An alternative vaccine prioritization approach in response to COVID-19 pandemic

Coronavirus Disease 2019 (COVID-19) is caused by a novel virus strain, Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) (CDC, 2020a), and has soon become a global pandemic. It is first detected by the end of December 2019, and shortly after, it has spread to nearly every country. By early April 2021, more than 150 million people have been infected with COVID-19, and it has caused over 3 million deaths worldwide. These numbers continue to rise every day (John Hopkins University, 2021). The primary mode of transmission is through respiratory droplets expelled from the mouth or nose of an infected person and inhaled by a healthy person (CDC, 2020b). The exponential growth of cases with COVID-19 is due to its high transmission rate. The basic reproductive number (R₀) of COVID-19, which is the average number of secondary cases generated from a single infectious case in a completely susceptible population, was estimated to be higher than 2 in early studies (Li et al., 2020), and later updated to be higher than 3.

By the end of 2020, the first COVID-19 vaccines started to be distributed. As of April 2021, one billion doses of vaccines have been administered. However, vaccine demand quickly surpassed the available supply, similar to all other medical supplies, i.e. Personal protective equipments (PPEs), testing kits and ventilators used in COVID-19 response (Thompson and Anderson, 2021). As a result, prioritizing different groups for vaccine access became inevitable. The most common approach was prioritizing the older population, front-line workers and people with severe health conditions (e.g. cancer patients) (Cylus et al., 2021). However, slow vaccine roll-out rates through age-based prioritization yielded infection hotspots among socially vulnerable communities (Dasgupta et al., 2020). There is a clear need in models that can identify these hotspots in early phases of pandemic to slow-down widespread infection rates.

Humanitarian logistics domain predominantly use equity measure when allocating critical limited resources to large population groups (Hasnain et al., 2021), but this approach does not reflect the vulnerabilities present within those populations. Indeed, in a recent systematic review, researchers review more than 300 studies and conclude that vulnerabilities of individuals or populations and appropriate pandemic response efforts for these communities have not been studied in the literature (Sokat and Altay, 2021). However, the pandemic actually further deepened the socio-economic inequalities (Perry et al., 2021). Accordingly, developing approaches that prioritize vulnerable communities especially for essential goods such as food (Breitbarth et al., 2021; Kumar, 2020) and vaccines (Degeling et al., 2021) become a research priority. The need for such studies in fast and effective vaccine scheduling is further amplified due to the extra-cold-chain requirements of vaccine distribution supply chains (Comes et al., 2018).

One particular source of vulnerability manifested itself in terms of food accessibility. Following the national guidelines, social distancing has become a general norm in communities, which affected the individual concerns in food access and stockpiling activities (Brizi and Biraglia, 2021). Food supply chains are also stressed with unprecedented shifts in demand from downstream to upstream (OECD, 2020). While food security is a significant concern during the pandemic, food access can also be a determinant of COVID-19 exposure risk (Siddiqi et al., 2020). This also raises questions about whether food insecurity has any influence on the dynamics of the community spread. Given these gaps in the literature, the research objectives (RO) are as follows:

In order to address RO1, we develop mathematical models that investigate the significance of several vulnerability measures on the COVID-19 outcomes, namely case prevalence and case mortality. To measure the vulnerability caused by the food accessibility, we use the Supplemental Nutrition Assistance Program (SNAP) beneficiary data and the food desert indicator as the two indicators for food access in the United States (US). In addition to food accessibility indicators, we also incorporate Centers for Disease Control and Prevention (CDC)'s social vulnerability index (SVI) variables in combination with the underlying chronic health conditions, such as diabetes and obesity. We have assessed the associations of these social determinants and health factors and their significance with the COVID-19 prevalence, case-fatality rates in the US, as shown in Figure 1. In our analyses, we classified the counties as urban and rural and ran the models separately to assess any differences and factors that contribute to these differences in rural and urban settings. We found that the food desert population proportion is correlated positively with the COVID-19 prevalence, and the SNAP population proportion correlated positively with the COVID-19 prevalence. While we did not see any statistically significant relationship between being in a food desert and COVID-19 related mortality, the proportion of the population on the SNAP has a positive correlation with the case fatality rates.

To achieve RO2, we convert the results obtained by our models to a COVID-19 Vulnerability Score. The CDC initially developed the SVI to help decision-makers to identify vulnerable communities that need support during disasters and disease outbreaks (CDC, 2018). However, this index is not encapsulating other factors that have a significant impact on COVID-19 outcomes. To this extend, we calculated the relative importance of each COVID-19 vulnerability factor identified by our models and then, combined them as a normalized score that can be used to compare and rank different counties in terms of their susceptibility of adverse COVID-19 outcomes. We validate the developed scores by comparing the actual COVID-19 progression in the counties and observe that the COVID-19 vulnerability index identify the potential vulnerabilities quite consistently. We also compare the rankings generated by proposed approach with the commonly used pro-rata prioritization strategy and discuss the benefits of vaccine prioritization through the proposed scores. We believe that the proposed methodology contributes to the literature by providing a data-driven COVID-19 vulnerability score which can be used in prioritizing potential disease hotspots which should be rapidly targeted by any public health program, e.g. testing, vaccination, food aid distribution, in response to the current pandemic. Furthermore, the proposed methodology demonstrates the significance of addressing these hotspots in early phases of future pandemics.

The remainder of this article is organized as follows: In Section 2, we summarize the related literature on food access measures and the effect of COVID-19 on supply chain management. In Section 3, we explain our modeling approach. In Section 4, we provide the results of the regression models and the implications of the developed COVID-19 vulnerability score. In Section 5, we validate our results using the metrics of two socio-economically diverse states: New York and Nebraska. Then, we compare our results with commonly used pro-rata approach. In Section 6, we conclude by discussing the implications for theory and practice, key lessons and limitations of the study.

While a pandemic possibility continuously poses global risks to public health systems and business continuity (Araz et al., 2020), federal and state health departments develop public health policies for mitigation and response (Ramirez-Nafarrate et al., 2019). Researchers and public health officials have used models and quantitative analyses to evaluate their effectiveness and costs in developing social distancing policies under possible pandemic scenarios. However, most of the time, unintended long-term consequences of these interventions have been ignored during the policy decision-making process.

Social determinants of health have been attracting great attention recently by the media as COVID-19 disproportionately has affected under-served populations worldwide, and more specifically in the US. Healthy food access has been identified as one of the predictors of certain chronic conditions, e.g. type 2 diabetes and obesity, especially in low-income communities (Widener et al., 2012). While policymakers and public health experts have been working on improving food access in these low-income populations (Widener et al., 2013), the COVID-19 pandemic has clearly shown that there are striking health disparities in the nation, which is still driven by food and healthcare access, as well as other social and economic indicators, e.g. insurance coverage. SNAP data also present great value not only on food security but also preventive health care services utilization across the US (Bronchetti and Christensen, 2019). Recent studies also indicate that participation in the SNAP has a positive effect on physical activity and physical health in the elderly population (Pak and Kim, 2020). The purchasing power of SNAP beneficiaries with a geographic variation on local food prices across the US is also investigated (Christensen and Bronchetti, 2020). This study suggests that the local food prices have significant effects on food access and related outcomes. A recent review finds that the negative impact of the COVID-19 pandemic on the food security of low-income children and families has the potential to magnify pediatric obesity later (Tester et al., 2020).

While food access and security has become more critical for low-income populations, the food supply chain and logistics have become a determining factor for the health outcomes (Loske, 2020). Another recent study presented a simulation-based scenario analysis on how the COVID-19 pandemic can affect the global supply chains (Ivanov, 2020). A decision support system is also developed specifically for the COVID-19 pandemic to manage demand in healthcare supply chains and reduce stress in communities and mitigate the effects of epidemics such as healthcare supply chain disruption (Govindan et al., 2020). Regarding the COVID-19 pandemic mitigation efforts and population access to related services, maximal coverage model is formulated to ensure COVID-19 testing access using the pharmacies in the US (Risanger et al., 2021).

These studies show that there is a great need for models and quantitative approaches for ensuring more equitable access to healthcare services and critical supplies, e.g. food, testing kits, vaccines, etc., during the pandemic. Our study contributes to the literature by empirically analyzing the effects of food access in the US by using the food desert classification of the counties, and the proportion of SNAP beneficiaries within each county, in addition to chronic disease conditions represented with the diabetes and obesity prevalence on the disease progression. Given that the data-driven approaches can help the decision-makers to design effective response programs during a rapidly evolving pandemic (Ordu et al., 2021), with this model, we aim to provide a tool to the decision-makers that they can use to identify and prioritize vulnerable communities as quickly as possible while determining social assistance and vaccination schedules.

In order to assess whether lack of food access has been playing an essential role in the COVID-19 cases and deaths across the US, we have used county-based data for each state and for a given time period at the beginning of the pandemic. In addition to the food desert indicator and SNAP program availability, we considered SVI such as age, employment, disability and minority population rates, and finally, diabetes and obesity prevalence. Before developing models to assess the highly associated factors and significant predictors of the COVID-19 related health outcomes, we first introduce all of these variables considered in the study. Then, we present correlations between specific vulnerability measures and the COVID-19 outcomes in the gathered data set.

In this section, we describe the dependent and independent variables used in our econometric models. These models aim to determine the factors that have a significant impact on COVID-19 related health outcomes.

As potential COVID-19 related health outcomes, we focused on case prevalence and crude mortality rates in our analysis. We collected the cumulative number of cases and deaths as of 08/06/2020, as well as the population estimates for each county from the USAFACTS database (USA Facts, 2020). USAFACTS reports county-level COVID-19 data after cross-validating with state- and county-level public health agencies, so it is considered a reliable source. We have excluded the counties that have less than 30 cases as of 08/06/2020 from our analysis. Then, we estimated our two dependent variables: Covid-19 Case Prevalence as the number of COVID-19 cases per thousand population and Mortality Rate as the number of COVID-19 related deaths per thousand populations.

To measure the food access-related vulnerabilities, we used the Food Access Atlas developed by United States Department of Agriculture (USDA)'s Economic Research Center. Food Access Atlas reports the populations that have limited access to healthy and affordable food retailers. For each census tract, the Atlas reports the number of people residing far (different distance markers are available, e.g. 0.5, 1, 10 miles) from a supermarket or a large grocery store. Based on this data, we calculated the percentage of the population in a given county with limited food access, or as commonly referred in the literature, living in a food desert. For the urban counties, the 1-mile distance marker and rural ones, the 10-mile distance marker, are considered food desert cut-off points. Besides the physical access to food sources, economic barriers can also impact health outcomes. To incorporate this factor, we also added another variable, SNAP Beneficiary that represents the percentage of the households receiving funds under the SNAP.

CDC developed SVI to identify and prioritize vulnerable populations in the event of a disaster (CDC, 2018). To achieve this, they have constructed and estimated 18 different variables representing the socioeconomic statistics, household composition and minority status of the households in each county. Among those variables, we included unemployment, age, disability and minority statuses in our models. For our analysis, we used the most recent estimates that are compiled in 2018.

It has been widely accepted that underlying health conditions can significantly impact the progression and the outcome of the COVID-19. Chronic conditions such as diabetes and obesity have been widely researched in the past, so data is available on these conditions. We used county-level age-adjusted percentage of adults that are diagnosed with diabetes and obesity in 2016 collected through the Diabetes Interactive Atlas by CDC (CDC, 2016).

Table 1 shows the description of the variables as well as their sources, and Table 2 shows the summary statistics for these variables with the means and standard deviations. While the food desert variable shows significant variation between urban and rural locations, the proportions of the population using the SNAP beneficiary program are very similar in both urban and rural counties. Table 3 shows the Pearson correlation computed for each pair in the used data set. Next, we present the empirical models developed to address the research questions stated above.

In order to understand the effect of social vulnerability index, i.e. unemployment rate, age distribution, disabled population and minority proportion, with the food access indicators, i.e. food desert and SNAP beneficiary variables, on COVID-19 related health outcomes, we developed cross-sectional multiple linear regression models with observations from 1,837 U S. counties out of 3,141. We did not include the counties with less than 30 COVID-19 cumulative cases at the beginning of June 2020. We tested four different models: for each independent variable (case prevalence and mortality), we developed one model for the rural counties and another for the urban ones. We have used regression analysis for couple of reasons: first to assess the explanatory association with the COVID-19 health outcomes (cases and deaths) with factors listed above in different communities. Due to its simplicity and the nature of the data used, regression analysis is a convenient methodology for this study. Second, the developed models would be useful for making simple predictions using changing values of model inputs. However, we also acknowledge that more complex systems level simulation models can be developed with the integration of epidemiological disease spread models. Our study would be useful to inform such future simulation-based studies.

The models developed for the COVID-19 prevalence and mortality are presented in Equations (1) and (2), respectively. The models are represented with the upper index m ∈ u, r denotes the models for the urban counties(u), corresponding models are also developed for rural counties (r).

All analyses are performed using Stata Standard Edition version 16.1. The results of these regression models inform the computation of a COVID-19 vulnerability score, which considers the food access as an input as well. Next, we describe the methodology for computing this score in detail and demonstrate an application case in the US.

While the multivariate regression analysis is helpful to identify the explanatory variables for the COVID-19 related health outcomes, focusing only on the significant variables can be misleading, as their coefficients signify the isolated impact if all other predictors are kept constant. In order to determine the relative importance of each COVID-19 vulnerability factor, we calculated the standardized regression coefficients using the estimates computed by the multivariate regression model. The standardized regression coefficient for an independent variable x_j is calculated as follows:

where $βjm$ is the regression coefficient, and y is the dependent variable. The standardized coefficients are unitless, and they are often used to assess the importance of each predictor in a regression model (Murray and Conner, 2009).

Here we would like to develop a score that can help prioritize the counties for COVID-19 vulnerability, which is more comprehensive than the SVI as it takes into account additional factors. In order to create a measure that ranks the counties from the most vulnerable to the least, for each county, we have multiplied the standardized coefficients with the percentile ranks of that county for each vulnerability measure. In this way the counties that appear in the highest percentiles for the measures that have larger relative effect would be ranked higher.

Let the percentile of the county k for an independent variable x_j is p_kj. Then, first, we compute their crude COVID-vulnerability score c_k as:

Then, we normalize these values among their respective group (i.e. rural or urban) so that each county k has a relative COVID vulnerability score $ck¯$ that takes a value between 0 and 1, where 0 is the least vulnerable and one being the most. While the counties can be prioritized solely based on their $ck¯$ values, another possible way to rank the counties is to rank them by their population-weighted COVID vulnerability scores: $popk*ck¯$⁠. Figure 2 demonstrates this process for estimating the COVID-19 vulnerability scores.

After presenting our methodology for computing the COVID-19 vulnerability score for each US county, we present our results in this section. First, we present the analyses of the regression models for both prevalence and mortality outcomes and then present the results for the COVID-19 vulnerability score.

We have developed two models, one has the prevalence as the dependent variable, and the other has the case fatalities. Both models are differentiated for rural and urban counties. In terms of the COVID-19 prevalence, we observe that both the rural and the urban counties with higher minority populations have reported higher prevalence rates. The other significant factor for rural counties is the proportion of the elderly population. Table 4 shows the results of the multivariate regression model for the COVID-19 prevalence rates. Counties with a higher proportion of the population with residents of age above 65 observe a lower COVID-19 prevalence. This may be explained by the lower mobility of the older population. For the urban counties, we observe a significant negative effect of food access vulnerability measures. Counties with larger population groups living in a food desert are expected to have higher COVID-19 prevalence. Similarly, having a higher proportion of SNAP beneficiaries in the population positively affects the case prevalence. Counties with higher unemployment rates are also likely to have higher prevalence effects. On the other hand, the counties with higher disability rates are likely to observe lower case prevalence. This can again be associated with the reduced mobility of disabled populations. Finally, diabetes has a weak positive correlation, and the young population has a weak negative correlation on the case prevalence of the urban counties.

Table 5 shows the results for the COVID-19 mortality effects. Two vulnerability factors have similar effects in rural and urban counties: Minority and older populations. As expected, counties with older populations seem to have higher mortality per population. Strikingly, counties with more minority groups are likely to have higher case prevalence and mortality rates. No other factors seem to have a significant effect on mortality in rural counties.

For the urban counties, we again observe the significance of food programs (p < 0.05). Counties with a higher proportion of SNAP beneficiaries have higher mortality rates. Similar to the prevalence effects, unemployment also has a significant positive relation with the mortality rates. The age effect is also in line with the current COVID-19 risk factors and observed related outcomes. Counties with older populations have reported higher mortality rates and the converse is also true: counties with a larger proportion of the population aged <17 observe lower mortality rates. This is an expected outcome. However, counties with higher rates of disability seem to observe lower mortality rates, which may not be explored in the literature before. In addition, counties with higher obesity rates have lower mortality rates, which is unexpected. This may be explained with the limited data reported, relative to the peak times of the pandemic, and fatalities would be observed and reported with 3–4 weeks of lags. Overall, the COVID-19 prevalence and mortality models have very similar results. Since the number of mortalities at the time of this study was significantly low, we find the case prevalence estimations as more robust, relative to the mortality model. Thus, we will focus on those estimates for the remainder of the paper. The standardized coefficients for both models are represented in Table 6.

In this paper, our main goal is to develop a more comprehensive COVID-19 vulnerability measure that can inform policies for food aid program prioritization and even be considered an alternative to some vaccination prioritization of communities and neighborhoods, etc. within a state or a county. In order to establish this measure, we have analyzed the predictors and their relations. The existing vulnerability measures in the literature are in additive formats, and the collinearity among the predictors is almost never considered (Murray and Conner, 2009). However, since we used them in a multivariate regression model, we need to investigate further whether the multicollinearity among the independent variables leads to estimation biases or not. As Table 3 demonstrates, most of these variables are correlated with each other. To this extend, we calculate the most widely used diagnostic for multicollinearity, that is, the variance of inflation factor (VIF) for both rural and urban models, and added the values in Table 7. As the results demonstrate, for most of the variables, the VIF values are very close to 1 (meaning that the variables are orthogonal). For all the variables under consideration, the VIF values are less than five, which is the typical threshold for reasonably low collinearity (James et al., 2017).

Using the results presented above with the described methodology, we apply the COVID-19 vulnerability scoring to the US counties. Table 8 shows the top-ranking US counties for both methods. The scoring metric is applied to all US counties; in the left-hand side of the table, we provide the ranks based on the COVID-19 vulnerability score counties are classified as Urban (U) and Rural (R). According to this measure, Queens, New York (NY) is the most vulnerable county in the whole country and Miami-Dade, Florida (FL), is a close second. Garza, Texas (TX) is the most vulnerable rural county according to this ranking. While using the COVID-Vulnerability Score by itself can be helpful for decision-makers to identify the most vulnerable population groups, e.g. such as Garza, TX and two Alaskan counties (Denali and Aleutians), it can also be beneficial to consider the most crowded counties with considerably high COVID-19 vulnerability scores. For this purpose, we also rank the counties based on population-weighted COVID-19 vulnerability scores. While this list is mostly dominated by the population size itself, we believe it still provides important information. For example, counties such as New York, NY, and Palm Beach, FL, scored higher. These three counties are ranked within the Top 20 counties with the largest vulnerable population list, even though their populations are considerably smaller than the counties such as Bexar, TX (population: 2.01 million), which is out of the list due to its lower COVID-19 Score (c_bexar: 0.4899). In  Appendix, we provide the Top 20 counties with the highest COVID-19 Score and compare it with the counties with the highest number of cases as of 13/10/20. In that table, we can see that the population-weighted COVID-19 vulnerability score accurately projected the case prevalence progress, even if the model parameters are trained with the data collected in early June.

In this section, we first explain the procedure that we followed to validate the COVID-19 Scores and then we numerically compare our results with the pro-rata based distribution policies.

We test the COVID-19 vulnerability score's validity based on early-onset data by comparing the actual COVID-19 prevalence observed in two states, i.e. New York and Nebraska. Since these two states have entirely different urban-rural compositions across the states, with differences in socio-demographic and economic attributes of the populations, we test the validity of the vulnerability metric throughout the pandemic by contrasting these two states. Informed by the model, first, we apply the vulnerability metric to these states and rank their counties based on the vulnerability score, and we compare these rankings with the COVID-19 prevalence reported as of 10/13/2020.

Since the pandemic has challenged the vulnerable populations across the country disproportionately and food access-related problems significantly contributed to this burden, several programs are considered to improve food access in communities. The Continuing Appropriations Act 2021 (CR PL 116-159) to extend federal appropriations through Dec. 11, 2020, was signed into law on Oct. 1, 2020. The resolution extends a number of the SNAP flexibilities under the Families First Coronavirus Response Act (FFCRA) and permits state SNAP agencies to adopt certain options without Food and Nutrition Service approval (USDA, 2020b). In the state of New York, emergency allotments have been approved from March to October (USDA, 2020c). Nebraska also has extended the certifications and adjusted their interview process (USDA, 2020a). Therefore, we can see both states have been affected by food access issues and considered policies to solve this problem. Therefore, a metric like the COVID-19 vulnerability score that considers food access measures can help prioritize communities for these kinds of public programs operated under limited funding. Figures 3 and 4 display the predicted COVID-19 vulnerability and the cases reported in mid-October. We observe that the scoring system is correctly projected which communities are more vulnerable to the COVID-19 in New York and Nebraska.

Tables 9 and 10 show the top 10 most vulnerable counties in both states based on the COVID-19 vulnerability scoring system and compare them with the counties that have the highest case prevalence as of 10/13/2020. In New York State, Westchester, Nassau, Richmond, Queens, Suffolk and Kings counties were predicted to be vulnerable by their COVID-19 score, and we see they turn-out vulnerable with the highest cases reported. In addition to overlaps presented in Table 9, New York County also has the 11th highest case prevalence (2.12%), and Putnam also has a relatively high prevalence (1.73%). The other counties with widespread prevalence also have high COVID-19 scores, e.g. Rockland 0.6229 and Bronx 0.660. The overlap is more evident in Nebraska, as shown in Table 10. More interestingly, the rural counties that are projected to be more vulnerable have reported case prevalence the highest by 10/13/2020. Colfax, Dakota, Saline, Dawson, Thurston turn out to be highly vulnerable, as predicted by their COVID-19 scores. The other counties that have a high COVID-19 score also have relatively high case prevalence: Buffalo (2.86%), Madison (2.81%), Dawes (1.98%) and Johnson (1.66%). Only Hayes has a very low prevalence, as it has 5 cases and a total population of 922. In Nebraska, the other counties that have high case prevalence also have high COVID-19 scores: Douglas (0.52), Platte (0.50) and Hall (0.44).

In this section, we extend our analysis by comparing the prioritization of counties based on the COVID-19 vulnerability score with the pro-rata approach. Pro-rata strategy is a common prioritization approach for different geographic locations (such as counties) when distributing emergency medical supplies (Araz et al., 2012). The available stocks of supplies are distributed simultaneously and proportional to the county's demand or population. This approach is widely adopted as it minimizes controversy for political authorities and reduces the burden of critical decision making. On the other hand, in case of an infectious disease, the progression is dynamic and unpredictable. Therefore, this policy may yield inefficient vaccine distribution outcomes.

Tables 11 and 12 shows the counties with the highest COVID-19 vulnerability scores and their pro-rata rankings for New York and Nebraska, respectively. For New York, we observe that the top three counties would not change under these two approaches; however, COVID-19 scoring would also enable the decision-makers to pay particular attention to vulnerable counties such as Tompkins and Putnam. In contrast to New York, for Nebraska counties, we find that if the pro-rata approach were to be used, counties with high COVID-19 prevalence such as Colfax, Dakota, Saline and Thurston would be overlooked entirely in distributing the limited resources.

At this point, we would like to remind the reader that the analyses conducted in this paper rely on data belonging to a relatively early time period of the disease spread. Nevertheless, the validation analyses conducted in this section indicate that even early data can be quite helpful in coordinating the response efforts that are conducted later. This potential benefit of such an ex ante analysis becomes rather evident when the cold-chain requirements of the first available vaccines are revealed. Both of the most widely distributed messenger ribonucleic acid (mRNA) vaccines require very strict and highly restricting storage and transportation conditions in very low temperatures (Ontario Ministry of Health, 2021). As a result, governments are required to invest in freezers and meticulously plan the distribution supply chain in advance, so that the vaccines can be distributed to broader populations in a short amount of time. A wait-and-see approach to identify the vulnerable populations can significantly delay the vaccine roll-outs.

Indeed, community vulnerability-based prioritization has started to gain attraction in public as well. A recent study exploring the public view on vaccine prioritization found that regardless of their political view, the US public supports the prioritization of the disproportionately affected racial groups (e.g. Black, Hispanic and Native American) as suggested by the National Academies of Science, Engineering and Medicine. However, the support is not as strong for aged-based prioritization (Persad et al., 2021). On the other hand, a prioritization approach solely based on race inevitably raises questions concerning the ethical and lawful implications. While US Supreme Court is hesitant to approve a race-based vaccine prioritization, other community-based approaches that include not only race but other vulnerability factors could more viable (Schmidt et al., 2020). We believe that more comprehensive metrics than SVI, such as the one developed in this study, can contribute to this discussion by providing statistical evidence and informing the decision-makers towards achieving social justice.

This study contributes to the literature by empirically analyzing the effects of food access on COVID-19 related health outcomes in the US. The food desert classifications of counties are taken into consideration with data of being on the SNAP in each county and chronic disease conditions represented with diabetes and obesity prevalence in each county. The analyses are performed for rural and urban classified counties for COVID-19 related health outcomes, including case fatalities, i.e. deaths per 1,000 population, and reported COVID-19 prevalence. In addition, a COVID-19 vulnerability score is developed based on these analyses, which is then applied and validated in New York and Nebraska with the more recently reported data in each state. The results show that the proportion of the population being in a food desert is correlated positively with the COVID-19 prevalence, similar to the proportion of the population on the SNAP. However, while we did not observe any statistically significant relationship between being in a food desert and COVID-19 mortality rates, the proportion of the population on the SNAP has a positive correlation with the fatalities.

Using the factors that are detected as significant indicators of COVID-19 outcomes, we have developed a normative COVID-19 vulnerability score for rural and urban states, which can be used to compare and rank different counties while prioritizing resources such as PPEs, testing kits but more importantly vaccines. We have also validated our results by comparing them with the actual COVID-19 hotspots in Nebraska and New York states. With the proposed methodology, the disease progression could have been halted quite rapidly before these counties become COVID-19 hotspots. To the best of our knowledge, this is the first study in the literature that proposes a data-driven vulnerability-based vaccine prioritization rule using the CDC's SVI, food accessibility metrics and chronic health condition prevalence.

This study also establishes important practical implications. Our analyses show that while applying pro-rata based prioritization policies in managing limited resources can be practical and equitable in more homogeneously populated states, e.g. state of New York, in highly rural states, pro-rata approach may contribute to health disparities and may not help mitigate the disproportional effects of COVID-19 pandemic across populations. Therefore, using more comprehensively developed metrics that take into account the chronic disease prevalence, food access, etc., can improve the equity concerns during pandemics and other public health emergencies. CDC originally developed the SVI to identify and prioritize vulnerable populations in the event of a disaster. However, this measure does not take into account other potential vulnerabilities such as food access or chronic health conditions.

Our study demonstrates a case in which the food access measured by the SNAP and the food desert status present important indicators of social vulnerability to disease outbreak and potentially to other public health crises. Therefore, the humanitarian aid organizations and public health agencies can consider our results to develop and support public campaigns for more equitable resource allocation in other type of crises. In case of extremely limited vaccine availability, the proposed scores can indicate where low-cost interventions such as social distancing, hand washing and mask mandates can be adopted widely through public health campaigns. While pro-rata (proportional to population) based resource allocations can be very convenient and less “politically” controversial, our study shows that population vulnerability indicators, such as proportion of the population in SNAP, chronic disease prevalence, food deserts, etc. would inform policies for resource allocations that would achieve broader social good.

In this subsection, we summarize the key contributions of the proposed scoring system by discussing its advantages over the existing strategies. First of all, our results show that more simplistic approaches such as age-based vaccination prioritization would not be sufficient. While age is one of the determinants of the COVID-19 outcomes in our analyses, it is not the sole determinant. We observe that racial and socio-economical parameters also heavily affect the outcomes. Second of all, our analyses show that CDC's SVI, a metric that is developed as a generic prioritization tool intended to identify vulnerable populations in the aftermath of any human-made or natural disaster, might be a good starting point. However, it can also be significantly improved to encapsulate other relevant factors for a particular emergency, especially pandemics. In this particular case, we observed that factors such as food inaccessibility and chronic underlying diseases – factors missing from the SVI – can also act as significant determinants of vulnerability. Finally, one can consider a wait-and-see approach for the disease progression and target the neighborhoods or communities affected the most from the disease rather than using our suggested approach. One commonly used wait-and-see method is using waste-water-based epidemiological (WBE) tracking of COVID-19 hotspots (Dharmadhikari et al., 2021). However, recent evidence from various countries show that (Mota et al., 2021; Chakraborty et al., 2022; Street et al., 2020), the results obtained through WBE converges with a vulnerability-based prioritization. Furthermore, WBE provides more accurate results in crowded residential areas, but performs poorly in less-dense regions (Zdenkova et al., 2022) Consequently, WBE or other ex-post approaches can cause the loss of valuable time in the fight against the disease progression and cost many lives, especially in rural areas with low population density. We believe the proposed approach in this manuscript overcomes the shortcomings of existing vaccine prioritization approaches.

In fact, most European countries such as UK, Germany, Austria, Portugal and Israel have adopted a simplistic age-based prioritization for their vaccination schedules with limited exceptions for healthcare providers and front-line workers (Cylus et al., 2021). However, the shortcomings of this approach became evident soon enough. First of all, the cut-off age for the vaccine roll-outs are not consistent; UK and Germany prioritized people older than 80, Portugal vaccinated people as young as 50 in their first roll-out. Another study finds that prioritizing middle-aged population is more beneficial in preventing mortality outcomes, if the vaccine availability is very limited (i.e. covering less than 25% of the population) or slow vaccine rollout rates (less than 0.2% of the population per day) (Bubar et al., 2021). Socially vulnerable communities became infection hot-spots with very high test-positivity rates (i.e. the proportion of COVID-19 positive cases per conducted tests) and case prevalence rates (Dasgupta et al., 2020), forcing public-health policymakers to reconsider their age-based strategies (Biden, 2021).

One notable example of reconsidering age-based strategy took place in Canada. The initial doses that arrived in the country were administered to healthcare workers and residents of senior care centers; realizing that indigenous communities residing in remote rural regions of the country with inadequate income and reduced access to basic needs, including healthcare, Canada decided to include those communities into the first wave of vaccination efforts (Public Health Agency of Canada, 2021). Even these efforts did not prevent the spread of the disease among vulnerable populations. Even with strict COVID-19 measures in effect, transmission rates remain persistently high in certain neighborhoods of British Columbia (BC) and Ontario. A closer look at such neighborhoods revealed that these are, in fact, the most racially diverse and lowest-income neighborhoods of the BC (Little and Hua, 2021). Eventually, the government of B.C. had to prioritize these neighborhoods, moving ahead of the age-based vaccination schedule (Ballard, 2021). Had the government adopted a measure similar to the one proposed in this article, the situation could have been prevented, and the response efforts could have been coordinated before these neighborhoods became infection hotspots.

Given that new information regarding the progression of the disease and different response approaches become available each day, it is clear that the public guidelines will continue to change with the emerging evidence. This work also has several limitations, which can be addressed in future studies. While we compared the COVID-19 vulnerability score-based ranking for prioritization with the well-accepted pro-rata policy, other geographic prioritization policies such as Area Deprivation Index (Schmidt et al., 2020) may also be considered. More comprehensive comparisons of these metrics can be beneficial in minimizing the injustice in distributing the limited resources and help minimize the health disparities more effectively.

Second of all, at the time of the study, the most recent county level obesity and diabetes data available belonged to 2016. However, the independent variables regarding the COVID-19 outcomes belong to year 2020. It may be beneficial to repeat the study when obesity and diabetes data become available for the year 2020.

Third, at the time of this study, county-level hospitalization duration was not available. Nevertheless, repeating this study when such data become available can also provide essential insights for capacity planning purposes. Furthermore, recent evidence shows that there is a clear link between smoking habits, especially in older patients, and the case severity (Zheng et al., 2020). It would be beneficial to include this measure if the metric is used for hospital capacity planning purposes. Finally, once the prioritization is established delivering vaccines through cold-chains these vulnerable communities rapidly may require advanced planning and technology (Comes et al., 2018). Accordingly, further studies on scheduling and transportation of the vaccines should be developed. In addition, complex epidemiological simulation methodologies could be used to integrate disease spread dynamics with complex interactions of the social vulnerability indicators. This would require more granular data, at the county and potentially individual level, and would require more computational time to simulate various scenarios. We believe our study can use for informing such model, e.g. a system dynamics model, which we plan to take this approach as a future study.

Besides these limitations, we believe that the results of this study can be beneficial for an optimal allocation of limited resources in developing an effective response plan for this pandemic (Engelbrecht et al., 2021) and can shed light on planning efforts for future pandemics.

This paper forms part of a special section “The COVID19 impact on humanitarian operations: lessons for future disrupting events”, guest edited by Bhavin Shah, Guilherme Frederico, Vikas Kumar, Jose Arturo Garza-Reyes and Anil Kumar.

The authors express their gratitude towards two anonymous referees, and the editorial team. Feyza G. Sahinyazan was funded by the Canadian Natural Sciences and Engineering Research Council (NSERC) under grant 2022-03668.

Table A1