{"id":"44dce544-2483-4ad1-adc3-5487220967f1","arxiv_id":"2511.21534","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A weighting-based bias decomposition lets practitioners bound the combined bias from ignored interference, unmeasured confounding, and non-transportability using interpretable sensitivity parameters.","lead":"This paper develops a sensitivity-analysis framework to quantify how much bias remains when a causal estimate ignores interference between units, unmeasured confounding, and differences between study populations. It provides formulas that turn a researcher's beliefs about these problems into upper bounds on the bias.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's sensitivity parameter β is used to bound σ̃_υ(g) via Popoviciu, but under its stated marginal definition the bound is false; σ̃_υ(g) can be 1 when β=0.","rationale":"The reader’s verdict is CONDITIONAL, and I agree that a condition is needed, but the most load-bearing issue is not the Assumption 2 A→A_N exclusion (which the paper explicitly acknowledges in Remark 4) or the absence of simulations. It is a concrete technical flaw in the sensitivity parameterization of Theorem 2. The authors state a bound on σ̃_υ(g) that does not follow from Popoviciu’s inequality under the literal marginal definition of β. The bias decomposition theorem itself may be correct, but the proposed sensitivity framework for transportability relies on this bound. If β is intended as a conditionally uniform bound, the concern is resolved by clarification. The counterexample is simple and would settle the dispute. Therefore I recommend CONDITIONAL rather than REJECT, because the core theorem can be fixed by redefining/clarifying β and the rest of the framework remains intact.","tokens_in":19170,"tokens_out":17675,"duration_ms":158099,"concrete_test":"Construct the counterexample above: S∈{1,2} equally likely, X∼Bernoulli(0.5) independent of S, and conditional probabilities p(G=1|S=1,X=0)=0, p(G=1|S=1,X=1)=1, p(G=1|S=2,X=0)=1, p(G=1|S=2,X=1)=0. Compute p(G=1|S=1)=p(G=1|S=2)=0.5 (so β=0), then compute υ(1) and σ̃_υ(1); the result is σ̃_υ(1)=1, contradicting σ̃_υ(g)≤β. Also compute T3 for any τ(1) that varies with X to show the covariance term is nonzero even when β=0. If the authors intended β as a pointwise bound on the conditional difference, the text must say so explicitly; the current wording does not.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section 4, after Theorem 2, the authors propose a sensitivity parameter β∈[0,1] such that 'for each value of g, p(G=g|S=1)−p(G=g|S=2)∈[−β, β]', and then assert that this 'can also be used to create an upper bound for the standard deviation of υ(g) through Popoviciu’s inequality: σ̃_υ(g) ≤ β.' This is not a consequence of Popoviciu’s inequality. Popoviciu’s inequality bounds the variance of a random variable by the squared half-range of its support; it does not bound the variance by the absolute value of its mean. Under the literal (marginal) reading, β bounds only the scalar difference of unconditional probabilities, while υ(g)=p(G=g|S=1,X_GY,U_GY)−p(G=g|S=2,X_GY,U_GY) is a random variable whose conditional values can vary widely even when its mean is 0. For example, let X be Bernoulli(0.5) independent of S, and set p(G=1|S=1,X=0)=0, p(G=1|S=1,X=1)=1, p(G=1|S=2,X=0)=1, p(G=1|S=2,X=1)=0. Then p(G=1|S=1)=p(G=1|S=2)=0.5, so β=0, but υ(1)∈{−1,1} and σ̃_υ(1)=1. Thus the claimed bound fails, and any transportability sensitivity bound that relies on σ̃_υ(g)≤β is invalid unless β is redefined as a pointwise bound on the conditional difference. The paper should clarify this definition; as written, Theorem 2’s T3 sensitivity parameterization is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sensitivity analysis framework for causal inference under interference, focusing on the bias of a naive inverse-probability-weighting (IPW) estimator that ignores interference and unmeasured confounding. The main result (Theorem 1) decomposes the bias into a sum of covariances between multiplicative error-in-weights (MEW) scores ε_a and baseline potential outcomes Y(a,0) and natural spillover effects γ(a). Theorem 2 extends the decomposition to transportability, adding a term T3 that captures differences between the reference and target populations. The framework is parameterized by interpretable sensitivity parameters (η_baseline, η_ε, η_γ, β, ρ terms). Section 5 extends the results to settings with undefined potential outcomes. The paper is largely theoretical, with proofs in the appendix and no simulation or empirical illustration.","tokens_in":19719,"tokens_out":7803,"duration_ms":72199,"significance":"If correct, this is a useful contribution to the sensitivity-analysis literature for interference. The bias decomposition is exact under the stated assumptions, and the framework allows practitioners to simultaneously assess unmeasured confounding, ignored interference, and lack of transportability without strong parametric assumptions. The MEW score formulation and the explicit covariance decomposition are elegant and generalize earlier weighting-based results (e.g., Shen et al. 2011). The extension to undefined potential outcomes in Section 5 addresses a real practical concern. However, the lack of any numerical illustration leaves the practical behavior of the sensitivity bounds unexamined, and the transportability parameterization contains a technical error that must be fixed before the framework can be used as stated.","major_comments":[{"comment":"The proof uses the independence S ⊥ X_GY,U_GY (to replace p(S=s|X_GY,U_GY) with p(S=s)), but this is not stated in Assumption 6. The bullet points only give Y(a,g) ⊥ A_N | X_GY,U_GY, S has direct causal effect on A and G, and Y(a,g) ⊥ S. The DAG in Figure 1 may imply the needed independence, but the assumptions as written do not. This is load-bearing for the derivation of T3. Please add the condition explicitly to Assumption 6 (or justify it from the existing conditions) and update the proof.","section":"Proof of Theorem 2 (Appendix A)"}],"minor_comments":[{"comment":"The expectation E[τ(g)] in T3 is not clearly defined with respect to which population. The proof shows it is an unconditional expectation over the pooled (S=1 and S=2) distribution. This should be stated explicitly, since the sensitivity parameter for E[τ(g)] would then refer to the pooled population, not the reference or target population.","section":"Section 4, Theorem 2"},{"comment":"Assumption 2 implicitly precludes a direct causal path from A to A_N, as Remark 4 acknowledges. This limits the framework to settings where interference acts only on outcomes, not on treatment adoption. This is a substantive limitation and should be highlighted earlier in the paper, not only in a remark.","section":"Assumption 2 and Remark 4"},{"comment":"The paper does not include any simulation study, data example, or even a small numerical demonstration of how the sensitivity parameters translate into bias bounds. This makes it difficult for a practitioner to calibrate the parameters or see the method in action. At least one illustrative example would strengthen the paper.","section":"Throughout"},{"comment":"The Bhatia-Davis bound in Remark 7 is correct for the case where all values of ζ(g) are known, but it is worth noting that this requires stronger information than the β parameter. The distinction between a pointwise bound and a mean bound should be made clear here, in light of the β issue.","section":"Remark 7"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the bias decomposition for the reference estimand (Theorem 1) is a real contribution: it cleanly extends Shen et al. to interference plus unmeasured confounding, and the split into baseline-outcome and spillover-effect covariances is both new and interpretable. Second, the transportability sensitivity parameterization in Theorem 2 has a load-bearing error: the claim that sigma_tilde_upsilon(g) <= beta follows from Popoviciu is false.\n\nThe reader's take is largely on target. The algebra in Theorems 1 and 2 is correct under the stated assumptions, and the no-interference special case reducing to Shen et al. is a nice sanity check. The sensitivity parameters are intuitive, and the paper is honest about its scope limitations in Remark 4. No simulations are a real gap for a methods paper, but not fatal for the theoretical core.\n\nThe beta issue is the main problem. As written, beta bounds the marginal difference p(G=g|S=1) - p(G=g|S=2). But upsilon(g) is a random variable that can have large conditional variation even when its mean is zero. The stress-test example is correct: if G depends on X within each population but the marginal probabilities match, beta=0 while sigma_tilde_upsilon(g) can be 1. So the proposed bound is simply unsupported, and any sensitivity analysis using it will understate the transportability bias. The fix is straightforward: either define beta as a pointwise bound on |upsilon(g)| or replace the Popoviciu step with Bhatia-Davis using the known mean, as the paper already does in Remark 7. But as written, the T3 sensitivity bounds are not valid.\n\nAssumption 2 precludes A->A_N feedback, which the paper acknowledges. That limits applicability to settings where interference acts only on outcomes, not on treatment adoption. It is a stated limitation, not a hidden one, but it deserves more emphasis than it gets.\n\nThe citation pattern looks honest, and the claimed novelty relative to VanderWeele et al. and Forastiere et al. is reasonable. This paper deserves peer review: the core decomposition is likely correct and useful, and the beta flaw is fixable. A good referee would catch it. I would encourage the editor to send it out rather than desk reject.","headline":"Theorem 1's bias decomposition is a genuine and useful contribution, but the transportability sensitivity bound in Theorem 2 is wrong as stated: sigma_tilde_upsilon(g) <= beta does not follow from Popoviciu and can fail badly.","tokens_in":20147,"tokens_out":2983,"would_cite":true,"duration_ms":32793,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The bias from ignoring interference and unmeasured confounding is a sum of interpretable covariances.","keywords":["causal inference","interference","spillover effects","sensitivity analysis","unmeasured confounding","transportability","inverse probability weighting","potential outcomes"],"falsifier":"Simulate a network where treatment is contagious (A_i affects A_j's treatment) with an unmeasured common cause of A and Y, then compare the empirical bias of the naive IPW estimator against the Theorem 1 expression computed from true potential outcomes. If the difference exceeds the bound implied by the true correlations and standard deviations, the decomposition is incomplete; if it is at most that, the bound is supported.","tokens_in":19087,"feed_emoji":"📊","tokens_out":3956,"duration_ms":40644,"temperature":0.7,"pith_summary":"The paper claims that when a researcher uses a naive inverse-probability-weighting estimator that ignores spillovers and unmeasured confounders, the resulting systematic bias is not an opaque quantity: it decomposes exactly into covariances between the estimator's errors-in-weights and two types of causal effects, baseline outcomes and natural spillover effects. Each covariance can be rewritten as a correlation times two standard deviations, which become sensitivity parameters a practitioner can bound using domain knowledge. A third term arises when transporting the estimate to a new population, capturing the contribution of non-transportability. The decomposition is meant to let researchers gauge how strong interference and confounding would have to be to overturn their conclusions, without specifying a full model of the data.","feed_headline":"Spillover bias is a sum of covariances","feed_subtitle":"New sensitivity framework turns ignored interference and confounding into bounds a researcher can set by hand.","key_machinery":"The machinery is the multiplicative error in weights (MEW) score, eps_a = p(A=a|X,U)/p(A=a|X), which collapses both unmeasured confounding and ignored common causes of A and A_N into one quantity. Alongside an exposure mapping G that summarizes neighbors' treatments, the paper expands Y(a) = Y(a,0) + gamma(a), splitting bias into baseline-outcome covariance and spillover-effect covariance. The transport term uses upsilon(g), the difference between populations in the conditional probability of neighborhood treatment level g.","core_discovery":"The central result (Theorem 1) is the identity Bias = sum over a in {0,1} of (-1)^{1-a}[Cov(Y(a,0), eps_a) + Cov(gamma(a), eps_a)] where eps_a is the ratio of the true propensity score to the pseudo-propensity score used by the naive estimator. Written as correlation-standard-deviation products, the first covariance captures confounding of treatment assignment with baseline outcomes and the second captures the interaction of confounding with heterogeneity in spillover effects. Theorem 2 adds T3 = sum_g [Cov(tau(g), upsilon(g)) + E[tau(g)](p(G=g|S=1)-p(G=g|S=2))] for transporting to a target population. The paper further extends the identity to settings where some potential outcomes are undef","pith_inferences":["The decomposition treats the pseudo-propensity model as correctly specified; if that model is itself misspecified, the MEW score would carry an additional term, so the framework's bounds are lower bounds on total misspecification bias.","Because T3 separates a covariance term from a marginal-imbalance term, the same framework could guide data collection in the target population: the imbalance term is estimable with lighter covariate data than the covariance term.","The discrete exposure mapping assumption could be relaxed to continuous G using density-ratio weighting, preserving the covariance structure with G-specific conditional effects.","The sensitivity parameters could be calibrated against observed proxies by computing partial correlations from auxiliary data, giving an empirical lower bound on the required correlation magnitude."],"forward_implications":["A practitioner who can bound the three standard deviations and two correlations obtains a worst-case bias interval for the naive IPW estimate under interference.","The T2 term implies that unmeasured confounding matters even when there are no common causes of A and A_N; it is driven by how much spillover effects vary across units.","If the outcome follows an additive model in A and G, the transport bias T3 vanishes (Proposition 2), so non-transportability only bites through interactions.","The extension to undefined potential outcomes (e.g., units with too few neighbors) keeps the same covariance logic with N-conditional covariances for spillover terms.","The framework applies to general interference structures beyond partial interference, since the exposure mapping only needs a scalar summary G."],"fun_headline_variants":["Spillover bias: a sum of covariances","New identity for bias from interference","Causal bias under interference: covariance identity","Set bounds on spillover and confounding bias","Sensitivity parameters for interference bias"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole bias identity depends on Assumption 2's consistency form, which fixes each unit's neighbors' treatments G when defining the natural potential outcome; if a unit's own treatment pushes neighbors to change treatment, the potential outcome should be Y(a,G(a)), and the decomposition no longer holds (a limitation the paper acknowledges in Remark 4).","fun_headline_variants_meta":{"raw":{"variants":["Spillover bias: a sum of covariances","New identity for bias from interference","Causal bias under interference: covariance identity","Set bounds on spillover and confounding bias","Sensitivity parameters for interference bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1114,"prompt_tokens":685,"completion_tokens":429,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":364}},"tokens_in":429,"tokens_out":429,"duration_ms":4995,"temperature":1.0,"reasoning_tokens":364,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:58:00.465578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a network where treatment is contagious (A_i affects A_j's treatment) with an unmeasured common cause of A and Y, then compare the empirical bias of the naive IPW estimator against the Theorem 1 expression computed from true potential outcomes. If the difference exceeds the bound implied by the true correlations and standard deviations, the decomposition is incomplete; if it is at most that, the bound is supported.","supporting_citations":[],"review_version":1}