, 1663--1704

  24. [32]

    G., Duflo, E., Jackson, M

    Banerjee, A., Breza, E., Chandrasekhar, A. G., Duflo, E., Jackson, M. O. and Kinnan, C. (2023) Changes in social network structure in response to exposure to formal credit markets, Review of Economic Studies

  25. [33]

    Banerjee, A., Breza, E., Chandrasekhar, A. G. and Golub, B. (2024) When less is more: Experimental evidence on information delivery during India's demonetisation, Review of Economic Studies, 91, 1884--1922

  26. [34]

    G., Duflo, E

    Banerjee, A., Chandrasekhar, A. G., Duflo, E. and Jackson, M. O. (2019) Using gossips to spread information: Theory and evidence from two randomized controlled trials, Review of Economic Studies, 86, 2453--2490

  27. [35]

    and Pande, R

    Barnhardt, S., Field, E. and Pande, R. (2017) Moving to opportunity or isolation? network effects of a randomized housing lottery in urban I ndia, American Economic Journal: Applied Economics, 9, 1--32

  28. [36]

    and Yang, D

    Barsbai, T., Licuanan, V., Steinmayr, A., Tiongson, E. and Yang, D. (2024) Information and immigrant settlement, Journal of Development Economics, 170

  29. [37]

    M., Freitas, P., Lima, G

    Batista, C., Costa, D. M., Freitas, P., Lima, G. and Reis, A. B. (2025) What matters for the decision to study abroad? a lab-in-the-field experiment in Cape Verde , Journal of Development Economics, 173

  30. [38]

    and Hansen, C

    Belloni, A., Chernozhukov, V., Fernandez-Val, I. and Hansen, C. (2017) Program evaluation and causal inference with high-dimensional data, Econometrica, 85, 233--298

  31. [39]

    and Mobarak, A

    BenYishay, A., Jones, M., Kondylis, F. and Mobarak, A. M. (2020) Gender gaps in technology diffusion, Journal of Development Economics, 143

  32. [40]

    and Mobarak, A

    Benyishay, A. and Mobarak, A. M. (2019) Social learning and incentives for experimentation and communication, Review of Economic Studies, 86, 976--1009

  33. [41]

    and Karaca, B

    Bjorkegren, D. and Karaca, B. C. (2022) Network adoption subsidies: A digital evaluation of a rural mobile phone program in Rwanda , Journal of Development Economics, 154

  34. [42]

    J., Gertler, P

    Bobonis, G. J., Gertler, P. J., Gonzalez-Navarro, M. and Nichter, S. (2022) Vulnerability and clientelism, American Economic Review, 112, 3627--3659

  35. [43]

    and Villar, P

    Boltz, M., Marazyan, K. and Villar, P. (2019) Income hiding and informal redistribution: A lab-in-the-field experiment in Senegal , Journal of Development Economics, 137, 78--92

  36. [44]

    and Chandrasekhar, A

    Breza, E. and Chandrasekhar, A. G. (2019) Social networks, reputation, and commitment: Evidence from a savings monitors experiment, Econometrica, 87, 175--216

  37. [45]

    G., McCormick, T

    Breza, E., Chandrasekhar, A. G., McCormick, T. H. and Pan, M. (2020) Using aggregated relational data to feasibly identify network structure without network data, American Economic Review, 110, 2454--2484

  38. [46]

    and Sadoulet, E

    Cai, J., De Janvry, A. and Sadoulet, E. (2015) Social networks and the decision to insure, American Economic Journal: Applied Economics, 7, 81--108

  39. [47]

    and Szeidl, A

    Cai, J. and Szeidl, A. (2018) Interfirm relationships and business performance, Quarterly Journal of Economics, 133, 1229--1282

  40. [48]

    Cai, S. (2020) Migration under liquidity constraints: Evidence from randomized credit access in China , Journal of Development Economics, 142, 10th AFD-World-Bank-Development Conference, CERDI, Clermont Ferrand, FRANCE, JUN 30-JUL 01, 2017

  41. [49]

    and Long, J

    Callen, M. and Long, J. D. (2015) Institutional corruption and election fraud: Evidence from a field experiment in Afghanistan , American Economic Review, 105, 354--381

  42. [50]

    and Yang, D

    Carter, M., Laajaj, R. and Yang, D. (2021) Subsidies and the African green revolution: Direct effects and social network spillovers of randomized input subsidies in Mozambique , American Economic Journal: Applied Economics, 13, 206--229

  43. [51]

    G., Kinnan, C

    Chandrasekhar, A. G., Kinnan, C. and Larreguy, H. (2018) Social networks as contract enforcement: Evidence from a lab experiment in the field, American Economic Journal: Applied Economics, 10, 43--78

  44. [52]

    J., Falco, P

    Chegere, M. J., Falco, P. and Menzel, A. (2024) Social ties at work and effort choice: Experimental evidence from Tanzania , Journal of Development Economics, 171

  45. [53]

    and Liu, E

    Colonnelli, E., Li, B. and Liu, E. (2024) Investing with the government: A field experiment in China , Journal of Political Economy, 132, 248--294

  46. [54]

    (2022) Identifying communication spillovers in lab-in-the-field experiments, Journal of Development Economics, 157

    Coutts, A. (2022) Identifying communication spillovers in lab-in-the-field experiments, Journal of Development Economics, 157

  47. [55]

    and Querubin, P

    Cruz, C., Labonne, J. and Querubin, P. (2017) Politician family networks and electoral outcomes: Evidence from the Philippines , American Economic Review, 107, 3006--3037

  48. [56]

    W., Lipscomb, M., Schechter, L

    Deutschmann, J. W., Lipscomb, M., Schechter, L. and Zhu, J. (2024) Spillovers without social interactions in urban sanitation, American Economic Journal: Applied Economics, 16, 482--515

  49. [57]

    and Vollenweider, X

    Di Falco, S., Feri, F., Pin, P. and Vollenweider, X. (2018) Ties that bind: Network redistributive pressure and economic decisions in village economies, Journal of Development Economics, 131, 123--131

  50. [58]

    and Traxler, C

    Drago, F., Mengel, F. and Traxler, C. (2020) Compliance behavior in networks: Evidence from a field experiment, American Economic Journal: Applied Economics, 12, 96--133

  51. [59]

    and Mill, W

    Evsyukova, Y., Rusche, F. and Mill, W. (2024) Linkedout? a field experiment on discrimination in job network formation, Quarterly Journal of Economics, 140, 283--334

  52. [60]

    and Vicente, P

    Fafchamps, M. and Vicente, P. C. (2013) Political violence and social networks: Experimental evidence from a Nigerian election, Journal of Development Economics, 101, 27--48

  53. [61]

    Flory, J. A. (2018) Formal finance and informal safety nets of the poor: Evidence from a savings field experiment, Journal of Development Economics, 135, 517--533

  54. [62]

    and Mansuri, G

    Gine, X. and Mansuri, G. (2018) Together we will: Experimental evidence on female voting behavior in Pakistan , American Economic Journal: Applied Economics, 10, 207--235

  55. [63]

    and Thornton, R

    Godlonton, S. and Thornton, R. (2012) Peer effects in learning HIV results, Journal of Development Economics, 97, 118--129

  56. [64]

    and McCasland, J

    Hardy, M. and McCasland, J. (2021) It takes two: Experimental evidence on the determinants of technology diffusion, Journal of Development Economics, 149

  57. [65]

    M., Namen, O., Naslund-Hadley, E

    Hernandez-Agramonte, J. M., Namen, O., Naslund-Hadley, E. and Biehl, M. L. (2024) Supporting early childhood development remotely: Experimental evidence from SMS messages, Journal of Development Economics, 166

  58. [66]

    and Schuendeln, M

    Hess, S., Jaimovich, D. and Schuendeln, M. (2021) Development projects and economic networks: Lessons from rural Gambia , Review of Economic Studies, 88, 1347--1384

  59. [67]

    and Ozier, O

    Jakiela, P. and Ozier, O. (2016) Does Africa need a rotten kin theorem? experimental evidence from village economies, Review of Economic Studies, 83, 231--268

  60. [68]

    and Baylis, K

    Kandpal, E. and Baylis, K. (2019) The social lives of married women: Peer effects in female autonomy and investments in children, Journal of Development Economics, 140, 26--43

  61. [69]

    (2024) Social signaling and childhood immunization: A field experiment in Sierra Leone , Quarterly Journal of Economics, 139, 2083--2133

    Karing, A. (2024) Social signaling and childhood immunization: A field experiment in Sierra Leone , Quarterly Journal of Economics, 139, 2083--2133

  62. [70]

    Kessler, J. B. (2017) Announcements of support and public good provision, American Economic Review, 107, 3760--3787

  63. [71]

    and Ramage, I

    Levine, D., Polimeni, R. and Ramage, I. (2016) Insuring health or insuring wealth? an experimental evaluation of health insurance in rural Cambodia , Journal of Development Economics, 119, 1--15

  64. [72]

    M., Mommaerts, C

    Meghir, C., Mobarak, A. M., Mommaerts, C. and Morten, M. (2022) Migration and informal insurance: Evidence from a randomized controlled trial and a structural model, Review of Economic Studies, 89, 452--480

  65. [73]

    and Sekhri, S

    Nagavarapu, S. and Sekhri, S. (2016) Informal monitoring and enforcement mechanisms in public service delivery: Evidence from the public distribution system in India , Journal of Development Economics, 121, 63--78

  66. [74]

    and Sands, E

    Pallais, A. and Sands, E. G. (2016) Why the referential treatment? evidence from field experiments on referrals, Journal of Political Economy, 124, 1793--1828

  67. [75]

    (2012) Limited insurance within the household: Evidence from a field experiment in Kenya , American Economic Journal: Applied Economics, 4, 140--164

    Robinson, J. (2012) Limited insurance within the household: Evidence from a field experiment in Kenya , American Economic Journal: Applied Economics, 4, 140--164

  68. [76]

    and Romagnoli, G

    Soldani, E., Hildebrandt, N., Nyarko, Y. and Romagnoli, G. (2023) Price information, inter-village networks, and ``bargaining spillovers``: Experimental evidence from Ghana , Journal of Development Economics, 164

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.