{"id":"304adf03-a163-48f6-90e8-377c9338cf80","arxiv_id":"1908.02377","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A step-detection and median-filter pipeline applied to individual distance-to-reference time series from call detail records identifies likely internally displaced persons after sudden-onset disasters without assuming destinations.","lead":"This paper describes a way to use mobile phone call records to identify people who likely left home after a disaster, by watching for a sudden jump in how far each person is from their usual location. Tested on three disasters, it flags tens of thousands of likely displaced people and shows more flags in harder-hit regions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Left-window rolling median in Section 3.7 delays every detected level shift by about three days, so the 'week after disaster' criterion is not anchored to actual displacement timing; a synthetic benchmark is needed.","rationale":"I read the paper as a methods proposal whose validity hinges on the distance curve being a faithful representation of location changes. The reader identified this as the weakest assumption; my stress-test found a concrete failure mode: the left-window rolling median mechanically delays step onset by three days. This is an internal property of the described algorithm, not a disagreement with consensus. It also explains why the post-hoc seven-day window appears to work: the anomaly in Figure 3 is smoothed and shifted, so a window chosen from that figure is measuring the filtered signal, not physical movement. The paper honestly acknowledges limitations and states that its IDP subset is not complete, but that caveat does not repair the timing bias: a pure subset requires that detected steps be true and correctly timed. The check is straightforward with synthetic data and does not require access to the proprietary CDR. I keep the reader's CONDITIONAL verdict because the method may be salvageable by centering the median or correcting for the lag and validating against controlled data, but the current claims are not established as presented.","tokens_in":14343,"tokens_out":8893,"duration_ms":101639,"concrete_test":"Generate 10,000 synthetic subscribers with known level shifts on known dates, including shifts on each of the seven days after the disaster, with daily call records and no missing days. Apply Sections 3.7-3.9 exactly: left rolling median with n=7, iterative median filter, and the diff rule. Record detected step dates. If the detected date is systematically t0+3 rather than t0, re-run the full IDP selection with a centered or right-aligned rolling median, or with a -3-day shift, and compare the Table 6 counts and Figure 3 step-date histograms. If these change materially, the reported IDP identification is partly an artifact of the imputation lag rather than a faithful record of displacement timing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the distance curve constructed in Sections 3.7-3.8 preserves the timing of real location changes. It does not. Step 1 of Section 3.7 replaces each daily distance with a left-window rolling median over n=7 days. For a subscriber whose distance jumps from 0 to C on day t0 and then stays at C, the rolling median on t0 is the median of six zeros and one C, which is 0; it is still 0 on t0+1 (five zeros, two C's) and t0+2 (four zeros, three C's), and only becomes C on t0+3. The iterative median filter of Section 3.8 preserves that delayed step, so the Section 3.9 rule reports the level shift at t0+3, not at t0. Consequently, the seven-day detection window in Section 3.10 is effectively shifted by about half the filter width: a true displacement on the disaster date is timestamped three days later, and displacements lasting fewer than roughly four days never leave the zero rolling median and are erased. The same construction can also turn a short trip in a noisy baseline into a flat artificial level (Appendix A, Example 2). Because the IDP counts in Table 6, the step-date curves in Figure 3, and the displacement distances in Figure 5 all use these step timestamps, the paper's central claim that detected level shifts mark disaster-induced displacement is not supported unless the lag is measured and corrected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method to identify internally displaced persons (IDPs) from call detail records after sudden-onset disasters. For each subscriber, the authors construct a daily 'distance curve' to a pre-disaster reference location, impute missing days, smooth the series with an iterative median filter, detect level shifts (steps) in the filtered signal, and then apply heuristic criteria to label individuals as IDPs. The method is applied to CDR data from the 2010 Haiti earthquake, the 2015 Nepal earthquake, and Hurricane Matthew in Haiti (2016). The paper reports the resulting IDP counts, spatiotemporal patterns of steps, regional distributions, and displacement distances, and it argues that the detected level shifts correspond to disaster-induced displacements.","tokens_in":14617,"tokens_out":5881,"duration_ms":59375,"significance":"The humanitarian application is important, and an individual-level trajectory method that can detect short-distance displacements would complement existing aggregate CDR studies and survey-based methods. The paper is clearly written and candid about several limitations, including the absence of false-positive/false-negative estimates and the post hoc choice of the seven-day post-disaster window. However, the central claim—that the detected level shifts actually mark disaster-induced displacements—is not validated against any ground-truth displacement data, and the signal-processing chain introduces a systematic timing error that biases the reported step dates. If the pipeline were corrected and externally validated, the method could be a valuable tool; as presented, the results in Tables 6 and 10 and Figures 3–5 cannot be interpreted as reliable IDP estimates.","major_comments":[{"comment":"The use of a left-window rolling median with n=7 shifts every level shift in the distance curve by approximately three days. For a jump from 0 to C at time t0, the rolling median remains 0 on t0, t0+1, and t0+2 and only becomes C at t0+3; the step rule in Section 3.9 then timestamps the event three days late. Consequently, the 'week after disaster' criterion in Section 3.10 is effectively a ten-day window for true displacements, and displacements lasting four days or fewer are erased before step detection. This delay propagates into the IDP counts in Table 6, the step-date curves in Figure 3, and the displacement distances in Figure 5, so the paper's central claim that detected level shifts mark disaster-induced displacement is not supported unless the lag is measured and corrected, for example with a synthetic benchmark or a centered window.","section":"Section 3.7, step 1; Section 3.8; Section 3.9"},{"comment":"The seven-day post-disaster window is chosen after examining the step-frequency figures in Section 4.1. Because the same data are used both to select the rule and to report the result, the observed concentration of IDP steps within the chosen window is partially an artifact of the selection. The text acknowledges this, but it remains a circularity. The lag described in the previous comment compounds the problem, since the 'anomalous week' visible in Figure 3 is itself shifted by the filter. The authors should justify the window independently, for instance from an out-of-sample disaster or from a pre-specified definition, or demonstrate that the qualitative conclusions are insensitive to the window length.","section":"Section 3.10, criterion 1"},{"comment":"The only empirical check of the IDP classification is a spatial correlation with reported affected regions. No ground-truth IDP data, survey-based displacement estimates, or independent IDP totals are used to compute precision, recall, or any error rate. The paper itself states in Section 6 that follow-up studies are needed to estimate true/false positive and negative rates. Without such validation, the counts in Tables 6 and 10 cannot be interpreted as estimates of displacement; at most they are the output of an unvalidated filter. The authors should add at least one external comparison, for example with IOM displacement-site records, survey estimates, or the earlier Haiti CDR study in reference [6], or explicitly reposition the paper as a method proposal with validation deferred to later work.","section":"Section 4.2 (Tables 7–9, Figure 4); Section 6"},{"comment":"The acknowledged processing error—implementing 'at least two call days away from the reference location' as 'two calls'—means the criterion can be satisfied by two calls made on the same day, weakening the intended assurance that the individual was away for longer than a single day. The authors should either correct the implementation and rerun the pipeline or quantify how many identified IDPs depend on this distinction.","section":"Footnote to Section 3.10, criterion 1"}],"minor_comments":[{"comment":"The first sentence reads 'whom we believe to be have been internally displaced'; the phrase 'to be have been' is a typo and should read 'to have been'.","section":"Abstract"},{"comment":"The listed pre- and post-disaster durations do not appear to add up to the stated study periods; for Haiti, 42 plus 182 days is 224 days, while the start and end dates span only about 196 days. Please clarify whether the durations include the disaster day and how the end dates are determined.","section":"Table 2"},{"comment":"The threshold for the median number of call days per week is said to be 'determined by looking at the distribution', but the actual cut-off values used for each dataset are not reported; please state them for reproducibility.","section":"Section 3.2.2"},{"comment":"The statement that the iterative median filter removes trips of less than (n+1)/2 days means that with n=7, trips shorter than four days are removed; this should be reconciled with the three-day minimum stay criterion in Section 3.10, since the filter may erase exactly the short displacements the criterion is meant to retain.","section":"Section 3.8"},{"comment":"The random tie-breaking for the modal reference location is described as 'relatively unimportant', but for subscribers with near-equal time at two locations in the pre-disaster period, the choice could affect whether a post-disaster step is detected; a sensitivity check would strengthen the claim.","section":"Section 3.4"},{"comment":"The explanation of the sharp drop followed by a peak in the Haiti panel refers to cell towers being out of service; marking the non-operational tower days directly on the plot would make the effect easier to verify.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's own acknowledgments of the post hoc window selection, the absence of validation, and the processing error in Section 3.10 are likely to be flagged by any careful reviewer. The left-window rolling-median delay identified in Section 3.7 is a technical flaw that affects all the central results and would need to be addressed before the paper is publishable as a validation study. If the authors can correct the timing issue and add an external or synthetic validation, a revised version may be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real step forward in applying CDR data to individual displacement, but the processing chain introduces a systematic lag in step timing that the paper never accounts for, and the seven-day window is chosen from the same data used to demonstrate the anomaly. The method is plausible but not yet supported.\n\nWhat's new: prior CDR disaster work mostly looked at aggregate flows; here they build a per-subscriber distance-to-reference curve, smooth it, and detect level shifts, which lets them catch short-distance displacements and requires no assumptions about destination. The pipeline is described in enough detail to re-implement, and the paper is honest about prioritizing precision over recall and about the SIM-as-person assumption.\n\nThe main problem is the signal processing. Section 3.7 builds the distance curve with a rolling median over a left window of n=7. If a subscriber jumps from 0 to C on day t0, the median stays at 0 until t0+3 because the window still contains more zeros than C's. The iterative median filter in Section 3.8 preserves that delay, and the step rule in Section 3.9 then timestamps the level shift three days late. So the 'week after the disaster' criterion in Section 3.10 is really catching steps up to ten days after the event, and displacements shorter than about four days are erased entirely. The IDP counts, step-date curves, and distance distributions all inherit that lag. The authors never measure or correct for it. This is separate from the circularity they admit: the seven-day threshold was chosen after looking at the very figures it is supposed to explain.\n\nThere's also no ground-truth validation. The only check is spatial correlation with reported affected regions, which is consistent but weak. The data are proprietary and not released, so external replication is impossible. A synthetic benchmark with known displacement times and durations would sort out the lag and the threshold sensitivity in a day's work, and I'd want to see that before trusting any of the numbers.\n\nThat said, the core idea is sound and the paper reads honestly. It's a good candidate for a rigorous review, not a desk reject; the flaws are fixable. I'd send it out, with the request that reviewers push hard on the timing analysis and threshold justification.","headline":"A useful individual-level CDR displacement method, but the left-window rolling median delays every detected step by ~3 days and the seven-day threshold is post hoc, so validation is needed before the counts are trusted.","tokens_in":15214,"tokens_out":2181,"would_cite":false,"duration_ms":23234,"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":"Internally displaced people can be identified from mobile phone records by detecting a level shift in each caller's daily distance-to-reference time series.","keywords":["call detail records","disaster response","internal displacement","mobile operator data","time series analysis","step detection","human mobility","piecewise-constant signal"],"falsifier":"Run the identical filtering and step detection on a pre-disaster control period of equal length in the Haiti dataset and count how many subscribers show a step in a randomly chosen seven-day 'disaster' window; if that number is of the same order as the 37,839 flagged after the earthquake, the step rule is detecting ordinary mobility rather than displacement.","tokens_in":14067,"feed_emoji":"📱","tokens_out":13111,"duration_ms":118232,"temperature":0.7,"pith_summary":"After a sudden-onset disaster, many people flee their homes but are not counted by traditional surveys. This paper claims that a reliable subset of those internally displaced people can be found directly in mobile phone data: each subscriber's movements are converted into a daily distance-to-reference curve, and a step (level shift) in that curve in the week after the disaster marks a change of stay location. The method needs no assumptions about where people fled to or how long they stayed away. The paper argues this can complement traditional survey methods by providing timely, large-scale assessments of displacement scale and characteristics.","feed_headline":"Phone records flag who was internally displaced","feed_subtitle":"A level shift in a caller's daily distance curve marks a change of stay location after a sudden disaster.","key_machinery":"The central machinery is the distance curve: the daily geodesic distance from a subscriber's pre-disaster reference location, treated as a piecewise-constant signal. Spatial clustering of nearby cell towers removes oscillation noise, an iterative median filter with a seven-day window smooths the curve into constant pieces while preserving sharp edges, and the step detection rule that a level shift occurs when the first difference of the filtered curve moves from zero to nonzero extracts the times of stay-location changes. This reduction turns a messy two-dimensional trajectory into a one-dimensional signal in which displacement is a single detectable edge.","core_discovery":"The paper's central claim is that disaster-induced internal displacement appears as a level shift in a piecewise-constant time series of an individual's distance to their pre-disaster 'reference location.' After filtering to frequent callers, clustering nearby cell towers, imputing missing days with rolling medians, and applying an iterative median filter with a seven-day window, the distance curve becomes a piecewise-constant signal. A step is declared at time $t$ when $\\mathrm{diff}(t-1)=0$ and $\\mathrm{diff}(t)\\neq 0$, meaning the distance changes abruptly from one constant level to another. Subscribers whose reference location lies in an affected region and who show such a step in the week after the disaster, with at least three days at the reference location in the prior week, are labeled internally displaced persons. The paper applies this to the 2010 Haiti earthquake, the 2015 Gorkha earthquake, and Hurricane Matthew in 2016, and finds the regional distribution of flagged IDPs aligns with reported severity.","pith_inferences":["If the step rule is valid, the same logic could date displacement onset more precisely than a one-week window, detecting secondary moves as repeated level shifts in the same subscriber's signal.","The method could be stress-tested by applying it to a control period with no disaster: a similar rate of flagged 'displacement' steps would imply the rule is capturing ordinary mobility, not disaster response.","The purity-focused filtering implies the flagged set is a censored sample, so any downstream estimate of total displacement would need adjustment for the known false-negative rate, a step the paper leaves to later work.","Because the distance curve collapses two-dimensional space to a scalar distance, two different destinations at the same distance are indistinguishable; adding directional information could separate moves toward or away from an epicenter or coastline."],"forward_implications":["Humanitarian responders can estimate the number and geographic origin of displaced people within weeks of a disaster, using data already collected for billing.","Short-distance displacements, under a few kilometres, that vanish in aggregate origin-destination studies become visible at the individual level.","Because no assumptions about destination or duration are needed, the method captures people who stay with relatives or in rented housing, not only those in formal camps.","The flagged IDP subset can be linked back to sequences of stay locations, enabling longitudinal studies of onward and secondary displacement.","The method transfers to other spatio-temporal datasets with a similar one-dimensional structure, a possibility the paper explicitly leaves open."],"supporting_citations":[{"why":"It supplies the finding that the location of the first or last call of a day is a good indicator of an individual's home location, justifying the daily sampling of the distance curve.","marker":"[5]"},{"why":"It demonstrates that CDR data from the 2010 Haiti earthquake can capture displacement and population movements, establishing the feasibility of the application and the individual-level approach.","marker":"[6]"},{"why":"It provides a prior CDR-based study of the 2015 Nepal earthquake, giving context and a comparison point for the method's output.","marker":"[7]"},{"why":"It supplies the standard trajectory-analysis stages that the method adapts to one-dimensional distance signals.","marker":"[19]"},{"why":"It provides the iterative median filter algorithm used to convert the irregular distance series into a piecewise-constant signal.","marker":"[26]"}],"fun_headline_variants":["Level shifts in call data flag displacement","Phone signals spot disaster-induced moves","Call records reveal who was uprooted","Detecting displacement from phone movement","When your phone shows you've been displaced"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline rests on the assumption that after clustering, imputation, and median filtering, a level shift in the distance curve really represents a person physically changing their stay location, rather than an artifact of missing days, tower outages, or the filter itself.","fun_headline_variants_meta":{"raw":{"variants":["Level shifts in call data flag displacement","Phone signals spot disaster-induced moves","Call records reveal who was uprooted","Detecting displacement from phone movement","When your phone shows you've been displaced"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1483,"prompt_tokens":958,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":464}},"tokens_in":574,"tokens_out":525,"duration_ms":6535,"temperature":1.0,"reasoning_tokens":464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:46:08.124334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical filtering and step detection on a pre-disaster control period of equal length in the Haiti dataset and count how many subscribers show a step in a randomly chosen seven-day 'disaster' window; if that number is of the same order as the 37,839 flagged after the earthquake, the step rule is detecting ordinary mobility rather than displacement.","supporting_citations":[{"cited_title":"Rapid and near real-time assessments of population displacement using mobile phone data following disasters: the 2015 nepal earthquake","cited_arxiv_id":null,"evidence_quote":"It provides a prior CDR-based study of the 2015 Nepal earthquake, giving context and a comparison point for the method's output."},{"cited_title":"Trajectory data mining: an overview","cited_arxiv_id":null,"evidence_quote":"It supplies the standard trajectory-analysis stages that the method adapts to one-dimensional distance signals."},{"cited_title":"Generalized methods and solvers for noise removal from piecewise constant signals","cited_arxiv_id":null,"evidence_quote":"It provides the iterative median filter algorithm used to convert the irregular distance series into a piecewise-constant signal."}],"review_version":1}