Lifestyle pattern analysis unveils recovery trajectories of communities impacted by disasters

The paper addresses how disasters disrupt community lifestyles—both essential (e.g., food, health care, education) and non-essential activities—and posits that the return of lifestyle activities is a critical, quantifiable milestone of community recovery. Existing recovery indicators are often qualitative and lack spatiotemporal granularity. The authors propose using location intelligence data to quantify lifestyle recovery at the census block group (CBG) level. The study focuses on Harris County, Texas, during and after Hurricane Harvey (Aug–Nov 2017). Objectives are to (1) assess the extent of disaster impacts on lifestyle activities, (2) identify recovery trajectory patterns, and (3) examine variations in recovery by flooding and demographic attributes. Research questions: (1) What distinctive patterns in typical lifestyles exist during normal periods? (2) To what extent do disasters impact lifestyle activities and how quickly do patterns recover? (3) How do recovery duration and trajectory vary with flooding impacts and demographics? The importance lies in providing quantitative, dynamic indicators of recovery to guide emergency management and resource allocation.

The paper synthesizes disaster recovery frameworks and indicators, noting recovery’s multidimensional and nonlinear nature. Prior frameworks (e.g., Cutter et al. 2008; Jordan & Javernick-Will 2013; Horney et al. 2017) span ecological, social, economic, infrastructure, institutional, and community competence dimensions, and emphasize indicators across housing, infrastructure, and planning. Literature stresses roles of social ties and inequities in recovery (e.g., Delilah Roque et al. 2020; Emrich et al. 2022; Peacock et al. 2014). Recovery has been measured via returns of population, infrastructure functionality, housing damage, and neighborhood poverty. However, few studies examine granular lifestyle activities at the CBG level. Location intelligence data has been used for evacuation, damage, impacts, and access to services (e.g., Wang & Taylor; Hong et al.; Podesta et al.; Esmalian et al.), yet leveraging it to specify and characterize community recovery remains limited. This gap motivates quantifying lifestyle recovery using high-frequency, granular human mobility and POI data.

Study context: Harris County, Texas during Hurricane Harvey (Category 4, Aug 2017) through Nov 2017. Data sources: (1) Spectus mobility data (anonymized GPS-based stops; device home CBGs; high accuracy/frequency; CCPA/GDPR-compliant); (2) SafeGraph POI data with NAICS codes (to classify POI types); (3) Microsoft US Building Footprints (to obtain POI polygons); (4) FEMA flood depth grid (Aug 29, 2017) aggregated to percent of CBG area flooded; (5) US Census CBG demographics (population, median income, percent White, percent elderly >65). Data processing: Identify device home_CBG from Spectus device matrix (consistent residence >1 day), link to stop table via user_id. Match Spectus POIs to building footprint polygons via place_id and coordinates; enrich with NAICS from SafeGraph (Spectus 2017 lacked NAICS). Classify POIs into eleven categories: essential (gasoline stations; grocery & merchandise; health & personal care; medical facilities; education) and non-essential (banks; stores & dealers; restaurants; entertainment; recreation & gym centers; beauty care), informed by FEMA community lifelines. Aggregate daily home_CBG-to-POI visit counts to weekly to reduce weekday variability. Baseline: mean weekly visits during first two weeks of Aug 2017 (pre-landfall). Recovery definition: a CBG is recovered in a given week if visits to essential/non-essential groups reach ≥90% of baseline. CBGs not reaching threshold set to 15 weeks (analysis window max). Sensitivity analysis for thresholds (90/80/70%) performed (Supplementary Fig. SI.4). Clustering: Two-stage k-means. Primary clustering (baseline lifestyle patterns): vectors are relative frequencies of visits to POI categories (each category’s visits divided by total POI visits) during baseline; features standardized (min–max). Number of clusters chosen by elbow method (optimal k=4). Secondary clustering (recovery trajectories) performed within each primary cluster on four features: maximum point of disruption for essential and non-essential (largest percent drop vs baseline) and duration of recovery (weeks to 90%) for essential and non-essential. Elbow method selected three sub-clusters per primary cluster. Statistical testing: ANOVA used to test differences across primary clusters (visit frequencies) and across secondary clusters (recovery features), with significance at p<0.05. Flooding and demographics: Associate CBG-level flooding (% area flooded ≥1% threshold) and demographic medians to secondary clusters to explore associations.

• Primary lifestyle clusters: Four baseline lifestyle clusters identified (k=4) with CBG distribution: Cluster 1 (60.26%), Cluster 2 (3.68%), Cluster 3 (27.94%), Cluster 4 (8.12%). ANOVA confirmed significant differences in relative frequencies across clusters (p<0.05). Distinctive dependencies: • Cluster 1 highest relative frequency in essential: health & personal care (77.40%) and education (0.32%); non-essential: stores & dealers (65.57%). • Cluster 2 highest in essential: medical facilities (31.97%) and gasoline (3.60%); non-essential: restaurants (34.27%), recreation & gyms (2.21%), entertainment (0.99%). • Cluster 3 showed no distinct high-frequency category. • Cluster 4 highest in essential: grocery & merchandise (28.40%); non-essential: beauty care (21.86%) and banks (13.37%). Non-essential categories ranked consistently across clusters (stores & dealers, restaurants, beauty care as top three).
• Recovery trajectories: Within each primary cluster, secondary clustering yielded three sub-clusters and four trajectory types: Immediate Recovery Duration (IRD, 0 weeks), Short (SRD, 1–6 weeks), Moderate (MRD, 7–14 weeks), Extreme (ERD, ≥15 weeks). ANOVA showed significant differences among recovery trajectories for both essential and non-essential across clusters (e.g., Cluster 1 essential F=4129.18; non-essential F=4474.89; all p≤0.05). Essential and non-essential recoveries were not strictly sequential; either could precede the other depending on the sub-cluster.
• Extreme recovery durations: Approximately 500 CBGs (Cluster 1-1/1-2, as reported), 4 CBGs (Cluster 2-1), 92 CBGs (Cluster 3-0), and 12 CBGs (Cluster 4-0) did not recover to 90% of baseline within 15 weeks (ERD). About 59% of CBGs with ERD did not experience ≥1% flooding in their home CBG.
• Flooding and spillover effects: Roughly 89% of non-flooded home CBGs still experienced at least 1 week of lifestyle disruption, showing impacts extend beyond directly flooded areas and reflect broader network and facility accessibility effects.
• Demographics: Across secondary clusters, ANOVA did not detect statistically significant differences in demographic means/medians among recovery trajectories within primary clusters. Exploratory contrasts suggest possible disparities: examples include sub-clusters with ERD associated with lower median incomes in some clusters (e.g., Cluster 4 ERD median income ~$37,283) versus immediate recovery in higher-income sub-clusters; however, opposite patterns also appeared (e.g., ERD with higher median incomes in some clusters), indicating complex relationships.
• Baseline impact magnitudes: Median maximum disruptions reached high levels (e.g., overall medians across all CBGs: essential 71.69%, non-essential 85.78% decrease), with recovery durations varying by sub-cluster (see Table 3 medians: e.g., SRD medians around 1–2 weeks; MRD around 7–14; ERD fixed at 15 weeks for analysis window).

The findings support the concept that lifestyle recovery—capturing population activities and access to facilities—serves as a quantifiable milestone of community recovery. Using granular location intelligence data enabled identification of four distinctive baseline lifestyle patterns and multiple recovery trajectories at CBG scale. Differential recovery rates within the same lifestyle cluster highlight heterogeneity in community capacities and service restoration, suggesting the need for targeted, proactive resource allocation. The observed spillover effects (substantial disruption in non-flooded CBGs and a majority of ERD CBGs not directly flooded) underscore that urban spatial structure, mobility patterns, and POI distributions extend flood impacts beyond inundated areas. While demographic analyses were not statistically conclusive, exploratory contrasts indicate potential equity concerns warranting further study. Overall, integrating mobility, POI, flooding, and demographic datasets offers actionable insights for emergency managers to prioritize restoration (e.g., medical facilities for medical-dependent clusters; groceries for grocery-dependent clusters), anticipate disproportionate recovery durations, and consider systemic network effects that impact accessibility even outside flooded zones.

The study introduces and operationalizes lifestyle recovery as a critical, quantifiable milestone in community recovery using privacy-enhanced mobility and POI data at CBG resolution. It identifies four baseline lifestyle clusters and four recovery trajectory types, revealing heterogeneous recovery within similar lifestyles and significant spillover of disruption to non-flooded areas. Recommendations for decision-makers: (1) characterize community dependencies on essential/non-essential services during normal periods to set baselines and prioritize restorations; (2) monitor for disproportionate recovery durations within similar lifestyles to proactively support vulnerable areas; (3) consider demographics and flooding extent jointly to anticipate extreme recovery durations. Future research should assess mechanisms of lifestyle recovery including trip distance and travel time to facilities, apply supervised learning to predict recovery from sociodemographics and hazard attributes, and leverage longer, multi-year datasets for seasonal control and finer spatiotemporal granularity.

• Baseline limited to two pre-event weeks (early Aug 2017); no multi-year or pre-2017 data to control for seasonality; future work could use longer historical baselines. - Representativeness concerns inherent to app-based mobility panels, though prior validations support Spectus data’s scale and demographic representativeness. - Recovery threshold set at 90% of baseline; although sensitivity analyses (80%, 70%) were conducted, threshold choice affects recovery timing. - Analysis window capped at 15 weeks; CBGs not recovered by then are censored at 15 weeks. - Flooding considered only at home CBGs; disruptions along travel routes or at destination POIs (accessibility) were not modeled. - Demographic-recovery relationships were exploratory and not statistically significant; potential disparities require further investigation. - Spectus mobility data are proprietary and accessed under license; full replication may require data access permissions.