Hidden COVID-19 deaths? Exploring the Spatial context of excess death rates during the COVID-19 pandemic

Abstract

Background

The COVID-19 pandemic caused substantial mortality in the United States with impacts unevenly distributed across the country. Official COVID-19-related death counts, however, almost certainly underrepresent the true impact of the pandemic due to underreporting, misclassification, and, particularly in the early stages of pandemic, limited testing and diagnosis [1]. Excess death rates, deaths above expected levels based on historical trends, arguably provide a more comprehensive measure of COVID-19 impacts by capturing both direct COVID-19 deaths and indirect fatalities related to pandemic disruptions. The goal of the study is to examine spatial and temporal disparities in COVID-19 excess mortality in 2020–2021 and 2021–2022 across the U.S., distinguishing between quantifiable sociodemographic influences and unmeasurable place-based factors through Multiscale Geographically Weighted Regression (MGWR).

Methods

Excess mortalities are examined in 2020–2021 and 2021–2022 to capture temporal and spatial shifts in COVID-19-related excess mortality patterns. MGWR is used to identify localized variations in the determinants of excess death rates using data on socioeconomic conditions, political affiliation, demographic factors, health status, and healthcare access.

Results

We present the results of calibrating both a global and a local model of excess death rates during two phases of the COVID-19 pandemic. In terms of the global results, in both time periods excess death rates were significantly higher in counties with high percentages of people below the poverty line, Republican-leaning residents, high proportions of elderly population, high levels of deprivation, high unemployment, and relatively high proportions of residents with diabetes. Rates were also significantly higher in counties with relatively high proportions of residents without health insurance, where there were more females than males, and where there were fewer younger adults, although these effects were not as strong as the previous associations. However, these macro-level conditioned associations can hide important local variations in the determinants of severe COVID-19-related health outcomes. Because COVID-19-related excess death rates exhibit strong spatial patterns, any covariate sharing a similar spatial distribution, even if coincidental, might spuriously be reported to have a significant impact on excess dates rates when examined globally. To examine this possibility, a local statistical model is calibrated which suggests some alternative views on the determinants of COVID-19-relates deaths. For instance, although excess death rates were strongly linked to Republican party support across the whole country in the first phase of the pandemic, this relationship was limited to the eastern seaboard and the Deep South in the second phase. There was a significant conditioned relationship between excess deaths and the elderly only across the southern half of the country in both phases of the pandemic. The impacts of being without health insurance were only severe in the western half of the country and only in the first phase of the pandemic. In contrast to the global finding, the positive association with diabetes was only found along the east coast and only in the first phase of the pandemic. In the first phase of the pandemic, excess mortality was only significantly positively associated with the proportion of Hispanics in the Southwest and was insignificant elsewhere, In the second phase of the pandemic, there were no significant positive relationships reported locally but there were significant negative relationships across the upper Midwest, the Northeast, and in Texas. In distinct contrast to the global results, the local conditioned relationship between excess death rates and percentage Black population was significantly positive across the country in both phases of the pandemic. In the first phase of the pandemic, conditioning on all the covariates in the model, excess deaths from COVID-19 were lower than expected in most parts of the country except for a cone-shaped set of states from Nebraska to Texas; in the second phase the unseen benefits of location were only experienced in the upper Midwest. The results support the use of local models to better understand the nature of pandemics and also that COVID-19 impacts arose from a complex interaction between both measurable factors and localized, often unobservable, influences.

Conclusions

Disparities in excess deaths during the COVID-19 pandemic reflect a combination of structural vulnerabilities and unmeasured local influences. To effectively reduce mortality gaps and strengthen preparedness for future health crises, public health interventions must be geographically tailored, targeting both region-specific risk factors and the contextual conditions that shape local outcomes.

Background

As of March 2025, the total number of COVID-19 deaths surpassed 1.2 million in the United States [2]. The COVID-19 pandemic amplified longstanding disparities in health outcomes shaped by demographics, socioeconomic status, and healthcare accessibility based on socioeconomic status, race and ethnicity, age, healthcare infrastructure, and political factors. A growing body of research suggests that COVID-19 mortality varied geographically, due in part to variations in socio-demographics. Counties with higher poverty levels, larger proportions of people of color, elderly individuals, women, low-income residents, and homeless populations appear to have been particularly susceptible to severe COVID-19 outcomes [3–7]. Despite increasing attention to these health disparities, inconsistencies and substantial gaps remain in literature. Some researchers, for instance, presumed that densely populated urban centers would experience elevated mortality rates [8]; however, other studies revealed that population density alone was not a consistent predictor of deaths [9, 10]. After accounting for metropolitan population size, higher density at county-level was, in fact, not associated with higher COVID-19 mortality rates. Instead, social, economic, and commuting connectivity emerged as more conspicuous contributors to higher death rates [9, 10].

Political affiliation also emerged as another distinct determinant of COVID-19 deaths, with areas having higher shares of Republican voters reporting higher death rates, possibly because of lower adherence to public health measures, such as mask-wearing, vaccination, social distancing, and stay-at-home orders [11–14]. However, other studies reported that the political affiliation of state governors showed no consistent association with death rates [15]. While certain policies such as protective mandates and gathering restrictions correlated with lower infection rates, they did not translate into significantly lower mortality at the state level [15].

Much of the existing research, however, has focused on national or state-level patterns. For example, Ruhm found that statewide restriction policies, such as mask requirements and vaccine mandates, were significantly associated with lower excess mortality rates [16]. Hu et al. showed that, between July and September 2021, diabetes prevalence and the number of older adults were positively associated with COVID-19 mortality across most counties [17]. Mollalo et al. used socioeconomic, environmental, behavioral, topographic, and demographic variables to model county-level COVID-19 incidence [18], although it is unclear whether correct inferential tests, with adjustments for multiple hypothesis testing were conducted.

Internationally, in Brazil, Raymundo et al. reported a positive association between economic diversity and COVID-19 incidence in the North of the country and nurse-to-inhabitant ratios were positively associated with incidence in the North and Northeast, whereas mortality ratios were negatively associated with incidence in Western and Southwestern regions [19]. In Oman, Mansour et al. found that larger elderly populations, higher population densities, and higher diabetes rates were positively associated with COVID-19 incidence [20]. In England, the proportion of minor ethnic groups, diabetes rate, deprivation, male population were found to have positive significant relationships with COVID-19 morbidity rate [21].

Another factor that instills uncertainty in the results of previous COVID-19-related research is that its focus has tended to be on disease incidence or mortality, rather than on excess mortality, which arguably offers a more comprehensive measure of the pandemic’s impact. The accuracy of officially reported COVID-19 deaths has been compromised by multiple factors, including limited testing capacity, case definitions, and reporting delays, particularly during the early stages of the outbreak. These limitations likely contributed to underreported or misclassified deaths [1]. Notably, research suggests that approximately 17% of COVID-19-induced excess deaths were not listed as COVID-19-related on death certificates [22]. Consequently, excess mortality has emerged as a realistic representation for gauging the pandemic’s true impact. Unlike official death counts, excess mortality captures not only confirmed COVID-19 fatalities but also deaths that were misdiagnosed as not COVID-19-related or deaths that were not actually from COVID-19 but were a result of the conditions during the pandemic. Examples of the latter are people who died from pre-existing health conditions but who had limited access to healthcare, or who were reluctant to seek medical treatment due to fear of infection [23, 24]. Socioeconomic hardship, including sudden job losses, housing issues, and food insecurity, may have also exacerbated health risks that led to COVID-19-related excess deaths [25]. Additionally, the pandemic saw a surge in drug overdoses, reflecting the intersection of economic hardship and mental health crises [26, 27]. This combination of factors prompts the need for a comprehensive approach to assessing pandemic-related mortality, taking into account both direct and indirect causes of death to better inform future public health policies.

To address the limitations in previous research on the effects of COVID-19, this study employs Multiscale Geographically Weighted Regression (MGWR) to account for local variations in the determinants of excess mortality occurring during the pandemic. By applying MGWR to county-level COVID-19-related excess death rates in U.S. for two critical time periods during the pandemic, 2020–2021 and 2021–2022, we attempt to identify and quantify spatial disparities and temporal shifts in the key determinants of COVID-19-related excess mortality. The findings provide insights, previously unseen, into the local dynamics of excess mortality, enhancing our understanding of how demographic, socio-economic, political, and healthcare-related factors shaped geographic patterns of COVID-19-related deaths. By identifying spatial variations in the relationships between COVID-19 excess mortality rates and key demographic and socioeconomic factors that exacerbated the risk of dying from COVID-19, we can learn to streamline future interventions designed to reduce the impact of pandemics. Similarly, evidence on spatial variations in the effectiveness of specific policy measures, such as early lockdowns and mandatory vaccination, can inform future decisions about which interventions might be most effective in different locations – a “one-size-fits-all” policy approach to pandemics might not be optimal. Consequentially, this study not only sheds light on the uneven excess mortality outcomes from COVID-19 across the United States but also provides actionable knowledge for strengthening public health responses by pinpointing where and why excess deaths occurred.

Methods

Data

In Table 1, we provide definitions of the two variables of interest. The first describes the excess mortality rates for each county for a two-year period from January 2020 to December 2021; the second describes similar data but for the period January 2021 to December 2022. To construct these two variables, we first retrieved county-level multiple causes of death rates per 100,000 population from 2015 to 2022 from the Centers for Disease Control and Prevention (CDC) via the CDC WONDER database [28, 29]. These data represent the national yearly final true mortality statistics based on death certificates for residents in each county rather than the estimated values. We follow a similar procedure outlined on the CDC website [30] and the approach used by Mostert et al. [31] to calculate the five-year average death rates from 2015 to 2019 to establish a baseline county-level death rate prior to the COVID-19 pandemic. This multi-year average smooths out anomalies and minimizes the impact of idiosyncratic events that may be more pronounced in any single year. Based on this baseline, we then computed two average crude death rates for each county for the two-year periods: (1) Jan 2020 –Dec 2021 and (2) Jan 2021–Dec 2022.

Table 1.

Dependent variables

Dependent Variables	Definitions
(1) ExcessMortality_Jan 2020-Dec 2021	Two-year average deaths divided by population times 100,000 in 2020 and 2021 minus the crude death rates from 2015 to 2019
(2) ExcessMortality_Jan 2021-Dec 2022	Two-year average deaths divided by population times 100,000 in 2021 and 2022 minus the crude death rates from 2015 to 2019

The excess mortality for each period is calculated using Eq. (1):

1

where = 2020–2021 or 2021–2022 and is the centroid of county .

County-level excess mortality for both 2020–2021 and 2021–2022 are visualized in Fig. 1. Excess mortalities are represented using a diverging color scheme with blue indicating fewer deaths than expected given the death rate over the previous five years while light yellow through dark red represent progressively higher than expected levels of excess mortality. The mortality map indicates significant geographic variations in excess death rates across the United States, with particularly elevated excess mortality observed in counties in the Southern and Midwestern regions relative to those in the West North Central and Mid-Atlantic regions. One underlying factor for these geographic disparities appears to be local political leaning. Specifically, counties which tend to be Republican-leaning generally exhibited higher excess death rates, particularly across the South; an association likely attributable to reduced adherence to public health guidelines such as mask-wearing and social distancing [11–14]. However, this will be explored statistically below.

Fig. 1.

Excess mortality at the county level in 2020–2021 and 2021–2022

The two time periods forming the basis of the results reported here capture distinct pandemic phases: Jan 2020–Dec 2021 encompasses the initial surge of infections with either no vaccine availability or vaccine shortages, while Jan 2021–Dec 2022 includes the subsequent variant waves amid vaccination rollouts [32]. According to the CDC, the first confirmed COVID-19 case in the U.S. occurred on January 20, 2020, and the World Health Organization (WHO) officially declared COVID-19 a pandemic on March 11, 2020 [32]. The FDA issued Emergency Use Authorizations (EUAs) for the Pfizer-BioNTech and Moderna vaccines on December 11 and December 18, 2020, respectively [32], falling within the 2020–2021 window. Toward the end of 2020 and into early 2021, several variants of the COVID-19 virus emerged, with the Alpha variant identified on December 29, 2020, and the Gamma and Beta variants detected in late January 2021 [32] which closely aligns with the start of the 2021–2022 period. We chose to examine the two overlapping periods separately as they represent different stages of the pandemic and provided greater robustness in the results compared to single-year analyses which suffered more from noise in the data. A combined analysis for 2020–2022 (three-year period) is also included in Appendix E; however, it is not the primary focus of this study. In Table 2, we list all the covariates and their definitions used in the study to inform on the causes of excess deaths related to the COVID-19 pandemic across the U.S. Summary statistics for these variables for 2020–2021 are provided in Table 6 and for 2021–2022 in Table 7, located in Appendix A.

Table 2.

Independent variables

Independent Variables
Political Preferences
PctRep	Percentage of votes for the Republican party in 2020
Demographics
PopDen	Population density
PctBlack	Percentage of the Black and African American population
LnPctHispanic	Natural log of percentage of the Hispanic population
Pct18to24	Percentage of the population aged between 18 and 24
Pct65Over	Percentage of the population aged over 65
SexRatio	The number of males per 100 females
Socioeconomics
LnMDI	Natural log of multidimensional deprivation index
PctPov	Percentage of the population that earns below poverty level
LnPctBach	Natural log of percentage of the population has at least bachelor’s degree
LnMedHhIncome	Natural log of median household income in US dollars ($)
LnPctUnemployed	Natural log of unemployment rate of labor force aged over 16
Medical-related
PctDiabetics	Percentage of the population that is diabetic
PctWoInsurance	Percentage of the population without health insurance

To model spatial variations in excess death rates, we selected independent variables across four primary domains: political preference, demographic factors, socioeconomic indicators, and health-related factors, based on the previous research on COVID-19 incidence and deaths [3–8, 11–15, 17, 20–22, 25, 33]. For political preferences, we gathered data from the MIT Election and Science Lab [34] to calculate the county-level percentage of all votes cast for the Republican party in the 2020 Presidential election. For the demographic, socioeconomic, and medical-related data, we obtained data on: population density; the percentage of Black population; the percentage of Hispanic population; the percentage of the population aged between 18 and 24; the percentage of the population aged over 65; sex ratio; the percentage of the population below the poverty level; the percentage of the population with at least a bachelor’s degree; median household income; unemployment rate; and the percentage of the population without health insurance from the United States Census Bureau’s American Community Survey (ACS) 5-year estimate datasets (2016–2020, 2017–2021, 2018–2022) to align with the periods for which excess death rates were calculated [35–37].

Data on the percentage of the population with diabetes were obtained from the 2022 County Health Rankings National Data [38]. We incorporate a Multidimensional Deprivation Index (MDI) into the model which gives an overall assessment of the standard of living, health, education, economic security, and housing quality in each county [39] as an alternative measure of MDI used by Morasae et al. [21]. Because MDI data are only available through 2019 at the county level, the 2019 values are used for each time period on the assumption that rates of MDI are slow to change over time [39].

To ensure consistency in our spatial analysis, we excluded the noncontiguous states of Alaska and Hawaii, as well as counties with populations below 5,000 which are shown as crossed out with light grey hatching in Fig. 1: the former because we calibrate local statistical models which demand contiguous data; the latter to minimize bias from small population sizes. The final dataset consists of 2,770 counties for models using data for the 2020–2021 period. For the models using data for the 2021–2022 period, data on the eight counties in Connecticut shifted to nine planning regions and data for those counties were unavailable [40]. Consequently, the final dataset for the 2021–2022 period consists of 2,762 counties.

MGWR

Multiscale Geographically Weighted Regression (MGWR) is a linear model with location-specific parameters which can be represented as:

2

where represent parameters to be estimated which are specific to location i;, , ,., represent bandwidths in covariate-specific distance-decay functions that determine the spatial proximity of the locations from which data are borrowed for the regression at location i; and x_1i, x_2i …x_ki denote covariates. The estimation of location-specific parameters is achieved by borrowing data from nearby locations through a spatial weights matrix, , which incorporates a kernel function that assigns decreasing weight to locations further from the regression point. The optimized bandwidth for each covariate determines the degree of distance decay in the spatial weighting function. This framework allows regression parameter estimates to be generated separately for each location, capturing spatial variations in the conditioned relationships between the dependent and independent variables [41].

MGWR is utilized here to uncover any spatially varying effects on excess death rates across the U.S. MGWR has been applied to several domains in public health such as obesity, food insecurity, diabetes prevalence, aging, opioid use disorder, female breast cancer mortality [42–46].

MGWR provides a mechanism to explore spatial non-stationarity and insights into how relationships evolve across different spatial scales. It captures both global and local spatial processes simultaneously and separates intrinsic contextual effects from behavioral contextual effects. The former are denoted by significant estimates of the local intercept; the latter by significant estimates of the local slope parameters [47]. As a flexible, albeit computationally intensive, modeling framework, MGWR can be applied to various research domains in which the researcher wants to uncover spatially heterogeneous relationships [48].

Figure 2 presents a structured workflow for modeling the spatial context of excess COVID-19-related death rates using MGWR. The workflow starts with data standardization so that all variables have a mean of 0 and a standard deviation of 1 to ensure comparability of estimated bandwidths across variables. Next, the Variance Inflation Factor (VIF) is calculated to check for multicollinearity among independent variables. Any variable with a VIF greater than 10 is removed from the dataset before proceeding. Once multicollinearity concerns are addressed, OLS regression is performed as a preliminary assessment of the model. Here, the OLS specification for excess death rates by county is:

3

Fig. 2.

MGWR workflow

where is the intercept; … are estimated coefficients for the corresponding covariates and is the residual error term. If the OLS results indicate a reasonable R² and significant variables, the workflow proceeds to MGWR calibration. We use the MGWR python package developed by Oshan et al. [49] to calibrate the model in Eq. (2). In the MGWR calibration, we used a bi-square weighting function with the number of the nearest neighbors defining the spatial extent of the bandwidth. Optimal bandwidths for each covariate are determined by the corrected Akaike Information Criterion (AICc). Additionally, independent variables are entered in the descending order of absolute influence in the OLS regression, and the minimum bandwidth is set to 15 as suggested by Kao and Fotheringham [50]. Following the MGWR calibration, a local multicollinearity check is conducted to ensure the robustness of local parameter estimates. A rule of thumb indicates that there might be an issue due to multicollinearity if the local condition number (CN) is higher than 30 [49].

A simple diagnostic test for possible nonlinear relationships is to plot the local estimates against the corresponding values of the covariate [51]. If a nonlinear pattern is detected, data transformation, such as applying the natural logarithm or squared transformations, is applied. These transformations help stabilize variance and improve the linear relationship between the independent and dependent variables and the model is recalibrated [51]. If no significant nonlinearity is present, we proceed to identify and map significant local parameter estimates using adjusted (for multiple hypothesis testing) critical t-values at a 95% significance level [41, 52] to visualize spatial variation in the determinants of excess death rates. Lastly, the Monte Carlo test evaluates the probability of whether spatial variation in the local parameter estimates can be attributed solely to noise or sampling variation [49]. This test statistically validates the presence of spatial heterogeneity by assessing whether the observed variation in parameter estimates is greater than that expected under a spatially homogeneous process.

Extracting knowledge on Spatial context via MGWR

MGWR offers a major advantage over traditional global models by enabling the separation of intrinsic contextual effects, which are often unobservable, from behavioral contextual effects through the estimation of local intercepts and slope parameters, respectively [41, 47]. After standardizing the variables in the model, MGWR can be rewritten as:

4

where represents the standardized value of the dependent variable for each county ; is the local intercept for county ; represents the local coefficient estimates for covariate at county ; represents the standardized values of covariate for county ; is the local residual term. Standardization of the variables is conducted using:

5

        

Rewriting Eq. (4) explicitly in terms of the unstandardized variables, we obtain:

6

where and are the mean and the standard deviation of , respectively, while and are the mean and the standard deviation of .

We can rearrange Eq. (6) to be in terms of :

7

From Eq. (7), we observe that three distinct components contribute to the observed value : The mean value of ; The intrinsic, unmeasurable, spatial contextual effect ; and The measurable contextual effect, , representing the combined influence of the covariates included in the model on the dependent variable.

We now present the results of our findings on the spatial variation in the determinants of COVID-19-related excess death rates.

Results

OLS calibration results

Table 3 presents the results of OLS regression models for excess death rates for the periods Jan2020–Dec 2021 and Jan 2021–Dec 2022.

Table 3.

OLS results of excess death rates.

OLS Results	Excess Mortality 2020–2021	Excess Mortality 2021–2022
Variables
Intercept	0.000	0.000
PctPov	0.220*	0.133*
PctRep	0.173*	0.160*
PctBlack	0.142*	−0.079*
LnPctBach	−0.131*	−0.001
Pct65Over	0.117*	0.163*
LnMDI	0.100*	0.068*
PctWoInsurance	0.088*	0.051*
LnPctUnemployed	0.087*	0.177*
PctDiabetics	0.075*	0.348*
SexRatio	−0.058*	−0.040*
Pct18to24	−0.056*	−0.076*
PopDen	0.022	−0.002
LnPctHis	−0.005	−0.123*
LnMedHhIncome	−0.005	0.034
Model Diagnostics
N	2770	2762
R 2	0.436	0.410
Adjusted R 2	0.433	0.407
AIC	6306.7	6411.9
AICc	6308.9	6414.1

*Significant at 95 percent level

These models provide global estimates of the relationships between excess death rates and socioeconomic, demographic, political, and health-related predictors. Since the data standardization procedure transforms all predictors into a same scale metric, the absolute values of the standardized coefficients can be compared directly to assess the relative magnitude of the conditioned relationships. For both models, VIF values are below 10 for all variables, indicating that multicollinearity is not a major concern. The complete OLS results for the two models are included in Appendix B: for 2020–2021 in Table 8 and for 2021–2022 in Table 9.

Table 8.

OLS results of excess mortality 2020–2021

(1) Excess mortality 2020–2021	Estimates	t-value	p-value	VIF
Variables
Intercept	−0.000	0.000	1.000	1.00
PctPov	0.220*	6.464	0.000	5.65
PctRep	0.173*	7.221	0.000	2.79
PctBlack	0.142*	6.628	0.000	2.25
LnPctBach	−0.131*	−4.485	0.000	4.15
Pct65Over	0.117*	5.941	0.000	1.88
LnMDI	0.100*	3.894	0.000	3.21
PctWoInsurance	0.088*	4.516	0.000	1.87
LnPctUnemployed	0.087*	4.970	0.000	1.49
PctDiabetics	0.075*	3.895	0.000	1.82
SexRatio	−0.058*	−3.664	0.000	1.21
Pct18to24	−0.056*	−2.970	0.003	1.76
PopDen	0.022	1.368	0.171	1.30
LnPctHis	−0.005	−0.252	0.801	1.76
LnMedHhIncome	−0.005	−0.121	0.904	6.91
Model Diagnostics
N	2770
R2	0.436
Adjusted R2	0.433
AIC	6306.7
AICc	6308.9

*Significant at 95 percent level

Table 9.

OLS results of excess mortality 2021–2022

(2) Excess mortality 2021–2022	Estimates	t-value	p-value	VIF
Variables
Intercept	0.000	0.000	1.000
PctDiabetics	0.348*	9.209	0.000	6.63
LnPctUnemployed	0.177*	9.932	0.000	1.48
Pct65Over	0.163*	7.979	0.000	1.95
PctRep	0.160*	6.596	0.000	2.73
PctPov	0.133*	3.734	0.000	5.93
LnPctHis	−0.123*	−6.013	0.000	1.94
PctBlack	−0.079*	−3.190	0.001	2.84
Pct18to24	−0.076*	−3.890	0.000	1.78
LnMDI	0.068*	2.573	0.010	3.28
PctWoInsurance	0.051*	2.496	0.013	1.95
SexRatio	−0.040*	−2.510	0.012	1.17
LnMedHhIncome	0.034	0.891	0.373	6.91
PopDen	−0.002	−0.113	0.910	1.31
LnPctBach	−0.001	−0.035	0.972	4.41
Model Diagnostics
N	2762
R2	0.410
Adjusted R2	0.407
AIC	6411.9
AICc	6414.1

*Significant at 95 percent level

The results indicate that several socioeconomic and demographic factors were significantly associated with excess death mortality. In both 2020–2021 and 2021–2022, poverty rate, Republican vote share, percentage of elderly residents, MDI, percentage without insurance, unemployment rate, and diabetes prevalence were significantly and positively associated with excess mortality. This indicates that throughout the pandemic, ceteris paribus, counties with higher levels of socioeconomic vulnerability, residents with pre-existing health conditions, and that were Republican-leaning, experienced significantly higher excess death rates than other counties. In contrast, counties with higher percentages of young adults and where there were more males than females, had significantly lower excess death rates. Although not significant in 2021–2022, education played a significant role in preventing fatal COVID-19 outcomes in 2020–2021. Interestingly, and counter to expectation, there was no significant association in either time period between excess death rates and population density, disputing the speculation that COVID-19 hit hardest in the more densely populated urban areas. Consequently, when conditioned on other variables representing ethnicity, age, political leaning, gender etc., there appear to have been no intrinsically detrimental effects from living in cities during the pandemic. There was also no significant association in either time period between excess death rates and household income although there was a strong positive relationship between poverty level and excess deaths.1 Therefore, while being poor was a major contributing factor to dying during the pandemic, beyond being above the poverty line, income had no impact on death rates.

Of further interest is that the impact of some variables switched signs or changed in intensity during the pandemic. For instance, the relationship between excess death rates and the percentage of Black population was significantly positive in the first half of the pandemic but significantly negative in the second half. The impact of being highly educated was, surprisingly, significantly negative during the first half of the pandemic but was insignificant during the second half, whereas the impact of being Hispanic was inconsequential during the first half of the pandemic but offered significant protection during the latter half. The relationship between COVID-19-related excess death rates and diabetes was significantly positive in both time periods but the conditioned association was much stronger in the second half of the pandemic.

Although these results are intriguing and provide some basis for further investigation, they assume the conditioned relationships being modeled are the same across all locations, whereas in reality they might vary geographically. To address this limitation, we calibrated the same models with MGWR which allows conditioned relationships to vary spatially, capturing any regional disparities and spatial heterogeneity in the determinants of excess mortality patterns [41, 47]. These results are now described.

MGWR calibration results

The calibration of the models by MGWR improves model performance by capturing spatial heterogeneity in the determinants of excess death rates as shown in Table 4 by the higher R² values and the lower AICc values.

Table 4.

Comparison of model performance between OLS and MGWR

	Excess Mortality 2020–2021		Excess Mortality 2021–2022
	OLS	MGWR	OLS	MGWR
R 2	0.44	0.58	0.41	0.62
Adjusted R 2	0.43	0.54	0.41	0.58
AIC	6307	5933	6412	5757
AICc	6309	5978	6414	5833

Both MGWR calibrations yield a maximum local condition number (CN) under 30, as shown in Table 5, suggesting acceptable local multicollinearity levels, and no serious nonlinear relationship were found, as shown in Figures 5 and 6 in Appendix C. A feature of MGWR is its covariate-specific bandwidth optimization, which determines how much spatial information is borrowed from neighboring data when optimizing local relationships. The optimal bandwidth for each predictor provides an insight into the spatial scale over which the conditioned association between that particular covariate and y is relatively constant: Smaller bandwidths indicate localized effects, while larger bandwidths suggest increasingly spatially homogeneous relationships. Those variables with small bandwidths tend to have local parameter estimates which exhibit significant spatial variation, as shown by the Monte-Carlo p value of the local estimates being drawn from the same population value in Table 5.

Table 5.

Optimal bandwidths and spatial test results

	Excess Mortality in 2020-2021 O ptimal B andwidth	Monte Carlo P -value	Excess Mortality in 2021-2022 O ptimal B andwidth	Monte Carlo P -value
Variables
Intercept	2769	0.952	474	0.000*
PctPov	55	0.000*	2761	0.970
PctRep	2343	0.086	466	0.000*
PctBlack	573	0.000*	2761	0.927
LnPctBach	2769	0.127	2761	0.904
Pct65Over	358	0.000*	180	0.000*
LnMDI	2769	0.900	2761	0.397
PctWoInsurance	1918	0.252	146	0.002*
LnPctUnemployed	2769	0.998	840	0.139
PctDiabetics	1826	0.269	50	0.000*
SexRatio	184	0.003*	132	0.006*
Pct18to24	896	0.001*	2231	0.597
PopDen	2769	0.952	2761	0.263
LnPctHis	310	0.002*	502	0.004*
LnMedHhIncome	2769	0.203	2758	0.098
Maximum CN	6.02		12.17

*Significant variation at 95 percent level

Fig. 5.

Nonlinearity check for excess mortality 2020–2021

Fig. 6.

Nonlinearity check for excess mortality 2021–2022

The results presented in Fig. 3 illustrate the spatial distribution of MGWR-estimated coefficients for several key predictors of excess mortality in the periods 2020–2021 and 2021–2022. Each panel shows the direction and strength of the local relationships with red areas indicating significant positive conditioned associations and blue areas indicating significant negative conditioned associations. Bandwidth values and corresponding OLS estimates are also provided for reference. Five of the conditioned relationships are relatively constant over space and so the local parameter estimates are not mapped. These are the relationships between excess death rates and education level, deprivation index, unemployment rate, population density and median household income, as shown in Figure 7 and discussed in Appendix D.

Fig. 3.

MGWR significant local estimates, optimal bandwidths, and corresponding OLS estimates

Fig. 7.

MGWR significant local estimates, optimal bandwidths, and corresponding OLS estimates

The locally varying parameter estimates shown in Fig. 3 show how some determinants of COVID-19-related excess death rates were not constant over space but exhibited significant spatial heterogeneity; a facet which is completely obscured in the results obtained from calibrating global models, which has been the traditional form of analyzing COVID-19 effects. For example, the previously unseen intrinsic influence of place, as a separate phenomenon compared to the influence of the attributes of place, which is denoted by the estimates of the local intercept from the MGWR calibrations, indicates that, ceteris paribus, in the early phase of the pandemic COVID-19 excess deaths were significantly lower, ceteris paribus, across all of the eastern US and across most of the western US compared to the southern Great Plains regions (Fig. 3a) [41, 47]. In the latter half of the pandemic excess mortality rates were significantly lower, ceteris paribus, only in the northern Great Plains and Upper Midwest. During this latter period, there were pockets of significantly higher excess mortality in South Carolina and in small pockets of southwestern New Mexico and southeastern Arizona. In the context of excess death rates during the pandemic, these variations likely reflect cultural-embedded and behavioral differences in responses to health advisories, public health guidelines, mask mandates, and social distancing enforcement.

In Fig. 3b, the influence of poverty levels exhibits geographic variability only in the early pandemic period, with significant positive relationships with excess death rates found in pockets across the country, particularly in Arizona, New Mexico, southern Texas, and North Dakota. Support for the Republican Party was significantly positively associated with excess mortality rates across the country in 2020–2021 as illustrated in Fig. 3c but, interestingly, this relationship disappeared from much of the country and was confined to the southeast and along the eastern seaboard in 2021–2022. These results suggest that vaccine take-up and compliance with social-distancing rules etc. increased in Republican-dominated areas across the western half of the country in the latter half of the pandemic but remained lower in similar areas throughout the south and east.

The impacts of COVID-19 through excess death rates fell unduly heavily on Black population throughout almost all the country in both periods of the pandemic as observed in Fig. 3d. Between 2020 and 2021, the only parts of the country where this relationship was not observed was in Ohio, Michigan and Indiana in the Midwest and in New Mexico and Arizona in the Southwest. In the second half of the pandemic these pockets had disappeared.

In Fig. 3e, there was a marked geographic variation in the susceptibility of the elderly population to the ravages of the COVID-19 pandemic which deserves further attention. In both periods, the elderly suffered significantly more across the South than in the North, possibly because the warmer climate allowed greater mixing of people and the elderly were more exposed to the virus. In North Dakota, in the latter half of the pandemic the elderly were significantly less impacted by COVID-19 than the general population; again, a finding that merits closer attention.

The local conditioned relationships between those without insurance and COVID-19-related death rates, as represented in Fig. 3f, shows a remarkably clear geographic pattern in the first half of the pandemic with the relationship being significantly positive all across the West and not significant in the Midwest and East. In the second half of the pandemic, as vaccines became more freely available, the relationship was insignificant across most of the country with some small pockets of significant positive association in the Southwest, between Oklahoma and Kansas, and between Arkansas and Missouri. Both maps, but particularly the former, prompt questions about the spatial variation in the vulnerability of the population during the COVID-19 pandemic. Was vaccine availability more restricted in the Western states during the first half of the pandemic? Was there a greater ignorance of, or lack of belief in, social distancing and mask-wearing in the West?

The negative impact of diabetes prevalence was greatest along the east coast in the first half of the pandemic, although overall the spatial variation in its impacts was insignificant as shown in Fig. 3g and Table 5. In the second half of the pandemic the spatial variability of the impacts of diabetes on death rates increased overall but locally significant effects were limited. In scattered parts of Arizona, southern Texas, North Dakota and the mid-eastern seaboard, higher excess deaths were associated with high rates of diabetes but in mid-Kansas, higher excess death rates were associated with lower rates of diabetes, a finding which seems spurious.

In Fig. 3h, the influence of gender showed a limited, but spatially distinct, negative association with excess mortality, concentrated primarily along the Eastern seaboard between North Carolina and Delaware, and in the Southwest, between Oklahoma and Kansas, indicating that in these areas males were less prone to death from COVID-19 related death. However, positive associations, contradictory to global OLS results, were observed in parts of South Dakota. This is an example of Simpson’s paradox in local versus global modeling; an issue explored in [53, 54].

Similarly, in Fig. 3i, although the global OLS results suggested that younger population appeared to be less vulnerable to severe COVID-19 related outcomes, locally there was little evidence of this effect and, in fact, the significant local estimates were positive in most parts of the Mid-Atlantic. Negative associations, on the other hand, were identified only along the border region of New Mexico, Northern Texas, and Oklahoma.

Lastly, in the 2020–2021 period, the percentage of the Hispanic population exhibited a locally positive spatial relationship with excess mortality across the southwest but is insignificant elsewhere, whereas in 2021–2022 the local relationships were predominantly negative across Texas, the Northeast and the upper Midwest and insignificant elsewhere. It is unclear what factors caused this reversal in this conditioned relationship between the first and second halves of the pandemic.

Extracted Spatial context of COVID-19 excess death rates

For both phases of the pandemic, Fig. 4 shows the observed excess death rates by county (4a); the predicted excess death rates (4b); the decomposition of the predicted excess death rates into three components as shown by Eq. (7): mean excess death rate, (4c); intrinsic, or place-based, contextual effects (4d); and sociodemographic impacts (4e). The mean excess death rate represents the average baseline effect across all counties, with values of 234 in 2020–2021 and 220.3 in 2021–2022. Figure 4d describes the spatial variation in the MGWR estimate of the local intercept transformed from standardized form. During the earlier phase of the pandemic, intrinsic contextual effects were relatively minor with minimal spatial variation. However, in the later phase, a distinct spatial pattern emerges, with positive intrinsic effects observed across counties in the South, while negative intrinsic effects dominate in counties across the Northeast and upper Midwest. This pattern suggests that unmeasured local factors, potentially including cultural attitudes, public health compliance, local policy enforcement, and the influence of local news outlets and social media significantly influenced the distribution of excess mortality in this phase so that the effects of the pandemic were greater across the South and less across the North, ceteris paribus. Figure 4e presents the measurable effects on COVID-19 excess death rates from variations in sociodemographic composition. Population composition played a significant role in shaping excess mortality across the United States during the COVID-19 pandemic. Much of the Southern U.S. experienced elevated excess death rates related to higher levels of poverty, diabetes prevalence, lack of health insurance, higher proportions of the elderly, and a higher share of politically right-leaning residents, all of which contributed to increased vulnerability. In contrast, states in the Northern U.S., particularly those in the Upper Midwest, Mid-Atlantic, and parts of the West Coast had lower rates of COVID-19 excess deaths because of their population compositions.

Fig. 4.

Contextual influences on excess death rates during the pandemic

The decomposition of excess death rates in this manner distinguishes the influence of population composition from that of unmeasured contextual effects and emphasizes the importance of targeted public health interventions not only focused on explicitly identified socioeconomic risk factors but also tailored to address underlying local cultural contexts to effectively mitigate future mortality risks during pandemics or other public health crises.

Discussion

This study demonstrates how MGWR can enrich our understanding of pandemic-related excess mortality by revealing both measurable sociodemographic effects and unmeasured contextual influences tied to place on COVID-19-related excess death rates. Consistent with prior studies [3–7, 12, 17, 20, 21, 25, 33], our findings confirm that counties with higher poverty rates, larger shares of people of color, older populations, fewer residents with health insurance, higher prevalence of chronic conditions, more females than males, lower educational attainment, greater socioeconomic deprivation, higher unemployment, and more people in poverty were more susceptible to severe COVID-19 outcomes. Educational attainment, in particular, showed a consistent and strong negative association with excess mortality nationwide, aligning with earlier findings by Stokes et al. [33]. However, the impact of several risk factors changed in their intensity over time. Deprivation, for instance, had a widespread positive influence on COVID-19 excess death rates in 2020–2021 but the association faded in 2021–2022. Similarly, unemployment rates had a nationwide positive relationship with COVID-19 excess death rates in the early phase of the pandemic but this relationship was only found in the western U.S. in the later phase. In contrast, income level served as a protective factor in many parts of the Southwest, South, Midwest, and Northeast in 2020–2021, though no such association appeared in 2021–2022. Population density, often debated in prior studies [8–10], had a more pronounced effect in the central and eastern regions of the U.S. in 2020–2021 but showed no influence thereafter.

In terms of variables showing geographical disparity, poverty, for example, exhibited strong localized effects in Arizona and New Mexico in 2020–2021 with no effects in 2021–2022. Political affiliation remained a consistent and strong predictor [12] of excess deaths in 2020–2021, with regional effects across Southeast and Mid-Atlantic in the subsequent year. Black population share showed a consistently positive association [33] with excess mortality across most regions, yet this relationship was absent in Ohio, Michigan, Indiana, New Mexico, and Arizona during 2020–2021.

Hispanic populations were particularly vulnerable in the Southwest in 2020–2021. The percentage of elderly residents was a significant mortality risk factor in the southern U.S. in 2020–2021and became more pronounced in the western states in 2021–2022. Lack of health insurance emerged as the prominent risk factor in the western counties in the early period before becoming more spatially concentrated in the Southwest in 2021–2022.

Some unexpected patterns also emerged. Diabetes prevalence had a broad impact along the East Coast in 2020–2021 but became sharply localized in 2021–2022, with strong influences in Arizona, South Texas, Southeast Ohio, and among Dakotas and Minnesota. One anomalous negative association occurred in Southern Kansas, warranting further investigation. Higher male populations were associated with lower mortality in parts of the Midwest and Southeast, though an unexpected positive association appeared in South Dakota. The share of early young adults serves as a protective factor against excess mortality in 2020–2021, particularly in counties across New Mexico, Texas, and Oklahoma but the influences disappeared in the following year.

Lastly, our findings underscore the importance of accounting for both structural determinants of health, temporal variation, and localized context. One of the innovative contributions of this study is its ability to decompose excess mortality into two distinct types of spatial effects. The first involves measurable sociodemographic factors such as income, health insurance coverage, or age which are captured through quantifiable covariates. The second includes intrinsic contextual effects which reflect unobserved influences such as cultural norms, neighborhood milieu, public health compliance, and even the role of local media. While sociodemographic risk factors significantly contributed to elevated excess death rates across much of the southern and southeastern United States, contextual effects became particularly salient in the southern counties during 2021–2022, highlighting how deeply embedded local behaviors and attitudes can shape health outcomes during a public health crisis in different periods.

Although this study provides local knowledge and insights into excess mortality spatially and temporally, this study faced several limitations related to data availability, the method used to calculate excess death rates, and the reliance on geographic units for which data were reported. Variables such as vaccination rates and lockdown mandates are only available at the state level, limiting our ability to capture their effects at finer geographic resolutions. Additionally, age-adjusted final national mortality data are not available at the county level at the CDC, which may influence excess death estimates in regions with aging populations [55]. However, while our significance tests show that the percentage of residents aged 65 and over was not significantly changing in 2020–2021, it became significant in 2021–2022, suggesting the need for future age-sensitive analyses.

Regarding the calculation of excess death rates, our calculation of excess mortality relied on finalized annual counts at county level, which, while robust, differ from alternative measures such as life expectancy, Age Standardized Years of Life Lost (ASYLL.) [56, 57]. As Bonnet et al. noted [57], such subnational analyses can yield more nuanced insights than national-level studies, yet they can also suffer from statistical uncertainty. Bonnet and Camarda recommend that statistical uncertainty can be reduced by aggregating small areas or by estimating excess mortality for both sexes combined [58], and we adopt the latter strategy in this study.

Although sex- and age-specific modeling might provide further detail, many of our key independent variables, such as Republican vote share, population density, MDI, and diabetes prevalence, are not disaggregated by sex or age; thus, excluding them would omit critical structural and contextual influences central to our spatial analysis.

Despite the mentioned constraints, MGWR framework offers a robust and replicable approach for identifying where and why mortality patterns vary, and it therefore provides actionable guidance for geographically targeted public-health interventions that reflect both measurable sociodemographic risks and localized contextual influences.

Conclusion

Since the outbreak of COVID-19, government officials, public health policymakers and researchers have struggled to understand the causes behind disparities in COVID-19-related mortality across the United States. Some contributing factors include the limitations of official COVID-19 records, which are often misrepresented by underreporting, testing constraints, and misclassification, and also the oversight of unmeasurable local contextual and sociodemographic effects across regions. While official death counts provide a basic gauge of the mortality, they often mask the true damage of the pandemic, especially in regions with limited testing or delayed, inaccurate, or incomplete reporting. To overcome these limitations, this study incorporates COVID-19-related excess mortality to capture both direct deaths caused by the virus and indirect fatalities resulting from the pandemic’s broader social, economic, and public health disruptions. We then apply MGWR to quantify how the effects of key risk factors vary spatially across U.S. counties in both the earlier phase of the pandemic, 2020–2021, and the latter phase, 2021–2022. MGWR enables the identification of the spatial scale at which each determinant operates, distinguishing between factors whose impacts are highly localized and those with relatively uniform effects across larger areas. By uncovering both observable and latent drivers of excess mortality, the approach offers actionable insight for designing targeted and equitable interventions. The workflow is also readily transferable to other national settings where spatial heterogeneity in health outcomes is suspected.

Our results reveal significant spatially varying effects of key determinants of excess mortality during the pandemic in 2020–2021 and 2021–2022: poverty, political affiliation, age structure, racial and ethnic composition, sex composition, pre-existing health conditions, and health insurance.

The policy implications are clear: Uniform, one-size-fits-all interventions will not be sufficient to address the uneven distribution of social vulnerability. Public health strategies must be geographically tailored, responsive not only to measurable spatial relationships of known risk factors but also to the place-specific norms that influence behavior and access. This study contributes to a growing body of spatial epidemiological research that not only quantifies disparities in health outcomes with two distinct spatial effects but also grounds them in different social structural conditions and underscores the urgent need for place-based public health planning, especially in preparation for future health crises.

Appendix A. Summary statistics

[Please See Tables 6 and 7 in Appendix A]

Table 6.

Summary statistics of variables used in excess mortality 2020–2021

(1) Excess Mortality 2020– 2021	Mean	Standard Deviation	Minimum	Maximum
Variables
ExcessDeathRate	234	126.8	−262	1390.7
PctPov	14.5	5.9	1.8	59
PctRep	63.9	15.6	8.7	92.4
PctBlack	9.6	14.7	0	87
LnPctBach	3.1	0.4	1.8	4.4
Pct65Over	18.8	4.3	3.1	57.6
LnMDI	2.6	0.6	−0.1	4.3
PctWoInsurance	9.4	4.8	1.1	44.9
LnPctUnemployed	1.6	0.4	−1.8	3.3
PctDiabetics	12.5	3.7	2.4	29.5
SexRatio	101.7	11.4	77.1	221.3
Pct18to24	8.7	3.4	2.7	48
PopDen	241.3	833.2	0.7	18457.8
LnPctHis	1.9	0.9	0.0	4.6
LnMedHhIncome	10.9	0.3	10.1	12

Table 7.

Summary statistics of variables used in excess mortality 2021–2022

( 2 ) Excess Mortality 202 1 −202 2	Mean	Standard Deviation	Minimum	Maximum
Variables
ExcessDeathRate	220.3	137.7	−400.7	1323.6
PctPov	14.5	5.9	1.8	59
PctRep	63.9	15.6	8.7	92.4
PctBlack	9.5	14.6	0	85.7
LnPctBach	3.1	0.4	1.7	4.4
Pct65Over	19.3	4.4	2.9	57.9
LnMDI	2.6	0.6	−0.1	4.3
PctWoInsurance	9.3	4.8	0.9	45.1
LnPctUnemployed	1.5	0.4	−0.9	3.1
PctDiabetics	10.9	2.3	5.5	21
SexRatio	102	11.1	78.2	223.5
Pct18to24	9	3.5	2.1	47
PopDen	240.5	830.3	0.7	18209.4
LnPctHis	1.9	0.9	0	4.6
LnMedHhIncome	11	0.3	10.1	12

Appendix B. OLS Models

[Please See Tables 8 and 9 in Appendix B]

Appendix C. Nonlinearity Check for the MGWR Models

[See Figures 5 and 6]

Appendix D. Excluded MGWR Significant Local Estimates Maps

In both 2020–2021 and 2021–2022, the share of population with at least a bachelor's degree shows a strong negative association with excess mortality across the country. MDI has a widespread and robust positive association with excess mortality in 2020–2021, but this effect largely disappears in 2021–2022. Unemployment rate is positively associated with excess mortality in both periods, though its spatial pattern changes notably. In 2020–2021, the effect is widespread across the whole country, while in 2021–2022, the strongest associations are concentrated in the Western half of the country and in North Carolina. Population density demonstrates an uneven positive relationship with excess mortality in 2020–2021, particularly in central and Eastern counties, but not along the West Coast. By 2021–2022, this association disappears entirely. Median household income is negatively associated with excess mortality across much of the Southwest, South, Midwest, and Northeast in 2020– 2021 but this spatial pattern vanishes in 2021–2022 (Fig 7).

Appendix E. Excess Mortality from 2020 to 2022

[See Tables 10, 11, 12 and 13 in Appendix E]

Table 10.

Summary statistics of variables used in excess mortality 2020–2022

Excess Mortality 2020–202 2	Mean	Standard Deviation	Minimum	Maximum
Variables
ExcessDeathRate	214.8	116.9	−300.9	1331.6
PctPov	14.5	5.9	1.8	59
PctRep	63.9	15.6	8.7	92.4
PctBlack	9.5	14.6	0	85.7
LnPctBach	3.1	0.4	1.7	4.4
Pct65Over	19.3	4.4	2.9	57.9
LnMDI	2.6	0.6	−0.1	4.3
PctWoInsurance	9.3	4.8	0.9	45.1
LnPctUnemployed	1.5	0.4	−0.9	3.1
PctDiabetics	10.9	2.3	5.5	21
SexRatio	102	11.1	78.2	223.5
Pct18to24	9	3.5	2.1	47
PopDen	240.5	830.3	0.7	18209.4
LnPctHispanic	1.9	0.9	0	4.6
LnMedHhIncome	11	0.3	10.1	12

Table 11.

OLS results of excess mortality 2020–2022

Excess Mortality 2020–2022	Estimates	t-value	p-value	VIF
Variables
Intercept	0.000	0.000	1.000
PctDiabetics	0.329*	9.037	0.000	6.63
PctRep	0.156*	6.692	0.000	2.73
Pct65Over	0.140*	7.087	0.000	1.95
PctPov	0.131*	3.811	0.000	5.93
LnPctUnemployed	0.131*	7.583	0.000	1.48
LnPctHispanic	−0.101*	−5.140	0.000	1.94
Pct18to24	−0.092*	−4.866	0.000	1.78
LnMDI	0.066*	2.589	0.010	3.28
SexRatio	−0.051*	−3.333	0.001	1.17
PctWoInsurance	0.040*	2.041	0.041	1.95
LnPctBach	−0.029	−0.983	0.326	4.41
LnMedHhIncome	−0.014	−0.382	0.702	6.91
PctBlack	−0.005	−0.190	0.849	2.84
PopDen	0.001	0.078	0.938	1.31
Model Diagnostics
N	2762
R2	0.450
Adjusted R2	0.447
AIC	6217.1
AICc	6219.3

*Significant at 95 percent level

Table 12.

Comparison of model performance between OLS and MGWR 2020–2022

	Excess Mortality 2020–2022
	OLS	MGWR
R 2	0.45	0.62
Adjusted R 2	0.45	0.58
AIC	6217	5726
AICc	6219	5796

Table 13.

Optimal bandwidths and spatial test results 2020–2022

	Excess Mortality 2020–2022 O ptimal B andwidth	Monte Carlo P -value
Variables
Intercept	2761	0.925
PctPov	2761	0.859
PctRep	1479	0.006*
PctBlack	2761	0.995
LnPctBach	2761	0.603
Pct65Over	184	0.000*
LnMDI	2690	0.148
PctWoInsurance	149	0.001*
LnPctUnemployed	898	0.122
PctDiabetics	48	0.000*
SexRatio	172	0.016*
Pct18to24	1857	0.148
PopDen	2761	0.142
LnPctHispanic	316	0.001*
LnMedHhIncome	2761	0.222
Maximum CN	9.32

*Significant variation at 95 percent level

[See Figures 8, 9, 10 and 11]

Fig. 8.

Excess mortality at the county level from 2020 to 2022

Fig. 9.

MGWR significant local estimates, optimal bandwidths, and corresponding OLS estimates

Fig. 10.

Contextual influences on the excess mortality during the pandemic

Fig. 11.

Nonlinearity check for excess mortality 2020–2022

Abbreviations

ACS

American community survey

AICc

Corrected Akaike information criterion

CDC

Centers for disease control and prevention

CN

Condition number

COVID-19

Coronavirus disease 2019

EUA

Emergency Use Authorization

ExcessMortality_Jan20–Dec 22

Three-year average crude death rates per 100k from 2020 to 2022 minus the average crude death rates from 2015 to 2019

ExcessMortality_Jan 21–Dec 22

Two-year average crude death rates per 100k from 2021 to 2022 minus the average crude death rates from 2015 to 2019

ExcessMortality_Jan 20–Dec 21

Two-year average crude death rates per 100k from 2020 to 2021 minus the average crude death rates from 2015 to 2019

IMD

Index of Multiple deprivation

LnMDI

Natural log of multidimensional deprivation index

LnMedHhIncome

Natural log of median household income in US dollars ($)

LnPctBach

Natural log of percentage of the population has at least bachelor’s degree

LnPctHispanic

Natural log of percentage of the Hispanic population

LnPctUnemployed

Natural log of unemployment rate of labor force aged over 16

MDI

Multidimensional deprivation index

MGWR

Multiscale geographically weighted regression

OLS

Ordinary least squares regression

Pct18to24

Percentage of the population aged between 18 and 24

Pct65Over

Percentage of the population aged over 65

PctBlack

Percentage of the Black and African American population

PctDiabetics

Percentage of the population that is diabetic

PctPov

Percentage of the population that earns below poverty level

PctRep

Percentage of votes for the Republican party in 2020

PctWoInsurance

Percentage of the population without health insurance

PopDen

Population density

SexRatio

The number of males per 100 females

VIF

Variance inflation factor

WHO

World Health Organization

Author contributions

C. Kao collected the data, conducted the analysis, and prepared all figures. A. S. Fotheringham conceptualized the study and provided methodological guidance. All authors contributed to the writing, editing, and revision of the manuscript.

Funding

This work is supported by a National Science Foundation award (#2117455) to Professor A. Stewart Fotheringham.

Data availability

The data sources used in this study are publicly available and listed below: Yearly final mortality data: https://wonder.cdc.gov/Election Data: https://doi.org/10.7910/DVN/42MVDX U.S. census data: https://www.census.gov/programs-surveys/acs.htmlHealth data: https://www.countyhealthrankings.org/health-data/methodology-and-sources/data-documentation/national-data-documentation-2010-2022.MDI data: https://www.census.gov/library/working-papers/2022/demo/SEHSDwp2022-19.html.The dataset and code supporting the conclusions of this article is available in the figshare.https://figshare.com/s/03ab81cf27ef2e584633.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.