key: cord-0196192-mbb0zk8y authors: Esmalian, Amir; Yuan, Faxi; Xiao, Xin; Mostafavi, Ali title: Characterizing Equitable Access to Grocery Stores During Disasters Using Location-based Data date: 2022-01-03 journal: nan DOI: nan sha: 738fec446157e6888bd228e4cd18c09f30429746 doc_id: 196192 cord_uid: mbb0zk8y Natural hazards cause disruptions in access to critical facilities, such as grocery stores, impeding residents ability to prepare for and cope with hardships during the disaster and recovery; however, disrupted access to critical facilities is not equal for all residents of a community. In this study, we examine disparate access to grocery stores in the context of the 2017 Hurricane Harvey in Harris County, Texas. We utilized high-resolution location-based datasets in implementing spatial network analysis and dynamic clustering techniques to uncover the overall disparate access to grocery stores for socially vulnerable populations during different phases of the disaster. Three access indicators are examined using network-centric measures: number of unique stores visited, average trip time to stores, and average distance to stores. These access indicators help us capture three dimensions of access: redundancy, rapidity, and proximity. The findings show the insufficiency of focusing merely on the distributional factors, such as location in a food desert and number of facilities, to capture the disparities in access, especially during the preparation and impact/short-term recovery periods. Furthermore, the characterization of access by considering combinations of access indicators reveals that flooding disproportionally affects socially vulnerable populations. High-income areas have better access during preparation as they are able to visit a greater number of stores and commute farther distances to obtain supplies. The conclusions of this study have important implications for urban development (facility distribution), emergency management, and resource allocation by identifying areas most vulnerable to disproportionate access impacts using more equity-focused and data-driven approaches. known as food deserts, are endemic to lower socioeconomic neighborhoods. The access inequality could be exacerbated during disasters. Disruptions in road networks during disasters (Akhavan et al. 2019; Weilant et al. 2019) , as well as residents' capabilities and lifestyle patterns hinder access (Barbosa et al. 2021; Deng et al. 2021) . Some areas face a greater impact during disasters from road closures and damage to the stores, which may disproportionally disrupt the people's access to the grocery stores (Cummins 2002; Esmalian et al. 2020; Walker et al. 2010 ). On the other hand, people of a lower socioeconomic status usually have fewer resources and capabilities to find alternatives allowing them to compensate for the disrupted access (Cutter 2017; Flanagan et al. 2011; Horney 2008; Larson and Shin 2018) . Most studies examining access have focused on analytical assessment using physical metrics. Some approaches define access based on the number of available stores Dong et al. 2006; Islam and Aktar 2011; Liu et al. 2014; Simini et al. 2021; Talen and Anselin 1998) or a combination of number of stores and distance from a specified area (Dong et al. 2020a; Faturechi and Miller-Hooks 2014; Logan and Guikema 2020) . However, these measures based on physical distance only partially characterize access to facilities. In particular, during disasters, distance alone is not a reliable measure. For example, in preparation for a hurricane, if residents may need to visit several grocery stores to find and purchase needed supplies, their access might be considered acceptable despite the increase in total accumulated travel distance. Another dimension of access is the duration of trips. Flooding exacerbates traffic congestion, significantly increasing the total trip duration, which affects access to grocery stores. Failure to account for multiple dimensions of access impairment limits characterization of access to facilities both during normal times, as well as during disasters (Dong et al. 2006; Islam and Aktar 2011; Liu et al. 2014 ). Hence, a combination of access metrics is needed to better understand the characteristics of access. For example, some residents might travel long distances to grocery stores, but they may have multiple stores to obtain their supplies. Therefore, a more nuanced characterization of access based on measures related to populationfacility network interactions reveals a more accurate picture of access. In addition to characterizing different dimensions of access to facilities, the current literature is rather limited in examining disparate impacts of disasters on sub-populations (Dong et al. 2006) . During disasters, a supply-demand imbalance occurs due to protective actions of residents and disruptions in infrastructure and supply chains. In preparation for an impending disaster, people who choose to shelter-in-place tend to stockpile supplies in anticipation of several days of disruptions with the effect of depleting grocery store inventories. This surge in demand for supplies leads to grocery stores run out of stock and people need to visit multiple stores at farther distances. Households in higher-income communities, however, would more readily have the means to travel further to obtain their needs from more distant stores when faced with local shortage of supplies. Thus, while the number of stores nearby and the physical distance could provide indications of levels of access during normal times, they are inadequate for understanding the preparedness behavior of different sub-populations when examining disparities in level of access during the preparedness stage. This limitation also exists in the short-term recovery stage. The supply-demand imbalance persists after a hazard event ends. Due to disruptions in road networks (Dong et al. 2020a ) and power outages, as well as supply chain impacts, the time span for returning to pre-event conditions can be a few weeks to more than a year. During the short-term recovery stage, it is important that residents are able to access facilities without significantly increasing trip distance and duration. While the extant literature recognizes importance of access to facilities for community resilience in disasters (Dong et al. 2020b; Gori et al. 2020; Qiang and Xu 2019) , limited empirical and observational insights exist to inform about the impacts of disasters on access to grocery stores (Lee et al. 2021; Podesta et al. 2021) . In this study, we examined equitable access to grocery stores during different phases of a disaster by harnessing and analyzing big location-based data. We constructed and analyzed the population-facility network for examining access disparities to grocery stores in the context of Hurricane Harvey in 2017. We characterized access to facilities based on three distinct but complementary dimensions: (1) redundancy in access (based on the number of unique stores visited); (2) rapidity of access (based on the duration of trips to stores); and (3) proximity of access (based on the distance to stores visited). Accordingly, this research aims to answer the following research questions: (1) What are the characteristics of access to grocery stores for different sub-populations at different stages of disasters? (2) What is the extent of disparity in access among different sociodemographic groups in different disaster phases? (3) What are the contributing factors to unequal access to grocery stores, and to what extent are facility distribution inequities, such as food deserts, indicative of access disparities? To address these research questions, we defined three distinct indicators for examining access and quantify these indicators based on location-based data related to people's visits to grocery stores in the context of Harris County, Texas. We examined disparities in access to grocery stores based on the structure and attributes of population-facility networks. The three steps of this methodology ( Fig. 1) were: (1) specifying distinct indicators for examining different dimensions of access, (2) analyzing variations in the access indicators among different sub-populations in three phases of the disaster, and (3) evaluating the factors influencing access disparity. First, we utilized node-level and link-level access metrics to capture different aspects of access in population-facility networks: (1) unweighted degree based on the number of visited grocery stores, (2) weighted degree based on the travel time, and (3) weighted degree based on trip distance. Second, we evaluated fluctuations in patterns of each measure across different spatial areas to find variations in access at different stages of the disaster. Third, we specified disparities in access to grocery stores in the face of a disaster and evaluated factors contributing to such disparities. This study examined the access to grocery stores in the context of Hurricane Harvey, which made landfall in Harris County, Texas, in August 2017. The storm dropped more than 60 inches of rain, triggering immense flooding and causing severe infrastructure disruptions (Blake and Zelinsky 2017; Federal Communications Commission 2017) . Harris County area is prone to flooding impact and is frequently impacted by service disruptions caused by flooding events such as transportation disruptions and power outages. The direct flooding impact to the grocery stores and the loss of access due to disruptions in infrastructure systems such as road inundations and power outages caused significant disturbance to people's access to grocery stores. Furthermore, Harris County, within which the city of Houston is located, is a metropolitan area encompassing populations of a diverse range sociodemographic characteristics, providing a representative testbed for examining the equitable access to grocery stores in the face of natural hazards. Data sources collected and analyzed in this study include location-based mobility data from Streetlight Data, points-of-interest (POI) visit data from SafeGraph, sociodemographic information from a 5-year estimate of the American Community Survey of the U.S. Census Bureau, Federal Emergency Management Agency (FEMA) Flooding Data, and the Food Access Research Atlas developed by United States Department of Agriculture. These data sources were aggregated at the census-tract level to construct the population-facility network models and to calculate access indicators and subsequent analyses. A detailed description of these data sources is provided below: Location-based mobility data: The mobility data are provided by StreetLight Data, a commercial platform that provides origin-destination (O-D) analysis data. Our analysis aggregated anonymized data from cell phones and GPS devices to create travel metrics, such as duration and distance (StreetLight 2021a). The O-D network of visits to grocery stores in Harris County was examined in this study from August 1 through September 30, 2017. Fig. 2 a shows the network of trips from traffic analysis zones (TAZs) to grocery stores during the second week of August 2017. These data incorporate the trips using different modes of transportation, including personal cars and public transit. By analyzing more than 40 billion anonymized location records across the United States in a month and enriching the analysis with other sources, such as digital road network and parcel data (StreetLight 2021b), StreetLight Data is capable of reaching approximately 23 penetration rate (InSight 2018), covering distinct census divisions in North America's road network. Thus, the data provide a proper sample of human mobility in Harris County for examination of grocery store access disparity in the face of Hurricane Harvey. Facility location data: POI data by SafeGraph were used to identify the location of facilities in this study. SafeGraph obtains location data by partnering with several location-based mobile applications. Data includes the basic information about the names, geographical coordinates, addresses, and North American Industry Classification System (SafeGraph 2021). In this study, the top facility categories related to grocery stores, specialty food stores, and general merchandise stores (including warehouse clubs and supercenters, restaurants, and other eating places) were considered as the grocery stores in Harris County. Then we manually filtered the facilities to ensure that these POIs are stores used by the Harris County residents to obtain grocery needs. Fig. 2b shows the distribution of these POIs in Harris County. Sociodemographic characteristics: Sociodemographic characteristics of census tracts were determined by collecting data from the demographic characteristics estimate over the 2015-2019 period of the American Community Survey. The sociodemographic characteristics of the TAZs were determined from their overlap with census tracts. Fig 2c shows the income level of the TAZs using the census tract data from the 5-year estimate of the American Community Survey. Sociodemographic characteristics, such as socioeconomic status, household composition, minority status, and transportation from the census data, were used to examine grocery store access disparity among different sub-populations. Flooding data: The Harvey Flood Depth Grid/Flood data from FEMA were adopted in this study to identify the flooded areas in examining the impacts of flooding on access (FEMA 2018). The dataset was developed using the gage points from the National Weather Service and the terrain data from USGS. In this study, the flood data were examined to determine the flooding extent in each TAZ. Then those areas with a flooding extent greater than 6.5% of the area, which is the 75th-percentile flooding level in Harris County, were marked as flooded to further examine the relationship between flooding status and disrupted access to grocery stores. Food access: The Food Access Research Atlas developed by USDA was used to identify areas labeled as having poor access to supermarkets in Harris County. According to the definition provided by the atlas (USDA 2021), a census tract is considered to have poor access to grocery stores if at least 500 or 33% of residents live more than 1 mile (in urban areas) or 10 miles (in rural areas) from the closest supermarket, supercenter, or a large grocery store. This information was implemented to interpret the findings and to understand the extent that poor access to grocery stores contributes to the variation of access in the face of a disaster. Method: First, we constructed the population-facility network models of the study area. Population nodes are TAZs, facility nodes are individual POIs, and the links represent trips. Three separate network models were created: (1) one with unweighted links; (2) one with weighted links representing trip duration; and (3) one with weighted trip distance links. Accordingly, network-based metrics were used for examining the spatial and temporal patterns of access to grocery stores. Then, the variations in access to grocery stores during the different disaster phases were analyzed to characterize access to grocery stores, to examine access disparities, and to identify the factors contributing to such disparities. Access indicators: In this study, we used network metrics for quantifying three distinct access indicators. These metrics, derived from the population-facility network models of visits to grocery stores, comprise topological and structural properties of the network of grocery visits. The three access metrics are: (1) the unweighted degree of TAZs (capturing number of unique POIs visited by the residents of a TAZ); (2) weighted degree of TAZs: trip duration (weighting indicates average trip duration; this indicator captures total average duration of residents at stores); and (3) weighted degree of TAZs: distance to stores (weighting indicates distance to stores; this indicator captures the total distance that residents of a TAZ take to access different stores). These metrics capture complementary characteristics of access, and are indicators of different access dimensions. The unweighted degree of TAZs captures redundancy for having access to grocery stores; the weighted degree based on duration captures rapidity of access and the weighted degree based on distance captures proximity. Table 1 summarizes the access indicators, dimensions of access, and the equations for measuring them based on the population-facility spatial networks. In these equations for calculating metrics for each TAZ ( ), are the elements of the adjacency matrix and _ and _ are the link weights based on the distance and duration of the trips from the TAZs, respectively. The access indicators ( Table 2 ) capture properties of access in different phases of the disaster. The number of unique visits to POIs informs about available options, which could provide redundancy in the face of disasters. Other access indicators (duration and distance), which capture the rapidity and proximity access dimensions, are influenced by human activities (lifestyle patterns) and infrastructure conditions (facility locations and road congestion). Table 2 summarizes the interpretation of these access indicators during the normal period, preparation, and impact/short-term recovery period. The normal period establishes the baseline access indicators. However, the preparation and impact/short-term recovery interpretations focus on the variation in access compared to the baseline (normal) period. This comparison with the baseline period provides insights regarding the effects of disturbance on residents' access to stores during the preparation and impact/short-term recovery period. Increased redundancy compared to baseline means greater recovery activity. Rapidity Duration (sec) Greater duration means a lower rapidity and longer commutes to grocery stores due to longer distance or traffic congestion. Reduced rapidity means a greater duration compared to baseline due to more time spent obtaining supplies as a result of longer commutes and traffic congestion. A lower rapidity (greater duration) compared to baseline means greater impact to access (due to store and/or road inundation). Proximity Distance (mile) Greater distance means a lower proximity and thus greater distance from stores. Closer proximity means greater distance compared to baseline, which means greater effort for obtaining supplies. A lower proximity (greater distance) compared to baseline means greater impact to access (due to store and/or road inundation). The examination of access to grocery stores based on the adopted access indicators was conducted across three disaster phases: prior to the event (without a disturbance to the network), during the period from the formation of Hurricane Harvey to prior to landfall (focus on preparedness activities), and during the landfall and recovery. The spatial and temporal patterns of impact and short-term recovery are evaluated to discover disparate access during different disaster phases. Then unsupervised learning approaches for time-series clustering characterize patterns of impact and short-term recovery on access indicators. Then, we performed statistical analysis to evaluate the influence of different factors (i.e., sociodemographic characteristics, location in a food desert, and flood status) on residents' access to grocery stores at different stages of disaster. The analysis steps are explained in the remainder of this section. Constructing the population-facility spatial network: First, we constructed the populationfacility network of commutes from TAZs ( ) to grocery stores ( ). This TAZ-POI network is a directed and bipartite network mapped based on the geographic coordinates (Li et al. 2021 ). The constructed network represents trips from TAZs to POIs as the links with weights ( ) based on distance and duration metrics. The three access indicators (i.e., visited POIs, duration, and distance) are determined based on a daily aggregated trip numbers on each link between TAZ and POI pairs. Calculating the percentage change of access indicators: The percentage change of each access indicator is calculated based on comparing the daily values with a defined baseline. The baseline period for each indicator at a TAZ is calculated considering each day as a unit, as a weekly pattern was observed in the data. The baseline period (August 1 through August 20, 2017) includes three weeks to define the weekdays' baseline values. Then the percentage change for each TAZ was calculated based on the following equation: where, , is the percentage change in the access indicator at TAZ ( ) in a date ( ). , is the access indicator, and , is the calculated baseline value for weekday corresponding to the date for determining percentage change. Then the resulting time series is used for examining the variations in access to the grocery store across different phases. Dealing with missing data: To deal with missing data on certain dates for the three indicators, we first filtered those TAZs which did not have data for 5 consecutive days. Then, the Kalman imputation method was applied to the data to deal with the missing data in the time series. The ImputeTS package in R was used to implement the algorithm on the univariate time-series for all the access indicators (Moritz and Bartz-Beielstein 2017) . The Kalman filter method uses the structural time series ideas where the system is outlined by a well-defined model with unknown parameters (Bianchi and Boyle 1999) . The maximum likelihood approach was implemented to determine the time-dependent model parameters (Alavi et al. 2006) . We implemented time-series clustering algorithms on the access indicators to characterize and examine the spatial and temporal variations. First, a 3-day moving average was applied to the time-series data of each access indicator to extract trend and limit noise. Then, a partitional clustering using dynamic time warping (DTW) distance in dtwclust in R (Package et al. 2019 ) was used to perform the time-series clustering. Partitioned procedures were considered to be optimization problems that maximize the inter-cluster distance and minimize the intra-cluster distance (Sardá-Espinosa 2019). DTW is a widely used approach for defining the distance in time-series data (Aghabozorgi et al. 2015) . In this study, we implemented a multivariable time-series clustering approach for classifying the TAZs. The 3-day moving average time series related to the access indicators for each TAZ were used to identify the TAZ clusters. Then, two cluster validity indices, COP and modified Davies-Bouldin, were used to determine the number of clusters (Arbelaitz et al. 2013) . The calculated access indicators for the TAZs (described in the Method section) were used to examine the access at three disaster phases: pre-disaster condition, during preparation, and during disturbance and recovery of the affected areas. The analysis covers a period between August 1 and September 15, 2017. The pre-disaster period (August 5 through August 16, 2017) covers the period immediately prior to landfall; however, no disturbance had occurred to the residents' access to grocery stores. This period captures access to grocery stores during normal conditions. Following the formation of Hurricane Harvey and the issuance of the hurricane watch on August 23, preparation activities were initiated, and some grocery stores were out of stock due to increased demand. Finally, Hurricane Harvey made landfall in Harris County on August 25, 2017, causing several road inundations and disrupting access to grocery stores, which affected Harris County residents' access. Pre-disaster normal period: To understand the key characteristics of the access in the normal condition, we analyzed the period before the Harvey landfall. Dynamic clustering was used to identify the distinct clusters of TAZs with similar access characteristics for each indicator. The results show the patterns of access to grocery stores for the different TAZs in Harris County. Fig. 3 depicts the identified clusters for the three access indicators together with the maps of these clusters (darker colors schemes show a greater value). These plots also show a spatial clustering pattern which suggests that areas in the proximate TAZs have similar access characteristics. The number of visited stores indicator, which captures the Redundancy dimension of access, shows a spatial clustering pattern in some areas. The TAZs in the East and South of Harris County show a lower number of visited stores, which translates to a low redundancy of access to grocery stores. The difference in the redundancy in access to grocery stores could be related to the number of nearby stores, the size of the stores, and households' lifestyles and shopping behaviors. On the one hand, the more stores available in the proximity of a TAZ, the more the visits would take place during normal conditions. On the other hand, the size of stores could also affect the frequency of visits to stores, as larger grocery stores with a larger, more varied inventory could meet diverse needs of the households; one visit would be adequate to satisfy the weekly needs for groceries and other supplies. Furthermore, some households rely on coupons or have lifestyles that require them to follow certain shopping behaviors that affect the number and type of grocery stores they visit. Therefore, the properties of the facilities, their distribution, and the lifestyle characteristics of the households could influence the access redundancy indicator. Proximity and rapidity indicators also show the presence of some spatial clusters, which suggests TAZs would have similar access indicators to their neighboring TAZs. Examining the three access indicators together, however, shows little overlap among the clusters for the three indicators. This result shows the spatial heterogeneity of TAZs in terms of their three access indicators. Fig 3d shows the aggregated map and clusters of TAZs based on the three indicators. In this map, the high and low clusters of the access indicators are used for characterization of access during normal times. The most frequently occurring categories are those with a low number of visits, low duration, and a low distance followed by those areas with a high number of visits and low duration and distance. Both of these clusters indicate good access. These categories (highlighted by light blue and green) with proper rapidity and proximity access are more concentrated in high-income areas west and southwest of downtown. The large TAZs in the north (highlighted in red) are the next most frequently occurring clusters with high levels of duration and distance and a low level of number of visits, which shows poor access in terms of proximity, rapidity, and redundancy. The association between the sociodemographic characteristics of the TAZs, as well as the facility distribution characteristics (i.e., number of stores in the TAZ and location in a food desert) with the access indicators, were examined to evaluate factors influencing spatial variations of access patterns. The results show that the identified clusters have distinguishable demographic and facility distribution characteristics. Table 3 shows the results of comparing the sociodemographic and facility distribution characteristics related to all access indicators for different clusters. ANOVA and chi-square tests were implemented to test if differences between the clusters are statistically significant at a 0.05 confidence level. The results for the number of visited POIs indicator show that the cluster with the highest number of unique stores visited (highest redundancy) had a lower socioeconomic status. These groups of people have diminished access to big supermarkets and are more likely to buy their needed supplies by using coupons from different stores to save money. In addition, the results show that this cluster, in fact, has more stores in the TAZs and also a lower chance of being located in a food desert (being far from a store) compared to other clusters. The availability of stores also partly explains a greater number of unique grocery stores visited by residents of TAZs in this cluster. The results related to the proximity and rapidity access dimensions do not show a clear disparity with respect to the sociodemographic characteristics. In fact, the results show that TAZs with a better socioeconomic status have lower proximity and rapidity access dimensions. This pattern could also be due to the fact that these TAZs are more likely located in residential areas, and their residents have a greater capability to commute longer distances to obtain their needed supplies from specific stores. In addition, these clusters are located in areas with a lower number of stores and a higher chance of being located in a food desert. Note: * shows the highest value for the comparisons, which are significant at a 0.05 confidence level through the ANOVA or chi-square test. Preparation: The percentage change in access indicators in comparison with the defined baselines was examined to assess the disparities in access to grocery stores in the preparation phase. In the time span between the identification of Hurricane Harvey until landfall in Harris County, residents attempted to store supplies, such as food and bottled water, to protect their households against hardship. Since there was limited mandatory evacuation by public officials, most residents sheltered in place; thus, storage of food and bottled water was critical to riding out the storm. Many grocery supplies were out of stock due to the high demand and the shelterin-place protective behavior of the residents, which disturbed the access patterns to grocery stores. We analyzed the percentage change in the access indicators to seek out disparities in access during the preparation phase. The percentage change of the access indicators on the day before the hurricane landfall (August 24, 2017) was chosen to examine disparities in access patterns. The examination of the daily patterns of access indicators showed that their greatest percentage change occurred on this date, and residents showed a high level of activity to obtain grocery supplies for their households. To explain the variations in the access indicators during the preparation period, we first examined the patterns in the pre-disaster level for the three indicators (as discussed in the previous section) together with income level. The examination of the association of these two factors with the percentage change in the indicators reveals that there is a statistically significant relationship between the access patterns in the normal condition with the percentage change of these indicators during preparation stage. The results show that those with access indicators of lower levels, namely, fewer unique POIs visited, shorter duration, and shorter distance, had a higher percentage change. We juxtaposed these results with the income level of the TAZs to better understand the effect of income and pre-disaster access level on the change in access indicators during the preparation. The TAZs were categorized based on their income (by considering high-income groups in the top 33% quantile and the low-income group in the bottom 33% of the income). The results show disparate access to grocery stores during the preparation phase of Hurricane Harvey. Fig. 4 shows the boxplots of the percentage changes in the access indicators. ANOVA and Tukey's tests show that there is a significant difference in the percentage change of the access indicators across income groups at 0.05 confidence level. The results (Fig. 5) show that the high-income group had a significantly higher increase in visits to unique stores compared to lower-income groups. This result indicates that high-income groups were able to improve their access redundancy to visit more unique stores during the preparedness stage to supply their households with adequate groceries before the hurricane made landfall in Harris County. A similar pattern exists in the rapidity dimension, as the TAZs with a longer trip duration (lower rapidity) in the normal condition showed a lower increase in the percentage change. However, trip duration disproportionately increased for the lower-income groups for those with long duration trips. While having a long trip duration to grocery stores in the normal condition severely impacts the access, low-income populations have a significantly larger increase in trip duration during the preparation period compared to that of populations in the high-income TAZs. The increase in the trip distance also shows that those TAZs which have shorter distance (better proximity) in the normal condition have a greater increase in the preparation period. However, the high-income groups show their capability to commute further to obtain their supplies and have the greatest increase in their distance in TAZs with a low distance to grocery stores during the normal condition. However, the increase in the percentage change of the distance is not significantly different across the income groups for TAZs with low proximity (long distance) in the normal condition. The findings of the effects of income and pre-disaster access condition on the access indicators during the preparation stage reveal the presence of an access disparity among income groups. These boxplots show the percentage change of the access indicators for the day before the hurricane landfall. The six groups categorized based on the income and pre-disaster access level show disparate access variations during the preparation phase: percentage change in the number of unique visited stores (a); the redundancy dimension of access also increase as this value increases; percentage change in the distance which has a negative relationship with the rapidity dimensions of access(b); percentage change of distance (c), the proximity access dimension decreases as the distance increases. In the next step, we examined the effect of facility distribution factors, namely the number of POIs in a TAZ and food desert status of a TAZ, on the access indicators. First, the correlation between the numbers of POIs in a TAZ with the access indicators (Table 4) shows that there is only a slight but significant association between the percentage increase in the visited POIs indicator and the number of POIs. TAZs with more POIs do not show a large increase in visits as they provide more options for the residents. However, the associations between the number of POIs and the duration and distance are not significant. The results indicate that change in access indicators during preparedness stage is not associated with the number of POIs in the TAZs where residents reside. This result shows that, during preparation stage and due to surge in demand for grocery supplies, residents need to take longer and further commutes to POIs outside their TAZs to obtain supplies regardless of the number of grocery stores in their proximity. Thus, those residents (low-income groups, as shown in earlier results) who could not increase their access distance would not be able to adequately prepare for the impending hazard. In the next step, we categorized TAZs into four categories based on income in conjunction with location in a food desert to understand the effect of these factors on access indicators Fig. 5 shows boxplots of the percentage change in the access indicators; the ANOVA test suggests a significant difference in access across the four categories. The results show an interaction between income and food deserts with access to grocery stores. High-income TAZs located in a food desert show a higher number of visited POIs, while the duration and distance are not significantly higher than that of low-income TAZs in a food desert. This result means that while being in a food desert and having lower redundancy in access, high-income TAZs increased their number of visited stores without having to increase their trip distance and duration while visiting more stores to obtain their grocery supplies in preparation for the hurricane. On the other hand, those TAZs which are not designated as a food desert and are high-income show the greatest increase in their access indicators. These TAZs have good access due both to not being in a food desert and also due to their capabilities to access farther POIs due to their higher income. Thus, being located in a food desert, while affecting people's access to the grocery stores to some extent, does not have an equal impact on the access of different income groups. Furthermore, when we integrate the results of food desert status with the sociodemographic information of the areas, the access inequalities reveal themselves. Also, this result shows that being located in a food desert is not adequate to evaluate residents' access to grocery stores during the preparation stage of disasters; access to grocery stores during preparedness stage is more influenced by the dynamics of human protective actions influenced by capabilities. Impact/short-term recovery: To understand the dynamic spatiotemporal patterns of impact and short-term recovery of access to grocery stores, we conducted multivariable temporal clustering on the three access indicators. The results indicate the presence of two clusters of access patterns to grocery stores in TAZs whose access was affected by flooding. Fig. 6 shows the dynamic pattern of access indicators for the identified clusters. TAZs in cluster 2, which are depicted in blue in Fig 6a, have better access and show a pattern that has a greater increase in the visited POIs indicator (increase in redundancy); however, the cluster shows a lesser increase for the distance (decrease in proximity) and duration (decrease in rapidity) indicators during the recovery period. In particular, the proximity indicator seems to show a lesser level of decrease (increase in distance) in cluster 2 compared to cluster 1. The patterns suggest that while cluster 2 visited more stores, their duration and distance did not increase significantly compared to cluster 1. This result shows that cluster 2 TAZs had better access, as they could meet their needs by visiting more stores without the need to significantly increase their trip distance. Increased redundancy without the need for decreasing proximity is an indication of better access for cluster 2 (compared with cluster 1, which did not have significant increase in access redundancy but had significant decrease in access proximity). (c) show that as duration and distance increase, rapidity and proximity decrease. The cluster identified as blue shows an increased level in visited POIs, while the distance and duration for this cluster show a better access pattern compared with the red cluster. Fig. 7 shows the comparison of the sociodemographic characteristics and the facility distribution characteristics in the two clusters. These boxplots (Fig 7 a-c) are the results of the comparisons among the TAZs in each cluster, which are significant at a 0.05 level of confidence. The results suggest that the TAZs in cluster 2 with better access have higher income and a lower percentage of the minority population. This result indicates that in the aftermath of disasters, the socially vulnerable population (low-income and minority groups) has lesser access to grocery stores. The lower degree of access is not particularly associated with the distribution of stores but could also be related to the capabilities of the population in cluster 1 in dealing with the impacts of disasters. While those TAZs in cluster 2 had a higher chance of being in a food desert and have a lower number of POIs in their TAZ, they have better access (in terms of improved access redundancy without decreased access proximity). The distribution of income level and location within a food desert in the two identified clusters was examined in a proportion test analysis. The results of the proportion test show that in cluster 2 with better access and a better socioeconomic condition, the proportion of highincome TAZs that are also designated as food deserts is significantly higher than that of lowincome TAZs in food deserts (0.53 compared to 0.25 with a p-value of less than 0.001). However, such a distribution did not exist in cluster 1, with proportions being 0.57 and 0.47 for the high-income and low-income TAZs, respectively, with p-value of 0.76. Thus, the high proportion of higher-income areas which are in food deserts allowed for better access in cluster 1 than in cluster 2 despite a higher level of being in a food desert. These results suggest that the food desert designation is not adequate to capture residents' access to grocery stores in the aftermath of disasters, and more data-driven insights (such as the results of this study) are needed to evaluate the access of different sociodemographic groups to critical facilities in times of disaster. Furthermore, the results from the comparison of the extent of flooding in the TAZs do not show a significant difference between the two clusters (p-value = 0.33). The insignificant difference in flooding shows that the diminishment of access to grocery stores goes beyond the extent of flooding; however, the sociodemographic characteristics and the facility distribution characteristics explain the disparate access to grocery stores for the two groups. and there are more grocery stores in the red cluster (e). Lastly, (f) shows that the level of flooding is not statistically different among the two groups. In this study, we analyzed high-resolution location-based data to characterize and uncover the disparities in access to grocery stores before and during disaster events. We developed and implemented access indicators based on the population-facility network structure and analyzed their spatial and dynamic patterns to characterize access to grocery stores and reveal the presence of disparities at different stages of a disaster. The high-resolution location-based data enabled capturing the patterns of trips of residents to grocery stores and facilitated the characterization of access based on observational data. The current approaches in examining access based on physical and distance-based metrics fail to provide adequate and reliable information about the properties of access during a disaster to enable decision-making for equitable access. The findings of this research show the importance of characterizing access based on the three dimensions of redundancy, proximity, and rapidity. The analyses of access dimensions during the normal time showed that areas with lower socioeconomic status have a higher redundancy, and their proximity and rapidity show no difference compared with areas of higher socioeconomic status. In fact, areas of lower socioeconomic status show lower proximity and rapidity for obtaining grocery supplies during the normal conditions; however, these groups have lesser access to the large supermarkets that provide a large variety groceries and other supplies, making them more dependent on visiting multiple stores. In addition, this subpopulation is more dependent upon public transportation and less capable of making longer commutes. The significance of the characterization of access is amplified by examining access indicators in a disaster context. A combined examination of the dynamic patterns of access indicators in the context of the 2017 Hurricane Harvey in Harris County, Texas, revealed that flooding disproportionately exacerbated access disruptions for socially vulnerable groups. Areas with higher income and a lower percentage of racial minorities have a higher redundancy with a lesser decrease in proximity and rapidity than areas with lower income and a larger minority population. Thus, these groups could visit more stores without longer and more lengthy commutes. The examination of access during the preparation period prior to the disaster also shows areas with a higher income show a greater increase in both the redundancy and the proximity, which reveals their higher capability to take protective actions compared to the lower-income TAZs in preparing for an impending hurricane. The examination of different dimensions of access in the context of disasters confirmed the inadequacy of physical measures of access, such as location in a food desert and the number of available stores in an area, for understanding access in disasters. The research findings show that while these factors are better capable of explaining the variations in access during normal times, they fail to explain access characteristics during different stages of a disaster. The multivariate analysis of access showed that the areas with higher income and a lower percentage of racial minorities have better access to grocery stores, while these groups, in fact, had a lower number of stores in their area and a higher likelihood of being in food deserts. Furthermore, the analysis of access during the preparation period showed that being located in a food desert connotes the lowest level of access, especially for low-income areas. Also, in areas that are not located in a food desert, the high-income TAZs have better access (i.e., more access indicators). These findings indicate that, while focusing on facility distribution characteristics is useful in understanding the disparities during the normal condition and for purposes of urban development, these measures are not sufficient for understanding access to facilities in the context of disasters. The data-driven evaluation of access and disparities in disasters inform future emergency response, preparedness, and mitigation plans and actions. For example, the placement of temporary distribution centers in areas with lowest levels of access during the preparedness stage could help with better and more equitable preparedness for shelter-in-place populations. In addition, infrastructure development and improvement plans could be informed by the findings of this study for equitable allocation of resources to enhance access to grocery stores. For example, roads which are critical for providing access to facilities can be identified and retrofits/enhancements could be implemented to mitigate flooding of those roads. In addition, the development of future facilities can be informed by the results of this study to enhance access in a more equitable manner both in normal times, as well as during disasters. The data that support the findings of this study are available from SafeGraph and Streetlight Data, but restrictions apply to the availability of these data, which were used under license for the current study. The data can be accessed upon request submitted on SafeGraph and Streetlight Data. Other data we use in this study are all publicly available. The code that supports the findings of this study is available from the corresponding author upon request. Time-series clustering -A decade review Accessibility inequality in Houston Filling gaps in evapotranspiration measurements for water budget studies: Evaluation of a Kalman filtering approach Reliable emergency service facility location under facility disruption, en-route congestion and in-facility queuing An extensive comparative study of cluster validity indices Household preparedness for the Aftermath of Hurricanes in Florida Uncovering the socioeconomic facets of human mobility Disaster preparedness and disaster response: Evidence from sales of emergency supplies before and after hurricanes A comparison of methods for trend estimation Capturing Disruptions to Food Availability after Disasters: Assessing the Food Environment following Hurricanes Florence and María Food deserts'---evidence and assumption in health policy making The Perilous Nature of Food Supplies: Natural Hazards, Social Vulnerability, and Disaster Resilience High-resolution human mobility data reveal race and wealth disparities in disaster evacuation patterns An Integrated Physical-Social Analysis of Disrupted Access to Critical Facilities and Community Service-loss Tolerance in Urban Flooding An integrated physicalsocial analysis of disrupted access to critical facilities and community service-loss tolerance in urban flooding An integrated physicalsocial analysis of disrupted access to critical facilities and community service-loss tolerance in urban flooding Moving from tripbased to activity-based measures of accessibility Susceptibility Curves for Humans: Empirical Survival Models for Determining Household-level Disturbances from Hazardsinduced Infrastructure Service Disruptions Travel time resilience of roadway networks under disaster Communication Status Report for Areas Impacted by Hurricane Harvey FEMA -Harvey Flood Depths Grid A Social Vulnerability Index for Disaster Management Accessibility and Recovery Assessment of Houston's Roadway Network due to Fluvial Flooding during Hurricane Harvey Factors associated with hurricane preparedness: Results of a pre-hurricane assessment Our Methodology and Data Sources Measuring physical accessibility to health facilities--a case study on Khulna City Fear During Natural Disaster: Its Impact on Perceptions of Shopping Convenience and Shopping Behavior Community-scale Big Data Reveals Disparate Impacts of the Texas Winter Storm of 2021 and its Managed Power Outage Keywords Disparate patterns of movements and visits to points of interest located in urban hotspots across US metropolitan cities during COVID-19 Uncovering patterns of inter-urban trip and spatial interaction from social media check-in data Reframing Resilience : Equitable Access to Essential Services imputeTS: Time series missing value imputation in R Quantifying community resilience based on fluctuations in visits to points-of-interest derived from digital trace data Empirical assessment of road network resilience in natural hazards using crowdsourced traffic data About Safegraph: Safegraph Docs Time-series clustering in R Using the dtwclust package A Deep Gravity model for mobility flows generation Location-based services data beats cellular on spatial precision Why streetlight: Our data Assessing spatial equity: An evaluation of measures of accessibility to public playgrounds USDA ERS -Documentation Disparities and access to healthy food in the United States: A review of food deserts literature Incorporating Resilience into Transportation Planning and Assessment Semiautomated social media analytics for sensing societal impacts due to community disruptions during disasters The authors would like to acknowledge the funding support from the National Science Foundation CAREER Award under grant number 1846069. The authors would also like to acknowledge SafeGraph and Streetlight Data for providing the points-of-interest and traffic data. Any opinions, findings, conclusions, or recommendations expressed in this research are those of the authors and do not necessarily reflect the view of the funding agencies.