{"id":"5c6b67a4-77c7-4056-babd-72da44bacb6e","arxiv_id":"1908.03867","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Granger causality test that conditions on the source series' estimated residual is shown to avoid spurious interactions caused by delayed latent common inputs.","lead":"Granger causality is a standard way to infer directed interactions from time series, but hidden common inputs can create false links. This paper proposes a variant that removes the influence of such confounders by conditioning on an estimate of the source series' own noise, and tests it with simulated data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim collapses if the network is symmetric: with x→y feedback, the first-stage residual for y is not a consistent proxy for the common input, so the F-test in Eq. (42) does not keep its null distribution.","rationale":"I considered whether the more serious weakness is the unproven F approximation after residual substitution or the restriction to single-lag cross-correlation. Both are real, but the asymmetry assumption is the most load-bearing because it is a precondition for the estimation strategy itself: the residual proxy ξ̂ is defined from y alone, and without exogeneity of y the entire likelihood factorization in Eq. (26) is invalid. The reader's weakest_assumption identifies exactly this point, and the paper's own Sec. IV confirms it. A simulation with feedback is the minimal decisive experiment: it directly tests whether the central claim survives when the stated assumption is relaxed. If the test shows size distortion, the paper's conditional acceptance is justified and the scope must be narrowed; if it does not, the concern would be weakened.","tokens_in":13676,"tokens_out":7420,"duration_ms":82208,"concrete_test":"Simulate the symmetric-network null: x_t = 0.9 x_{t-1} − 0.5 x_{t-2} + ε_t; y_t = 0.5 y_{t-1} − 0.2 y_{t-2} + 0.3 x_{t-1} + ξ_t, with (ε_t, ξ_{t-1}) jointly Gaussian with correlation 0.4, and no y→x interaction (c=0). Apply the proposed two-step procedure and compute the empirical rejection rate of the F-test (Eq. 42) at nominal 5% over 10,000 trials with T=1000. If the rejection rate is substantially above 5%, the asymmetric-network assumption is load-bearing; if it remains near 5%, the method may tolerate this form of feedback.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's load-bearing assumption is stated in Sec. II.B and repeated in Sec. III: 'we assume that the network structure is asymmetric, i.e., the interaction from xt to yt is absent.' The two-step estimator depends on this. Eqs. (35)–(36) and (45)–(46) replace the unobserved common-input component ξ_t with the residual from a univariate AR fit to y_t. That replacement is only consistent if y_t is generated by y_t = Σ b_i y_{t-i} + ξ_t with ξ_t independent of the x process. If x_{t-k} enters the y equation, the univariate AR for y is misspecified: b̂ is inconsistent and ξ̂_t is a prediction residual that contains the x-driven component of y, not the structural innovation. The likelihood factorization in Eq. (26) also fails, so the proposed F statistic in Eq. (42) is not comparing nested conditional models with exogenous regressors. Consequently, under the null c=0 with feedback x→y present, conditioning on ξ̂ does not remove the delayed common input; the null distribution need not be F, and false positives can appear. The paper acknowledges this only as a closing limitation ('the extension to the case where the networks structure may be symmetric will be in future works'), yet the title and abstract claim robustness to latent common inputs without this caveat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new Granger causality measure intended to be robust to latent common inputs with delayed effects. The model is a two-variable AR system with an asymmetric interaction (y→x only) and a lagged cross-noise correlation between ε_t and ξ_{t-1}. The author reparametrizes the noise, conditions on the source AR innovation, estimates the innovation in a first step from a univariate AR fit to y, and constructs an F statistic for testing c=0. The paper claims that this statistic approximately follows an F distribution under the null, verifies this by simulation, and develops a BIC-based stepwise procedure for the case where the lags are unknown. Numerical experiments compare the proposed method with normal Granger causality and report that the proposed method avoids spurious interactions under the simulated latent-common-input scenarios.","tokens_in":13963,"tokens_out":5830,"duration_ms":58649,"significance":"If the approximate F claim were rigorously established, the measure would be a useful alternative to partial Granger causality: it targets the source-target noise correlation rather than the target-conditioning correlation, avoids the negative-test-statistic issue, and the simulations are suggestive. The author is honest about the two-step approximation and lists limitations in the conclusions. However, the central distributional result is an unproven assumption, and the robustness claim is explicitly restricted to asymmetric networks; these gaps prevent acceptance in the current form.","major_comments":[{"comment":"The validity of the two-step estimator in Eqs. (35)–(36) and (45)–(46), and hence of the F test in Eq. (42), depends critically on the asymmetric-network assumption stated in §II.B. If the true system contains x→y feedback, the univariate AR fit to y is misspecified, ξ̂ is not a consistent estimate of the structural innovation ξ, and the conditioning argument underlying the F statistic does not hold. The manuscript acknowledges this only in the concluding future-work sentence, while the title and abstract claim robustness to latent common inputs without this caveat. The claims should be scoped explicitly, or the method should be extended to symmetric networks.","section":"§II.D and §IV"},{"comment":"The statement that F approximately follows the F distribution after substituting estimated residuals is assumed, not derived. Standard GLM deviance theory does not cover generated regressors; the text below Eq. (34) says 'we assume that the test statistic obeys the F-distribution approximately even if we substitute ξ̂_t.' The numerical support in §III.A is limited to one null model (model (b)) with known lags, and it does not vary ρ, the AR coefficients, or the lag structure. A theoretical justification, or a substantially broader simulation study (including misspecified lags, varying feedback strengths, and a range of ρ values), is needed to support the central claim.","section":"§II.D, Eq. (42)"},{"comment":"The proposed practical procedure performs sequential tests with BIC-based lag selection and a Bonferroni correction, but no analysis is given for the familywise error rate of the full sequential procedure. Figure 6 reports accuracy for specific simulation settings; it does not establish control of false positives across the multiple testing decisions. The text should either provide a formal FWER statement or report simulation results that specifically give false-positive rates under the null for the entire procedure.","section":"§III.B"}],"minor_comments":[{"comment":"The second summation is written as ∑ c_i x_{t-i}; it should be ∑ c_i y_{t-i} to match the model specification and all subsequent equations.","section":"Eq. (37)"},{"comment":"The summation in Eq. (29) starts at t=2, whereas the likelihood in Eq. (27) and the final statistic in Eq. (42) start at t=3 or lmax+1; the index ranges should be made consistent.","section":"Eq. (29)"},{"comment":"The caption states sample sizes 300, 1,000, and 3,000, but the panel labels read T=75, T=300, and T=3000; this mismatch should be corrected.","section":"Fig. 4"},{"comment":"The accuracy labels on the bars, e.g., '0.9710.914', are printed without separation, making the reported values difficult to read.","section":"Fig. 6"},{"comment":"The text cites Ref. [19] for the negative-value criticism of partial Granger causality but does not mention the later multivariate extension's treatment of the same issue; a brief remark would help place the contribution in the literature.","section":"Sec. II.A"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely publishable after the distributional gap is addressed and the claims are scoped. The asymmetry limitation is known to the authors, but it should appear in the abstract. The novelty is modest: the main contribution is the corrective focus on source-target noise correlation rather than target-conditioning noise correlation, and the numerical experiments support the method within the assumed model family."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: 1908.03867 is worth a careful read, but it's a conditional result, not the general fix the abstract implies. The core idea is a real improvement over partial Granger causality. Guo et al. remove the noise correlation between target and conditioning variable; Arai points out that the correlation causing spurious directed interactions is between target and source, so he conditions on the source's estimated noise. Concretely, estimate y's AR model, take the residual ξ̂ as a proxy for the latent common input, and test the coefficient c in x_t = Σ a_i x_{t-i} + Σ c_i y_{t-i} + η ξ̂_{t-l} + ω_t. If ξ were observed, the likelihood-ratio F-test is textbook and the paper's derivation is standard. The numerical experiments are honest: the null distribution is approximately F, and power grows with sample size. The paper also proposes a practical BIC/Bonferroni procedure and explicitly lists its limitations: latent variables are out of scope and the network structure is assumed asymmetric.\n\nThat last limitation is the soft spot. The whole two-step estimator depends on y being a univariate AR. If x→y feedback exists, the first-stage residual for y contains the x-driven component and is not a consistent proxy for the common input; the likelihood factorization in Eq. (26) fails and the F statistic in Eq. (42) need not hold under the null. The paper does acknowledge the asymmetric assumption in Secs. II and IV, but the abstract and title claim robustness to latent common inputs without that caveat. Also, the central distributional claim with estimated residuals is assumed rather than derived — the generated-regressor problem is waved off as an approximation, and the econometric residual-inclusion literature isn't cited. Simulations are all within the assumed model family, so they can't challenge the boundary.\n\nNone of this kills the paper. The idea is clear and could be useful for applied researchers who can credibly assume an asymmetric network. But it's not a general-purpose tool, and the missing proof (or at least a serious treatment of the estimated-residual step) is a genuine gap. No code and no data, though the simulations are easy to reproduce. I'd send it to peer review with a referee who knows the generated-regressor literature. A Monte Carlo under symmetric feedback should be requested, and the abstract should carry the asymmetry caveat.","headline":"A genuine twist on partial Granger causality, but the central F-claim rests on an unproven generated-regressor step and the abstract overstates the robustness.","tokens_in":14436,"tokens_out":2671,"would_cite":false,"duration_ms":28462,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62F03"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new Granger causality measure removes the confounding influence of latent common inputs and keeps its F-distribution under the null.","keywords":["Granger causality","latent common inputs","partial Granger causality","F-distribution","autoregressive model","spurious causality","confounding","multiple testing"],"falsifier":"Simulate the symmetric feedback model $x_t=a x_{t-1}+c y_{t-1}+\\varepsilon_t$, $y_t=b y_{t-1}+d x_{t-1}+\\xi_t$ with delayed correlated noise, run the proposed F-test, and compare the null histogram and false-positive rate to the nominal F-distribution; if they depart, the no-feedback assumption is essential.","tokens_in":13484,"feed_emoji":"📊","tokens_out":7592,"duration_ms":77362,"temperature":0.7,"pith_summary":"The paper proposes a Granger causality test meant to stay valid when two observed time series share a latent common input that reaches one series with a delay. The author argues that ordinary Granger causality mistakes such delayed common inputs for directed interaction, producing spurious links that become more certain as sample size grows. The new measure removes the noise correlation between source and target by conditioning on the source series' own autoregressive residual, which acts as a proxy for the common input. The central technical claim is that the resulting F statistic approximately follows the F-distribution under the null hypothesis of no interaction, and numerical experiments support this and show the test can detect genuine interactions once the sample is large enough. A practical lag-selection procedure based on BIC and Bonferroni-corrected multiple tests is included for cases where model orders are unknown.","feed_headline":"New Granger F-test stays valid under latent common inputs","feed_subtitle":"By removing the delayed noise correlation between source and target, it avoids spurious directed links.","key_machinery":"The load-bearing object is the conditional-variance reparameterization of the noise covariance: writing $\\varepsilon_t=\\eta\\,\\xi_{t-l_\\eta}+\\omega_t$ isolates $\\tau^2=\\Sigma_{xx}-\\Sigma_{xy}\\Sigma_{yy}^{-1}\\Sigma_{yx}$, the variance of the target noise after the correlated component with the source noise is removed. The estimation machinery is two-step: first fit the source series $y_t$ by its own autoregressive model, compute residuals $\\hat{\\xi}_t$, then include those residuals in the regression for the target $x_t$ and test the interaction coefficients $c$ with the scaled deviance difference that yields the F statistic in Eq. (42). This construction ensures the test statistic is nonnegative and gives it a known null distribution under the paper's assumptions.","core_discovery":"At the paper's core is the claim that delayed latent common inputs—noise shared by $x_t$ and $y_t$ at lag $l_\\eta$—make ordinary Granger causality report a directed interaction that is not there, and that the spurious link can be removed by conditioning on the source series. The author reparameterizes the noise as $\\varepsilon_t=\\eta\\,\\xi_{t-l_\\eta}+\\omega_t$, so that $\\tau^2=\\Sigma_{xx}-\\Sigma_{xy}\\Sigma_{yy}^{-1}\\Sigma_{yx}$ is the part of the target's noise uncorrelated with the source's noise. Assuming the source $y_t$ receives no feedback from the target $x_t$, the residuals from fitting $y_t$ alone estimate $\\xi$, and including them in the target regression makes the F statistic of Eq. (42) approximately $F$-distributed under the null $c=0$. Numerical experiments show the statistic tracks the null distribution when no interaction exists and separates from it under a true interaction, while ordinary Granger causality detects spurious interactions with high probability as the sample grows.","pith_inferences":["A practical diagnostic the paper implies but does not develop: the delay structure of the noise, not the mere presence of correlation, determines when ordinary Granger causality goes wrong, so checking whether cross-correlation at nonzero lags exists could flag data sets needing the new test.","The two-step residual-substitution idea should extend to any asymmetric pair of autoregressive processes, including nonlinear or point-process versions where one channel's 'residual' plays the role of the common input; the paper lists point-process extension as future work.","In symmetric systems the residual from $y$ alone is contaminated by feedback from $x$, so a testable extension would fit both directions, compare the two one-sided residuals, or iterate the estimation until the residuals stabilize.","Because the proposed test has lower power than ordinary Granger causality at small samples, it may be most useful as a confirmatory filter after ordinary Granger causality flags candidates, rather than as a replacement in small-sample studies."],"forward_implications":["In the asymmetric setting, practitioners can use standard F-distribution tables for the new statistic, with degrees of freedom $(p_1-p_0,\\,T-l_{\\max}-p_1)$, without bootstrapping.","When latent common inputs arrive with delay, ordinary Granger causality will keep producing spurious directed links as sample size grows, whereas the new procedure keeps its false-positive rate under control.","The proposed BIC-plus-Bonferroni procedure detects true directed interactions with accuracy approaching one as the sample grows, at the cost of lower power than ordinary Granger causality at small samples.","The method applies to latent common inputs whose effect appears as cross-correlated AR noise; latent variables that create autocorrelated noise are explicitly outside its scope and require a moving-average or state-space treatment."],"supporting_citations":[{"why":"Defines Granger causality as predictive improvement in autoregressive models, the concept the new test operationalizes.","marker":"[2]"},{"why":"Provides the toy-model setup used in the numerical experiments.","marker":"[7]"},{"why":"Introduces partial Granger causality, whose conditional-variance idea the new measure modifies.","marker":"[17]"},{"why":"Raises the negative-statistic and validity objections that motivate the revised estimation and test statistic.","marker":"[19]"},{"why":"Extends partial Granger causality to multivariate systems and supplies the joint maximum-likelihood estimation that the paper contrasts with its two-step estimator.","marker":"[20]"},{"why":"Supplies the generalized-linear-model deviance theory from which the F-distribution claim is derived.","marker":"[22]"},{"why":"Supplies the deviance and scaled-deviance distribution theory used to construct the F statistic.","marker":"[23]"},{"why":"Another source of the simulated toy models used to benchmark directed-interaction detection.","marker":"[24]"}],"fun_headline_variants":["Granger causality without spurious links from hidden drivers","Condition on source noise to block latent Granger confounders","Latent-input-proof Granger test avoids false directed interactions","Robust Granger test ignores hidden common inputs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the source series receives no feedback from the target, so its autoregressive residual can be estimated from the source alone and used as a clean proxy for the latent common input.","fun_headline_variants_meta":{"raw":{"variants":["Granger causality without spurious links from hidden drivers","Condition on source noise to block latent Granger confounders","Latent-input-proof Granger test avoids false directed interactions","Robust Granger test ignores hidden common inputs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000521,"raw_usage":{"total_tokens":2493,"prompt_tokens":886,"completion_tokens":1607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":1542}},"tokens_in":502,"tokens_out":1607,"duration_ms":11196,"temperature":1.0,"reasoning_tokens":1542,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:59:15.820002+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the symmetric feedback model $x_t=a x_{t-1}+c y_{t-1}+\\varepsilon_t$, $y_t=b y_{t-1}+d x_{t-1}+\\xi_t$ with delayed correlated noise, run the proposed F-test, and compare the null histogram and false-positive rate to the nominal F-distribution; if they depart, the no-feedback assumption is essential.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces partial Granger causality, whose conditional-variance idea the new measure modifies."},{"cited_title":"Roelstraete and Y","cited_arxiv_id":null,"evidence_quote":"Raises the negative-statistic and validity objections that motivate the revised estimation and test statistic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends partial Granger causality to multivariate systems and supplies the joint maximum-likelihood estimation that the paper contrasts with its two-step estimator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the generalized-linear-model deviance theory from which the F-distribution claim is derived."},{"cited_title":"Mccullagh and J","cited_arxiv_id":null,"evidence_quote":"Supplies the deviance and scaled-deviance distribution theory used to construct the F statistic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Another source of the simulated toy models used to benchmark directed-interaction detection."}],"review_version":1}