{"id":"ae84dd0e-29cd-4735-bedd-4d30fd4091bb","arxiv_id":"2412.02183","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A framework with asymptotic theory separating direct treatment effects from network-mediated indirect effects in randomized experiments where the social network responds to the treatment.","lead":"This econometrics paper shows how to separate the direct effect of a treatment from effects that spread through social networks that the treatment itself rewires, in randomized trials with two waves of network data. It provides formal conditions under which ordinary regressions, shift-share instruments, and a spectral denoised instrument are consistent, and illustrates the machinery on a savings-account experiment in Nepal.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 2.1(4) asserts β2 equals the average spillover effect, but with the paper's own 0/0=0 convention the mediator difference is 1{degree>0}, not 1; in bounded-degree networks SE(t) = β2·P(degree>0).","rationale":"I focused on the spillover equality because it is the narrowest point at which the central causal claim breaks while all maintained assumptions hold. The reader's concern about nonlinearity is valid but external: it requires leaving the paper's model. The isolated-degree issue is internal: under the paper's own 0/0=0 rule and Assumption 1's bounded-degree regime, SE(t) ≠ β2. This invalidates the statement 'β2 captures the spillover effect' and the corresponding empirical magnitudes, though it does not invalidate the consistency theorems for β2 as a regression coefficient or the indirect-effect decomposition IE(t)=β2·ΔE[M]. The error is readily fixable by adding the P(degree>0) factor and adjusting empirical claims, so I keep the reader's conditional verdict rather than escalating to reject. The concrete Monte Carlo check directly computes Definition 1 and settles whether the equality holds.","tokens_in":88331,"tokens_out":9960,"duration_ms":101818,"concrete_test":"Simulate Design 1 with q_pre = q_post = n^{-1}, n=500, β2=0.5, and compute the exact average spillover effect from Definition 1 by Monte-Carlo-integrated counterfactuals M_i(t,1)−M_i(t,0) under the DGP. If the resulting SE(t) equals β2 times the sample fraction of units with positive post-treatment degree rather than β2, Corollary 2.1(4)'s equality is refuted. An equivalent data check: in the Prina sample, compute the mean of 1{Σ_j A^post_ij>0}; the reported 'spillover' estimates in Table 5 should be multiplied by this fraction to match Definition 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Corollary 2.1(4) states that with M_i = (Σ_j A^post_ij T_j)/(Σ_j A^post_ij), E[M_i | T_i=t, T_-i=1_{n-1}] − E[M_i | T_i=t, T_-i=0_{n-1}] = 1, so that β2 equals the spillover effect SE(t). This is false under the paper's own convention 0/0=0 (Section 3.1). For any unit i, the counterfactual mediator is 1 if the unit has at least one post-treatment friend and 0 if it has none; hence the difference equals 1{Σ_j A^post_ij > 0}, not 1. Consequently SE(t) = β2 · P(Σ_j A^post_ij > 0 | T_i=t), not β2. Assumption 1 explicitly allows bounded-degree networks with q_post ≍ n^{-1}, where a non-vanishing fraction of units is isolated; even in the paper's own application, the post-treatment network has ~329 undirected links among 915 households, an average degree of ~0.72, so roughly half the households are isolated and the true spillover effect is about half of the reported β2. The empirical interpretation in Section 6 therefore overstates the average spillover by the factor 1/P(degree>0). This is an internal inconsistency in the causal interpretation of β2, not an external misspecification; it does not depend on nonlinearity or on violations of Assumption 2(d).","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies randomized experiments with post-treatment network interference. It defines direct, indirect, and spillover effects within a linear model whose mediator is the fraction of treated friends, and it develops OLS, shift-share IV, and eigendecomposition-based denoised IV estimators. The main theoretical results are consistency and asymptotic normality theorems under various sparsity regimes for pre- and post-treatment random graph models, supported by Monte Carlo simulations and an empirical reanalysis of Prina (2015).","tokens_in":88572,"tokens_out":6212,"duration_ms":63754,"significance":"If the results are correct, the paper makes a useful contribution by extending shift-share IV ideas to network-mediated causal effects in RCTs with two waves of network data, and by providing a denoising modification that restores consistency in denser networks. The paper ships formal proofs, clearly stated asymptotic regimes, and simulation evidence that track the predicted boundaries, including the SSIV break-down at q_n = n^{-1/5} and the restoration of consistency by the denoised estimator. The interpretation of β2 as the spillover effect is, however, currently compromised by an internal inconsistency related to the 0/0=0 convention for the mediator, and the OVB formulas in Section 3.2 are incorrect as stated. These issues do not necessarily invalidate the estimation machinery for β2 itself, but they affect the causal interpretation and the empirical claims in Section 6, so the paper requires substantial revision before the central claims can be accepted.","major_comments":[{"comment":"The claim that with M_i as in (1), E(M_i | T_i=t, T_-i=1_{n-1}) - E(M_i | T_i=t, T_-i=0_{n-1}) = 1 is false under the paper's stated convention 0/0=0. For a unit with no post-treatment friends, the mediator equals 0 regardless of the treatments of others, so the counterfactual difference is 1{Σ_j A^post_ij > 0}, not 1. Consequently SE(t) = β2 · P(Σ_j A^post_ij > 0 | T_i=t), not β2. This is an internal inconsistency, not an external misspecification: Assumption 1 explicitly allows bounded-degree networks where a non-vanishing fraction of units is isolated, and in the paper's own application with roughly 329 links among 915 households (average degree about 0.72), the true spillover effect is approximately β2 · P(degree>0), which is about half of the reported β2. The empirical interpretations in Section 6, including the statement that assigning others to treatment increases fish consumption by Rs. 252.91, therefore overstate the spillover effect by the factor 1/P(degree>0). The causal interpretation section and the empirical discussion need to be corrected, e.g., by redefining the parameter or by reporting the scaling factor.","section":"Corollary 2.1(4) and the paragraph following it"},{"comment":"The omitted variable bias formulas for the pre-network regression are incorrect. The paper states that in a regression with X_pre = (1, T_i, M_pre_i), the coefficient on T_i is β1 + β2 · Cov(T_i, M_i)/Var(T_i) = ToE. That formula applies only when M_i is omitted from the regression entirely. When M_pre_i is included as a regressor, the coefficient on T_i is β1 plus β2 times the partial regression coefficient of M_i on T_i given M_pre_i, which is generally not Cov(T_i, M_i)/Var(T_i) and is not equal to ToE; similarly, the coefficient on M_pre_i is not β2 · Cov(M_pre_i, M_i)/Var(M_pre_i) unless T_i and M_pre_i are uncorrelated. A concrete counterexample is M_pre_i = M_i, in which case including M_pre_i yields β_pre_1 = β1, not ToE. The claims in this section about what pre-network regressions recover should be corrected or qualified.","section":"Section 3.2"}],"minor_comments":[{"comment":"The text says 'I use Designs 1 and 2 to represent Case (a) with non-degenerate ξi, and Designs 3 and 4 to represent Case (a) with constant ξi,' but Section 3.1 defines Case (a) as Var(ξ_i)>0 and Case (b) as Var(ξ_i)=0. The intended reference for Designs 3 and 4 appears to be Case (b); please correct this labeling to avoid confusion.","section":"Section 5.1, text after Table 1"},{"comment":"In the definition of the IV variance estimator ˆV_iv_num, the (2,3) element is written as π(1-π) Σ_i Σ_j A^pre_ij ˆu^iv_j ˆu^iv_j; the second residual index should be i, i.e., A^pre_ij ˆu^iv_i ˆu^iv_j, matching the population quantity in (13).","section":"Equation (14)"},{"comment":"The condition q_pre ≻ log(n)/log(log(n))/n is written ambiguously; it should be clarified as q_pre ≻ log(n)/(n log log n) or with explicit parentheses so that the intended sparsity threshold is unambiguous.","section":"Notation, Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is suitable for the journal's scope in econometrics and has solid asymptotic machinery, but the causal interpretation of β2 is invalid in its current form because of the 0/0 convention, and this directly affects the headline empirical numbers in Section 6. The OVB formulas in Section 3.2 are also incorrect and need to be fixed. I would be willing to look at a revised version once these issues are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know upfront. This is a serious, technically strong paper on estimating direct and network-mediated effects in RCTs with two waves of network data, and its main estimation results look right. But it has an internal inconsistency in the causal interpretation of β2, the coefficient on the fraction of treated friends, and the empirical section overstates a few things.\n\nThe genuinely new material is the asymptotic characterization: OLS consistency under two regimes (non-degenerate ξi and degenerate ξi with q_post not too dense), the shift-share IV relevance threshold at max{q_pre, q_post} ≺ n^{−1/2}, the failure of SSIV in denser networks, and the denoised SSIV that restores consistency under a low-rank graphon assumption. The simulations track the claimed boundaries, and the paper is honest about its own limits—Assumption 2(d) cannot be validated, the modified estimator fails when both networks are dense (Remark 4.6), and panel network data is required. Credit is earned there.\n\nThe load-bearing soft spot is Corollary 2.1(4). The paper claims that with the fraction mediator and its 0/0=0 convention, E[M_i | T_i=t, T_{−i}=1] − E[M_i | T_i=t, T_{−i}=0] = 1, so β2 equals the spillover effect. But under 0/0=0, M_i is 1 under all-others-treated only when i has at least one post-treatment friend; under all-others-control, M_i is 0. So the difference is 1{degree>0}, not 1. The correct statement is SE(t) = β2 · P(degree>0 | T_i=t). In bounded-degree networks, including the Prina application (average degree ≈ 0.72, so roughly half of households are isolated), the reported spillover effect is overstated by about a factor of two. This is an internal inconsistency in the causal interpretation, not an external robustness concern.\n\nOther soft spots are real but less severe. Table 5 claims SSIV and normalized SSIV point estimates are 'closely aligned' when they clearly are not—e.g., fish expenditure: 252.9 vs 70.1, total expenditure: 5726 vs 2128. The rank r for the denoised SSIV is unspecified; applied readers need guidance for choosing it. Pooling 19 villages into a single network deserves a robustness check. The linearity and additivity in Assumption 3 is strong, but the paper acknowledges that the causal parameters flow through it.\n\nWho gets value: applied researchers with two-wave network data from an RCT, and econometricians working on network interference or shift-share relevance in graphon settings. It deserves a serious referee. The referee should require a fix to Corollary 2.1(4) and the interpretation of β2 as a spillover parameter, plus a more careful empirical discussion. I would not cite it in its current form.","headline":"Serious, technically strong paper with a real internal inconsistency in the spillover interpretation of β2 due to the 0/0=0 convention, plus some overstated empirical claims, but the estimation framework is new and deserves a serious referee.","tokens_in":89198,"tokens_out":5757,"would_cite":false,"duration_ms":52759,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62F12","62P20","91D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that in randomized experiments where treatment rewires social networks, direct and network-mediated causal effects can be separated and consistently estimated using two waves of network data, with shift-share instruments…","keywords":["causal inference","interference","network mediation","shift-share instrument","peer effects","random graphs","endogeneity","treatment-induced network change"],"falsifier":"Run the same estimation on data generated from a nonlinear response model, for example $Y_i=\\beta_0+\\beta_1T_i+\\beta_2M_i+\\beta_3M_i^2+\\lambda(w_i)+\\varepsilon_i$, or in real data add $M_i^2$ and $T_iM_i$ to the fitted regression; if the added coefficients are significantly nonzero in large samples, the linear decomposition of Corollary 2.1 fails and the claimed $\\beta_2$ is not the indirect effect.","tokens_in":88021,"feed_emoji":"📊","tokens_out":6305,"duration_ms":62924,"temperature":0.7,"pith_summary":"This paper claims that in a randomized experiment where the treatment can rewire the social network, the total treatment effect can be decomposed into a direct effect and an indirect effect that runs through the fraction of treated friends, and that both parts can be consistently estimated from two waves of network data. The key difficulty is that the post-treatment network is endogenous: latent traits shape both who is friends with whom and the outcome, and the treatment itself changes the network. The paper shows that ordinary least squares is consistent when no unobserved confounder is present, and that a shift-share instrumental variable built from the pre-treatment network and the random treatments of others is consistent when the pre- and post-treatment networks are sufficiently sparse. For denser networks, an eigendecomposition-based 'denoised' version of the instrument restores consistency. If correct, the results give applied researchers a concrete recipe for separating direct from network-mediated causal effects instead of estimating only the total intention-to-treat effect.","feed_headline":"Two network waves separate direct from mediated treatment effects","feed_subtitle":"New estimators use pre-treatment friendship structure as an instrument, with a denoising fix for dense networks.","key_machinery":"The central object is the 'fraction of treated friends' mediator $M_i=\\sum_j A^{post}_{ij}T_j/\\sum_j A^{post}_{ij}$, which the paper decomposes as $M_i=\\xi_i+r_i^*$, where $\\xi_i$ is the conditional probability that a connected friend is treated given $i$'s latent trait and treatment, and $r_i^*$ is a remainder. The argument works through this decomposition: $\\xi_i$ is i.i.d.-like but uncorrelated with the shift-share instrument, so it contributes only noise to the first stage; the signal comes from $r_i^*$. The shift-share instrument $Z_i=\\sum_j A^{pre}_{ij}(T_j-\\pi)$ combines random treatment shocks $(T_j-\\pi)$ with non-exogenous pre-treatment exposure weights $A^{pre}_{ij}$, and its relevance decays as networks densify. The denoised instrument removes the component of $Z_i$ along the leading eigenvectors of $A^{pre}$, which carry the latent variable information, thereby shrinking the noise term while preserving the signal.","core_discovery":"Under a linear potential-outcome model $Y_i(t_i,m_i)=\\beta_0+\\beta_1 t_i+\\beta_2 m_i+\\lambda(w_i)+\\varepsilon_i$, where $m_i$ is the fraction of treated friends measured in the post-treatment network, the paper establishes that $\\beta_1$ is the direct effect of treatment and $\\beta_2$ is the coefficient driving the indirect and spillover effects. With random treatment assignment and the assumption that unobserved covariates $w_i$ are the only source of confounding between the network mediator and the outcome, the paper proves consistency and asymptotic normality of OLS estimators when endogeneity is absent. When confounding is present, the paper constructs the shift-share instrument $Z_i=\\sum_j A^{pre}_{ij}(T_j-\\pi)$ and shows the corresponding IV estimator is consistent when $\\max\\{q_{pre},q_{post}\\}=o(n^{-1/2})$; when the network is too dense for this to hold, projecting the instrument onto the leading eigenvectors of the pre-treatment adjacency matrix removes the noise that kills relevance and restores consistency. The empirical application to a savings-account experiment in Nepal illustrates that the direct and network-mediated channels can have different signs and significance.","pith_inferences":["The same noise-versus-signal mechanism likely threatens any instrument built from pre-treatment network structure in dense networks, including peer-of-peer instruments, so the sparsity threshold may offer insight beyond shift-share designs.","The linear and additive response in Assumption 3 is doing heavy lifting; if true effects are nonlinear in the fraction of treated friends or the confounder interacts with treatment, the estimated $\\beta_2$ is not the indirect effect, so applied work should report robustness checks that add a quadratic term in $M_i$ or an interaction $T_iM_i$.","The denoising recipe suggests a general empirical strategy: before using any network-share instrument in a dense network, regress the instrument on leading eigenvectors of the adjacency matrix and use the residual.","When only one wave of network data is available, the identification strategy fails entirely; the paper's reliance on two waves motivates further work on recovering pre-treatment shares from aggregated relational data."],"forward_implications":["Researchers with pre- and post-treatment network data can recover the direct effect of an intervention separately from the effect mediated by network rewiring, rather than only the total effect.","Using only the pre-treatment network to measure peer exposure recovers the total effect but not the direct/indirect decomposition; post-treatment networks are needed for the mediation channel.","The shift-share IV is reliable only in relatively sparse networks ($\\max\\{q_{pre},q_{post}\\}=o(n^{-1/2})$); denser networks require the denoised eigenvector version.","Standard heteroskedasticity-consistent variance estimators are valid for OLS, and the paper provides variance estimators for the IV versions that account for cross-unit dependence induced by the instrument.","Treatment-induced changes in the network can increase the variation of the mediator and therefore improve convergence rates relative to a fixed network."],"supporting_citations":[{"why":"Supplies the shift-share instrument construction that combines exogenous shocks with non-exogenous exposure weights, which the paper adapts to the network setting.","marker":"Borusyak and Hull (2023)"},{"why":"Provides the random-graph asymptotics and the principal-component balancing idea that the paper uses for its eigendecomposition-based denoised SSIV.","marker":"Li and Wager (2022)"},{"why":"Establishes the anonymous interference framework under which the exposure mapping used as the mediator is defined.","marker":"Hudgens and Halloran (2008)"},{"why":"Justifies the fraction-of-treated-friends form as an anonymous interaction assumption and connects it to the linear-in-means model.","marker":"Manski (2013)"},{"why":"Provides the classical linear mediation formulas that give $\\beta_1$ and $\\beta_2$ their causal interpretations as direct and indirect effects.","marker":"Baron and Kenny (1986)"},{"why":"Supplies the treatment-effect-with-network-changes setting and the panel network data used in the paper's empirical illustration.","marker":"Comola and Prina (2021)"}],"fun_headline_variants":["Direct vs friend-mediated: new estimators for network experiments","Shift-share instruments separate direct and mediated effects","Denoising dense networks restores treatment effect estimates","Endogenous interference: direct vs peer-mediated effects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire causal interpretation rests on the potential outcome being exactly linear and additive in the fraction of treated friends, together with a cross-world independence condition that the paper itself says can never be validated; if the true response is nonlinear or the unobserved confounder interacts with treatment, $\\beta_2$ no longer measures the indirect or spillover effect.","fun_headline_variants_meta":{"raw":{"variants":["Direct vs friend-mediated: new estimators for network experiments","Shift-share instruments separate direct and mediated effects","Denoising dense networks restores treatment effect estimates","Endogenous interference: direct vs peer-mediated effects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000961,"raw_usage":{"total_tokens":4113,"prompt_tokens":988,"completion_tokens":3125,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":3064}},"tokens_in":604,"tokens_out":3125,"duration_ms":21514,"temperature":1.0,"reasoning_tokens":3064,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:47:25.494740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same estimation on data generated from a nonlinear response model, for example $Y_i=\\beta_0+\\beta_1T_i+\\beta_2M_i+\\beta_3M_i^2+\\lambda(w_i)+\\varepsilon_i$, or in real data add $M_i^2$ and $T_iM_i$ to the fitted regression; if the added coefficients are significantly nonzero in large samples, the linear decomposition of Corollary 2.1 fails and the claimed $\\beta_2$ is not the indirect effect.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the anonymous interference framework under which the exposure mapping used as the mediator is defined."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the treatment-effect-with-network-changes setting and the panel network data used in the paper's empirical illustration."}],"review_version":1}