{"id":"b5e0dc45-b120-4d7d-b3b0-73b0a63aa4e0","arxiv_id":"1908.03107","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A multivariate peaks-over-thresholds model is coupled with counterfactual causation probabilities to attribute heavy rainfall extremes to anthropogenic climate change, using an optimal linear projection to maximize the estimated necessary-causation probability.","lead":"This paper combines multivariate extreme-value statistics with Pearl's counterfactual causality to estimate whether human-caused climate change made heavy winter rainfall events more likely. It models joint extremes as a multivariate generalized Pareto distribution and proposes a data-driven linear projection that maximizes the estimated probability of necessary causation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4.1's 'optimal weight' maximizes (4.3), which equals PN only if dependence is unchanged; the CNRM analysis and Figure 6 assume changing dependence, so the reported optimal-weight PN gains may be artifacts of the approximation.","rationale":"The reader's weakest assumption correctly identifies the load-bearing flaw. The paper's contribution has two parts: (i) embedding multivariate GPDs into Pearl's PN, and (ii) constructing a weight vector that maximizes PN. The first part is sound: Eq. (4.1) and the estimator (4.4) coherently estimate PN when the dependence enters through the first factor. The second part breaks down under the paper's own motivating scenario. Proposition 4.1 optimizes (4.3), which drops the dependence-sensitive factors P[w^T Z^(i)>0]. The paper is candid in noting the equality condition, but it does not reconcile this with Figures 8–9, where the two worlds are expected to have different spatial dependence. Figure 6 demonstrates that dependence changes matter for PN with equal weights, but no simulation validates the optimal-weight procedure under changing dependence. Other concerns (in-sample optimization, the typo in (4.4), lack of code/data) are real but secondary: in-sample optimization would bias PN upward even under constant dependence, but the constant-dependence issue undermines the very definition of 'optimal.' The proposed test on simulated bivariate GPDs with changing dependence would settle whether the approximation error is large enough to change the maximizer. The verdict should remain conditional: the paper is a useful contribution if the optimal-weight claim is either restricted to constant dependence or replaced by numerical maximization of the full estimated PN.","tokens_in":18065,"tokens_out":10385,"duration_ms":99589,"concrete_test":"Reproduce the bivariate simulation of Figure 4 (dotted line: χ^(0)=0.3, χ^(1)=0.5, σ^(0)=(1,1), σ^(1)=(2,2), γ=0) using the supplementary formulas for P[w^T Z^(i)>0] and the GPD survival H. For v at the 99% quantile of w^T Z^(0), compute the true PN(v,w) from Eq. (4.2) on a fine grid of w in (0,1) and locate its maximizer w*. Compare w* with w_opt from Proposition 4.1 maximizing Eq. (4.3). If |w* - w_opt| > 0.1, or if the true PN at w_opt is more than 0.05 below the true PN at w*, then the 'optimal weights' claim fails in the changing-dependence regime the paper motivates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central methodological innovation is the claim that a linear projection w can be chosen to maximize the probability of necessary causation PN(v,w) = max(1 - P[w^T X^(0)>v]/P[w^T X^(1)>v], 0). Proposition 4.1 derives an explicit w_opt, but the objective it actually maximizes is expression (4.3), which is 1 - H(v; w^T σ^(0), γ)/H(v; w^T σ^(1), γ). For exact multivariate GPDs, PN(v,w) = max(1 - [P[w^T Z^(0)>0] H0] / [P[w^T Z^(1)>0] H1], 0). The factors P[w^T Z^(i)>0] depend on the dependence structure of the i-th world and on w. Thus (4.3) equals PN only when P[w^T Z^(0)>0] = P[w^T Z^(1)>0], i.e., when the dependence structure is identical in the two worlds. The authors state this immediately after (4.3). However, the paper's motivation is explicitly the opposite case—spatial dependence differs between factual and counterfactual worlds (Section 2.1, Figure 4, and the CNRM application)—and Figure 6 uses changing dependence (χ^(1)>χ^(0)) to show multivariate modeling raises PN. In the application, 'optimal weights' are selected via Proposition 4.1 under the constant-dependence approximation, while the reported PN values are computed from (4.4), which includes the empirical first term P[w^T X^(i)>w^T u^(i)] that carries the dependence factor. If the dependence correction is material (Figure 4 suggests it is at finite v), the weights chosen by Proposition 4.1 need not maximize the PN actually reported, so the observed increases in PN with 'optimal weights' (Figures 8–9) may be an artifact of optimizing a misspecified objective. The claim 'maximizes such causation probabilities' in the abstract and Section 4.2 is therefore not supported in the changing-dependence regime that motivates the method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a multivariate peaks-over-thresholds approach to extreme event attribution. It models the joint tail of a spatial precipitation vector by a multivariate generalized Pareto distribution and uses Pearl's counterfactual probabilities of necessary causation, PN = max(1-p0/p1, 0), for events of the form {w^T X > v}. The main methodological contribution is Proposition 4.1, which derives an explicit bivariate weight vector maximizing an approximation of PN based on the univariate GPD behavior of linear projections, together with an inference strategy that estimates PN through a combination of an empirical exceedance factor and a fitted GPD tail. The method is illustrated on weekly winter precipitation maxima from the CNRM CMIP6 model, where cluster-specific multivariate PN estimates are compared with univariate estimates. The authors conclude that accounting for spatial dependence can increase PN and reduce uncertainty relative to univariate analyses.","tokens_in":1834,"tokens_out":2059,"duration_ms":91898,"significance":"If the main claims hold, this would be a useful contribution to climate event attribution: it connects modern multivariate extreme-value theory with causal attribution, gives an explicit formula for choosing a projection that increases causal evidence, and demonstrates the approach on a real CMIP6 model with clustering based on tail dependence. The use of Proposition 2.1 from Rootzen, Segers and Wadsworth is well grounded, and the paper is honest about several scope restrictions: common shape parameter, asymptotic dependence, stationarity, and the fact that Eq. (4.3) equals PN only when dependence is unchanged between the two worlds. The simulations and the CNRM application are valuable even if the optimal-weight claim needs qualification.","major_comments":[{"comment":"The paper's central 'optimal weight' claim is only proved for a simplified objective. For exact multivariate GPDs, PN(v,w) = max(1 - [P(w^T Z^(0)>0) H0] / [P(w^T Z^(1)>0) H1], 0), while Proposition 4.1 maximizes (4.3), namely 1 - H0/H1, where H_i = H(v; w^T sigma^(i), gamma). The factors P(w^T Z^(i)>0) depend on the dependence structure of world i and on w. The paper states after (4.3) that the equality holds only when P(w^T Z^(0)>0) = P(w^T Z^(1)>0), i.e., when dependence is unchanged. However, the paper's motivation is precisely the opposite case: Section 2.1 argues that spatial dependence can differ between factual and counterfactual worlds, Figure 4 illustrates PN changes under changing dependence, and Figure 6 uses chi^(1)>chi^(0). In the application, the 'optimal' weights are selected under the constant-dependence approximation, while the reported PN values in Figures 8-9 are computed from (4.4), which retains the empirical first factor. There is therefore no guarantee that w_opt maximizes the PN actually reported, and the observed increases in PN and apparent uncertainty reduction may be artifacts of optimizing a misspecified objective. Please either restrict the optimality claim to the constant-dependence setting and present w_opt as a heuristic in changing dependence, or assess the bias by simulation and by evaluating the full PN at w_opt versus other weights under changing dependence.","section":"Section 4.2, Proposition 4.1, Eqs. (4.1)-(4.4), Figures 8-9"},{"comment":"The optimal weights are estimated from the same data on which the PN values and their confidence intervals are computed. It is not stated whether the bootstrap procedure in Figures 8-9 re-estimates w_opt for each bootstrap sample or conditions on the estimated weights. If the weights are held fixed, the reported intervals understate the sampling variability of the full procedure and the comparison of uncertainty with univariate analyses is not on equal footing. Please clarify the bootstrap scheme and, if weights are not re-estimated, either re-estimate them in each bootstrap replication or add a sensitivity analysis showing that the conclusions are robust to the selection step.","section":"Section 4.3, Section 5, Figures 8-9"},{"comment":"Eq. (4.3) also omits the threshold normalization that appears in the paper's own inference formula (4.4): for general X^(i), the ratio p0/p1 involves P(w^T X^(i) > w^T u^(i)) in addition to the GPD tail terms. Even if the extremal dependence is constant, these threshold exceedance probabilities can depend on w and may differ between worlds. The current wording after (4.3) only mentions the equality for multivariate GPDs with P(w^T Z^(0)>0)=P(w^T Z^(1)>0). Please add a remark that, in the thresholded-data approximation, the first factors in (4.1) are not generally cancelled, and that Proposition 4.1 therefore maximizes an approximation that discards part of the dependence information in the data.","section":"Section 4.2, Eqs. (4.1)-(4.3)"}],"minor_comments":[{"comment":"The second case in Eq. (4.4) has the condition 'if v <= w^T u^(i)', which is identical to the first case and must be a typo; it should read 'if v > w^T u^(i)'.","section":"Section 4.3, Eq. (4.4)"},{"comment":"The sentence 'it is equal to the PN when X^(0), X^(0) are multivariate GPDs' contains a typo: the second subscript should be X^(1).","section":"Section 4.2, after Eq. (4.3)"},{"comment":"The notational definitions of PN, PS, and PNS before Eq. (2.1) appear garbled or missing overbars and do(C) notation, making them hard to parse; please clarify them to match Hannart et al. (2016).","section":"Section 2.1, Eq. (2.1) context"},{"comment":"The definition of the empirical exceedance factor writes the first term as a function of v, but it is later evaluated at w^T u^(i); please define it with a generic argument and then specialize, to avoid notational confusion.","section":"Section 4.3, Eqs. (4.1)-(4.4)"},{"comment":"The condition gamma = gamma 1_d is used to apply Proposition 2.1; please define the notation 1_d explicitly at first use.","section":"Section 4.1, Eq. (4.1)"},{"comment":"The sentence 'calculated based on a pre-simulation run of sample size 10^6 and held fixed' is unclear about which quantity is held fixed; if the weights are equal weights and fixed, say so explicitly.","section":"Section 4.3, Figure 6"},{"comment":"The phrase 'only zero or unit weights maximize this ratio' should specify that the maximization is of the ratio in (4.3), not of the full PN in (4.2), to avoid overstating the result.","section":"Proposition 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the external grounding in multivariate GPD theory is solid. The main issue is not novelty or correctness of the probabilistic building blocks, but the scope of the optimality claim: the 'optimal linear projection' result is proved for a constant-dependence approximation, yet the applications and headline figures emphasize changing dependence. I think this can be fixed with careful rewriting, additional simulations evaluating the actual PN at the proposed weights, and explicit treatment of selection uncertainty in the bootstrap. I would not recommend rejection, because the underlying methodology is defensible and transparently presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something new that was worth doing: it puts Pearl's necessary-causation probability into a multivariate GPD framework and shows how a linear projection can be chosen to maximize it. Proposition 4.1, giving explicit optimal weights for the bivariate constant-dependence case, is original and likely useful. The writing is clear, the link to existing multivariate GPD theory is careful, and the application to CMIP6 precipitation is a reasonable demonstration of intent. Credit where due: the authors state the key limitation right after (4.3) -- expression (4.3) is only equal to the true PN when the dependence structure is unchanged between worlds.\n\nThe soft spot is exactly where the stress-test note lands. The paper motivates the whole multivariate exercise by arguing that dependence can change between factual and counterfactual worlds (Section 2.1, Figure 4), and Figure 6 uses changing dependence to claim that multivariate modeling raises PN. Yet Proposition 4.1's w_opt maximizes (4.3), not the PN defined in (4.2). When dependence changes, (4.3) is a potentially poor approximation, and the weights chosen by Proposition 4.1 need not maximize the PN that the application actually reports. The reported increases in PN with \"optimal weights\" in Figures 8-9 may therefore be artifacts of optimizing a misspecified objective. This is not a fatal flaw for the constant-dependence theory, but it is a serious gap between the claim and the evidence.\n\nMinor issues: Eq. (4.4) contains an obvious typo (the second branch should be v > w^T u^(i)); the weights are selected on the same data used to estimate PN, so the bootstrap intervals understate selection uncertainty; and the paper gives no code or data, which would help the reader judge the clustering and the PN estimates. None of these are showstoppers, but they need attention.\n\nThe paper deserves a serious referee -- the combination is novel and the theoretical core is sound under an explicitly stated assumption. But the applied claims about optimal weights in a changing-dependence world need either to be withdrawn or supported by a numerical optimization of the full PN objective, with appropriate treatment of selection effects. I would send it to review, and I'd ask for major revision.","headline":"A genuine bridge between multivariate GPDs and Pearl's causation probabilities, but the optimal-weight claim only holds under constant dependence, and the application assumes the opposite.","tokens_in":19077,"tokens_out":2753,"would_cite":false,"duration_ms":29831,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G32","62P12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Modeling the full spatial field of extreme rainfall with a multivariate generalized Pareto distribution yields stronger and less uncertain climate attribution statements than univariate analyses.","keywords":["extreme event attribution","multivariate generalized Pareto distribution","counterfactual theory","necessary causation","peaks-over-thresholds","heavy rainfall","CMIP6","tail dependence"],"falsifier":"On simulated bivariate GPD data with known marginal parameters and different tail dependence coefficients in the two worlds, compute the true $PN$ by Monte Carlo and compare it with the optimal-weight $PN$ from Proposition 4.1; if the proposition's maximizer yields a lower necessary-causation probability than equal weights or than the true optimum, the claim that the proposed weights maximize $PN$ under changing dependence is falsified.","tokens_in":17867,"feed_emoji":"🌧","tokens_out":7561,"duration_ms":68663,"temperature":0.7,"pith_summary":"The paper claims that climate extreme event attribution can be made more informative by modeling the entire spatial vector of precipitation as a multivariate generalized Pareto distribution rather than reducing it to a single regional average. It connects this extreme-value model to Pearl's counterfactual theory, defining the probability of necessary causation as $PN = \\max(1 - p_0/p_1, 0)$, and shows how to choose the linear weights on grid points that maximize this probability. Under the approximation that a linear projection of a multivariate GPD is a univariate GPD, the method produces explicit optimal weights in the bivariate case and a general inferential recipe. Applied to weekly winter maxima of French CNRM CMIP6 precipitation, it yields higher necessary-causation probabilities and narrower confidence intervals than univariate approaches, particularly around northern Italy. If correct, this gives attribution studies a principled way to extract causal signal from high-dimensional climate model output.","feed_headline":"Multivariate tail modeling strengthens climate attribution evidence","feed_subtitle":"Modeling the full rainfall field yields stronger, less uncertain attribution statements.","key_machinery":"The central object is the multivariate generalized Pareto distribution (MGPD) with the stochastic representation $Z^* = E + T - \\max_{1\\le j\\le d} T_j$, where $E$ is unit exponential and $T$ is an arbitrary random vector independent of $E$. The load-bearing property is Proposition 2.1: when all shape parameters equal $\\gamma$, the linear projection of an MGPD vector, conditioned on being positive, is a univariate GPD with scale $w^T\\sigma$ and shape $\\gamma$. This property reduces high-dimensional threshold exceedances to univariate GPD tails, enabling the approximation (4.1), the optimal-weight Proposition 4.1, and a simple estimation scheme based on probability weighted moments. Tail dependence coefficients $\\chi$ drive the spatial clustering used to delimit homogeneous regions.","core_discovery":"The central claim is that for events of the form $\\{w^T X > v\\}$, where $X$ is a $d$-dimensional vector of precipitation at grid points and $w$ are non-negative weights summing to one, the probability of necessary causation can be approximated by modeling the tail of $X$ as a multivariate generalized Pareto distribution and using the linear-projection property $[w^T Z \\mid w^T Z > 0] \\sim GPD(w^T \\sigma, \\gamma)$ when all shape parameters are equal. This leads to the approximation $P[w^T X^{(i)} > v] \\approx P[w^T X^{(i)} > w^T u]\\, H(v - w^T u; w^T \\sigma^{(i)}, \\gamma)$, from which the paper derives, in the bivariate case, an explicit optimal weight $w_{opt}(v)$ that maximizes the necessary-causation probability (Proposition 4.1). The paper further claims that the optimized multivariate PN is larger than the univariate PN and has smaller uncertainty, and demonstrates this on simulated data and on CNRM CMIP6 weekly winter precipitation maxima after clustering grid points into homogeneous regions via tail-dependence-based partitioning around medoids.","pith_inferences":["Because the paper's optimal-weight formulas drop the ratio $P[w^T Z^{(0)} > 0]/P[w^T Z^{(1)} > 0]$, a direct testable extension is to maximize the full PN expression allowing dependence to change; on the CNRM data, where dependence does change, the reported optimal-weights PN may be a biased estimate of true necessary causation.","The method could be applied to compound events where an event is defined by several variables exceeding thresholds jointly; the same projection machinery would then identify the linear combination of variables that maximizes causal evidence.","The clustering step uses only the counterfactual run; re-clustering in the factual world, or using a dependence distance that varies with the forcing, might reveal regions where a change in dependence itself is part of the causal story."],"forward_implications":["A univariate analysis that aggregates grid points with equal weights will generally understate necessary causation when tail dependence is stronger in the factual world than in the counterfactual world.","The optimal-weight procedure gives an objective, data-driven choice of the spatial projection, so attribution statements can be tied to the grid locations that carry the strongest causal signal.","The method requires homogeneous regions with a common shape parameter; clustering by tail dependence makes this assumption workable for precipitation fields.","For the CNRM CMIP6 data, the multivariate approach reports necessary-causation probabilities above 0.5 for most of central Europe at the fifty-year return level, with near-one values around northern Italy.","The same inferential recipe transfers to other threshold-exceedance events and other climate model outputs, provided the stationarity and asymptotic-dependence conditions hold."],"supporting_citations":[{"why":"Supplies the stochastic representation and the linear-projection property that turns multivariate GPD projections into univariate GPDs.","marker":"Rootzén, Segers and Wadsworth (2018a)"},{"why":"Introduces the multivariate generalized Pareto distribution that the paper uses to model joint extremes.","marker":"Rootzén and Tajvidi (2006)"},{"why":"Defines the counterfactual probabilities of necessary and sufficient causation used throughout.","marker":"Pearl (2000)"},{"why":"Derives the simplified formulas $PN = \\max(1-p_0/p_1,0)$ and related quantities under monotonicity and exogeneity.","marker":"Hannart et al. (2016)"},{"why":"Proposes maximizing causation probabilities via linear projections in the Gaussian case, which the paper extends to multivariate GPDs.","marker":"Hannart and Naveau (2018)"},{"why":"Provides the peaks-over-thresholds inference methodology for multivariate GPDs used in estimation.","marker":"Kiriliouk et al. (2019)"},{"why":"Defines the original univariate event-attribution comparison that the paper's multivariate analysis refines.","marker":"Stott, Stone and Allen (2004)"},{"why":"Motivates the equal-shape regional modeling assumption and the regional frequency analysis approach.","marker":"Carreau, Naveau and Neppel (2017)"},{"why":"Gives the template for clustering spatial extremes by dependence, adapted here via tail-dependence distances.","marker":"Bernard et al. (2013)"},{"why":"Supplies the partitioning around medoids algorithm used to form homogeneous clusters.","marker":"Kaufman and Rousseeuw (1990)"}],"fun_headline_variants":["Optimal projection boosts climate event attribution certainty","Multivariate tails sharpen counterfactual climate attribution","New projection maximizes causal evidence for heavy rainfall","Tail modeling improves attribution of extreme climate events","Counterfactual approach strengthens multivariate event attribution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument requires that the spatial dependence among extremes stays the same between the counterfactual and factual worlds, because only then does the probability of a projected sum being positive cancel out and the maximized expression equal the true necessary-causation probability.","fun_headline_variants_meta":{"raw":{"variants":["Optimal projection boosts climate event attribution certainty","Multivariate tails sharpen counterfactual climate attribution","New projection maximizes causal evidence for heavy rainfall","Tail modeling improves attribution of extreme climate events","Counterfactual approach strengthens multivariate event attribution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000441,"raw_usage":{"total_tokens":2258,"prompt_tokens":987,"completion_tokens":1271,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":1204}},"tokens_in":603,"tokens_out":1271,"duration_ms":11049,"temperature":1.0,"reasoning_tokens":1204,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:24:33.291810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On simulated bivariate GPD data with known marginal parameters and different tail dependence coefficients in the two worlds, compute the true $PN$ by Monte Carlo and compare it with the optimal-weight $PN$ from Proposition 4.1; if the proposition's maximizer yields a lower necessary-causation probability than equal weights or than the true optimum, the claim that the proposed weights maximize $PN$ under changing dependence is falsified.","supporting_citations":[{"cited_title":"( 2000 )","cited_arxiv_id":null,"evidence_quote":"Defines the counterfactual probabilities of necessary and sufficient causation used throughout."},{"cited_title":", Pearl , J J","cited_arxiv_id":null,"evidence_quote":"Derives the simplified formulas $PN = \\max(1-p_0/p_1,0)$ and related quantities under monotonicity and exogeneity."},{"cited_title":"Naveau , Philippe P","cited_arxiv_id":null,"evidence_quote":"Proposes maximizing causation probabilities via linear projections in the Gaussian case, which the paper extends to multivariate GPDs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the peaks-over-thresholds inference methodology for multivariate GPDs used in estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the original univariate event-attribution comparison that the paper's multivariate analysis refines."},{"cited_title":", Naveau , P P","cited_arxiv_id":null,"evidence_quote":"Motivates the equal-shape regional modeling assumption and the regional frequency analysis approach."},{"cited_title":", Naveau , Philippe P","cited_arxiv_id":null,"evidence_quote":"Gives the template for clustering spatial extremes by dependence, adapted here via tail-dependence distances."},{"cited_title":"Rousseeuw , Peter J P","cited_arxiv_id":null,"evidence_quote":"Supplies the partitioning around medoids algorithm used to form homogeneous clusters."}],"review_version":1}