Community-scale big data reveals disparate impacts of the Texas winter storm of 2021 and its managed power outage

The study investigates whether the impacts of the February 2021 Texas winter storm (Winter Storm Uri) and associated managed power outages, burst pipes, and food inaccessibility were experienced equitably across subpopulations in Harris County, Texas. Against the backdrop of climate change-driven increases in extreme weather and well-documented disproportionate disaster impacts on vulnerable groups (low-income, racial/ethnic minorities, elderly), the paper asks if managed outages and subsequent service disruptions were uneven by income, race, and ethnicity. Due to lack of granular outage data, the study leverages community-scale digital trace data to proxy for outage extent and duration and to assess related impacts. Objectives are to categorize responses to infrastructure failure and assess impacts, and to reveal potential disparate impacts by income, race, and ethnicity using community-scale big data.

Prior research shows disasters disproportionately affect vulnerable populations. Surveys after Hurricane Katrina found strong race- and class-based differences in impacts; housing recovery studies after Hurricanes Andrew and Ike showed low-income areas suffer more damage and slower recovery; after Hurricane Harvey, less advantaged socioeconomic and minority households reported decreased ability to withstand disruptions. For winter storms, literature is scarce despite severe events in 2014 and 2021. Studies of managed outages and equity are limited; during Hurricane Irma, social vulnerabilities and utility differences affected outage duration. Social media has been used for impact sensing but suffers from geocoding sparsity and demographic biases. Recent work demonstrates the potential of digital trace data (mobility, telemetry, POI visits) for more reliable, fine-grained assessment of community impacts and resilience.

Study area and period: Harris County, Texas (Houston metro). Data from January–February 2021 were analyzed; January served as baseline (averaged by weekday), and February was assessed for impacts. All data were aggregated to the census-tract level (786 tracts). Data sources: (1) Telemetry-based population activity (Mapbox) aggregated from mobile app SDK location data, anonymized and normalized to a baseline (Jan 11–17, 2021). Activity Index A for 100 m x 100 m grid cells at 4-hour intervals was scaled by the 99.9th percentile of baseline device counts. For each census tract ct and time t, Activity Density Da(ct,t) = (1/N) Σ A_{u,t}, averaging across N grid units in the tract. Due to insufficient nighttime data in many tracts, analyses used 8:00–20:00 only. Two outage-proxy features were computed: greatest negative change in Da (impact extent) and recovery duration (time to return to baseline; capped at 20 for non-recovery by period end). (2) POI visits (SafeGraph) to grocery stores (NAICS 445110) and restaurants (NAICS 72251), daily counts. To account for cross-tract patronage near boundaries, a 1-mile buffer around each tract was used when aggregating visits to the tract level. (3) 311 Houston Service Helpline records for water-related issues (e.g., leaks, flooding, potable water, drainage). Daily counts per tract for February 2021 were used as indicators of burst-pipe impacts; because of low baseline volume, raw counts were analyzed. To compare tracts fairly, case peaks were normalized by tract population and area. (4) Demographics from 2019 ACS 5-year estimates at tract level. Tracts in top/bottom 25% were labeled high-/low-income; minority classifications were based on higher ratios of Black and Hispanic populations (top 25%) versus nonminority (bottom 25%). Statistical analysis: The Kruskal–Wallis test (nonparametric ANOVA on ranks) assessed differences in medians between minority/nonminority groups (income, Black, Hispanic) for outage proxies and normalized 311 peaks. Spatial autocorrelation was assessed with Global Moran’s I using queen contiguity weights and Monte Carlo permutation (999 permutations) to test significance. Trend classification: Agglomerative hierarchical clustering (Euclidean distance, Ward’s linkage) grouped tract-level time series of restaurant and grocery visits into four trend classes (a–d) from most to least impacted to evaluate spatiotemporal patterns of food inaccessibility and their relationship to demographics.

- Power outages (proxy via Da): A representative tract showed ~50% Da drop on Feb 15, 2021, with intermittent recovery Feb 17–19 and near-normal by Feb 20–22, aligning with resident experience. Across tracts, greater negative Da changes indicate severe outage impact; recovery duration measured persistence (non-recovered assigned 20). Kruskal–Wallis tests showed low-income tracts had significantly greater negative Da changes (significant at p ≤ 0.01; figure indicates p < .001). Hispanic-ethnicity minority tracts had longer recovery durations (p = 0.091). Although differences in greatest negative changes by race/ethnicity were not always statistically significant, minority medians were lower (worse) than nonminority. 13.5% of tracts (106/786) were significantly impacted (very low activity). Compared to all county tracts, these significantly impacted tracts had lower median income ($40,853 vs. $56,429), higher median Black population ratio (20.43% vs. 12.01%), and higher median Hispanic ratio (43.58% vs. 38.01%); group comparisons were significant at p ≤ 0.10, with low-income and racial minority comparisons p < 0.05. This indicates inequality in outage impacts across socioeconomic and racial/ethnic lines, despite some exceptions. - Burst pipes (311): 67% of tracts reported water-related 311 cases in February; case numbers and number of tracts peaking surged Feb 18–19. After normalizing case peaks by tract population and area, low-income and racial minority tracts had higher normalized peaks than high-income and racial nonminority tracts (p < 0.05). Twice as many high-income and racial nonminority tracts had zero 311 calls compared to low-income and racial minority tracts, indicating disproportionate burst-pipe impacts. - Food inaccessibility (POI visits): Clustering of restaurant and grocery visit time series produced four classes (a most impacted to d least impacted). Restaurant visits decreased starting Feb 11 with troughs Feb 16–17; grocery visits showed pre-storm surges (preparedness) and a large dip on Feb 18 before recovery. Spatial autocorrelation was strong (Moran’s I: restaurants 0.83; grocery 0.81; both p = 0.01). Most impacted areas (class a) clustered near downtown Houston; secondary impacted areas (class b) tended to surround class a. Class b tracts generally had lower median household income and higher median Black and Hispanic population ratios than other classes, indicating greater food accessibility challenges for low-income and minority communities.

The analyses address the research question by demonstrating that managed power outages and related storm impacts were not uniformly experienced. Using digital trace proxies, the study shows low-income and minority communities experienced greater outage severity and, for Hispanic-majority areas, longer recovery, more frequent and intense burst-pipe issues, and higher likelihood of food access disruptions. These findings underscore the need for integrating equity considerations into infrastructure operations during extreme events. The work illustrates the utility of community-scale big data (telemetry-derived activity, POI visits, 311 calls) for rapid, fine-grained assessment when provider outage data are unavailable, enabling identification of spatial patterns and vulnerable areas to inform more equitable preparedness and response strategies.

The study demonstrates that community-scale big data can rapidly quantify disparate impacts of extreme weather across power outages, water infrastructure failures, and food access. In Harris County during Winter Storm Uri, low-income and racial/ethnic minority communities experienced greater outage severity, higher burst-pipe burdens, and more food accessibility challenges. Contributions include: a proxy-based framework for assessing managed outage impacts using mobility telemetry; integration of 311 and POI data for multi-infrastructure impact assessment; and evidence of inequities to guide equitable resilience planning. Future directions include obtaining fine-resolution outage datasets to validate proxies, addressing scale effects (MAUP), incorporating additional infrastructure interdependencies, and developing operational tools for pre-planned equitable outage rotations and communications (e.g., predetermined outage windows), pipe protection measures, and ensuring access to critical services during extreme events.

- Lack of fine-grained, high-resolution power outage data; population activity served as a proxy and may be influenced by factors beyond outages (e.g., evacuation, charging behavior), with nighttime data often insufficient, limiting analyses to 8:00–20:00. - No causal inference is claimed; observed disparities may also be influenced by unmeasured factors (e.g., aging infrastructure, building stock, local grid topology). - Aggregation at the census-tract level may introduce Modifiable Areal Unit Problem (MAUP) effects; alternative scales were not investigated. - Potential sampling and coverage biases in digital trace and POI datasets; anonymization and thresholds may omit low-activity areas; 311 data reflect reporting behaviors and access to services. - Generalizability limited to Harris County and this event; different regions/utilities may have different outage management practices.