Cohort geospatial inventory

In 2021, three study team members (JAH, LMB, MPJ) drafted a geospatial inventory form to capture essential geospatial methods and measures within each C4R cohort (S4 Appendix in S4 File). The form documents each cohort’s: (1) format and frequency of collecting and geocoding addresses, as well as the types of addresses collected (e.g., residential, work, school) and level of geocoding (e.g., latitude-longitude or Zip + 4); (2) types of neighborhoods assessed, such as administrative data appended to geocodes (e.g., city, state, county) and the availability of buffers (e.g., Euclidian/straight line or network/street buffers); and (3) measures of the environment collected, created, or attached to participant geocodes. Specifically, the form assessed available neighborhood-level measures (i.e., already linked to participant locations) of the natural environment (e.g., greenspace, air quality), built environment (e.g., food stores, social and walking destinations, housing density), and social or economic environment (e.g., racial and ethnic composition, neighborhood residents’ socioeconomic status). Within each of these three broad categories, the form asks about specific neighborhood metrics, the method of collection or derivation (e.g., survey, GIS, audit), the data source and year, the level of measurement (e.g., census tract or Euclidian buffer), and the time period(s) in which the measure was linked to locations (e.g., baseline, visit year). Lastly, the form collects COVID-specific neighborhood measures: number of cases, number of deaths, vaccination rates, policies, and disparity indices (e.g., race- or income-based differences). For each COVID-specific neighborhood metric, the form asks for the specific metrics available, the source of the data (e.g., survey, government agency), the level of measurement (e.g., city, zip), and the timing of measurement (e.g., by years, months). The draft inventory form was reviewed for clarity and completeness by three investigators and was piloted to assess usability in one of the C4R cohorts before finalization for use with the remaining 13 C4R cohorts.

Cohort study design

The C4R Study design, as previously described [20], followed a cohort ancillary studies model. Researchers in each cohort study were directly responsible for accomplishing data collection in accordance with the standard protocols and under the supervision of their own observational studies monitoring board, steering committee, institutional review board (IRB), and any other applicable regulatory authorities. Columbia University served as the Data Coordination and Harmonization Center (DCHC) for C4R (Columbia University Institutional Review Board, IRB-AAAT3035). A full list of cohort IRBs supervising implementation of the C4R protocols is provided in S2 Table. For the purposes of the work presented in this manuscript, data were accessed at Drexel University from February 28, 2022, to April 26, 2024. Drexel IRB determined that work within this manuscript was not human subjects research. Cohort data inventories were conducted beginning February 28, 2022, and finalized January 9, 2023. Study data were previously recorded as part of the C4R Study and were handled in accordance with all applicable confidentiality and data protection regulations.

Cohort geospatial data acquisition process and ensuring confidentiality

We (JAH, LMB, MPJ, STD, TC) contacted representatives from each of the 14 C4R cohorts requesting data manuals for each cohort’s geospatial data from which information was extracted using the geospatial inventory form. We also looked for published manuscripts using measures of interest. This was done to minimize effort for each individual cohort. When each inventory was complete, we confirmed with the cohort that our assessment was accurate.

To provide data on the geographic coverage of C4R, we requested a list of geospatial units for participants’ residential locations. We requested cohorts to provide the geographic administrative units to which participant addresses were geocoded (e.g., census tract, ZIP code, county, state) which were the smallest spatial resolution shareable without potential participant identification. While select cohorts have workplace locations, we did not include these. Geocoding data were provided to us in an anonymized format (i.e., as a list of administrative geographies) and no individual participant data were linked to these locations.

Public Data Sources Linked to Geographic Coverage of C4R

Utilizing the American Community Survey (ACS) 2015−2019 5-Year estimates and the geospatial data acquired for cohorts, we derived relevant population, social, and economic context information for a comparative analysis between areas where C4R participants have lived and those in which C4R participants have not lived. We used census tracts that match the 2015−2019 ACS boundaries, derived from the National Historical Geographic Information System (NHGIS) 2019 Census Tract Boundary File: United States and Puerto Rico [21]. These measures included total population density (pop/km²), median household income, as well as percentages of older adults (≥65 years, ≥ 85 years), non-Hispanic Whites, foreign-born residents, home ownership (% rent, % own), residents below poverty, residents with formal educations (% with high school (HS) degree, % with college degree), residents employed, commuting by car, and crowding. Index of Concentration at the Extremes (ICE) was calculated from ACS data using standard methods [22] for both the high-income White households versus low-income Black households and the high-income White non-Hispanic households versus low-income Hispanic households. ICE is scaled from −1 to +1 where −1 indicates that 100% of the population in the given area is concentrated among the most deprived group, and +1 means that 100% of the population is concentrated in the most privileged group.

Data regarding urbanicity were derived from the McAlexander et al. community type classification [23]. We operationalized urbanicity as the percent of total tracts that were classified as “high density urban”, “low density urban”, and “small town/suburban”, which make up all categories other than “rural” (and any undesignated tracts) as classified by McAlexander et al. To measure hospital access, we used staffed hospital bed counts per county from the Area Health Resources Files 2018–2019 to create a number of beds per 100,000 population density measure [24]. Data on greenness was derived from the 2019 National Land Cover Database (NLCD) [25] by including six categories of land cover previously used to generate a greenness measure (proportion of each category) at the census tract level [14]. We reclassified the categories into one single “green” category and used zonal statistics to get the percent areal coverage per census tract.

We captured metrics of neighborhood COVID burden from the Johns Hopkins COVID-19 dashboard (cases per quarter) and the Centers for Disease Control (CDC) COVID Data Tracker (vaccination rates) [26,27]. We aggregated daily COVID-19 case counts per county from the Johns Hopkins COVID-19 Dashboard to quarterly counts then used population to create a quarterly cases per 100,000 population density measure. The CDC publishes data on COVID-19 vaccination rates per county, which we used to indicate completed primary series vaccination rates at the end of each quarter (Jan-Dec). We examined case rates from Q1 of 2020 until Q4 of 2021, and vaccination rates from Q1 of 2021 (when a vaccine was available to a majority of the population) to Q4 of 2021. Despite access to data for 2022, we did not examine these rates due to a lack of reporting that likely makes these data not representative of overall COVID levels.

Analyses

Inventory data on neighborhood measures within each cohort were grouped and summarized based on overall availability and construct. For example, since many cohorts did not obtain neighborhood housing type measures (i.e., age of housing, mix of multi-family vs. single family), but more collected home ownership (i.e., % within neighborhood who own or rent), we combined these into a single category (i.e., “Home ownership/housing conditions”). Similarly, many cohorts measured greenspace through park data or through satellite images, making it harder to differentiate between greenness measures and park access (which are related), so these were combined into “Greenspace/Parks.” Within each category, measures were counted and summarized based as 0 measures, 1 measure, 2–5 measures, 6 or more measures. For example, for “Access to Food Stores/Food Environment,” a cohort might have data on densities of fast-food establishments, restaurants, supermarkets, convenience stores, wholesale markets, and fruit and vegetable stands, as well as ratios of types of stores (e.g., healthy to unhealthy sources). This would count as 6 + measures. Similarly, data were summarized as being derived from GIS (e.g., the densities in the prior example) or derived from surveys to participants (e.g., self-reported access to fruit and vegetables within their neighborhood). We did not receive information from the cohorts indicating that they have audited any environmental characteristics using in-person audit tools. Finally, cohorts either had measures from only one time point (e.g., baseline or during an ancillary study) or multiple time points longitudinally (e.g., continuously, at each exam); therefore, we summarized data within each category using that dichotomy. All neighborhood measures were visualized in a heatmap that could illustrate data timing (color family), number of measures (intensity of color), and type of measures (pattern). Cleaning and summarization of inventory data was done in Excel, and the heatmap was created in R (ggplot2 package, RStudio 2023.06.1 with Base R version 4.3.0).

We received participant locations from each cohort in different aggregated geographic administrative levels and vintages. For example, one cohort provided counties in which participants had lived in 2015, while another provided tracts from four separate decennial censuses, ranging from 1980 to 2010. To harmonize these disparate boundaries, we merged and dissolved all files into one dataset and then overlaid it with 2019 census tracts from NHGIS. Tracts with at least 5% of their area overlapped by C4R participant geographies were included as “C4R areas” in the final tract-level datasets for comparison. All counties that these tracts fell in were used in final county-level datasets. Remaining census tracts and boundaries in the contiguous US were considered “non-C4R areas.” Only the contiguous US was considered for any aspect of this study, meaning that “national” measures refer to data for that subset only. All spatial processing was performed in ArcGIS Pro 2.9 (ESRI, Redlands, CA).

All public data sources on social, economic, and other contexts were joined to either the county- or tract-level geographies on matching Federal Information Processing Standards (FIPS) codes. These were then aggregated nationally, for C4R coverage area and for non-C4R coverage area. To calculate representativeness and coverage, total population and land area measures were summed in the aggregation. All other ACS 2015–2019 measures were aggregated by means. Therefore, final aggregated measures are means of tract-level measures, not percentages or indices of that level. For example, the HS Degree (%) measure is the mean of all tract-level percentages of persons with at least a high school diploma, rather than the total percentage of persons with a high school diploma. We used this method of aggregation for all additional measures. We described means and standard deviations (SD) for population, sociodemographic, economic, and environmental context data for C4R locations, non-C4R locations, and nationally. We used t-tests to calculate p-values comparing the means of C4R locations to means of non-C4R locations. All aggregation and analyses were done in Python 3.9 (Wilmington, DE).

Address and geolocation data

COPDGene was the only cohort that did not centralize recorded addresses on inception of the cohort (i.e., address information kept within the 21 sites), instead centralizing addresses during a subsequent exam year once informed consent language was updated. Most cohorts updated their participants’ address information each time an exam was performed, with some cohorts updating information more frequently. For example, REGARDS recorded an initial address at the beginning of the study, then asked participants to update their addresses annually and confirmed addresses every six months during a follow-up call.

While collection and maintenance of address data was (and still is) done by all 14 cohorts, fewer cohorts (n = 12) have processed these participants’ addresses to identify their geographic location (i.e., geolocation or geocoding to a latitude-longitude for linkage to other geographic layers). Twelve of the cohorts (ARIC, CARDIA, COPDGene, FHS, HCHS/SOL, JHS, MASALA, MESA, NOMAS, REGARDS, SHS, SPIROMICS) have derived geospatial measures for administrative boundaries (e.g., tracts, zip codes, or counties), and six cohorts (ARIC, CARDIA, HCHS/SOL, JHS, MESA, REGARDS) have derived measures for Euclidian or network buffers (e.g., 1 km around residence). Cohorts varied in use of Euclidean or Network buffers and size of buffers around residential addresses; some cohorts calculated buffer sizes using metric (1 km, 3 km, 5 km) and others used imperial (0.5 mi, 1 mi, 3 mi, 5 mi). Buffers used were also specific to certain GIS measures (e.g., some cohorts measured distance to nearest roadway, which is not a buffer, while others calculated densities of road network within a buffer around a participant’s residence).

C4R geospatial coverage

C4R cohort participants covered 28% of the US contiguous land area, which includes 52% of the US total population (Table 1). Generally, compared to the US, C4R locations are more urban, with higher population densities, lower proportion non-Hispanic White residents, and higher foreign-born populations. The higher urbanicity of C4R locations can also be seen in higher proportions renting, a lower proportion commuting by car, and slightly lower greenness. Several locations within the US have participants from multiple C4R cohorts (“cohort overlap”) (Fig 1). These areas include San Francisco, CA, Minneapolis, MN, Chicago, IL, Birmingham, AL, Winston-Salem, NC, Baltimore, MD, and New York, NY. These cities represent potential opportunities to harmonize and attach raw environmental GIS data available in one cohort to another cohort that may not have these data.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Population, sociodemographic, economic, and environmental characteristics for coverage of the C4R cohort geographies compared to national and areas not covered by C4R cohorts.

https://doi.org/10.1371/journal.pone.0352170.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Geographic Extent of C4R Participants.

Geographic coverage obtained from cohorts anonymized and aggregated by spatial unit (e.g., county, census tract) and time (e.g., baseline, most recent exam, all exams). Since participant locations were provided at varying time periods and geographies, the map does not necessarily represent locations during COVID-19 pandemic. Geographic boundaries shown in Fig 1 come from U.S. Census Bureau's TIGER/Line shapefiles.

https://doi.org/10.1371/journal.pone.0352170.g001

Neighborhood measures

In our inventory of the C4R cohorts’ geospatial data, twelve cohorts (ARIC, CARDIA, COPDGene, FHS, HCHS/SOL, JHS, MASALA, MESA, NOMAS, REGARDS, SHS, SPIROMICS) indicated that they have measured neighborhood geospatial data for their study participants (Fig 2). Two cohorts (FIP/PrePF, SARP) indicated they have not collected, processed, or linked any neighborhood geospatial data for their participants (blank columns, Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Neighborhood data in the C4R cohorts by timeframe, measure type, and number of measures.

Abbreviations: ARIC, Atherosclerosis Risk in Communities; BCL, Biorepository and Central Laboratory; C4R, Collaborative Cohort of Cohorts for COVID-19 Research; CARDIA, Coronary Artery Risk Development in Young Adults; CONNECTS, Collaborating Network of Networks for Evaluating COVID-19 and Therapeutic Strategies; COPDGene, Genetic Epidemiology of COPD; FHS, Framingham Heart Study; HCHS/SOL, Hispanic Community Health Study/Study of Latinos; JHS, Jackson Heart Study; MASALA, Mediators of Atherosclerosis in South Asians Living in America; MESA, Multi-Ethnic Study of Atherosclerosis; NOMAS, Northern Manhattan Study; PASC, Post-Acute Sequelae of SARS-CoV-2 Infection; PrePF, Prevent Pulmonary Fibrosis; REGARDS, REasons for Geographic and Racial Differences in Stroke; SARP, Severe Asthma Research Program; SPIROMICS, Subpopulations and Intermediate Outcome Measures in COPD Study; SHS, Strong Heart Study.

https://doi.org/10.1371/journal.pone.0352170.g002

Built environment data

Built environment data were collected by nine cohorts (ARIC, CARDIA, FHS, HCHS/SOL, JHS, MASALA, MESA, NOMAS, REGARDS). The most common data collected by cohorts was access to food stores or food environment. These cohorts measured neighborhood amenities for retail food stores, such as wholesale clubs, grocery and convenience stores, chain stores, co-ops, fast-food, specialty, meat and fish, and supermarkets. Three cohorts (JHS, MESA, REGARDS) utilized data from the National Establishment Time Series (NETS) dataset to calculate changes in the business environments of their cohorts over time [28]. This dataset allows for cohorts to track the locations and types of businesses dating back to 1990, enabling longitudinal neighborhood change analysis. CARDIA, HCHS/SOL, JHS, MESA, and REGARDS all retained data on commercial buildings, land use, walkability, recreation/physical activity facilities, and alcohol/tobacco/marijuana outlets (counts per buffer or density per land area). These data were often derived from the NETS dataset, InfoUSA/ReferenceUSA, or similar dataset offerings from ESRI (i.e., ArcGIS Business Analyst). Traffic and road infrastructure, often primarily calculated using distance to arterial roads, were measured by eight cohorts (ARIC, CARDIA, HSCS/SOL, FHS, JHS, MESA, NOMAS, REGARDS); this was disparate from street connectivity, measured using both density of intersections, ratios of Network to Euclidean buffers, and other measures of connectivity.

Social environment data

Social environment data were the most plentiful data type connected to the C4R cohorts. Most cohorts have collected aggregated data on neighborhood socioeconomic status and data on urbanicity and population density, with all cohorts in possession of neighborhood-level data reporting some level of socioeconomic status. Most commonly, cohorts linked participant data to neighborhood employment data, median household income, education levels of residents, and poverty derived from the US decennial census and ACS. Cohorts that measured urbanicity of their participants’ communities often relied on rural-urban commuting area (RUCA) codes, a tract-based classification that uses standard census measures of population density, levels of urbanization, and journey-to-work commuting to characterize all US census tracts with respect to their rural/urban status [29].

Five cohorts reported measuring levels of neighborhood social engagement and cohesion (CARDIA, JHS, HCHS/SOL, MASALA, MESA). As expected for these types of constructs, these measures were assessed using survey tools administered to cohort participants. Uniquely, MESA measured these constructs simultaneously at the community level through administration of the survey to nearby, non-MESA participants and then created neighborhood estimates within a set buffer of MESA participant residences using Bayesian estimation techniques [30]. Five cohorts incorporated levels of violence and crime (HCSC/SOL, JHS, MASALA, MESA, REGARDS). Sometimes perceptions of violence or crime were measured (e.g., in MESA and their community survey), while other times crime data were GIS-based from local municipalities or the CrimeRisk Index (provided through ESRI from Applied Geographic Solutions) [31].

Natural environment data

Natural environment data were the least available construct of geospatial data available among cohorts. Nine measured air quality (ARIC, CARDIA, HCHS/SOL, FHS, JHS, MESA, NOMAS, REGARDS, SPIROMICS), five measured greenspace/parks (CARDIA, HCHS/SOL, JHS, MASALA, MESA, REGARDS), three measured bluespace (CARDIA, HCHS/SOL, REGARDS) and three measured temperature/weather (CARDIA, FHS, MESA), and only two measured hilliness/elevation (CARDIA, MESA). Cohorts that measured air quality included data on multiple pollutants, especially nitrogen dioxide, ozone, particulate matter less than 10 um or 2.5um (PM₁₀, PM_2.5), and nitrous oxide, measuring them in many different ways (e.g., modeling, averages during specific time periods). Some cohorts collected air quality data from EPA air pollution sensors [32], while many used more complex modeling techniques to estimate fine-scale air pollution [33]. Cohorts that measured greenspace/parks with park access generally quantified park density, size, and characteristics frequently sourced from local parks departments, city governments, or ParkServe [34]. When measuring natural vegetation, cohorts often used the National Land Cover Database from the USGS [25] or the Normalized Difference Vegetation Index (NDVI), a measure of vegetation calculated using ArcGIS from Landsat or Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery [35]. Three cohorts measured some type of bluespace data (CARDIA, HCSC/SOL, REGARDS), measuring distances to the nearest shoreline and surface land use that is coded as water or ice. The three cohorts (CARDIA, FHS, MESA) that measured temperature (e.g., average variation, extreme temperature events) and weather (e.g., precipitation levels) tended to source data from the National Oceanic and Atmospheric Administration (NOAA) local climatological data archive [36] or local monitors (e.g., Logan International Airport). Almost no cohorts measured light pollution or noise, so these were not included in Fig 2.

COVID Pandemic contextual data

To date, none of the surveyed cohorts have linked neighborhood-level COVID data to participants’ addresses. While the C4R project gathered self-reported COVID within participants’ friend and family networks, a large gap remains in connecting participants to the COVID levels or pandemic responses (e.g., policies) they experienced around their residences.

Forthcoming geospatial or neighborhood data

On or before January 2024, several cohorts reported geospatial measures being developed by investigators (through funded projects, ancillary studies, or approved manuscript proposals). These measures do not appear in this manuscript as they are not yet readily available for outside researchers through the coordinating centers. Examples include new segregation measures in ARIC, new disaster and climate measures in MESA, and new segregation, redlining, and neighborhood disinvestment measures in JHS (S3 Table).

3.8.1. Sociodemographic, economic, environmental, and COVID representativeness of C4R geographics.

Areas where C4R participants have lived or presently reside are more dense, urban, affluent, formally educated with a higher proportion foreign born residents and less car-dependent, with lower proportions employed and green than the non-C4R locations and the US overall (Table 1). Variability exists across cohorts (not shown). However, unsurprisingly, since many cohorts are based at institutions located in major cities, the geographic coverage of C4R is more dense (2613.08 population/km²) and urban (82.69) than the entire US (2087.58 and 75.61, respectively) and non-C4R locations (1575.32 and 68.72, respectively, p < 0.001). Relatedly, C4R areas have higher percentages of residents with a college degree and higher median household income, but also have a higher proportion of residents renting, below the poverty line, and unemployed. Tracking with their higher urbanicity, C4R locations have lower percentage of residents commuting by car and more racial and ethnic diversity as represented by both higher proportion foreign-born residents and lower non-Hispanic White residents. Interestingly, using ICE comparing only high-income White households to low-income Black households showed C4R locations have less concentrated privilege, while using ICE comparing high-income non-Hispanic White households to low-income Hispanic households showed C4R locations to have more concentrated privilege than non-C4R locations or the national average.

Estimates for the community COVID context of C4R participant locations showed that they generally experienced higher rates of infection and slightly higher rates of vaccination (Table 2). For example, in Q2 and Q3 of 2020, when the pandemic first peaked, the case rates for C4R locations were 536.74 cases per 100,000 and 1498.62 cases per 100,000 people, respectively, while the national rates were 489.68 cases per 100,000 people and 1435.37 cases per 100,000 people, respectively. We observed lower case rates in C4R locations in Q4 of 2020 and Q4 of 2021 than the case rates of the nation and non-C4R areas. Vaccination rates were always higher in the C4R locations than for the nation and non-C4R areas.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. COVID Experience for the C4R cohort geographies².

https://doi.org/10.1371/journal.pone.0352170.t002

Application of findings to research questions

Our inventory provides a comprehensive overview of environmental data available across C4R cohorts and highlights opportunities for future harmonization relevant to COVID research. The comprehensive characterization of pre- and mid-pandemic neighborhood conditions with sufficient geospatial variability and participant diversity can enable interdisciplinary research. Specifically, these pre- and mid-pandemic data allow the examination of environmental factors’ contribution to COVID risk and disparities, including on infection, morbidity, mortality, persistent symptoms, and vaccination uptake. Place-based differences in COVID impact by urbanization or city characteristics have been demonstrated across countries [37,38] and within the US [39]. Examinations of how neighborhoods impacted COVID outcomes informed previous prevention efforts, such as opening up park space or closing streets to create protected pedestrian spaces [40,41]. This work can also inform the creation of environments that are resilient to future public health disasters [39,42,43]. The geographic representation and sociodemographic diversity of the C4R cohorts, along with their combined sample size, permits research into how policies or neighborhoods contribute to COVID and related health disparities. Emergency measures in other domains implemented by some locales, such as eviction moratoriums, could be helpful to public health and used as a basis for future policies to protect populations [39,44–47].

The geographic and individual diversity of C4R could also be leveraged to understand how neighborhoods and policies differentially impact COVID and other health outcomes by participant characteristics such as race and ethnicity, sex/gender, age, and socioeconomic status. This may help identify the conditions and populations for which policies are most effective. Ultimately, while the COVID pandemic intensity wanes, lessons learned about neighborhoods during the COVID pandemic and recovery period can be applied to enhance future disaster preparedness [48]. However, this can only happen if efforts are made to integrate individual-level risk information to avoid both ecological bias [48] and confounding by prior conditions. Longitudinal data from C4R helps to overcome some of these limitations, allowing for better control of changing ecological conditions over time.

C4R cohorts’ neighborhood environment and contextual data could also be leveraged to understand critical shifts that occurred in built, social, and natural environments post-pandemic and their subsequent health impacts. These may be primary shifts that occurred early during the pandemic (e.g., policies around eviction, stay-home orders, pedestrianized streets) or enduring changes to retail environments, physical spaces, economic conditions, or population distributions. Evidence suggests that early pandemic prevention measures lowered pollution levels, limited access to public spaces, closed retail establishments, and produced lost income [49–51]. A number of cities made intentional health-supportive changes to their environments (e.g., pedestrianized streets, parklets), with varying success retaining these as the pandemic has progressed [52–54]. Similar evidence has begun to emerge regarding shifts in health-related contexts and social environments, including crime [55–57], neighborhood trust or social cohesion [58,59], and movement patterns [60–62]. As the country and world ease into a “new normal,” it is essential to understand long-term impacts of the pandemic period on our lived experience, including neighborhood conditions. Further, the addition of novel forthcoming geospatial measures (e.g., segregation, resilience to climate disasters) could be integrated into existing datasets to encourage additional inquiries as conditions evolve. Connecting C4R neighborhood data to the extensive pre-pandemic (1975–2020) and planned (2020-present) deep phenotyping in the cohorts on physical function, cardiac measures, brain measures, sleep, biomarkers, and genotyping, among others, overcomes confounding by lack of knowledge regarding previous health behaviors and trajectories. Linking neighborhood data with these cohorts would provide an opportunity to understand how the COVID pandemic and related neighborhood or community changes affected other health conditions commonly assessed across the cohorts, such as cardiovascular risk, aging, and related diseases.

Future harmonization

Despite the myriad opportunities afforded by C4R neighborhood data, challenges remain for harmonization and future work. First, the pandemic disrupted data collection and infrastructure systems across the country and world. Specific to the US, there are known concerns about the quality and viability of Census 2020 and American Community Survey data that include the year 2020 (i.e., all five-year estimates from 2016–2025). This is similar for any other measure traditionally collected and processed in-person (e.g., building inspection measures), but is less likely to impact remote sensing measures (e.g., greenness). Furthermore, the validity and meaning of these measures may have changed due to pandemic restrictions (e.g., land uses designated for commercial purposes were closed and their public functions suspended). Second, although many cohorts have measures within a domain, these measures use different metrics, were collected at different levels of resolution (e.g., counties vs. census tracts), or were based on different underlying administrative and feature data. For example, some cohorts may source land use measures from parcel data collected in specific jurisdictions while others use national satellite imagery. Similarly, some may calculate intersection density for street networks while others use block length or a ratio of Network buffers to Euclidean buffers. Existing survey data are rich and comprehensive within each cohort but very sparse across cohorts, and without it, we may lack the ability to understand individuals’ perceptions of neighborhood amenities or culture pre-, mid-, and “post”-pandemic. Even within cohorts that collect survey data on perceived safety, social cohesion, and belonging, there are no signs of convergence toward a single standard set of tools. These inconsistencies make harmonization of existing data challenging. In addition, in several cohorts, neighborhood and spatial data were created, stored, and managed by ancillary studies rather than centrally within the coordinating center that currently interfaces with C4R. Harmonization is further hindered by the sensitive nature of spatial data; there are human subjects concerns when sharing geospatial data such as addresses and the various cohorts and university IRBs have different restrictions about the identifiability of participant latitudes-longitudes.

Despite privacy concerns, options for harmonization exist that circumvent some of these obstacles. One option is to create national-level metrics at a small administrative scale (i.e., block groups or tracts) that could be linked by each cohort to their participant locations. Alternatively, the creation of a national set of gridded metrics (i.e., as GIS rasters) would facilitate linkage to locations across a smooth, interpolated surface. These options create a balance between consistency across cohorts and ensuring data are backwards compatible, especially for measures where there is little to no ability to create historic metrics. While these workarounds may help with privacy concerns, some cohorts have limited internal resources, experience, and trained staff for managing geographic or spatial data to do the final linkage. Although harmonization is important to study neighborhood impact on COVID, pandemic impacts on everyday contexts, and their subsequent effects on health, these are only one arm of the many pressing scientific lines of inquiry possible for these cohorts. Ultimately, all cohorts are balancing different competing priorities and must remain accountable to their primary missions (e.g., cardiovascular, pulmonary), existing ancillary studies, and participants. Harmonization will require considerable financial resources across an extended timeline, and execution by a committed and experienced team of diverse researchers.

Limitations and strengths

This paper provides a comprehensive description of the neighborhood and spatial data across cohorts but is not without limitations. First, many cohorts were missing some of the details on software for geocoding and, as in many transdisciplinary studies, we needed to establish common language around geography terminology and concepts (e.g., administrative boundary, uncertain geographic context problem, modifiable areal unit problem). Missing technical details may have occurred when neighborhood measures were compiled for ancillary studies but not distributed through the broader network of cohort researchers. When inventorying neighborhood data, the use of a survey and self-report by staff within cohort coordinating centers may introduce recall bias. We minimized these issues by collecting data documentation to confirm reported survey data and looking for published manuscripts using measures of interest. To facilitate harmonization, future work should prioritize documentation of software details and centralization of ancillary study datasets, code, and documentation at coordinating centers. In addition, as stated in our results, the inventory is a snapshot of what cohorts had in February-November 2022 and does not represent projects underway. For example, some cohorts had recently funded grants to incorporate new measures that were not yet calculated at the time of this inventory. We have tried our best to capture these forthcoming measures in the Supplemental files but may have missed upcoming research.

Our calculation of spatial extent has limitations which may introduce ecological bias and related issues, including modifiable area unit problem. Due to differences in human subject restrictions or levels of geocoding, cohorts sent residential location data at different spatial scales (e.g., cities, counties, census tracts) and across different spatial vintages (e.g., census 2000 tracts, census 2010 tracts). For anonymity, we only present data aggregated across all cohorts (i.e., counts of cohorts in a location, not the named cohort) and some cohorts additionally aggregated their data across time, providing all locations that a participant has lived across any exam. Additionally, cohorts with older adults (e.g., ARIC, MESA) have some participants who live in the southern US during the wintertime, and those wintertime locations are not necessarily captured in our presented data. Therefore, the data we presented on spatial representation combines time, spatial unit, and cohort. As such, the map of the geographic extent of C4R participants does not necessarily represent participants’ locations during the COVID-19 pandemic and the comparison of this extent to national demographics may include some spatial and/or temporal misclassification that could introduce bias. For research on specific exposure-outcome pairs, future work should try to gain approval and access to participant data including their COVID-19 specific outcomes and data collected at a spatial and temporal resolution best fitting the study’s specific aims, which would help to minimize ecological bias (e.g., modifiable area unit problem, ecological fallacy). The map also represents residential areas of C4R without weighting to account for differential counts of participants for given geographies (e.g., some census tracts may have one participant, some may have many). Similarly, our calculations of sociodemographic and COVID measures for the geographic extent required reaggregation of the spatial scale at which those data were provided (e.g., census tract for census, county for COVID data). This assumes an even distribution of each characteristic across the areal unit. Furthermore, we were limited to counties for national COVID data, and different reporting issues occurred at different points in the pandemic (lack of access to testing early in the pandemic, use of at-home tests later in the pandemic that may have not been reported to the city, etc.). Finally, to ease interpretation, all data presented are summaries and do not represent harmonized measures across the cohorts. More specific details on the availability of geospatial data for individual cohorts should be requested from the authors or directly from each cohort; access to these data should be requested through the approved processes of each cohort.

Despite these limitations, several strengths enhance the utility of this paper to guide future research on how environmental context shapes health outcomes and disparities. First, we used a centralized, systematic method for cataloging available geospatial data through a uniform data collection tool and process across all cohorts. Our graphical display of these data allows researchers to see potential synergies that would otherwise be difficult (or impossible) to recognize. In addition to sparking novel research questions and enabling well-powered analyses across diverse samples and regions, our clear summarization could be used as a model for other projects. Our method could be applied to those seeking to harmonize data across cohorts or as a guide for the creation of new measures (e.g., noise) that could be developed for one or more cohorts. Second, capturing key elements across these metrics, including differences in how data were collected (e.g., survey data vs. GIS data), allows researchers interested in utilizing these data to be informed about the inconsistencies across cohorts. This points researchers toward potential limitations of their analyses and provides information to interpret disparate study results across cohorts. Third, our systematic approach provides a roadmap for researchers who would like to collect their own geospatial data. For example, our collection process has already led to discussions across cohorts regarding best measures for specific environmental features. We believe that this summarization will encourage selection and operationalization of measures and methods that increase the coherence of the evidence created. This would enable even more scientific rigor and lead to the creation of larger and more diverse datasets that can answer highly complex questions about the intersection of environment and human health. Finally, our diverse and interdisciplinary team of national leaders in neighborhood health demonstrates best practices for team science. Our process ensured that multiple viewpoints were represented, ultimately encouraging shared methods and promoting collaborations with a wide range of disciplines with great potential to improve public health.