{"id":"84747677-cb71-4aff-910d-43a7bf5624e5","arxiv_id":"1908.07384","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Facebook posting rates of local businesses can be used to estimate how long a region's businesses stay closed after a natural disaster.","lead":"This paper measures business recovery after natural disasters by tracking how often local businesses post on Facebook. The method could give disaster managers a simple, fast way to see whether a city's businesses are reopening.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated proxy: reduced Facebook posting after a disaster may reflect power/Internet loss or evacuation rather than business closure; a control-page test is needed.","rationale":"The single most load-bearing concern is the semantic link between posting activity and business open/closed status. Without validating that link, no amount of signal processing can establish the central nowcasting claim. The paper has genuine strengths: a simple, parameter-light method; public social media data; three distinct disasters; and sensitivity analyses. Those strengths make the work plausible, but they do not eliminate the confound of infrastructure loss and population displacement. The reader's conditional verdict is appropriate: the paper should not be rejected outright, but it should not be accepted until the proxy is tested against a control group that experiences the same external conditions without 'closing.' My recommended verdict therefore remains CONDITIONAL, i.e., UNCHANGED from the reader's verdict, with the added condition of the control-page test.","tokens_in":16635,"tokens_out":6081,"duration_ms":71186,"concrete_test":"Apply the exact four-step pipeline (Sects. 3.3-3.7) to a matched set of Facebook pages in the three cities that are not businesses—e.g., schools, churches, community organizations—using the same collection method and time window. These pages face the same power/Internet/evacuation conditions but do not 'close' for economic reasons. If control pages show comparable dips in r_U(t) and detected events of about 50/118/52 days after the Gorkha earthquake, Hurricane Maria, and the Chiapas earthquake, the signal is digital silence and the central claim fails. If control pages show no such events while business pages do, the proxy is supported. This requires only data the authors already hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim depends on the assumption, stated in the Discussion, that 'businesses tend to publish more posts when they are open and fewer when they are closed.' This premise is never independently calibrated at the business level. In all three case studies, the detected 'downtime' is a period of reduced aggregate posting, but that period coincides with strong confounders: in Puerto Rico the main text reports 45% of islanders without power for three months, so even open businesses could not post; in Kathmandu the cited mobile-phone study reports roughly 60 days of net population outflow, so posting could be absent because owners/employees were displaced rather than because businesses were closed. The validations do not remove this ambiguity: Kathmandu surveys come from the authors' own prior field mission and show large standard deviations; San Juan 'ground truth' is obtained by applying the same elbow method to a different proxy (cruise tourism); Juchitán rests on 16 survey responses. Consequently, the reported 50, 118, and 52-day values could measure digital silence or infrastructure loss rather than economic downtime. A control set of non-business pages subject to the same outages would disambiguate this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework to estimate the post-disaster 'downtime' of small businesses by analyzing the aggregated Facebook posting activity of business pages in three affected regions: Kathmandu (2015 Gorkha earthquake), San Juan (2017 Hurricane Maria), and Juchitán de Zaragoza (2017 Chiapas earthquake). The method has three phases: data collection via the Facebook Graph API, data processing that transforms each business's daily post counts through a Probability Integral Transform, aggregation, shift/rescale to a standard normal, and a final PIT to a uniform series, and downtime detection using a rolling weekly mean below a threshold selected via an 'elbow method.' The authors report downtimes of 50, 118, and 52 days for the three events, validate against field surveys, tourism data, and a small Facebook survey, and provide a real-time simulation that crops data at successive weeks. The central claim is that downtime can be nowcast in real time from aggregated public posting rates without NLP.","tokens_in":16862,"tokens_out":3124,"duration_ms":32115,"significance":"If the method holds, it would offer a low-cost, scalable, and near-real-time indicator of post-disaster economic recovery, filling a gap in current practice where recovery is measured through slow and expensive surveys. The paper contributes a clearly specified, reproducible pipeline for turning raw Facebook timestamps into a normalized anomaly score, and it is commendable for collecting and processing large datasets (11,818 businesses in Kathmandu, 10,894 in San Juan, 1,728 in Juchitán) and for including sensitivity analyses on business subsets and posting-rate filters. The three-case-study design across different disaster types and socioeconomic contexts strengthens the plausibility that the method generalizes. However, the empirical support for the core proxy assumption—that posting activity equals business open/closed status—is indirect and confounded, and the real-time validation is internally circular, so the headline numerical claims should be treated with caution until stronger external validation is provided.","major_comments":[{"comment":"The threshold T* used to define a downtime event is selected from the same data used for event detection, via an 'elbow method' whose statistical properties are not established. The detected event durations (50, 118, 52 days) are therefore conditional on a threshold that has no stated uncertainty or null-model justification. To make the central claim load-bearing, the authors should demonstrate on synthetic data with known downtime periods that the elbow method recovers the true threshold and duration with bounded error, and they should report a sensitivity analysis over a range of T values for the three case studies.","section":"Section 3.7, Figure 5e, Tables 2-4"},{"comment":"The real-time validation is circular: the 'ground truth' d_GT(t) is defined as min(t, d*), where d* is the method's own final estimate using all data. The RMSD therefore measures how quickly the real-time estimate converges to the batch estimate, not how accurately it tracks actual business reopening. The authors should instead compare d_RT(t) against independent ground truth (e.g., the survey and tourism data used later) at each t, or at least report the range of d* consistent with those external anchors (e.g., 40-56 days for Kathmandu). Without such a comparison, the claim of 'accurate real-time estimates' is unsupported.","section":"Section 1.5, Eq. for RMSD, Figures 4, 10, 11"},{"comment":"The framework's foundational premise—that businesses post more when open and less when closed—is never independently calibrated at the business level. The paper's own context introduces major confounders: in Puerto Rico, 45% of islanders were without power for three months, and in Kathmandu, mobile-phone data show roughly 60 days of net population outflow. Both would suppress posting even for businesses that remain physically open, so the detected 'downtime' could measure digital silence or infrastructure loss rather than closure. The authors should provide a control test using non-business pages (or pages of businesses independently known to be open) in the same regions, or a per-business comparison of posted status against a small ground-truth survey, to disentangle the proxy from these confounders.","section":"Section 2 and Section 3.1-3.3"},{"comment":"The single-business PIT uses the empirical CDF of the 'current year,' which includes days after the disaster. This means the transformation for post-event days is contaminated by the anomaly the method is designed to detect: a business that posts zero for many days will have those zeros mapped through a CDF inflated by the very absence of posts, biasing the mid-quantiles upward and potentially shortening the detected downtime. The authors should use a pre-event reference period (e.g., the previous calendar year) or a causal expanding-window CDF that excludes the evaluation window, and show that the reported downtimes are robust to this choice.","section":"Section 3.3, Step 1"},{"comment":"The Juchitán validation rests on 16 survey responses out of 52 contacted, with no description of how the 52 were selected, no non-response analysis, and no confidence interval on the reported 63-day average closure time. Similarly, the Kathmandu survey means have very large standard deviations (e.g., Area A: mean 29, SD 44), so the claimed agreement between the method and surveys (e.g., 33 days vs. 28 days) is not statistically meaningful as reported. The authors should provide response-bias considerations, confidence intervals, and, if possible, a comparison at the business level rather than only at the regional level.","section":"Section 3.8 (Facebook surveys) and Supplementary Table 1"}],"minor_comments":[{"comment":"The abstract states the method estimates downtime 'in real time' but the real-time section simulates with cropped data rather than truly deploying in real time; consider clarifying this in the abstract and text.","section":"Abstract and Results"},{"comment":"The paper reports 52 days for Juchitán while also mentioning 66 days if the 12-day All Saints' Day period is included; please state explicitly which number is the headline figure and how holidays contiguous with an event are treated in the method.","section":"Section 1.3 and Table 4"},{"comment":"The linear trend removal in Step 3 is ad hoc and will also remove any slow recovery trend in the signal; if reopening is gradual rather than abrupt, this step could truncate the estimated downtime. The authors should discuss this possibility and test with simulated gradual-recovery data.","section":"Section 3.5"},{"comment":"The statement that mobile-phone data show 'roughly 60 days of net population outflow' from Kathmandu is presented as a validation anchor, but population displacement is a different quantity than business closure; the text should clarify why this is an upper bound for downtime and not a direct validation.","section":"Section 3.1 and Discussion"},{"comment":"The sensitivity tables show that downtime estimates fluctuate with business filters (e.g., Juchitán drops to 10 days at 400 posts/rate 0), yet the main text claims results are 'not affected by the filtering.' Please reconcile these numbers and state conditions under which the estimate is stable.","section":"Supplementary Tables 5-7"},{"comment":"Since the raw data cannot be shared due to Facebook's Platform Policy, the authors should consider depositing the derived aggregate time series and the code used for the entire pipeline in a public repository to enable reproducibility; currently only 'available upon request' is offered.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a novel and potentially impactful idea, and the data collection and preprocessing are substantial. However, the validation strategy has a clear circular component in the real-time evaluation, and the core proxy assumption is not independently tested, which in this field (disaster recovery) is a correctness-risk issue, not merely an editorial preference. I would like to see a revision that addresses the threshold-dependence, the causal PIT normalization, and a control-page or business-level validation; if the authors can provide those, the paper could become a strong contribution. The work fits the journal's scope as a data-driven sociophysical study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper has a genuinely new idea — using aggregated Facebook business-page posting rates as a real-time proxy for business downtime after disasters, with no NLP required. The three case studies (Kathmandu, San Juan, Juchitán) are well chosen, and the reported downtimes align roughly with independent anchors: Kathmandu mobility data, San Juan tourism data, and a small Facebook survey in Mexico. The normalization pipeline (per-business PIT, shift/rescale, aggregate PIT) is standard but carefully applied, and the sensitivity analysis with business subsampling is a plus. The authors are also candid about their core assumption: businesses post more when open.\n\nThe soft spots are real, and one is load-bearing. The “real-time” validation in Section 1.5 compares the real-time estimate to d_GT(t) = min(t, d*), where d* is the method’s own final estimate from all data. That is circular; it only shows the real-time estimate converges to its own final answer, not that the answer is correct. Second, the threshold T* is chosen from the same data via an elbow heuristic, so the detected events are not independent of the detection rule. Third, the proxy is unvalidated at the business level: reduced posting could reflect power loss, internet outage, or evacuation rather than closure. The text itself notes 45% of Puerto Rico was without power for three months, and Kathmandu had roughly 60 days of population outflow. A control set of non-business pages subject to the same outages would disambiguate, but none is included. The Juchitán validation rests on 16 survey responses; San Juan’s tourism data is processed with the same elbow method, so it is consistency, not independent confirmation.\n\nAlso minor: the Step 1 PIT uses the CDF of the current calendar year, which for pre-event days includes post-event days in the full-data analysis — a lookahead that could slightly distort the transformation. No code or raw data is provided (Facebook policy), only derived data on request, which limits reproducibility.\n\nBottom line: the central idea is plausible and worth testing further, but the current evidence does not establish that the detected dips are business closures rather than digital silence. I would send it to peer review — a good referee can push for a control-page calibration or external ground truth — but expect major revision. It is a solid contribution for disaster-response researchers, not a definitive method.","headline":"A plausible new proxy for business downtime from Facebook posting rates, but the real-time validation is circular and the proxy is not independently calibrated.","tokens_in":17360,"tokens_out":2062,"would_cite":false,"duration_ms":21293,"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":"A region's post-disaster business downtime can be nowcast from aggregated Facebook posting rates alone, without reading any post contents.","keywords":["social media","disaster recovery","downtime estimation","nowcasting","Facebook","probability integral transform","time series anomaly detection","business closure"],"falsifier":"Compare posting rates against a known-answer set of businesses with independently recorded open/closed dates in an area where connectivity was unaffected: if posting rates stay high while businesses are closed, or drop while businesses are open, the proxy fails.","tokens_in":16419,"feed_emoji":"📉","tokens_out":10082,"duration_ms":89732,"temperature":0.7,"pith_summary":"Even though post contents are never read, the timestamps of public Facebook posts by local businesses carry a measurable recovery signal: after a disaster, aggregate posting rates drop and stay low for a window that tracks business closures. The paper claims this window — the 'downtime' of small businesses — can be estimated automatically and in near real time from the posting time series alone, with no surveys, interviews, or semantic analysis. The claim is demonstrated on three very different events: the 2015 Gorkha earthquake in Kathmandu (50 days downtime), Hurricane Maria in San Juan (118 days, 103 attributable to the hurricane and 14 to the Christmas–New Year period), and the 2017 Chiapas earthquake in Juchitán (52 days). Such a method matters because traditional recovery indicators are expensive, slow, and hard to scale, whereas this one runs from public data and can be updated weekly as the event unfolds.","feed_headline":"Facebook post counts reveal how long disaster closures last","feed_subtitle":"Timestamps alone, with no text analysis, measure business downtime in three disaster zones.","key_machinery":"The load-bearing machinery is a four-step normalization that turns a messy, non-stationary count series into a stationary signal in which anomalous low-activity periods stand out uniformly. (1) Each business's daily post count is replaced by its mid-quantile within that calendar year — a per-business Probability Integral Transform — which makes businesses with very different posting rates comparable and suppresses outlier days. (2) The daily sum of these mid-quantiles is shifted and rescaled by the mean and standard deviation of the Irwin-Hall distribution (the distribution of a sum of independent Uniform(0,1) variables, $n(t)/2$ and $\\sqrt{n(t)/12}$), so that under normal conditions the aggregated series is approximately standard Normal. (3) A linear trend is removed and the variance is rescaled to one, correcting for residual correlation among businesses and for errors in estimating how many businesses are active. (4) The standard Normal CDF maps the series to Uniform(0,1), giving $r_U(t)$. A downtime event is defined as a period longer than seven days where the weekly rolling mean of $r_U(t)$ lies below a threshold $T^*$, and $T^*$ is fixed by the 'elbow' of the curve counting how many events are detected at each threshold. This design is what lets the same pipeline work across languages, posting rates, and calendar cultures, and it is also what makes the method independent of post content.","core_discovery":"The central discovery is that the aggregate volume of Facebook posts by local businesses drops sharply after a natural disaster and remains depressed for a measurable period, and that this period matches independent estimates of when businesses were actually closed. By converting each business's daily post count into a mid-quantile of that business's own posting history within the calendar year, summing across businesses, and then shifting and rescaling the sum to the mean and variance expected for a sum of independent uniform variables, the authors obtain a detrended, normalized signal in which normal fluctuations are uniform on [0,1]. Sustained low values of the weekly rolling mean of this signal, below a threshold chosen by an 'elbow' rule, are counted as downtime events. The method reports 50 days for Kathmandu, 118 days for San Juan, and 52 days for Juchitán, and in a real-time simulation that uses only data available up to week t after the event, the estimates track the final downtime with root-mean-square errors of 2.75, 0.88, and 3.5 days respectively.","pith_inferences":["A natural next test is to apply the pipeline to non-disaster closures whose dates are known exactly (holiday shutdowns, strikes, renovations) to see whether the posting dip's length and timing reproduce those dates; if yes, the proxy is about operation, not trauma.","The per-business quantile normalization is platform-agnostic in principle, so the same idea could be tried on Google Business profiles, Yelp check-ins, or Uber pick-up volumes, but the 'open means posting' correlation would have to be re-established for each new data source.","In regions where disasters also knock out internet and mobile networks, the method may overestimate downtime because it cannot distinguish 'closed' from 'cannot post'; combining the posting signal with mobility or satellite data would let a user separate those two causes.","A fully causal real-time deployment would need a threshold fixed before the disaster strikes, whereas the paper's real-time simulation still estimates the threshold from data cropped after the event; testing with a pre-specified threshold, or with a threshold updated only from pre-disaster data, is a direct next step."],"forward_implications":["Downtime can be nowcast during an unfolding disaster: using only posts published up to week $t$, the method tracks the final downtime with RMS errors of 2.75 days (Kathmandu), 0.88 days (San Juan), and 3.5 days (Juchitán), all within the stated plus/minus one week accuracy.","The signal needs no text: timestamps alone carry the information, so the approach transfers across languages and does not require storing or reading post contents.","The same anomaly-detection pipeline is explicitly general: any aggregated time series built from many individual components — mobile phone call volumes, web-page visits, perhaps foot-traffic counts — could in principle be processed the same way to expose anomalous low-activity periods.","Downtime estimates remain stable when the number of businesses is large (thousands), and are still accurate with a few hundred businesses that post more than about once per week, though very small samples tend to underestimate the true downtime.","Cultural closures appear in the same framework (Dashain in Nepal, Christmas–New Year in Puerto Rico, All Saints'/Revolution Day in Mexico), which both supports the method's validity and shows that disaster-related closures can be separated from calendar-driven ones by looking at the timing of the detected events."],"supporting_citations":[{"why":"Supplies the Kathmandu field-survey downtime data used as the lower-bound validation (40 days for the city, 28 days for Region A) and motivates business-downtime measurement.","marker":"[11]"},{"why":"Provides the mobile-phone displacement estimate that supplies the upper-bound comparison (~56 days) for Kathmandu downtime.","marker":"[22]"},{"why":"Supplies official cruise-passenger statistics for Old San Juan used to validate the San Juan downtime estimate (102 days tourism vs 118 days Facebook).","marker":"[26]"},{"why":"Defines 'downtime' within the Performance-Based Engineering framework, the concept the paper adapts to small businesses.","marker":"[5]"},{"why":"Frames the recovery definitions used to decide what 'downtime' measures after a disaster.","marker":"[6]"},{"why":"Provides the Probability Integral Transform used in steps 1 and 4 of the data-processing pipeline.","marker":"[38]"},{"why":"Documents the large number of local business pages on Facebook (65 million), the data source that makes the method scalable.","marker":"[37]"}],"fun_headline_variants":["Facebook silence reveals disaster recovery time","Posting lulls measure business downtime after disasters","Downtime from Facebook post gaps in disaster zones","How long shops close? Check their Facebook silence","Disaster recovery measured in Facebook post pauses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole framework rests on the premise that businesses post more when they are open and less when they are closed, meaning the measured drop in posting activity is a proxy for closure rather than for loss of internet access, evacuation, or a shift to another platform.","fun_headline_variants_meta":{"raw":{"variants":["Facebook silence reveals disaster recovery time","Posting lulls measure business downtime after disasters","Downtime from Facebook post gaps in disaster zones","How long shops close? Check their Facebook silence","Disaster recovery measured in Facebook post pauses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1148,"prompt_tokens":901,"completion_tokens":247,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":179}},"tokens_in":517,"tokens_out":247,"duration_ms":3495,"temperature":1.0,"reasoning_tokens":179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:19:37.288870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare posting rates against a known-answer set of businesses with independently recorded open/closed dates in an area where connectivity was unaffected: if posting rates stay high while businesses are closed, or drop while businesses are open, the proxy fails.","supporting_citations":[{"cited_title":"De Luca , D","cited_arxiv_id":null,"evidence_quote":"Supplies the Kathmandu field-survey downtime data used as the lower-bound validation (40 days for the city, 28 days for Region A) and motivates business-downtime measurement."},{"cited_title":"Wilson, E","cited_arxiv_id":null,"evidence_quote":"Provides the mobile-phone displacement estimate that supplies the upper-bound comparison (~56 days) for Kathmandu downtime."},{"cited_title":"Accessed: 2019-02-11","cited_arxiv_id":null,"evidence_quote":"Supplies official cruise-passenger statistics for Old San Juan used to validate the San Juan downtime estimate (102 days tourism vs 118 days Facebook)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines 'downtime' within the Performance-Based Engineering framework, the concept the paper adapts to small businesses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Frames the recovery definitions used to decide what 'downtime' measures after a disaster."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Probability Integral Transform used in steps 1 and 4 of the data-processing pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the large number of local business pages on Facebook (65 million), the data source that makes the method scalable."}],"review_version":1}