Urban design attributes and resilience: COVID-19 evidence from New York City

1.1 Urban density and crowding

The COVID-19 pandemic arrived at a time when there is a lack of consensus about the effect of urban density on the spread of infectious diseases (Teller 2021). Prior published research contended that densely populated urban areas in cities provide optimal conditions for the spread of pathogens (Neiderud 2015). Thus, one strategy to decrease epidemic risk is to promote a more horizontally distributed urban development (Megahed & Ghoneim 2020). Other studies have countered this narrative, most of which addressed the negative relation between density and COVID-19 in terms of death rate, which is likely confounded by factors as varied as the increased availability of healthcare resources and higher income in urban areas. Fewer studies showed the insignificance of the population density on the infection rate (Hamidi et al. 2020), or its negative effects on the spread of COVID-19 in the early stage of the epidemic (Liu 2020). Jamshidi et al. (2020) highlighted the complexity of this association by showing that urban density may have multiple temporal roles in the spread of the virus, being a promoter of spread initially but then diminishing in importance as the pandemic grew.

There are many ways to measure density. In addition to the urban density that is commonly measured as the ratio between total population and the total area of a city or neighborhood, crowding has been used as a separate metric in housing studies (Blake et al. 2007; Forsyth 2003). Overcrowding refers to the perception that there are too many residents per building, per room, or per floor area. Studies such as the American Community Survey (ACS) define overcrowding as more than one person living in a single room (Blake et al. 2007). Density and crowding are related yet distinct concepts because highly populated urban districts are not necessarily crowded (Angel et al. 2021). Several pandemic studies have used crowding to identify epidemic hotspots (Emeruwa et al. 2020; Rader et al. 2020). However, further investigations are needed about different density and crowding indicators and their relationship to COVID-19 case rates or pandemic resilience in general.

1.2 Urban mobility and commute patterns

Another controversial topic is the epidemic risk associated with urban mobility and different commute modes in particular. There is a notion that public transport systems are major disseminators of viruses and ‘have seeded the massive coronavirus epidemic’ in NYC (Harris 2020). But some critics argue against this (Gordon 2020; PedestrianObservations 2020), stating that public transit is not necessarily a health hazard in the pandemic as many cities that rely heavily on public transit do not show higher infection rates than others. Scientific evidence supports that successful distancing and the use of face masks can effectively reduce infection risks in public transportation (Tirachini & Cats 2020).

Meanwhile, changes in urban commute patterns are already occurring. Public transit ridership and revenue plummeted around the world during the COVID-19 outbreaks (United Nations 2020). In large cities such as NYC, these trends can be quantified using large and openly shared data, such as the Citi Bike systems data (Citi Bike NYC 2020) and Metropolitan Transportation Authority’s (MTA) turnstile data (MTA 2020). Between January and July 2020, the total count of Citi Bike ridership has risen by 58%; the average duration of Citi Bike ridership has increased by 98%; and subway turnstile-use counts have shrunk by 73%. Further investigation is needed to understand how these commute modes are related to pandemic resilience, and which mobility strategies are suitable for the short-term rapid pandemic response and long-term mobility planning.

1.3 Urban activities and points of interests (POIs)

Population visits to POIs have been used to represent urban activities responding to the pandemic risk and the associated policies (Google 2020; Li et al. 2021). However, the impacts of the density and spatial distribution of POIs have been rarely considered in pandemic-related research. One reason for this may be that it is very difficult to capture the spatial distribution of urban activities and translating them into straightforward statistics. To the best of the authors’ knowledge, the only study that has integrated POIs’ visiting data into a Susceptible-Exposed-Infectious-Removed (SEIR) model proved that neighborhoods with more populated POIs are likely to have higher infection rates (Chang et al. 2020). To understand the relation between POIs, urban activities, and pandemic resilience, and to effectively inform urban design and planning practice, further investigations are needed that directly consider the diversity, density, and spatial distribution of POIs in the urban fabric.

1.4 Urban inequality

In addition, many of the denser US cities face a considerable amount of income inequality and related spatial segregation (Glaeser et al. 2008). This segregation has led to spatial disadvantages, including problems associated with the previously described KUDAs density and crowding, accessibility to mobility services, and allocation of diverse POIs. These problems result in systematically disproportionate pandemic vulnerabilities, as shown in the substantial variation of COVID-19 hospitalizations and deaths between different boroughs in NYC (Wadhera et al. 2020).

To consider pandemic resilience in future urban design practice, a deeper understanding of the spatial distribution of social inequalities and their impact on public health is needed. The current pandemic has brought health equity to the forefront of discussions in the design, planning, and science communities (Forsyth 2020; World Bank 2020), and the emerging urban data sets that record public health indicators and COVID-19 cases in a high spatial resolution provide a unique opportunity to understand and tackle this question.

1.5 Research questions

The aim of this paper is to understand the relationship between the four previously described KUDA categories and the COVID-19 case rates. From the perspective and scale of urban design and planning, the research goal is to leverage the zip code tabulation area (ZCTA) data for NYC to answer four main research questions in the context of NYC:

• How does urban density and crowding affect the infection rate, and what are their differences?

• How does the commuting environment relate to pandemic resilience?

• How do allocations of POIs impact the infection rate?

• How do evident inequalities in a city influence vulnerability during a pandemic?

2.1 Data collection and aggregation

Publicly available urban data from the US Census Bureau, NYC Department of City Planning, NYC MTA, US Environmental Protection Agency (USEPA), NYC OpenData, SafeGraph, and Google Maps Platform were collected. These data were aggregated them modified zip code tabulation areas (MOD-ZCTA) resolution that concords with COVID-19 data from the NYC Health Department. The data at the Census Block Group (CBG) level are aggregated into the ZCTA tile that contains the geometrical center of the CBG. Table 1 summarizes all urban variables used throughout the study, including the COVID-19 daily case rates, the sociodemographic variables, and the explanatory variables grouped by topics of density, crowding, mobility, and activity. The full sample size (i.e. the total number of MOD-ZCTAs in NYC) is 177, but some categorical groups have fewer observations because of missing data in certain tiles.

Daily new COVID-19 case rate (DCR) is used as the dependent variable and is retrieved separately for five phases in the timeline of the ‘first wave’ of the NYC pandemic. Phase 1 (P1, January 20–March 22) is the initial outbreak; Phases 2 (P2, March 23–April 5) and 3 (P3, April 6–June 7) represent the first two-week lag and the rest of ‘NY on PAUSE’ lockdown period (New York State 2020); Phase 4 (P4, June 8–July 19) is when NYC started to reopen; and Phase 5 (P5, July 20–August 26) is when the reopening process concluded and the city entered the ‘new normal’. March 27 is the earliest available date for the COVID-19 data from the NYC Health Department (2020). Thus, Phase 1 is discarded from all following analyses due to the complete lack of case data. It is also worth noting that the data collection ends on August 26 because the DCR has remained to be consistently low by then (an average of 3.34 daily new cases per 10,000 population), indicating a temporary halt in the case increase. Sociodemographic variables, including age, income, household size, rent, and health insurance, are key indicators to discuss social equity and its spatial differences in this study. Density variables, including building densities of the residential, office, and retail use, describe the urban construction density. Contrarily, crowding variables examine the people density within a confined space at the architectural level, which consists of building-level crowding (number of units per building), and room-level crowding (number of occupants per room). Mobility variables include population-level mobility behavior recorded before the pandemic (e.g. census data). This includes commute modes (percentage of the work commute by public transport (PT), walk, bike, car), commute time (work commute time by PT), and commute destinations (number of jobs within 45 min by PT). In addition, the mobility context during the pandemic as control variables (percentage of the population working from home, turnstile usage count, percentage of reduced turnstile usage) is included. The percentage of the population working from home is derived from SafeGraph (2020), a platform that provides mobility data aggregated from cell phones. It considers mobility patterns that are derived from typical full- or part-time work behavior, i.e. spending a specified amount of time at one location other than home within the time window from 8 a.m. to 6 p.m. The turnstile data come from the NYC MTA (2020).

2.2 Activity data generation

Activity variables aim to describe the spatial allocation of key POIs, including supermarkets, shopping malls, and parks. The common metric, such as the per capita number of POIs in each ZCTA tile, was not used. This type of simple aggregation measure entails a misrepresentation problem as the travel behavior of urban dwellers seldomly adheres to zip code boundaries and thus will be different for each location within a tile, especially when the location is close to the tile boundary. This data problem is particularly vital for studies using small-scale spatial tiles, such as ZCTAs, because the human activity radius often surpasses a single tile, as shown in Figure 1a.

Figure 1

Thus, the accessible POI area per capita was generated using the mobility simulation tool Urbano.io (Dogan et al. 2020). For each CBG i, its residential population p_i can go to all accessible POIs on the map with more people using the closer ones and fewer people choosing the distant ones. The population going from CBG i to POI j is denoted as w_ij, which follows the distribution of a distance-decay function F_ij (Figure 1a) (equations 1 and 2). After sending the populations of all CBGs to their POI destinations, the total spatial catchment of user count for each POI v_j can be estimated (Figure 1b) (equation 3). Assuming each POI’s area is known as s_j, how much area of individual POI destination j it occupies as w_ijs_j/v_j is then derived for each CBG i. By summing all occupied areas of CBG i and dividing it by the CBG population p_i, one can obtain the accessible POI area per capita denoted as x_i (equation 4):

where m is the number of all accessible POIs for the CBG i; F_ij is the distance-decay function referenced from National Cooperative Highway Research Program (NCHRP) Report No. 716 (National Academies of Sciences, Engineering, and Medicine 2012); t_ij denotes the network walking time between ij; and n is the number of CBGs that have residents going to the POI j.

It is worth noting that the same unit area s_j = 1 is assumed for all supermarkets and shopping malls because there are no data specifying the individual area of these POIs from Google Maps Platform (2018), and there is a limited range of differences in the area of these indoor places. However, for parks, the real area s_j is used. This is based on the shapefile from NYC Open Data (2018) as the outdoor green areas vary considerably and thus are more relevant. Note too that this method only considers spatial catchments of total users but ignores the temporal variances such as visiting frequency and visiting schedule. Although the latter factors are important to evaluate actual visitor density in POIs, there are no data that could inform detailed spatiotemporal POI visiting patterns at CBG or zip code resolution.

2.3 Statistical data analysis

The primary outcome of this study is the average daily new COVID-19 case rate (DCR) for each phase. Means and standard deviations are computed for each variable in each phase as descriptive statistics. Pearson’s correlation coefficient is obtained to determine pair-wise correlations among explanatory variables. Multiple linear regression (MLR) is performed in two steps to assess the contributions of the exploratory variables to DCR outcomes. Step 1 is to construct and examine DCR prediction models including all exploratory variables within each categorical group and for each phase. Step 2 is to construct and examine DCR prediction models including selected variables based on the result of Step 1 for each phase and by using a backward elimination stepwise regression approach (level of significance for staying in the model = 0.1) (Kohavi & John 1997). Potential multicollinearity is also eliminated in the MLR analysis.

All descriptive data analysis and the MLR analysis are performed using Stata 16.1 (StataCorp LLC 2019) and SPSS 26 (IBM 2020). Data processing and Pearson correlation were conducted in Python (Python.org 2020). The mobility simulation for activity data and the spatial plots for case studies were created in Rhino 3D and Grasshopper (Robert McNeel & Associates 2020). For all statistical tests, a two-sided type 1 error probability ≤ 5% is considered as the level of significance.

4.1 How do urban density and crowding affect infection rates?

This study clarifies how urban density and crowding represent two related but different concepts in pandemic-resilient design and planning theory (Forsyth 2003, 2020; Batty 2009) by identifying the specific effects of density and crowding on NYC’s 2020 COVID-19 epidemic.

High residential building density appears to be consistently advantageous for pandemic resilience in all phases in NYC. One reason is that high residential density often entails other beneficial urban features such as closer jobs, shorter commute time, and thus better mobility conditions as shown in the correlation matrix. After considering all relevant variables in Table 5, residential building density still has significant negative coefficients, which may be explained by other possible advantages mentioned in the literature such as effective pandemic prevention measures, well-informed dwellers, as well as good access to the internet and communication technologies (Forsyth 2018). Office and retail building density does not show significance in the present MLR analyses, which could be attributed to the high work-from-home percentage throughout the studied period as well as the less active business environment during the pandemic (Table 3). However, it is worth noting that high office and retail density in NYC are extremely concentrated in a few areas, as shown in Figure 4 (plots 7 and 8), which is concerning because it may entail high interzonal transmissions after in-person work and activities entirely resume.

Results further reveal the need to investigate crowding indicators at the architectural level. Building-level crowding appears to be an unfavorable urban feature in almost all phases in the pandemic, probably because many residents sharing the confined circulation space and facilities in buildings (e.g. elevators and hallways) can facilitate disease transmissions. Room-level crowding also appears to seed the surge of the infection in vulnerable communities such as The Bronx according to the present case study. However, statistical evidence is needed in future studies for a more profound understanding about the effects of these crowding indicators and their potential interaction with other factors such as income and age (Ghosh et al. 2021).

Given the evidence, developing compact cities with well-mixed land use remains a desirable planning paradigm as opposed to sprawl, as it provides a high-density living environment that can support overall better infrastructures for pandemic response, and sustain advantageous urban features such as proximities of jobs and services. However, to avoid potential ramifications of overcrowding, architects and developers should develop more resilient apartment building typologies, e.g. by offering flexible apartment layouts with enough rooms to facilitate occupant isolation, by providing safer congregate space through spatial thresholds and physical barriers (CDC 2020), or by offering alternative, well-ventilated and sun-exposed circulation routes. Especially in vulnerable, low-income communities such as The Bronx, such design strategies could help to mitigate existing overcrowding. For example, a newer development in The Bronx, Via Verde (Kimmelman 2011), with such resilient and affordable housing typologies, is located in the area that has a relatively lower case rate than the rest of The Bronx. Meanwhile, better data are needed to relate the findings to architectural parameters that it is believed are impactful, such as building entrance type, circulation space, ventilation, and daylighting conditions, so that more specific evidence can be provided to guide future design practices.

4.2 How does the commuting environment relate to pandemic resilience?

To address concerns about public transit systems facilitating the spread of airborne pathogens, the PT-related mobility variables were investigated. Both MLR models and case studies show that—in the current COVID-19 pandemic—a long commute time by PT is a more concerning urban attribute than the percentage of people commuting by PT. Numerous studies on indoor disease transmissions can help with understanding this result. The latest research shows that for closed environments such as PT vehicles, adherence to physical distancing, high-frequency cleaning, and proper use of face masks can successfully reduce the probability of contagion (Noakes 2020; Tirachini & Cats 2020), but the total viral load and the risk of inhaling infectious aerosols both increase with time spent in a confined space (EMG-SAGE 2020; Prather et al. 2020). Additionally, short-distance travelers have more flexible mobility choices and are more likely to transition to alternative modes, as shown in the correlation matrix that shorter PT commute time correlates with a higher percentage reduced turnstile.

Thus, to ensure the long-term viability of public transit in dense cities, urban designers should consider promoting planning schemes with dense and well-mixed urban zoning to reduce commute time in general, such as the decentralized neighborhood centers (Chu 2020), the walkable city (Claris & Luebkeman 2016), and the 20-min city (Capasso Da Silva et al. 2020). Mixing land uses and creating more jobs in proximity of homes can be an effective way of realizing this goal. Besides, low-income neighborhoods should be prioritized in future PT development as recent studies stressed that PT is more critical infrastructure to them during COVID-19 compared with the high-income population (Liu et al. 2020).

There is no definite evidence from the present MLR analyses proving the benefits of active transportation modes because the current data at hand account only for work commutes and they cannot broadly represent generalized biking and walking behaviors. However, considering that Figure 4 demonstrates areas with good walkability and bikeability to be more resilient in the current COVID-19 crisis, as well as the current drop in PT ridership and the emerging trend of transitioning to active transportation (Table 3), there is a great momentum to advance biking and pedestrian-oriented mobility infrastructures.

4.3 How does the allocation of points of interest within a city impact the infection rate?

In MLR analyses, higher accessible shopping mall area per capita relates to a lower DCR both before and after the lockdown period (P2 and P4). In the correlation matrix, it is found that low-income and low-rent communities have less accessible POI area per capita in general. These findings correspond well with Chang et al. (2020) who used the SEIR model to simulate that lower income CBGs have higher infection rates in part because they tend to visit more congested, and therefore higher risk, POIs. The present paper is one of the first to study the direct link between the spatial allocations of POIs and the infection rate. The given evidence calls for the considerate planning for more essential services such as grocery stores, pharmacies, food services, and other retailers in future city development.

As for parks, it is found that higher accessible park area per capita particularly relates to a lower DCR during the lockdown period. A similar finding was reported by Liu (2020) who found a negative coefficient for the per capita green space area when predicting city-level COVID-19 case numbers in China. Besides, the present correlation matrix shows that denser residential areas tend to have a less accessible park area per person, which is probably because the compact built environment has restricted space for green infrastructure development. Prior studies pointed out that residents in dense urban centers visit parks more often and reduce their trips to parks much less during the lockdown because densely populated areas offer smaller housing units and fewer other outdoor spaces such as private yards to help residents mitigate anxiety and stress (Hamidi & Zandiatashbar 2021). Thus, outdoor green parks, though commonly referred to as ‘non-essential’ POIs, still play a critical role for the public well-being during the pandemic response.

4.4 How do evident inequalities in a city influence vulnerability during a pandemic?

Spatial inequalities for the vulnerable (e.g. low income, low coverage of health insurance) and high-risk (e.g. large household size) population are perpetuated in the pandemic. These communities tend to live in more crowded houses in areas with poor accessibility to jobs and other urban services. As a result, these communities have longer commute times and fewer transportation alternatives.

To help alleviate the current spatial disparities for vulnerable communities, several approaches from the perspectives of previous discussions can be considered. Developing more affordable housing with resilient building typologies in dense areas can help accommodate more heterogeneous populations in urban centers. Bringing more job opportunities near these neighborhoods and advancing more active transportation supports can lower their travel-related risks as well as costs. Providing adequate indoor and outdoor POIs in proportion to the resident density and demands can not only mitigate their risks in interpersonal transmissions when visiting these public places but also increase their quality of life in general.