pith. sign in

arxiv: 2604.14657 · v1 · submitted 2026-04-16 · 📊 stat.AP

Evacuation destination choices during Hurricane Ian: A direct demand modeling approach

Alessandra Recalde , Luyu Liu , Xiaojian Zhang , Sangung Park , Shangkun Jiang , Xilei Zhao This is my paper

Pith reviewed 2026-05-10 09:42 UTC · model grok-4.3

classification 📊 stat.AP
keywords evacuation behaviorhurricane evacuationmobile device datasocial vulnerability indexdirect demand modeldestination choicetravel impedance
0
0 comments X

The pith

Mobile location data reveals that vehicle availability, English proficiency, and road density significantly shape hurricane evacuation destination choices.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a zone-to-zone direct demand model at the census tract level using mobile device location data from Hurricane Ian. It directly incorporates Social Vulnerability Index variables to measure how they affect both the decision to evacuate from an origin tract and the attractiveness of destination tracts. The model shows that vehicle availability, residence in group quarters, road density, and English proficiency have statistically significant effects, while longer travel distances strongly reduce the likelihood of choosing a particular destination. This replaces survey-based methods that suffer from recall bias with observed movement patterns. Understanding these links matters because it can inform targeted emergency planning that accounts for social barriers to evacuation.

Core claim

Using mobile device location data from Hurricane Ian, a zone-to-zone direct demand model at the census tract level is estimated that includes Social Vulnerability Index variables. Vehicle availability, residence in group quarters, road density, and English proficiency exert significant effects on evacuation demand by influencing both origin departure rates and destination attractiveness. Travel impedance measured by distance substantially reduces the probability of longer trips.

What carries the argument

A zone-to-zone direct demand model that uses mobile device location data to estimate evacuation flows between census tracts while incorporating Social Vulnerability Index variables as predictors of both origin and destination effects.

If this is right

  • Emergency plans should prioritize vehicle access programs in tracts with low car ownership to increase evacuation rates.
  • Public communications during hurricanes need to address language barriers in tracts with low English proficiency to improve destination choices.
  • Higher road density tracts act as stronger attractors, suggesting infrastructure investments could influence where evacuees go.
  • Distance remains a dominant deterrent, so pre-positioning shelters closer to vulnerable origins could raise compliance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same modeling approach could be tested on other hurricane events to check whether the same social vulnerability factors consistently dominate.
  • If the mobile data sample under-represents certain groups, policies based on these results might overlook the most vulnerable populations.
  • Combining this model with real-time traffic data could enable dynamic destination recommendations during an event.

Load-bearing premise

Mobile device location data provides an unbiased and representative sample of evacuation movements for all population groups.

What would settle it

If an independent dataset such as post-storm traffic sensor counts or household surveys shows evacuation rates and destination distributions that differ substantially from the model's predictions after controlling for the same variables, the estimated effects would be undermined.

Figures

Figures reproduced from arXiv: 2604.14657 by Alessandra Recalde, Luyu Liu, Sangung Park, Shangkun Jiang, Xiaojian Zhang, Xilei Zhao.

Figure 1
Figure 1. Figure 1: Method overview. 4 Case Study and Data Our case study focuses on Hurricane Ian, which ranks as the third costliest and one of the deadliest hurricanes in United States history (NOAA Climate.gov, 2022). This devastating Category 5 hurricane landed in Lee County, which accounted for nearly half of the deaths and reportedly issued evacuation orders less than 36 hours before Ian’s landfall (Fleischer, 2022). T… view at source ↗
Figure 2
Figure 2. Figure 2: The chart flow of the evacuation behavior inference algorithm. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The chart flow of the evacuation destination inference algorithm. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Inter- and intra-county origin–destination (OD) flow maps of evacuee travel during Hurricane Ian. The left panel illustrates inter-county movements across Florida, with lines representing trips between census tracts, most of which were to nearby areas. The right panel zooms in on Lee County, highlighting dense intra-county flows. 4.3 Social vulnerability data In addition to mobile location data, we incorpo… view at source ↗
Figure 5
Figure 5. Figure 5: Spatial distribution of evacuees. 5.2 SVI Variables in our model We estimated a multiplicative direct-demand model in log-linear form using evacuee counts between each tract pair as the re- sponse variable. Predictors were drawn from the CDC SVI component measures, the built environment, and distance between census tract centroids (m), with a selective transformation applied to each predictor (natural log … view at source ↗
read the original abstract

Hurricanes are causing unprecedented damage to the natural environment, infrastructure, and communities. Understanding evacuation behavior is essential for improving emergency preparedness. Past studies have relied on surveys and interviews, which are prone to recall bias. Additionally, they urge incorporating social vulnerability in evacuation research, emphasizing its impact on evacuation capability and destination choice. This study addresses these gaps by analyzing evacuation behavior using mobile device location data from Hurricane Ian, one of Florida's deadliest hurricanes, and directly incorporating variables from the Social Vulnerability Index (SVI) into a zone-to-zone (census tract level) evacuation demand model. We find that vehicle availability, residence in group quarters, road density, and English proficiency have significant effects on evacuation demand, shaping both the ability to evacuate from origin tracts and the attractiveness of destination tracts. Travel impedance, measured as distance, also plays a significant role, with evacuees substantially less likely to travel longer distances.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper develops a zone-to-zone direct demand model for evacuation flows during Hurricane Ian at the census-tract level, using mobile-device location data to estimate origin and destination effects. It incorporates SVI-derived covariates (vehicle availability, group-quarters residence, road density, English proficiency) and a distance impedance term, reporting statistically significant coefficients that shape both the propensity to evacuate from an origin tract and the attractiveness of destination tracts.

Significance. If the data and specification are shown to be robust, the work supplies one of the first large-scale, spatially disaggregated empirical quantifications of how social-vulnerability factors influence evacuation destination choice. This could inform targeted emergency messaging and resource allocation, and it demonstrates the feasibility of replacing survey-based approaches with passively collected location data for disaster-behavior research.

major comments (3)
  1. [Data and Methods] Data and Methods sections: the central claim that SVI variables exert significant effects on evacuation demand rests on the assumption that the mobile-device sample is representative across the full range of SVI characteristics. No post-stratification weights, coverage validation against ACS 5-year estimates, or comparison with official evacuation counts are described; differential under-sampling of low-vehicle, limited-English, or group-quarters populations would directly bias the origin and destination coefficients reported for those variables.
  2. [Results] Results section: the abstract and summary tables report “significant effects” for vehicle availability, group quarters, road density, and English proficiency, yet no model-fit statistics (pseudo-R², log-likelihood ratio test against null, or out-of-sample prediction error), standard-error clustering, or robustness checks (e.g., restricting to high-coverage tracts or alternative impedance specifications) are provided. Without these, it is impossible to assess whether the reported significance survives reasonable alternative specifications.
  3. [Model] Model specification: the direct-demand formulation treats observed flows as the dependent variable and SVI covariates as exogenous, but the paper does not discuss potential simultaneity (e.g., road density may be correlated with unobserved infrastructure quality that also affects evacuation) or selection into the mobile-data panel. A concrete test—such as an auxiliary regression of coverage rates on SVI variables—would be needed to support the causal interpretation implied by the headline findings.
minor comments (2)
  1. [Model] Notation: the distance impedance variable is referred to interchangeably as “distance” and “travel impedance” without a clear definition or units in the equation or table captions.
  2. [Results] Tables: the coefficient tables lack sample size (number of tract pairs), number of observed evacuations, and any indication of whether the model is estimated by OLS, Poisson, or negative binomial.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have strengthened the manuscript. We address each major comment below and have revised the paper to incorporate additional analyses, discussions, and robustness checks where feasible.

read point-by-point responses

  1. Referee: Data and Methods sections: the central claim that SVI variables exert significant effects on evacuation demand rests on the assumption that the mobile-device sample is representative across the full range of SVI characteristics. No post-stratification weights, coverage validation against ACS 5-year estimates, or comparison with official evacuation counts are described; differential under-sampling of low-vehicle, limited-English, or group-quarters populations would directly bias the origin and destination coefficients reported for those variables.

    Authors: We agree that sample representativeness is critical. In the revised Data and Methods section we have added a new subsection comparing mobile-device coverage rates against ACS 5-year estimates for vehicle availability, group-quarters residence, English proficiency, and other SVI components. We report coverage proportions by tract and discuss implications for potential bias. Post-stratification weights are not available from the data provider, and tract-level official evacuation counts for Hurricane Ian are not publicly released, so direct validation against those counts is not possible; we now explicitly note this limitation and its consequences for coefficient interpretation. revision: partial

  2. Referee: Results section: the abstract and summary tables report “significant effects” for vehicle availability, group quarters, road density, and English proficiency, yet no model-fit statistics (pseudo-R², log-likelihood ratio test against null, or out-of-sample prediction error), standard-error clustering, or robustness checks (e.g., restricting to high-coverage tracts or alternative impedance specifications) are provided. Without these, it is impossible to assess whether the reported significance survives reasonable alternative specifications.

    Authors: We accept that model-fit and robustness information should have been included. The revised Results section now reports pseudo-R², log-likelihood ratio tests against a null model, and standard errors clustered at the origin-destination pair level. We have also added two robustness checks: (1) re-estimation restricted to tracts above the median coverage rate, and (2) an alternative logarithmic distance impedance specification. Both checks preserve the statistical significance and sign of the key SVI coefficients. revision: yes

  3. Referee: Model specification: the direct-demand formulation treats observed flows as the dependent variable and SVI covariates as exogenous, but the paper does not discuss potential simultaneity (e.g., road density may be correlated with unobserved infrastructure quality that also affects evacuation) or selection into the mobile-data panel. A concrete test—such as an auxiliary regression of coverage rates on SVI variables—would be needed to support the causal interpretation implied by the headline findings.

    Authors: We have expanded the Model section to discuss simultaneity and selection explicitly, noting that road density may proxy unobserved infrastructure quality and that our estimates are best interpreted as conditional associations. We performed the suggested auxiliary regression of tract-level coverage rates on the SVI variables and report the results in a new appendix table; the coefficients are small and mostly insignificant, indicating limited selection bias on observables. We have also clarified the language throughout to avoid implying strict causality. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical demand modeling chain

full rationale

The paper estimates a zone-to-zone direct demand model from external mobile device location data and SVI covariates at the census-tract level. Significant effects on origin and destination attractiveness are obtained via regression on observed flows and distance impedance; these coefficients are not equivalent to the input variables by construction, nor do any predictions reduce to fitted parameters or self-citations. The derivation relies on standard gravity-style modeling applied to independent data sources and contains no self-definitional loops, uniqueness theorems imported from prior author work, or renaming of known results.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim depends on the accuracy of mobile location data as a proxy for behavior and on the appropriateness of SVI variables as direct inputs, with model coefficients fitted to the specific dataset.

free parameters (1)
  • Demand model coefficients
    Coefficients for SVI variables and impedance are estimated from the observed evacuation flows in the mobile data.
axioms (2)
  • domain assumption Mobile device location data accurately captures evacuation movements without substantial demographic bias
    Invoked by using the data to estimate origin-destination flows in the model.
  • domain assumption SVI variables directly measure factors affecting evacuation capability and destination attractiveness
    Used to incorporate social vulnerability into the zone-to-zone demand equations.

pith-pipeline@v0.9.0 · 5469 in / 1383 out tokens · 42660 ms · 2026-05-10T09:42:39.670504+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

  1. [1]

    My Kids are my Priority

    Anyidoho, P., Ju, X., Davidson, R., et al. (2023). A machine learning approach for predicting hurricane evacuee destination location using smartphone location data.Computational Urban Science, 3:30. Bayraktar, N. and Dal Yılmaz, ¨U. (2018). Vulnerability of elderly people in disasters: a systematic review.The Turkish Journal of Geriatrics, 21(3):467–482. ...

    work page 2023

  2. [2]

    Fleischer, P. (2022). Records show Lee County delayed evacuation order for 12 hours ahead of Ian. Fu, H., Wilmot, C., and Baker, E. (2006). Sequential logit dynamic travel demand model and its transferability.Transportation Research Record, 1977(1):17–26. 26 Fussell, E. (2012). Help from family, friends, and strangers during hurricane katrina: Findings fr...

    work page 2022

  3. [3]

    Huang, S.-K., Lindell, M., and Prater, C. (2015). Who leaves and who stays? a review and statistical meta-analysis of hurricane evacuation studies.Environment and Behavior, pages 1–39. Huang, S.-K., Lindell, M., and Prater, C. (2017). Multistage model of hurricane evacua- tion decision: Empirical study of hurricanes katrina and rita.Natural Hazards Review...

    work page 2015

  4. [4]

    How to use Gravy forensic flags for retail and transportation analysis

    Mryan (2022). How to use Gravy forensic flags for retail and transportation analysis. National Hurricane Center (2023). Tropical cyclone report: Hurricane ian (al092022), 23–30 september

    work page 2022

  5. [5]

    NOAA Climate.gov (2022)

    Technical report, National Oceanic and Atmospheric Administration. NOAA Climate.gov (2022). 2022 U.S. billion-dollar weather and climate disasters in historical context. Olivo, A., Hawkins, D., Oakford, S., and Dance, S. (2022). Lee County’s hurricane Ian evacuation timeline. Ort´ uzar, J. and Willumsen, L. (2001).Modelling Transport. Wiley, West Sussex, ...

    work page 2022