Exploring income and racial inequality in preparedness for Hurricane Ida (2021): insights from digital footprint data

Preparedness for adverse events is critical to building urban resilience to climate-related risks. While most extant studies investigate preparedness patterns based on survey data, this study explores the potential of big digital footprint data (i.e. population visits to points of interest (POI)) to investigate preparedness patterns in the real case of Hurricane Ida (2021). We further investigate income and racial inequality in preparedness by combining the digital footprint data with demographic and socioeconomic data. A clear pattern of preparedness was seen in Louisiana with aggregated visits to grocery stores, gasoline stations, and construction supply dealers increasing by nearly 9%, 12%, and 10% respectively, representing three types of preparedness: survival, mobility planning, and hazard mitigation. Preparedness for Hurricane Ida was not seen in New York and New Jersey states. Inequality analyses for Louisiana across census block groups (CBGs) demonstrate that CBGs with higher income have more (nearly 8% greater) preparedness in visiting gasoline stations, while CBGs with a larger percentage of the white population have more preparedness in visiting grocery stores (nearly 12% more) in the lowest income groups. The results indicate that income and racial inequality differ across different preparedness in terms of visiting different POIs.


Introduction
Hurricane Ida landed near Louisiana on 29 August 2021, as a category four hurricane with 150 mph sustained winds, leading to 91 total casualties across nine states, among which New Jersey, Louisiana, and New York have the most casualties with 32, 28, and 18 deaths respectively (Hanchey et al 2021).Though Hurricane Ida hit Louisiana the hardest from the climatology perspective, Louisiana has the least casualties directly caused by Hurricane Ida among the three states.Louisiana has a total of 28 casualties out of which ten died of hyperthermia due to unprecedented power outages, six died of preexisting conditions and two died of drowning, while New Jersey has 28 died of drowning and one died of electrocution and all 18 casualties in New York were due to drowning (Hanchey et al 2021).The inconsistent casualty outcomes with the influence of Hurricane Ida on three states motivated us to investigate social factors and behavioral drivers in addition to climatological factors (e.g. the maximum sustained wind speed) that contribute to hurricane preparedness (Aerts et al 2018, Noll et al 2022).In this paper, we focus on the investigation of patterns related to preparedness for Hurricane Ida in Louisiana, New York, and New Jersey states.Furthermore, we have a more detailed exploration in Louisiana to evaluate social inequality in hurricane preparedness, addressing both income and race.
Preparedness for future adverse events is a critical part of building resilience (Adger et al 2005, Paton et al 2006).Preparedness could include longterm planning (e.g.land use regulations, evacuation plans, insurance purchase, mitigation funds) and short-term mitigation strategies (e.g.installing hurricane shutters, storing food and water, fixing electric appliances) right before the adverse events (Paton andJohnston 2001, Godschalk 2003).Extant studies investigate preparedness for adverse events mainly based on collected survey data (Yoshida andDeyle 2005, Horney et al 2008), which highly depend on the quality of developed survey instruments and usually have restricted sample sizes.For example, Peacock (2003) developed a survey to investigate the factors affecting shutter and envelope coverage adopted by single-family homeowners for hurricane events.Mozumder et al (2015) administered a stakeholder survey to investigate their preferences for a mitigation fund.Josephson et al (2017) conducted phone interviews to assess the preparedness of small businesses for hurricane disasters.An international survey was conducted to investigate cross-national social and behavioral drivers for household adaptation activities for climate change in Noll et al (2022).In this paper, we use actual digital footprint data on visits to points of interest (POIs) for three states: New York, New Jersey, and Louisiana in August 2021 to investigate preparedness for Hurricane Ida.Visits to POIs not only provide information on mobility footprints but also encode human needs and social interactions, providing more insights into social drivers and constraints for preparedness.
Many factors can influence household and individual-level preparedness for adverse events, including demographics and socioeconomic status, risk experiences and perceptions, institutional regulations and policies, and media coverage (Paton et al 2006, Noll et al 2022).Among these factors, demographics and socioeconomic status are widely studied, the inclusion of which is critical to help identify and target populations that fail to adopt the preparedness (Peacock 2003).Extant studies have consistent conclusions regarding the roles of race and socioeconomic status in preparedness for adverse events: income is related to the capability of individuals/households to take preparedness and ethnic minorities have limited access to essential resources for preparedness (Peacock 2003, Bamberg et al 2017).Many studies have investigated racial and socioeconomic inequality in hazard preparedness regarding one of the most influential hurricane events in Louisiana: Hurricane Katrina (2005).For example, Elder et al (2007), Messias and Lacy (2007), and Thiede and Brown (2013) find that ethnic minorities and immigrants are less likely to evacuate before Hurricane Katrina due to the language barrier and distrust of the authority.Curtis et al (2007) and Elliott and Pais (2006) find that income is a strong determinant of pre-Katrina evacuation because of the cost and required access to vehicles.
However, extant studies that investigated racial and income inequality in preparedness usually do not account for the strong correlation between race and income (e.g.census block groups (CBGs) with larger white populations having higher median incomes, see (Tong et al 2021)).In this paper, we investigate social inequality in preparedness for Hurricane Ida at a fine spatial scale, employing a method recently developed in Tong et al (2021) to unpack income and racial effects.Specifically, we explore inequality in preparedness across income strata, and by race within income strata to unpack the effect of race while controlling for income.The method has been applied previously to evaluate inequality in energy use intensity; here we address preparedness.Such inequality analyses can help guide the policy related to the distributional aspect of social equity, wherein 'equity' refers to the allocation of resources and benefits across social strata, with the goal of reducing disparities for the most disadvantaged (Braveman et al 2010, Ramaswami 2020, Clark et al 2022).
Overall, the goals of the paper are twofold: (a) to investigate preparedness activities in the three states affected by Hurricane Ida (2021) using digital footprint data of population visits to POIs, and (b) to investigate social inequality based on demographics and socioeconomic status in preparedness.The inequality analyses are focused on Louisiana state (given the only clear preparedness pattern for Hurricane Ida), and utilize demographic and socioeconomic data at the CBG level, the finest spatial scale wherein data on both income and race are reported by the US Census.This data enables unpacking both income inequality and the impact of race within income groups, wherein the impact of racial inequality within the lowest income group is particularly relevant from a vulnerability and equity perspective.The current knowledge about preparedness for adverse events mainly derives from survey instruments.Surveys are time-consuming, subject to recall bias, and difficult to implement over large regions.In contrast, our paper introduces a data-driven approach that can be used to investigate actual behaviors in real time over large areas.Furthermore, we applied a methodology to unpack racial and income inequality in preparedness.Such insights cannot be readily available through surveys because they need large data to stratify and unpack income and race.

Materials and method
Compared with current literature mainly using survey instruments to investigate preparedness for adverse events (Peacock 2003, Mozumder et al 2015, Josephson et al 2017, Noll et al 2022), we use digital footprint data of the population visiting POIs provided by SafeGraph company (SafeGraph 2020).A point of interest is a specific physical location that someone may find useful or interesting, such as restaurants, hospitals, grocery stores, and gasoline stations.This POI dataset was widely used in many studies for human behavior dynamics analysis, see (Benzell et al 2020, Chang et al 2021, Li et al 2021).We propose a novel approach to investigating the preparedness pattern for Hurricane Ida.We construct the baseline of POI visits based on data three Q Li et al weeks before Hurricane Ida and test the significance of changes in POI visits before Hurricane Ida using statistical analysis.The approach could provide statistical evidence for preparedness, and the big data provides more statistical power compared with the traditional survey instruments.We use the disparity ratio (P75/P25) by race and income, which is widely adopted (see (Braveman et al 2010)), to investigate the inequality in preparedness in Louisiana.We follow a novel approach to unpacking the race and income effects in Tong et al (2021).We elaborate on the details of the method in the following paragraphs.

Investigate preparedness and related POIs in three states
The adopted POI data includes nearly 1.5 million unique traced devices of more than 140 million visits to POIs in August 2021 to investigate preparedness for Hurricane Ida.Detailed information related to data is provided in the supplementary materials.
New York state has approximately a total of 244k POIs, and New Jersey and Louisiana states have approximately 119k and 72k POIs respectively.Each POI has a category determined by (i.e.North American Industry Classification System (NAICS) developed by the US Census Bureau) according to its primary business activity (e.g.restaurants, grocery stores, and gasoline stations).We group POIs based on their categories according to their NAICS codes.New York state has 168 categories of POIs, while New Jersey has 161 categories and Louisiana has 163 categories.There are two benefits to grouping POIs based on their categories: (1) grouping POIs could remove potential noise in data because visits to single POIs are easily affected by random local events, (2) grouping POIs could reflect a general pattern of a large population, and (3) visiting different categories of POIs demonstrates population needs for hurricane preparedness.
We assume that preparedness for Hurricane Ida will lead to significant increases in POI visits (Dargin et al 2021).Therefore, we use the seven-day-rolling average POI visits from August 1 to 14 August 2021, which is two weeks before the week of Hurricane Ida as a baseline to reflect the normal situation of visiting POIs.Then we calculate the daily Z scores of rolling average POI visits for each day during the week ending with Hurricane Ida's landfall from August 21 to 28 August 2021.The Z score is calculated as x i −µ σ , where µ and σ are the mean and standard deviation of the baseline (i.e. two weeks prior to the hurricane week).The Z score larger than 3 shows a statistically significant increase in POI visits on that day.
When the data show an increasing trend (the baseline is not flat), we use the first difference of rolling average POI visits.Only a few POIs that had non-linear increasing trends (e.g.visits to colleges, showing the increasing trend in the summer, see the supplementary materials for the examples) still have large Z scores.We manually check these POIs and remove them from the analysis.

Investigate income and racial inequality influence on preparedness
Percentage changes in weekly visits to POIs were explored at the CBG scale to unpack income and racial inequality in hurricane preparedness in Louisiana.We used the ratio of the number of residents visiting POIs (visit-POIs) in the week before and during hurricane landfall (week of 23 August 2021, to the average visit-POIs three weeks before Hurricane Ida (i.e. an average of the week of 2 August, 9 August, and 16 August 2021)), as an indicator of preparedness for Hurricane Ida.Weekly aggregated visit-POI data are used because daily data are not reported at the CBG scale, unlike at the state scale.We study inequality in hurricane preparedness by comparing the ratio of the number of residents who visit different POIs before and after Ida.We focus on two identified POIs-grocery stores and gasoline stations related to preparedness (see materials and methods for detail).The POI, building material and supplies dealers (NAICS code: 4441) is not covered because it showed the least aggregated response among the three POIs.
We used CBG-level demographic and socioeconomic data provided by the United States Census Bureau's American Community Survey 5-year Estimates.The latest version is the 2015-2019 version.We calculate the Spearman correlation between the median income and white per total population at the CBG level in Louisiana.The results show that race and income have a significant positive correlation: 0.6 at a 99% confidence level (see figure S7 in the supplementary materials).To unpack the effects of race and income on preparedness, we stratify the data based on four quartiles of the median income of CBGs in Louisiana.Then we compare the ratio of the number of residents who visit POIs before and after Ida across income (across CBGs clustered by income quartiles), and then by race within each of the income strata.
There are 3471 CBGs in Louisiana and 3221 CBGs that have demographic and socioeconomic data reported by the United States Census Bureau.We further filter the data because certain CBGs have very low footprint data coverage.According to the statistical sample requirements for the population in Kotrlik et al (2001), we determined the minimum number of data in each CBG based on the margin of error equal to 0.03 and the level of risk that the true margin of error exceeding 0.03 is less than 0.1.As a result, 1530 CBGs meet the statistical requirement of sample size.Therefore, each income stratum has around 382 CBGs in Louisiana.Following the methods in Tong et al (2021), we calculate the disparity ratio (i.e.P75/P25-the median for CBGs with income or white percentage population more than 75th percentile divided by the median for CBGs with income or white percentage population less than 25th percentile) to investigate income and racial inequality across and within income strata.We use the Mann-Whitney U test to assess the statistical difference between the two groups.

Three POIs reflecting three types of preparedness
We find a statistically significant increase in visits (by nearly 9%, 12%, and 10%) to three POIs: grocery stores (NAICS code: 4451), gasoline stations (NAICS code: 4471), and building material and supplies dealers (NAICS code: 4441) on 27 and 28 August 2021, in Louisiana.This implies that the preparedness happened two days before Hurricane Ida hit Louisiana (figure 1).Visiting the three POIs demonstrate three types of hazard preparedness discussed in the literature: preparedness for survival, preparedness for plans (e.g.evacuation, travel, commute), and preparedness for hazard mitigation (e.g.secure and reinforce the structure) (Dooley et al 1992).

Disparate patterns of preparedness in three states
For New York state, we find statistically significant increased visits to the same POIs on 21 August 2021, right before Hurricane Henri (2021), and there is no statistical evidence for increased POI visits before Hurricane Ida (figure 1).This result implies preparedness happened one day before Hurricane Henri and there was very limited preparedness for Hurricane Ida in New York State, probably due to the short interval between two hurricanes.For the State of New Jersey, we did not find any statistically significant increase in visits to POIs, suggesting no preparedness for either Hurricane Henri or Ida reflected by the POI visits (figure 1).
Visits to POIs could partly reflect human needs and the characteristics of cities.For example, POI Restaurants and Other Eating Places (NAICS code: 7225) has the most monthly visits in all three studied states, while in New York, POI Museums, Historical Sites, and Similar Institutions (NAICS code: 7121) ranks second.In Louisiana, gasoline stations have the second most visits and POI Lessors of Real Estate (NAICS code: 5311) ranks second in New Jersey.Grocery stores and gasoline stations both rank within the top six in the three states.Building material and suppliers ranks 15th in Louisiana and 18th in both New York and New Jersey.The ranks of visits to the three POIs identified for preparedness are similar in the three states (see supplementary materials for the detailed rank of the top 10 POI visits in each state).The disparate increased number of visits to the three POIs before Hurricane Ida could reflect disparate patterns of preparedness in the three states and may explain why Louisiana had the least direct deaths, while New Jersey had the most direct deaths, although Ida hit Louisiana the hardest from the perspective of climatology.For example, Ida made landfall at Port Fourchon, Louisiana with a wind speed of 150 mph while the Ida remnants hit New Jersey and New York with a wind speed of around 52 mph (Beven et al 2022).

Income inequality in preparedness
We find disparate results in preparedness by income and by POI in Louisiana.Table 1 demonstrates the disparity ratio in hurricane preparedness by income: there is a statistically significant difference between the highest income quartile CBGs and lowest income quartile CBGs for visits to gasoline stations (no statistically significant difference was seen across income groups for grocery stores).The highestincome quartile CBGs have 8% more resident visits to gas stations than the lowest-income quartile CBGs, possibly reflecting social differences in preparing for Ida.

Racial inequality in preparedness
We investigated the disparity ratio of preparedness for Ida by race in each income stratum (table 2).CBGs with more than the 75th percentile of the white percentage population have statistically significantly more residents visiting grocery stores (∼12% more) than the ones with less than the 25th percentile of the white percentage population in the lowest-income quartile CBGs (income less than the 25th percentile).Exploring race effects within the lowest income quartile is important from an equity perspective, as they often represent the most vulnerable and disadvantaged in society.For gasoline stations, the disparity ratios in preparedness by race within all the income strata are not statistically significant, indicating no observable racial inequality in visiting gasoline stations in preparedness for Hurricane Ida.

Discussion
This paper shows the potential of digital footprint data to provide key insights into preparedness for upcoming urban disruptions.Furthermore, footprints from CBGs to POIs (e.g.restaurants, grocery stores, gasoline stations, amusement facilities), combined with the demographic and socioeconomic data, shed new light on human needs for different services in preparedness as well as the influence of income and racial inequality.
The extant studies regarding hazard preparedness mainly adopt survey instruments (Horney et al 2008, Mozumder et al 2015, Bamberg et al 2017), which usually have limited sample sizes (less than 500), subject to recall bias, and are difficult to implement over large regions.Studies regarding hazard preparedness for Hurricane Katrina largely focus on  The results of our study have overall consistent conclusions with existing studies regarding racial and income effects on preparedness.However, our findings are novel as well and are a good complement to existing works.Our findings reveal race effects unpacked from income only in the lowest income strata and for certain POIs (i.e.grocery stores).This does reinforce the vulnerability of lowincome minorities (non-white populations) to a critical basic need.Our data-driven approach quantitatively demonstrates this and can be used in the future to compare preparedness in different states or in the same state over time.Other literature has identified the presence of food deserts (i.e. the lack of grocery stores and healthy food) in low-income minority neighborhoods (Walker et al 2010), meaning they have to travel a long distance to grocery stores which can inhibit preparedness.On the other hand, our results reveal income effects were found only significant in visiting gasoline stations.This could be a number of reasons for low-income groups not to refill, such as financial constraints or lower ownership of personal vehicles.Such differences between POIs are not found in traditional preparedness literature and reflect the contribution of this data-driven POI footprint methodology combined with an approach to unpacking income and race.
The study has limitations and some future directions could be pursued.First, we only investigate income and racial inequality in preparedness in this study.Other potential explanatory factors affecting preparedness such as media coverage, emergency planning coverage, language barriers, and ownership of vehicles are worth investigating in a future study.Second, we did not investigate why New York state and New Jersey state did not show preparedness for Hurricane Ida, which could be an interesting research direction to pursue.For New York state, we did find that a short interval between Hurricane Henry and Hurricane Ida's landfall (less than one week) may explain the absence of preparedness for Hurricane Ida.However, the speculation needs to be confirmed by more data on historical sequential landfall events.Lastly, Hurricane events have strong physical characteristics (e.g.track, rainfall, storm surge) and affect areas differently.Future studies investigating inequality could take this pre-existing condition into consideration.

Conclusion
Hurricane Ida landed on 29 August 2021, affecting nine states with the top three casualties in Louisiana, New Jersey, and New York.We use digital footprint data of more than 140 million visits to 164 categories of POIs based on 1.5 million unique traced devices in August 2021 in these three states to study preparedness for Hurricane Ida.Based on the daily total number of visits (at the state level) to different categories of POIs, we find disparate preparedness patterns for Hurricane Ida in the three states.In Louisiana, we find statistically significant increases in visits to three POIs two days before Hurricane Ida: Grocery Stores (NAICS code: 4451), gasoline stations (NAICS code: 4471), and building material and supplies dealers (NAICS code: 4441).Visiting these three POIs shows three types of hazard preparedness at the individual and household levels: survival preparedness, mobility planning preparedness, and hazard mitigation preparedness (Dooley et al 1992).In New York, we find statistically significant increases in visits to the same POIs one day before Hurricane Henri's landfall on 22 August 2021, one week before Hurricane Ida, and we do not find statistical evidence showing preparedness for Hurricane Ida.In New Jersey, we do not find statistically significant increased visits to any POIs.The results may partly explain why Hurricane Ida hit Louisiana the hardest while New Jersey has the most direct deaths.
Further, we explore the income and racial inequality in preparedness for Hurricane Ida in Louisiana, using weekly visits to POIs in 1530 CBGs in Louisiana combined with demographic and socioeconomic data of CBGs.The results reveal that income and race have significant influences on preparedness for Hurricane Ida and racial effects vary in different income strata and POIs.Importantly, our results show that the influence of race and income need to

Q
Figure 1.Identified three points of interest with statistically significant increased visits.In each sub-figure, the left y-axis shows the absolute number of visits and the right y-axis shows the normalized value by the total number of visits in one month.The horizontal grey band shows the boundary of Z scores of seven-day-rolling means from −3 to 3. The vertical dark grey band illustrates the landing stage of Hurricane Henri (22-23 August 2021), and the vertical shallow grey band illustrates the landing stage of Hurricane Ida (29-30 August 2021).
a Significant at 99% confidence level, the difference between the two groups was assessed using the Mann-Whitney U test.b Median income for 383 CBGs with income <25th percentile is $ 31.910/yr.c Median income for 383 CBGs with income >75th percentile is $ 86 667/yr.

Q
Li et alDarginJ S, Li Q, Jawer G, Xiao X and Mostafavi A 2021Compound hazards: an examination of how hurricane protective actions could increase transmission risk of COVID-19 Int.J. Disaster Risk Reduct.65 102560 Deng H, Aldrich D P, Danziger M M, Gao J, Phillips N E, Cornelius S P and Wang Q R 2021 High-resolution human mobility data reveal race and wealth disparities in disaster evacuation patterns Humanit.Soc.Sci.Commun.Josephson A, Schrank H and Marshall M 2017 Assessing preparedness of small businesses for hurricane disasters: analysis of pre-disaster owner, business and location characteristics Int.J. Disaster Risk Reduct.23 25-35 Kotrlik J, Higgins C and Bartlett J E 2001 Organizational research: determining appropriate sample size in survey research appropriate sample size in survey research Inf.Technol.Learn.Perform.J. 19 43 (available at: www.opalco.com/wpcontent/uploads/2014/10/Reading-Sample-Size1.pdf)(Accessed 4 November 2023) Li Q, Bessell L, Xiao X, Fan C, Gao X and Mostafavi A 2021 Disparate patterns of movements and visits to points of interest located in urban hotspots across US metropolitan cities during COVID-19 R. Soc.Open Sci. 8 201209 Messias D A K H and Lacy E 2007 Katrina-related health concerns of Latino survivors and evacuees J. Health Care Poor Underserved 18 443-64 Mozumder P, Chowdhury A G, Vásquez W F and Flugman E 2015 Household preferences for a hurricane mitigation fund in Florida Nat.Hazards Rev. 16 04014031 Noll B, Filatova T, Need A and Taberna A 2022 Contextualizing cross-national patterns in household climate change adaptation Nat.Clim.Change 12 30-35 Paton D and Johnston D 2001 Disasters and communities: vulnerability, resilience and preparedness Disaster Prev.Manag. 10 270-7 Paton D, McClure J and Bürgelt P T 2006 Natural hazard resilience: the role of individual and household preparedness Disaster Resilience: An Integrated Approach (Charles C Thomas Publisher, Ltd.) pp 105-27 Peacock W G 2003 Hurricane mitigation status and factors influencing mitigation status among Florida's single-family homeowners Nat.Hazards Rev. 4 149-58 Ramaswami A 2020 Unpacking the urban infrastructure nexus with environment, health, livability, well-being, and equity One Earth 2 120-4 SafeGraph 2020 Weekly pattern (Safegraph) (available at: https:// docs.safegraph.com/docs/weekly-patterns)Thiede B C and Brown D L 2013 Hurricane Katrina: who stayed and why?Pop.Res.Policy Rev. 32 803-24 Tong K, Ramaswami A, Xu C, Feiock R, Schmitz P and Ohlsen M 2021 Measuring social equity in urban energy use and interventions using fine-scale data Proc.Natl Acad.Sci.USA 118 e2023554118 Walker R E, Keane C R and Burke J G 2010 Disparities and access to healthy food in the United States: a review of food deserts literature Health Place 16 876-84 Yoshida K and Deyle R E 2005 Determinants of small business hazard mitigation Nat.Hazards Rev. 6 1-12