{"id":"615193ac-56b1-4806-ba19-6135a06bc57e","arxiv_id":"1908.02381","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Displaced populations after three disasters resettled at rates well described by the sum of two exponential decays, with half resettled in four to five weeks.","lead":"Researchers used mobile phone records to estimate when people displaced by three disasters returned to their normal routines. They found the return rate follows a simple two-part exponential pattern, with half of displaced people resettling within about a month.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pre-disaster self-validation in Appendix A shows the Section 2 estimator counts people as 'resettled' even with no disaster, so the post-disaster decay curves and their two-exponential fit are contaminated by a threshold-crossing artifact.","rationale":"The reader identified the unvalidated operational definition of resettlement as the weakest assumption, which is close to this concern. My stress-test sharpens it using the paper's own Appendix A: the estimator, when applied to a period with no disaster, still classifies most individuals as 'resettled' within a short time. That is not merely a missing ground-truth check; it is internal evidence that the definition conflates noise with recovery. If the pre-disaster null curve decays as fast as or faster than the post-disaster IDP curves, the two-exponential fit and the 'half within four to five weeks' headline could be produced by the threshold-crossing procedure alone, applied to stationary mobility noise. I do not claim the authors are wrong that post-disaster mobility changes; the control/vs-IDP differences in Figure 1 show some signal. But the central quantitative claim—the two-exponential decay law and its universality—is not supported unless the background first-passage artifact is quantified and subtracted. This preserves the conditional verdict but adds a specific, internally checkable reason for requiring revision before acceptance.","tokens_in":14864,"tokens_out":5807,"duration_ms":64554,"concrete_test":"Fit the same two-exponential model used in Section 3.2 to the Appendix A pre-disaster decay curves, at least for the control group and for Sstep, for each disaster. Compute the pre-disaster false-resettlement rate: the fraction of controls whose rolling mean first crosses the baseline within the first 5 pre-disaster weeks. If this null curve decays to zero with a half-life comparable to the 4–5 week post-disaster IDP half-lives, or if the two-exponential model fits the null curve as well as it fits the post-disaster IDP curves, then the post-disaster curves cannot be attributed to resettlement without subtracting this background artifact.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing premise is that the decay curves measure resettlement, not an artifact of the estimator. In Section 2, step 4, 'resettlement' is defined as the first post-disaster week in which the four-week rolling mean of a mobility metric is at or below the pre-disaster mean. Because that mean is computed from a noisy weekly series, any individual whose metric fluctuates will eventually cross this threshold even in the complete absence of displacement. The paper's own validation in Appendix A demonstrates this: running the same steps on the pre-disaster period produces decay curves that fall from 100% to near zero for both the IDP and control groups during an undisrupted period. The authors call this 'expected', but for a meaningful resettlement measure the undisrupted survival curve should remain near 100%—nobody is resettling in the pre-disaster baseline. The fact that it decays implies a high false-positive 'resettlement' rate driven by noise. Post-disaster curves are therefore the convolution of true return-to-normal times with this noise-driven first-passage process. Fitting f(t) = α1 exp(-β1 t) + α2 exp(-β2 t) to these curves and interpreting the parameters as two IDP subpopulations may be fitting the artifact. The parameter differences in Table 3 (Haiti β1 = 0.63 vs 0.22 for the other two disasters) are then hard to interpret substantively. Section 5 acknowledges that no ground-truth validation has been performed, but Appendix A provides internal evidence that the artifact is real and quantifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a method to estimate individual resettlement times after sudden-onset disasters using mobile phone call detail records. It defines resettlement as the first week after the disaster in which the four-week rolling mean of a mobility metric falls to or below the individual's pre-disaster mean, and applies this definition to three disasters (Haiti 2010, Nepal 2015, Hurricane Matthew in Haiti 2016) using four mobility metrics. The central claim is that the fraction of IDPs remaining disrupted decays as a sum of two exponentials, f(t) = α1 exp(−β1 t) + α2 exp(−β2 t), and that the decay rates are similar across disasters, with half of the displaced resettled within four to five weeks. The paper also compares the performance of the four metrics, argues that radius of gyration is unsuitable, and includes a control group and a pre-disaster self-validation as checks. The authors explicitly acknowledge the lack of ground-truth validation and several other limitations.","tokens_in":15181,"tokens_out":5841,"duration_ms":60443,"significance":"If the reported two-exponential decay law and cross-disaster similarity are real, the paper would make a valuable contribution to disaster-resilience measurement, potentially enabling near-real-time estimates of IDP numbers from CDR data. The use of actual operator data for three substantial disasters, the inclusion of a control group, and the attempt at internal validation are strengths; the paper also makes falsifiable predictions (parameter values and half-times). However, the significance is currently conditional: the pre-disaster self-validation in Appendix A shows that the resettlement estimator labels large fractions of the population as 'resettled' even in the absence of a disaster, which threatens the interpretation of the post-disaster decay curves. The lack of ground-truth validation, the absence of goodness-of-fit statistics, and the post hoc selection of the step-entropy metric further weaken the central claim.","major_comments":[{"comment":"The pre-disaster validation curves in Figure 8 decay from essentially 100% to near zero for both the IDP and control groups during an undisrupted period. Since the resettlement date is defined as the first week the four-week rolling mean falls at or below the pre-disaster mean (Section 2, step 4), this decay demonstrates that stochastic fluctuations alone cause a large fraction of individuals to be classified as 'resettled' even with no disaster. The post-disaster decay curves used in all subsequent analysis are generated by the same first-passage rule, so the two-exponential fits in Section 3.2 are likely contaminated by this threshold-crossing artifact. The authors describe the pre-disaster recovery as 'expected' (Section 3.1) but do not quantify the false-positive rate or correct for it; without such a correction, the central claim that the decay curves measure resettlement dynamics is not supported.","section":"Section 2 and Appendix A, Figure 8"},{"comment":"The claim that the curves 'fit very well' to f(t) = α1 exp(−β1 t) + α2 exp(−β2 t) is not backed by any goodness-of-fit statistic (R², residuals, or model comparison). The reported 1σ parameter errors are not a substitute for a fit-quality measure. In addition, the assertion of 'similar' decay rates across disasters is based on visual inspection of Figure 3; no statistical test for the equivalence of the β parameters is provided, and Haiti's β1 = 0.63 is roughly three times larger than the other two (0.22), which weakens the abstract's summary claim.","section":"Section 3.2, Table 3"},{"comment":"The decision to present only the step-entropy results after Section 3.1 is made after examining all four metrics, with one justification being that the step-entropy curve lies 'in the middle' of the others. This post hoc metric selection creates a risk of selection bias in the reported two-exponential parameters and half-times; the paper does not report the equivalent fits for the other metrics or apply a multiple-testing correction. To make the central result robust, the metric choice should be justified a priori or confirmed on held-out data.","section":"Section 4.2.2"},{"comment":"The method's key assumption is that 'disaster-induced disruption manifests as an increase only (not decrease) in the value of the mobility metric.' Section 4.2.1 then shows that for many IDPs disruption appears as a decrease in long-distance travel and an increase in short-distance travel, so that the radius-of-gyration metric can indicate recovery before true recovery. The paper acknowledges this for RoG but does not reconcile it with the original 'increase only' assumption for the remaining metrics. If other metrics are also sensitive to the mix of increases and decreases, the estimated resettlement dates are structurally biased.","section":"Section 2 vs. Section 4.2.1"},{"comment":"The paper acknowledges that no ground-truth validation has been performed and that individuals in camps or temporary accommodation may exhibit normal activity. This limitation, combined with the artifact documented in Appendix A, means that 'resettlement' as operationalized here is a proxy for the return of a mobility metric to its pre-disaster average, not an externally validated measure of resettlement. The central claim requires either an external validation or a redefined estimator whose pre-disaster false-positive rate is explicitly modeled and subtracted.","section":"Section 5"}],"minor_comments":[{"comment":"The displayed formula for the radius of gyration is incomplete: the summand should be (r_i − r_c)^2, not (r_i − r_c); the text also says 'mean absolute deviation' after presenting a standard-deviation-like expression, which is inconsistent.","section":"Section 2.1.1"},{"comment":"The solid/dotted line distinction between IDP and control groups in Figure 1 (and similar figures) is described only in the caption; labeling the lines directly or adding an in-figure legend would make the plots much easier to read.","section":"Figure 1"},{"comment":"The sentence 'half of the displaced residents have resettled back at their home after four months' is imprecise because Table 4 reports 17 weeks for Haiti; please state the number of weeks as well as or instead of months.","section":"Section 4.3.3"},{"comment":"The abstract's statement that half of the displaced resettled within four to five weeks should be explicitly qualified as referring to the step-entropy metric, since Figure 1 shows that other metrics give different decay curves.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and important problem, and the data are unique, but the pre-disaster self-validation in Appendix A is a serious internal threat to the central claim. The authors should be asked to reanalyze with a resettlement definition that accounts for noise-driven threshold crossings, to report goodness-of-fit and model comparisons, and to justify the metric-selection procedure. This is fixable within the manuscript's scope, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is worth reading but the headline regularity—displaced populations resettle as a sum of two exponentials with similar rates across three disasters—is not established by the analysis. The per-person baseline-return definition of resettlement time and the new step entropy metric are real additions, and the paper is clearly written and unusually honest about its limitations. The double-exponential description is an interesting empirical observation, not yet a usable prediction.\n\nWhat is actually new: previous CDR disaster work tracks aggregates or uses coarse 'return to admin region' criteria. Here, resettlement time is defined for each individual as the first post-disaster week in which the four-week rolling mean of a mobility metric returns to or below the pre-disaster mean. Step entropy is a sensible way to capture changes in travel frequency and short-distance moves, and it is a genuine methodological contribution. The comparison with [5] and [6] is fair and useful.\n\nThe soft spot is not small. The pre-disaster validation in Appendix A shows the same estimator produces decay curves that drop from 100% to near zero during an undisrupted period. That means 'resettled' is being assigned to people whose metric merely fluctuated below its own baseline. The post-disaster curves are therefore a convolution of true return-to-normal times with a noise-driven first-passage process. The authors call the pre-disaster decay 'expected', but for a displacement measure the undisrupted survival curve should stay near 100%; the fact that it decays shows a high false-positive rate. They never subtract this baseline or quantify it, so the two-exponential fit and the parameter values in Table 3 cannot be interpreted as describing two IDP subpopulations. This also undermines the 'similar across disasters' claim, because the estimator's crossing-time distribution may be similar across datasets for reasons unrelated to resettlement.\n\nOther, smaller issues: no goodness-of-fit statistics for the exponential model, the choice of step entropy was made after seeing the results, and 'similar' is judged by eye from three curves. These would be fixable if the artifact were handled.\n\nThe paper is for researchers working on CDR-based displacement estimation and humanitarian data analytics. It is a good source for the metric definition and a useful case study in why pre-disaster null expectations matter.\n\nMy recommendation: send it to peer review, but require a baseline-subtracted or matched-control survival analysis, explicit fit statistics, and an attempt at ground truth for at least a subset. Until then, the empirical regularity is not believable.\n\nBest,\n[Your name]","headline":"Useful applied paper with a new per-person resettlement metric, but the pre-disaster validation exposes a false-positive artifact that likely contaminates the two-exponential decay claim.","tokens_in":15703,"tokens_out":4627,"would_cite":true,"duration_ms":47334,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Displaced people after three disasters resettle as a two-exponential decay, with half resettled in four to five weeks.","keywords":["call detail records","disaster response","internal displacement","mobile operator data","mobility metrics","resettlement","resilience","exponential decay"],"falsifier":"Compare CDR-derived resettlement dates against field-survey data for a known disaster cohort: if substantial numbers of people still living in camps or temporary shelters are classified as resettled by the mobility threshold, the proxy and the two-exponential decay claim would be falsified.","tokens_in":14677,"feed_emoji":"📱","tokens_out":3429,"duration_ms":38292,"temperature":0.7,"pith_summary":"The paper uses mobile phone call detail records to estimate when internally displaced people (IDPs) return to their normal mobility after a sudden-onset disaster. It finds that the fraction of IDPs who have not yet resettled decays over time as the sum of two exponential curves, one fast and one slow, and that this pattern is similar across the 2010 Haiti earthquake, the 2015 Nepal earthquake, and Hurricane Matthew in 2016. Half of the identified IDPs are estimated to have resettled within four to five weeks. If this shape holds for other disasters, the number of people still displaced at any time could be inferred from an initial estimate of displacement immediately after the disaster, and recovery rates could be compared across events without relying on field surveys alone.","feed_headline":"Half of disaster-displaced people resettle within five weeks","feed_subtitle":"Mobility data from call detail records fit a two-exponential recovery curve across three disasters.","key_machinery":"The central object is the two-exponential decay curve fitted to the weekly fraction of IDPs who have not yet resettled. It is computed from a mobility threshold: an individual is considered resettled in the first week after the disaster when their four-week rolling mean of a mobility metric (radius of gyration, logarithmic radius of gyration, temporal-uncorrelated entropy, or step entropy) drops to or below the pre-disaster average. The step-entropy metric is used for the main results because it is consistent with the stay-location method used to detect IDPs, sits between the other metrics, and responds to travel frequency and short-distance moves.","core_discovery":"The paper claims that the resettlement rate of a disrupted population can be modelled very well by f(t) = α1 exp(−β1t) + α2 exp(−β2t), where t is weeks after the disaster and the two exponentials represent a faster-recovering group and a slower-recovering group. Applying four mobility metrics to call detail records from three disasters, the authors define an individual's resettlement date as the first post-disaster week in which the four-week rolling mean of the mobility metric falls to or below its pre-disaster mean. The resulting decay curves for the IDP group are clearly distinct from a control group, and for all three disasters half of the displaced persons are resettled within four to five weeks. The paper also argues that radius of gyration is less suitable because disaster disruption often appears as a decrease in long-distance travel and an increase in short-distance travel, which RoG can misread as recovery.","pith_inferences":["Because the two-exponential fit collapses many individual trajectories into two rates, a natural next test is whether the fast and slow groups correspond to observable factors such as home damage severity, socioeconomic status, or whether people returned home versus resettled elsewhere.","The same rolling-mean threshold could be applied to other sudden disruptions — disease outbreaks, conflict displacement, or climate evacuations — to test whether a universal recovery curve exists, though the paper does not claim this.","A sharper validation would re-derive resettlement dates with different window lengths and a sustained-below-baseline criterion; if the two-exponential shape and the four-to-five-week half-life persist, the conclusion is much stronger.","The control group likely contains some true IDPs, so the reported IDP–control contrast is conservative; a cleaner unaffected-region control would probably show an even larger separation."],"forward_implications":["If the two-exponential shape is universal, the number of IDPs still displaced at any time can be estimated from an initial displacement count alone, without knowing resettlement locations.","The fitted parameters give a quantitative, comparable measure of disaster resilience across events and across administrative regions, highlighting which areas recover slowest.","Mobility-based monitoring can complement field surveys with near-real-time, interview-free estimates, and can capture displaced people who avoid official camps.","The metric comparison warns that relying on radius of gyration alone underestimates the population still needing support because it misses increased short-distance travel frequency.","The method provides a way to compare return-to-home versus resettle-elsewhere rates, suggesting that recovery of mobility and finding a new home take roughly similar times."],"supporting_citations":[{"why":"Supplies the IDP-detection method and the distance curves from which all mobility metrics and resettlement dates are computed.","marker":"[1]"},{"why":"Provides the earlier Haiti earthquake mobility findings using radius of gyration and entropy, which this paper extends and partially revises.","marker":"[5]"},{"why":"Provides the Nepal earthquake displacement dataset and return-rate estimates used for qualitative comparison.","marker":"[6]"},{"why":"Defines the original entropy measures (random, temporal-uncorrelated, true) from which the paper adapts its entropy metrics.","marker":"[12]"},{"why":"Supplies Hurricane Katrina return-migration rates used to benchmark how much slower resettlement can be in a different disaster context.","marker":"[13]"}],"fun_headline_variants":["Call records show half of displaced resettle in five weeks","Two-exponential recovery curve fits disaster displacement data","Mobility metrics reveal half of IDPs resettle by week five","CDR analysis: half of disaster IDPs resettle within five weeks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an individual has resettled exactly when their four-week rolling average mobility metric first falls to or below its pre-disaster average, an operational proxy that the paper cannot validate against ground-truth resettlement records.","fun_headline_variants_meta":{"raw":{"variants":["Call records show half of displaced resettle in five weeks","Two-exponential recovery curve fits disaster displacement data","Mobility metrics reveal half of IDPs resettle by week five","CDR analysis: half of disaster IDPs resettle within five weeks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1295,"prompt_tokens":1035,"completion_tokens":260,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":190}},"tokens_in":651,"tokens_out":260,"duration_ms":3529,"temperature":1.0,"reasoning_tokens":190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:45:40.269462+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare CDR-derived resettlement dates against field-survey data for a known disaster cohort: if substantial numbers of people still living in camps or temporary shelters are classified as resettled by the mobility threshold, the proxy and the two-exponential decay claim would be falsified.","supporting_citations":[{"cited_title":"Detecting individual internal displacements following a sudden-onset disaster using time series analysis of call detail records, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the IDP-detection method and the distance curves from which all mobility metrics and resettlement dates are computed."},{"cited_title":"Predictability of population displacement after the 2010 haiti earthquake","cited_arxiv_id":null,"evidence_quote":"Provides the earlier Haiti earthquake mobility findings using radius of gyration and entropy, which this paper extends and partially revises."},{"cited_title":"Rapid and near 22 real-time assessments of population displacement using mobile phone data following disasters: the 2015 nepal earthquake","cited_arxiv_id":null,"evidence_quote":"Provides the Nepal earthquake displacement dataset and return-rate estimates used for qualitative comparison."},{"cited_title":"Limits of predictability in human mobility","cited_arxiv_id":null,"evidence_quote":"Defines the original entropy measures (random, temporal-uncorrelated, true) from which the paper adapts its entropy metrics."},{"cited_title":"Race, socioeconomic status, and return migration to new orleans after hurricane katrina","cited_arxiv_id":null,"evidence_quote":"Supplies Hurricane Katrina return-migration rates used to benchmark how much slower resettlement can be in a different disaster context."}],"review_version":1}