{"id":"9ca92c39-f07a-4ec6-ae16-0c63bb219435","arxiv_id":"2509.02653","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A standardized pipeline translates customer-weighted outage-days into monetized deprivation costs for four hurricanes, showing regressive burdens and restoration duration as the key driver.","lead":"This paper converts utility outage data from four U.S. hurricanes into dollar estimates of household welfare loss, and finds that lower-income areas carry a larger cost relative to their income. It offers cities and utilities a template for comparing outages across storms and targeting restoration where the social cost of delay is highest.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated transfer of the deprivation cost function (Eq. 1) from ref [32] to all four storms makes the monetized social-cost totals and cross-event comparisons load-bearing but unsupported; the paper's own Discussion concedes this limitation.","rationale":"The reader identified the transfer of the deprivation cost function as the weakest assumption, and I agree. The paper's central contribution is a monetized, cross-event accounting of outage costs; those dollar values are produced by applying Eq. 1 unchanged across four storms and thousands of ZCTAs. The only support for that transfer is the statement that the function is 'empirically derived' in ref [32], which is not available in this manuscript and is authored by an overlapping group. The paper itself flags the limitation in the Discussion, so the concern is acknowledged by the authors rather than manufactured. No independent validation, uncertainty propagation, or sensitivity analysis is provided. If the coefficients were estimated in a different population or outage context, the Table 2 totals and the SHAP driver ranking would not support the claimed quantitative conclusions, even though the qualitative regressive pattern might remain. I do not see an internal logical error that would justify rejection; rather, the appropriate response is to condition acceptance on validation or on a substantial reframing of the claims. Because the reader already issued a CONDITIONAL verdict for essentially this reason, my stress-test does not move the verdict.","tokens_in":13193,"tokens_out":4513,"duration_ms":59760,"concrete_test":"Obtain from ref [32] the full covariance matrix of the DCF coefficients and construct 95% confidence intervals for DC(t) at t = 1, 3, 7, and 14 days. Compare these intervals with independent US willingness-to-pay or resilience values for comparable outage durations (e.g., Baik et al. 2020, Nature Energy) and with any available event-specific damage proxies for Ida and Beryl. Predefine a tolerance—for example, if the DCF interval misses the independent estimate by more than 50% at any matched duration, the transfer is unsupported. As a complementary check, fit event-specific scaling of the DCF to the AIC-selected duration/cost curves in Figures 7–8; if implied DC-to-duration relationships differ by more than a factor of two across events, fixed-coefficient cross-event dollar comparisons are not justified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every monetized result in Table 2 flows through Eq. 1: DC = 35.95 t^2 + 107.84 t + 71.89. Those coefficients were estimated in a separate stated-preference study by overlapping authors (ref [32]) and are applied here unchanged to four hurricanes, all ZCTAs, and all income strata. No uncertainty intervals, no calibration to observed outage outcomes, and no external validation are reported. The paper's Discussion explicitly states that the function 'may not fully reflect variations in willingness-to-pay across diverse cultural, health, and social contexts.' If the true deprivation valuation varies by region, income, storm intensity, or outage duration beyond what the fixed coefficients capture, then the headline totals ($1.50B, $1.26B, $629M, $411M), the per-capita burdens, and the cross-event rankings are not comparable in dollar terms. The qualitative regressive pattern could survive, but the central claim of 'comparable monetized social costs' and the SHAP 'dominant driver' conclusion lose their quantitative force. This is not an internal contradiction; it is an unvalidated transfer of an external parameter that the authors themselves flag as a limitation. Conditional acceptance should require either demonstrating transfer validity or reframing the dollar figures as scenario-based illustrations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a standardized, cross-event framework for monetizing the social cost of hurricane-induced power outages. It combines EAGLE-I outage traces with ACS ZCTA demographics to compute customer-weighted average outage-days, restoration duration, and relative restoration rate for Hurricanes Beryl, Helene, Milton, and Ida. These metrics are converted into dollar 'deprivation costs' using a quadratic function DC = 35.95 t^2 + 107.84 t + 71.89 (Eq. 1), imported from the authors' companion stated-preference study (ref. [32]). The analysis reports total deprivation costs of $1.50B (Ida), $1.26B (Milton), $629M (Beryl), and $411M (Helene), with per-capita costs of $1,757, $387, $674, and $285. It then examines regressivity by income, fits DC-versus-restoration curves, runs SHAP-based random forest models, and applies K-means clustering to define recovery typologies. The paper claims these are the first cross-event, fine-scale welfare-based outage cost estimates, with restoration duration as the dominant driver and consistently regressive burdens.","tokens_in":13594,"tokens_out":4892,"duration_ms":57038,"significance":"If the quantitative claims are reliable, the framework would be a useful decision-support tool for equity-informed restoration prioritization and resilience investment, going beyond SAIDI/CAIDI by monetizing welfare losses at ZCTA scale. The strengths are the uniform pipeline across four events, the explicit welfare-economics framing, the use of high-frequency observational data, and the combination of explanatory modeling with unsupervised clustering. However, nearly every headline dollar figure and cross-event comparison flows through Eq. 1, a fixed-coefficient function transferred from a companion preprint without uncertainty propagation or transfer validation. The paper's own Discussion concedes that the function 'may not fully reflect variations in willingness-to-pay.' The qualitative regressive pattern may well survive even if the dollar values are treated as illustrative, but the current manuscript does not establish the quantitative transferability it claims. The contribution is potentially significant but is presently conditional on external parameter validity and on a few non-standard methodological choices.","major_comments":[{"comment":"The monetized outputs in Table 2 are all computed through DC = 35.95 t^2 + 107.84 t + 71.89, with coefficients taken from the authors' companion preprint (ref. [32]). No uncertainty intervals are attached to these coefficients, no transfer validation is provided for applying a single stated-preference function to four different states, income strata, and storm contexts, and the Discussion explicitly concedes the function may not reflect WTP variation across cultural, health, and social contexts. Because the total costs, per-capita burdens, cross-event rankings, and the regressivity measures all scale with Eq. 1, this is a load-bearing unsupported transfer. The authors should either demonstrate transfer validity (e.g., sensitivity to plausible coefficient ranges, calibration to observed outage outcomes, or evidence that WTP is stable across regions) or reframe the dollar figures as scenar","section":"Methods, 'Deprivation cost function' (Eq. 1); Discussion"},{"comment":"The average outage-days metric is computed by dividing total customer-days by the maximum daily customer count, and the total deprivation cost is then computed as 'average DC times total affected customers.' Since Eq. 1 is convex in t (t^2 term), DC(E[t]) * N is not equal to the sum of individual deprivation costs; the aggregation is biased and its direction depends on the outage-duration distribution. The manuscript does not define 'total affected customers' unambiguously in the total-cost formula. This affects every total in Table 2. The authors should either compute the sum of per-customer DCs using the full distribution of outage durations, or provide bounds and state clearly that the reported total is an approximation based on the mean duration.","section":"Methods, 'Average power outage duration (days) per customer' and 'Deprivation cost function'"},{"comment":"The affected-area classification is not fully specified. For Helene and Milton, the text says a ZCTA is affected if its impact median 'exceeded the threshold,' but no threshold value is reported. For Ida, the baseline is taken from a post-event recovery window (September 18–24, 2021) and the affected-area rule is '80% of the baseline,' which is qualitatively different from the rule for the Florida events. These choices determine which ZCTAs enter the analysis and therefore drive the totals and per-capita values in Table 2. The authors should report the exact threshold(s), justify the event-specific differences, and provide a robustness check (e.g., varying the threshold) to show that the cross-event comparisons are not artifacts of the inclusion rule.","section":"Methods, 'Affected Area Identification'"},{"comment":"The claim that restoration duration is the 'dominant driver' of deprivation cost and that the DC–duration relationship is 'mechanistic' is overstated. Restoration duration and relative restoration rate are computed from the same outage accumulation/restoration curves that produce average outage-days, which is the direct input to Eq. 1. The positive association between DC and restoration duration and the negative association with relative restoration rate are therefore partly mechanical. The SHAP analysis does not control for average outage-days or for the functional form of Eq. 1, so the dominance of restoration duration may reflect collinearity rather than an independent causal mechanism. The authors should control for average outage-days (e.g., partial dependence or SHAP with t included) and soften the causal language, presenting the SHAP results as explanatory associations within the","section":"Results, Figures 7–8; Methods, 'Random Forest and SHAP'"},{"comment":"Several quantitative claims rest on curve fits and clusters whose quality is not reported. The curve fitting selects functional forms by AIC but no R^2, residual diagnostics, or confidence bands are given for the DC-versus-income, DC-versus-restoration-rate, or DC-versus-duration fits. The K-means analyses report k=4 without any model-selection criterion (e.g., elbow or silhouette) or validation. In addition, the text and Figure 11 caption contradict each other on whether Cluster 3 is 'lowest income' or 'high income.' These issues do not necessarily invalidate the qualitative findings, but they prevent the reader from assessing the strength of the reported regressive patterns and typology claims.","section":"Methods, 'Curve fitting' and 'K-means Cluster'; Figures 6–8, 11–14"}],"minor_comments":[{"comment":"'per capital' should be 'per capita'; the terms 'restore duration' and 'restoration duration' are used inconsistently throughout.","section":"Abstract and text"},{"comment":"The text refers to the deprivation cost table as 'Table 1' while the table is labeled 'Table 2'; Table 1 is the data sources list. Please renumber or fix the in-text reference.","section":"Results, Table 2"},{"comment":"Equation 2 is rendered with garbled symbols; the definitions of outage rate and restoration rate should be written out explicitly so the ratio is unambiguous.","section":"Methods, Eq. 2"},{"comment":"The data section attributes outage data to EAGLE-I at Oak Ridge National Laboratory, while the Acknowledgment mentions the I-EAGLE dataset from NREL; please clarify the exact data source and access terms.","section":"Data and Acknowledgment"},{"comment":"The random forest implementation lacks key reproducibility details: number of trees, hyperparameters, train/test split, cross-validation, and whether the target DC was transformed. Please report these.","section":"Methods, 'Random Forest and SHAP'"},{"comment":"The curve fits in Figures 6–8 would be much more informative with fit statistics and shaded confidence intervals; as shown, the fitted lines are not accompanied by any measure of scatter around the trend.","section":"Figures 6–8"},{"comment":"No code or data availability statement is provided. Given the reproducibility-oriented claims of a 'standardized pipeline,' the authors should make the processing scripts and derived data available.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central monetary results depend on coefficient estimates from a companion preprint by overlapping authors (ref. [32]). The current manuscript does not establish that this transfer is valid, and the Discussion concedes exactly this limitation. The editor may wish to consider whether the journal should publish dollar-valued social cost estimates that inherit uncertainty from an unreviewed companion paper without confidence intervals. The contribution is potentially useful, but the claims need to be reframed or the transfer validated before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: the paper gives the first cross-event, ZCTA-level monetized comparison of power outage costs for four U.S. hurricanes, and the regressive pattern—lower-income areas bearing a larger share of deprivation cost relative to income—is plausible and worth taking seriously. But the dollar totals are not as solid as the abstract implies. The deprivation cost function (DC = 35.95 t^2 + 107.84 t + 71.89) comes from a separate stated-preference study by overlapping authors and is applied here without validation or uncertainty bounds. The authors acknowledge this in the Discussion, saying the function “may not fully reflect variations in willingness-to-pay.” That means the $1.50B for Ida, the $1.26B for Milton, and the cross-event rankings should be read as scenario-style illustrations under a common valuation, not robust estimates of social cost.\n\nWhat the paper does well: the pipeline is clear and reproducible in principle, the EAGLE-I data are real, and the decision to compare four storms with one standardized metric is genuinely new. The clustering analysis is a nice addition—it shows that similar outage durations can have very different welfare impact depending on income and restoration dynamics. The SHAP analysis is suggestive, but I'd be careful calling restoration duration the “dominant driver” when the cost metric is derived from the same outage curves; the correlation is real but not independent evidence.\n\nSoft spots beyond the DCF transfer: the affected-area identification thresholds are not fully reported, and for Ida the text says “80% of baseline,” which is confusing and may be a typo. Also, computing deprivation cost from the average outage-days per customer and multiplying by customers ignores the convexity of the DCF; if outage durations vary within a ZCTA, that underestimates total cost. That's a real issue, though it mostly biases levels, not the qualitative equity pattern.\n\nThe authors do not appear to have released code or data, and the thresholds should be documented. I'd recommend sending this to peer review with the expectation that the authors either validate the DCF transfer with sensitivity analysis or reframe the dollar figures as illustrative. The framework is useful, the qualitative results are likely to survive, and the paper deserves serious referee time.","headline":"Useful cross-event outage-cost comparison, but the dollar figures rest on an unvalidated deprivation-cost function and underreported thresholds; worth refereeing with revisions.","tokens_in":13990,"tokens_out":3386,"would_cite":true,"duration_ms":40449,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Converting outage-days into dollar losses shows hurricane blackouts cost up to $1.5 billion per event and weigh most heavily on low-income communities.","keywords":["deprivation cost","power outage","hurricane resilience","energy equity","restoration duration","welfare economics","SHAP","K-means clustering"],"falsifier":"Re-estimate the deprivation cost curve from willingness-to-pay data collected specifically in Hurricane Ida-affected parishes and compare it to the fixed curve; or test the framework by predicting total outage losses for one of the four events and comparing against observed insurance claims, business-interruption losses, or utility compensation payouts. A systematic gap between predicted and observed dollar losses would falsify the transferability of Eq. 1.","tokens_in":13107,"feed_emoji":"⚡","tokens_out":6242,"duration_ms":72745,"temperature":0.7,"pith_summary":"This paper tries to put a welfare price tag on hurricane-driven power outages. Rather than reporting how many customers were out and for how long, it multiplies customer-weighted outage-days by a deprivation cost function and reports social loss in dollars. Applied to four hurricanes—Ida, Milton, Beryl, and Helene—the method yields comparable event-level costs (from $411 million for Helene to $1.50 billion for Ida) and shows per-capita burden falling hardest on lower-income areas. The authors argue restoration duration is the main lever on that loss, with faster relative restoration cutting costs, and that clustering reveals recovery patterns ordinary reliability indices miss. If the framework holds, utilities gain a dollar-denominated way to rank communities by avoidable social loss when restoring power.","feed_headline":"Hurricane outages carry a dollar cost—up to $1.5B a storm","feed_subtitle":"Welfare-based outage accounting shows restoration speed drives social loss and the poor bear the brunt.","key_machinery":"The load-bearing object is the deprivation cost function (DCF), an empirically estimated convex mapping from outage duration to monetary welfare loss, originally estimated in a stated-preference study. It is what turns engineering outage counts into additive, comparable dollar losses. Around it are three recovery metrics: customer-weighted average outage-days (the exposure variable), restore duration (the active recovery phase), and relative restoration rate (restoration speed divided by outage accumulation speed). SHAP analysis on a Random Forest model attributes cost drivers, and K-means clustering groups ZIP codes into recovery typologies; both are secondary tools that translate the dolla","core_discovery":"The central claim is that a single quadratic function, DC = 35.95·t² + 107.84·t + 71.89 with t in average outage-days per customer, converts outage exposure into a monetary deprivation cost, and that total cost equals this per-customer cost times affected customers. Combining customer-weighted outage-days computed from sequential high-frequency outage observations with ZCTA-level census demographics, the paper produces the first cross-event, fine-scale accounting of outage costs. Its central empirical results are: Ida imposed about $1.50B in total deprivation costs with $1,757 per capita; Milton $1.26B ($387 per capita); Beryl $629M ($674 per capita); and Helene $411M ($285 per capita). Acro","pith_inferences":["The paper does not test whether the same quadratic cost curve transfers across states and income groups; re-estimating the curve on independent data for one of these events would confirm or break the dollar scale, even if the qualitative regressive pattern survives.","Because the cost function is convex in duration, shortening the longest outages saves more social loss per day than shortening shorter ones; this suggests the equity and efficiency arguments for prioritizing the longest, low-income outages align.","The dollar-additive design invites a natural extension to compound events: summing deprivation costs from consecutive storms (e.g., Helene followed by Milton) would quantify cumulative seasonal burden, which the paper notes but does not measure.","The same normalization logic—restoration rate over outage rate—could be applied to water, communications, or transit interruptions if separate willingness-to-pay functions were estimated, producing comparable deprivation ledgers across infrastructure sectors."],"forward_implications":["Utilities can compute the social loss avoided by cutting restoration duration in a given area, turning restoration speed into a dollar-denominated investment criterion.","Regulators can supplement SAIDI/CAIDI with a welfare-based, distribution-sensitive metric, creating a basis for equity-informed performance targets.","Because total costs are additive dollars, event-to-event comparisons become benefit-cost inputs for resilience spending across different storms and regions.","The consistent regressive pattern implies restoration plans that ignore income will systematically undervalue losses in low-income communities.","Recovery typologies from clustering can direct targeted aid to high-burden, slow-recovery communities even when their aggregate outage statistics resemble better-off areas."],"supporting_citations":[{"why":"Companion study that supplies the deprivation cost function coefficients in Eq. 1, the load-bearing valuation.","marker":"[32]"},{"why":"Econometric contingent-valuation basis for deprivation cost functions, the theoretical source of Eq. 1.","marker":"[29]"},{"why":"Discrete-choice foundation for estimating deprivation costs in humanitarian logistics.","marker":"[28]"},{"why":"Estimation of deprivation level functions, contributing to the functional form used here.","marker":"[31]"},{"why":"Independent estimate of US residential willingness to pay for resilience to long outages, used as a valuation benchmark.","marker":"[13]"},{"why":"Equity-focused metric for infrastructure disruption burden, the conceptual anchor for regressive-impact measurement.","marker":"[14]"},{"why":"ACS 5-year ZCTA-level demographic and income data used for all equity and clustering analyses.","marker":"[26]"},{"why":"Prior Hurricane Beryl outage-disparity analysis that supplies the event's outage dataset and motivates the case study.","marker":"[21]"}],"fun_headline_variants":["Ida outage social cost hits $1.5B, poor areas hardest","Restoration speed drives hurricane outage social costs","Four hurricanes: outage costs hit poor areas disproportionately","Ida's outage social cost: $1.5B, but restoration pace matters"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The dollar numbers stand on one cost curve—DC = 35.95t² + 107.84t + 71.89—estimated in a separate survey and applied unchanged to every ZIP code in four storms; if that curve misprices a day without power for a particular population, the totals and income comparisons change, though the qualitative regressive pattern could remain.","fun_headline_variants_meta":{"raw":{"variants":["Ida outage social cost hits $1.5B, poor areas hardest","Restoration speed drives hurricane outage social costs","Four hurricanes: outage costs hit poor areas disproportionately","Ida's outage social cost: $1.5B, but restoration pace matters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2692,"prompt_tokens":768,"completion_tokens":1924,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":1852}},"tokens_in":512,"tokens_out":1924,"duration_ms":15848,"temperature":1.0,"reasoning_tokens":1852,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:34:52.124747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-estimate the deprivation cost curve from willingness-to-pay data collected specifically in Hurricane Ida-affected parishes and compare it to the fixed curve; or test the framework by predicting total outage losses for one of the four events and comparing against observed insurance claims, business-interruption losses, or utility compensation payouts. A systematic gap between predicted and observed dollar losses would falsify the transferability of Eq. 1.","supporting_citations":[{"cited_title":"Estimating Deprivation Cost Functions for Power Outages During Disasters: A Discrete Choice Modeling Approach","cited_arxiv_id":"2506.16993","evidence_quote":"Discrete-choice foundation for estimating deprivation costs in humanitarian logistics."}],"review_version":1}