Our Take Care Texas project includes 3 geographically distinct sites within Texas, each with unique demographic characteristics (Multimedia Appendix 1). Southeast Texas, represented by the Houston area within Harris County, has an ethnically diverse population (N=4.73 million), covering both urban and suburban communities. This region serves as a significant residence for various racial and ethnic minority groups and is recognized for having one of the largest metropolitan economies in the United States. The South Texas site, represented by Cameron County (population: N=423,029) in the Rio Grande Valley, is primarily Hispanic, with a considerable proportion of the population (22.6%) living below the federal poverty line, exceeding the national average (11.5%) [29]. The Northeast Texas (NETX) region covers an extensive area, including the Smith, Gregg, Henderson, Van Zandt, Anderson, Wood, and Rains counties. NETX is characterized by a mix of small cities and rural communities, with a lower population density when compared to other sites. Anderson County was excluded from this study due to insufficient SARS-CoV-2 infection data coverage at the time of analysis. The demographic distribution across these regions reflects a wide socioeconomic and cultural spectrum, offering a comprehensive understanding of the diverse population within Texas. There are 2144 CBGs in Harris County, 222 CBGs in Cameron County, and 349 CBGs from the NETX region, according to the 2010 census data.

The process for identifying priority CBGs (PBGs) in this study is presented in Figure 1. The process began with the identification of geographic areas experiencing significant COVID-19 burden, using SARS-CoV-2 surveillance data and population estimates at the ZCTA level. We then identified CBGs within these areas that exhibited high socioeconomic disparities, using the community disparity index we constructed to achieve the desired spatial granularity. Detailed descriptions for determining disease burden and community disparities are provided in the Persistent and Recent COVID-19 Burdens and CBG-Level Community Disparity Index sections.

In response to the dynamic nature of the COVID-19 pandemic, our identification and selection process was conducted monthly with a near–real-time data feed from the local public health department for timely and effective intervention planning. Table 1 illustrates the stepped planning and intervention calendar. Between May 2021 and October 2022, a total of 10 rounds of CBG selection were conducted for each of the three study sites. In each round, among the CBGs within the identified high–disease burden ZCTAs, we selected the most vulnerable CBGs based on the community disparity index (ie, CBGs in the top 25% for this metric). We then excluded previously randomized CBGs, their neighbors (ie, CBGs that share a common geographic boundary), and any CBGs with small populations (total population<500) to form the PBG list, which served as the sampling universe. From the PBGs, we sampled geographically nonadjacent sets of CBGs (3 CBGs as a set), ensuring suitable geographic separation to mitigate potential intervention “spillover,” since the study goal was to compare 2 distinct intervention strategies to a control arm. We sought feedback from the community health workers and staff in the three regions, to determine if the potential CBGs presented safety, geography, and access concerns, by showing maps of the areas and then discussing the areas and documenting concerns. When there was lack of clarity about an area, we reached out via phone calls and emails to community experts in our advisory groups (health department leaders, city leaders, and local nonprofits serving the areas) to obtain their feedback as well. CBGs that posed challenges in reaching community members (such as CBGs with primarily gated communities or business districts), which would prevent implementing the intervention protocol (eg, door-to-door visits), were excluded from the sampling frame. This collaborative approach ensured that our selection process was informed by both data-driven analysis and community and implementer insights. The final selected CBGs within each set were then assigned to 1 of 2 intervention groups (multilevel intervention and just-in-time adaptive intervention) or a control group through baseline covariate adaptive randomization, balancing disparity index and population size, as well as vaccination coverage when such data were available. These CBGs then underwent a 1-month planning phase, followed by a 2-month intervention period. In each round, this study selected and randomized 6 CBGs (2 sets) for Harris County due to its larger geographic area and population, 3 CBGs for Cameron County, and 3 CBGs for NETX counties.

The persistent and recent high COVID-19 burdens were quantified by using SARS-CoV-2 surveillance data obtained from local public health departments. We computed a 7-day moving average for infection rate, which was defined as the daily new cases per 100,000 population. For each ZCTA, the persistent disease burden was evaluated based on the number of days on which ≥10 new cases per 100,000 population were reported, as per the common COVID-19 risk level threshold [30], starting from March 1, 2020, up to the respective analysis date for each selection round. To evaluate the recent disease burden, we first calculated the 80th percentile of daily new cases per 100,000 population across all ZCTAs within each study site, using data from the 30 days preceding each analysis date. We then counted the number of days on which the infection rate for each ZCTA exceeded this 80th percentile threshold. In each selection round and for each site, we ranked the ZCTAs, in descending order, based on their number of days with (1) a persistent high burden and (2) a recent high burden. The ranking was performed for each metric separately, and the ZCTAs in the top 50% for either metric were identified as high–disease burden ZCTAs.

We developed the CBG-level community disparity index due to the lack of spatial granularity in existing indices, such as the SVI, which did not meet our study’s needs. Using 12 SDOH measures from the 2014‐2018 American Community Survey 5-year estimates [31], our community disparity index covered the socioeconomic, demographic, and housing dimensions of SDOH (Multimedia Appendix 2). It integrated economic stability factors (employment and poverty rates, per capita income, education level, and health insurance coverage), demographic characteristics (percentages of individuals younger than 18 years, individuals older than 65 years, individuals in single-parent households, and racial minority individuals), and housing factors (percentage of renter-occupied units, rent burden, and crowded housing).

We used principal component analysis (PCA) to construct the community disparity index, similar to the approach used by Kolak et al [32]. PCA is a data-driven approach that synthesizes multiple pieces of information into one combined measure, while simultaneously handling high correlation and minimizing information loss [32-35]. We kept the principal components with eigenvalues exceeding 1, following the Kaiser criteria. The scores of these selected principal components were weighted according to the variance they accounted for and then summed, with higher scores representing a greater community disparity. The resulting indices were then ranked and scaled from 0 to 1 within each study region, providing a relative ranking of CBGs. All data were processed and analyzed with R statistical software (v4.1.2; R Foundation for Statistical Computing) [36].

We developed ArcGIS (Environmental Systems Research Institute Inc) and R Shiny (Posit PBC) data dashboards that provided key geospatial information and relevant COVID-19 data for this study. Feedback was received from the dashboard’s end users, including community health workers. The dashboards included maps that showed community disparity indices, PBGs, intervention-prioritized CBGs, COVID-19 testing sites, road maps, and SARS-CoV-2 infection and testing rates at the CBG and ZCTA levels. These tools enabled community health workers to identify the boundaries of the priority areas and access the relevant COVID-19 data for effective on-ground intervention messaging and referral to needed services. Both dashboards were designed to be interactive to facilitate strategic planning efforts and were password protected for internal team use only. The dashboard was especially useful during the planning phase and facilitated the creation of a community profile document, which community health workers used to share real-time data and deliver messaging tailored to the communities.

The research protocol was approved by the UTHealth Committee for the Protection of Human Subjects (HSC-SPH-20‐1372). Data user agreements were established between UTHealth and the public health departments of Harris County, Cameron County, and NETX counties. This study did not involve direct contact with participants, as data were aggregated at the CBG level, so no individual-level informed consent was required. Data were deidentified, and all analyses followed data privacy guidelines.

In this study, we used geospatial data science methods to develop a protocol for the adaptive selection of 120 CBGs in 3 regions of Texas, which occurred across 10 rounds over a 17-month period (May 2021 to October 2022). This protocol, which was developed for the Take Care Texas project, integrated real-time dynamic data on SARS-CoV-2 infection rates and contextual SDOH to identify and prioritize the CBGs most in need of tailored activities and resources to improve COVID-19 testing and vaccination. The adaptive selection of CBGs prioritized the most vulnerable communities and allowed for the rigorous evaluation of community-based interventions in a multilevel trial.

Our protocol demonstrated several unique strengths in integrating public health surveillance data and geospatial data science. First, our study offered spatial granularity at the level of CBGs, which has rarely been seen in other studies [11,16,37]. Such granularity could potentially result in more tailored interventions that are best suited to the local communities for which they were developed. Second, we incorporated both the reported SARS-CoV-2 infection rates and the community SDOH, reflecting the socioeconomic and housing conditions that contributed to health disparities. This allowed us to prioritize communities experiencing not only high disease burden but also high social vulnerability that could potentially impact community resilience and the ability to adapt to challenges presented by the pandemic. Third, the dynamic nature of our protocol, particularly the initial selection of high–disease burden ZCTAs, allowed for real-time adaptation of the selection strategies in response to the changing pandemic, thereby enabling us to prioritize the most in-need and most vulnerable communities effectively. Finally, the protocol was strengthened by the input from and collaboration among community members and frontline community health workers, which were critical in designing and implementing interventions. The final CBG selection was made by community partners considering on-the-ground intervention feasibility. In the context of this study and site selection, the input from our community partners and community health workers supplemented the data-driven approaches [38]. Although multicriteria decision-making methods have been previously used for site selection problems [39-42], including the integration of objective criteria with experts’ qualitative input to determine site suitability, some multicriteria decision-making techniques have been critiqued for arbitrary and irrational rankings of alternatives [43] and questionable mathematical validity [44]. By contrast, previous work has shown that PCA is a robust approach for integrating various vulnerability and health indicators and providing reliable unequal weighting [33,34] to help guide the selection of local communities for health care interventions. Future studies should continue to investigate best practices for engaging communities in trial design and implementation.

Our protocol also presented several limitations. First, the COVID-19 pandemic highlighted significant challenges in data reporting due to the absence of established surveillance data infrastructures for COVID-19 and other rapidly spreading communicable diseases. The state reporting requirements for data elements and definitions changed over time, complicating the consistent use of data across study regions. In our study, we directly collaborated with individual county public health departments and processed and harmonized data across different study regions, which were time-consuming processes. Second, the time lag commonly seen in surveillance databases posed challenges in accurately identifying current hot spots, and incomplete demographic data hindered the assessment of health equity throughout the COVID-19 pandemic by delaying the analysis of COVID-19 health outcomes by subgroups and across more granular geographic scales [45,46]. The Congressional Research Service’s 2020 report highlighted the need to modernize public health reporting systems to improve the timeliness and accuracy of public health data reporting [45]. However, differences in data standards, technical capacity, and regional policy priorities, particularly in the early stages of the COVID-19 pandemic, limited effective data sharing across study regions. Third, our protocol required accurate and complete address information for geographical granularity in identifying the high–disease burden priority areas, and when address information was missing, we may not have been able to include the SARS-CoV-2 case records in determining the PBGs. Finally, finding geographically separated CBGs in later rounds of selection could be difficult in small counties with fewer CBGs, such as Cameron County. Traditionally, administrative boundaries are used to define spatial units, yet alternative approaches may better capture actual community layouts. For example, Tuson et al [47] developed the Spatial Targeting Algorithm to generate flexible priority regions not restricted by administrative boundaries. However, while flexible boundaries may better capture neighborhoods, they may not provide SDOH information, which is typically accessible only at the administrative level. Additional work is needed to develop more flexible spatial units, with or without administrative boundaries as a constraint, for population-based research.

Despite these challenges, our study provided valuable protocols for data processing and analytical support to an intervention team working with local public health departments and organizations. It demonstrated a useful approach to ensuring that the most vulnerable communities participated in a multilevel intervention trial for increasing COVID-19 testing and vaccination uptake. Our proposed protocol could be adapted to other population-based randomized trials for which rapid assessment and allocation of finite resources are imperative. For example, geospatial approaches have been used to assess cancer screening rates among a vulnerable population in Ontario, Canada [48] and determine priority clinics to allocate funding for improving up-to-date colorectal cancer screening rates [49]. The integration of PCA and the real-time nature of our protocol can also be adapted to other population-based trials. When tackling multifaceted and complex problems, our protocol demonstrates that PCA is effective at robustly incorporating a wide range of health indicators. Further, the real-time nature of our protocol allows for ongoing, continuous assessment and adjustment in allocation of resources, provided that accurate and timely data are available. This could be applied to study recruitment protocols or other intervention-based trials.

The selection of optimal sites for health care and public health interventions, especially under limited resources, is a crucial challenge. Our protocol contributes to existing research by integrating disease surveillance and SDOH data, the robust PCA method, and community engagement to identify locations where COVID-19 testing interventions for vulnerable populations are most needed. The proposed data-driven approach could be adapted to other disease control and prevention programs to improve population health and reduce disease burden in areas with disparities.