Spatial nonstationarity and the role of environmental metal exposures on COVID-19 mortality in New Mexico

Abstract

Worldwide, the COVID-19 pandemic has been influenced by a combination of environmental and sociodemographic drivers. To date, population studies have overwhelmingly focused on the impact of societal factors. In New Mexico, the rate of COVID-19 infection and mortality varied significantly among the state’s geographically dispersed, and racially and ethnically diverse populations who are exposed to unique environmental contaminants related to resource extraction industries (e.g. fracking, mining, oil and gas exploration). By looking at local patterns of COVID-19 disease severity, we sought to uncover the spatially varying factors underlying the pandemic. We further explored the compounding role of potential long-term exposures to various environmental contaminants on COVID-19 mortality prior to widespread applications of vaccinations. To illustrate the spatial heterogeneity of these complex associations, we leveraged multiple modeling approaches to account for spatial non-stationarity in model terms. Multiscale geographically weighted regression (MGWR) results indicate that increased potential exposure to fugitive mine waste is significantly associated with COVID-19 mortality in areas of the state where socioeconomically disadvantaged populations were among the hardest hit in the early months of the pandemic. This relationship is paradoxically reversed in global models, which fail to account for spatial relationships between variables. This work contributes both to environmental health sciences and the growing body of literature exploring the implications of spatial nonstationarity in health research.

Introduction

The impact of the COVID-19 pandemic worldwide has been influenced by a combination of environmental and sociodemographic drivers (Wang et al. 2022). In New Mexico, the rate COVID-19 infection and mortality varied significantly among the state’s geographically dispersed, and racially and ethnically diverse populations who are exposed to unique environmental contaminants related to resource extraction. While many studies focus on the relationship between social vulnerabilities and COVID-19 mortality, somewhat fewer explore the role of environmental pollution (e.g., Yu et al. 2018; Meng et al. 2021, 2022). However, a knowledge gap remains in understanding how environmental and social drivers together coproduce increased COVID-19 mortality and none have been systematically conducted in rural western states. One such state, New Mexico, is home to a potentially high exposure burden, as well as geographically and racially/ethnically diverse populations.

In previous environmental health research, members of our team conducted immune response modeling using estimated annual consumption of toxicants via drinking water sources among study participants – uranium, arsenic, nickel, mercury, and total radium (Ra-226 and −228). These metal contaminants were selected due to their prevalence in New Mexico water supplies (principally arsenic, uranium, and radium) (Hoover et al. 2017) and on a priori knowledge of their potential immunotoxic effects (principally arsenic, mercury, and nickel) (Aranyi et al. 1985; Haley et al. 1987; Ahmed et al. 2014). For example, arsenic has also been shown to be immunosuppressive for the development of autoantibodies. In addition, women may convert arsenic to particularly immunosuppressive metabolites, and long-term exposure of Navajo men to arsenic during aging may suppress these autoantibodies. These findings are consistent with the association between urinary arsenic and decreased protective antibody response to the mumps vaccine (Raqib et al. 2017) and against tuberculosis among Bangledesh children (Ahmed et al. 2014), and that sex-specific differences may complicate this effect (Lindberg et al. 2008). In direct tests arsenic trioxide inhibited interferon-γ synthesis in vitro by lymphocytes from patients with systemic lupus erythematosus (Hu et al. 2018) and prolonged survival and decreased anti-nDNA production in lupus-prone mice (Xia, Lin, and Zhou 2007), possibly by inhibiting inflammasome activation (Maier et al. 2014). Despite significant knowledge gaps in the immunotoxicity of arsenic and other metals/metalloids, these findings strongly suggest that such environmental exposures should be further investigated in the context of COVID-19 disease severity and death in New Mexico.

Impacts on the immune system may be especially pronounced and diverse among New Mexico populations based on proximity to mining- and fracking-related environmental exposures. For example, highly remote, rural populations such as those living on the Navajo Nation have documented urinary uranium concentrations at the 99th percentile of the National Health and Nutrition Examination Survey (NHANES) among non-occupationally exposed community members across the lifespan (Harmon et al. 2017; Hoover et al. 2020; Scammell et al. 2020; Erdei et al. 2022), indicating significant exposure risk to uranium and other metals through multiple environmental pathways. Resource extraction processes disrupt the natural distribution of soils, increasing the dispersion of particle-bound metals into environmental media (e.g., air, water, and soils) and the exposure risk to communities. Exposure to environmental metals has demonstrated impacts on immune system, which may render exposed individuals and communities at greater risk for COVID-19 mortality (Erdei et al. 2020; Wu et al. 2020; Faruk et al. 2023).

Sociodemographic factors are implicated in higher SARS-CoV-2 transmission rates, particularly among minority communities and people of color (Erdei et al 2020). The relationship between social vulnerability and COVID-19-related hardships in the US is well-documented (Dasgupta et al. 2020; Gaynor and Wilson 2020; Karaye and Horney 2020; Khazanchi et al. 2020; Nayak et al. 2020; Tieskens et al. 2021). In New Mexico, social vulnerabilities including household crowding, diminished access to running water or telecommunications infrastructure, and structural mechanisms of institutionalized racism have all been implicated in disproportionate adverse impacts from COVID-19 on communities of color, particularly among Indigenous populations in the western half of the state (Yellow Horse, Deschine Parkhurst, and Huyser 2020; Gorris et al. 2021; Rifat and Liu 2022). Rural New Mexican populations are also disproportionately underserved by healthcare and other crucial services (Soto Mas et al. 2019). We argue that the simultaneous comparison of environmental exposure and social determinants of health (SDOH) may shed light on their relative contribution to COVID-19 mortality.

The extent to which social and environmental drivers of COVID-19 morbidity and mortality are locally compounded has not been established and commonly applied ecologic models may be insufficient for capturing how multiple factors heterogeneously interact in localized contexts to exert influence on COVID-19 mortality. Because environmental and social contexts vary across space, the drivers of COVID-19 infection and death may be locally compounded. The heterogeneous distribution of health determinants and outcomes are thus spatially nonstationary. Spatial nonstationarity is the variability of causal relationships and statistical associations across space (Kwan 2021). Model coefficients are subject to spatial nonstationarity, meaning that global regression models may not adequately capture the complexity and nuance of spatial associations because they frame health-environment relationships as unchanging across space (Brunsdon, Fotheringham, and Charlton 1996).

Spatial regression models like geographically weighted regression (GWR) have been widely adopted in environmental health and justice research to account for spatial nonstationarity. For example, using GWR, Gilbert and Chakrborty (2011) demonstrated significant spatial patterns of socioeconomic disparities and cumulative cancer risks in Florida, and Grineski et al. (2015) revealed spatial associations between PM2.5 and social characteristics with children’s respiratory health in Texas. Other research has demonstrated the spatial heterogeneity of determinants of obesity (Oshan, Smith, and Fotheringham 2020) and associations between green spaces and COVID-19 transmission (Huang et al. 2020).

We hypothesize that spatial heterogeneity of predictor variables indicates spatial associations between potential exposure to environmental contaminants and COVID-19 mortality when simultaneously considering localized sociodemographic population characteristics. To that end, we expect to see the relationship between documented social and environmental exposures and COVID-19 mortality to manifest in different directions and magnitudes across the state, with significant positive associations in underserved regions. Those drivers include potential social vulnerabilities derived from the ATSDR-CDC Social Vulnerability Index (SVI) and the American Community Survey (ACS), as well as potential environmental exposure to environmental toxicants stemming from sites of resource extraction.

We specifically aim to address how $\beta$ coefficients reflecting the relationship between potential predictors and COVID-19-related mortality may be subject to Simpson’s paradox when spatial stationarity of model terms is assumed. In spatial models, Simpson’s paradox may occur when models calibrated on the same aggregated data contradict each other when population samples are subset into logical subgroups at the local level (Sachdeva and Fotheringham 2023). Our methods apply combinations of geographically weighted regressions (GWR) and multiscale GWR (MGWR) modeling techniques to account for local heterogeneity in explanatory environmental and social variables and COVID-19 mortality, examining linear relationships within spatial neighborhoods. The application of our methodology provides a simultaneous comparison of environmental exposure and SDOH on their relative contribution to COVID-19 mortality. Our analysis supports the ongoing planning of state and local public health response by promoting New Mexico and place-specific understanding of environmental risks and COVID-19 disease associations. These are critical considerations in a rural and underserved state in the Southwestern U.S. By asking where and to what degree do complex and complementary social and environmental processes related to COVID-19 mortality exist across the state, the purpose of the study is to demonstrate the analytical nuance provided by local regression modeling and increase understanding of the health challenges faced by diverse communities in New Mexico.

Additionally, we aim to address the effect of potential underreporting of COVID-19 cases on regression results. The State of New Mexico Department of Health (NMDOH) collected and compiled all available data on SARS-CoV-2 infection diagnoses since the beginning of the pandemic. Medical facts-based public health measures were implemented swiftly in New Mexico in the first week of March 2020 and significantly curbed the virus spreading primarily at hospital and primary health care settings. However, this process became increasingly challenging due to actions taken by US Centers for Disease Control and Prevention (CDC) and other governmental agencies via the distribution of household-based, free of charge, individual testing kits. The actual reporting of cases was potentially hindered by these diverse testing opportunities and COVID-19 reporting requirements (Castro et al. 2021). People who tested positive but who were unable or unwilling to seek medical care might be underrepresented in state datasets (Katzman et al. 2021; Sood et al. 2021). Furthermore, at the beginning of the pandemic as health care provider capacities were shut down limited, reporting may not reflect the actual numbers of cases.

Materials and Methods

Our exploratory study design develops and evaluates models at contrasting spatial specificity to test our hypothesis. Given that our data was available in a range of spatial units (e.g., zip code tabulation area [ZCTA] and census tract), we started by spatially harmonizing environmental and sociodemographic data related to environmental exposure and COVID-19. Next, we developed and compared linear and logistic regression models at both global and local levels. Finally, we rely on a priori knowledge of the study area to evaluate how sociodemographic and environmental factors differentially play out and coalesce as lurking variables in global regressions in rural and underserved regions of New Mexico compared to urban centers.

Study Area

New Mexico is a minority-majority state located in the southwest US (Figure 1) with 50.2% Hispanic or Latino and 11.2% American Indian/Alaska Native populations (U.S. Census Bureau 2021). The population in 2020 was approximately 2.1 million (U.S. Census Bureau 2021). The state is highly rural – an estimated 12% of the total population resides in areas classified as either rural or small towns according to the USDA Rural-Urban Commuting Area (RUCA) Codes, which account for approximately 52% of total land area. Significant natural resource production, refinement, and waste disposal activities occur in predominantly rural areas across the state, potentially rendering local communities disproportionately exposed to environmental contaminants. Figure 1 shows the locations of uranium and other hard rock mining sites, and unconventional oil and natural gas wells across the state. Many of the hard rock mine features include ore tailings disposal sites. Also included on the map is the Waste Isolation Pilot Project (WIPP) – one of three large underground storage facilities for radioactive waste in the world, which is located in the southeastern corner of the state.

Figure 1.

Study area map

Outcomes

A total of 3,863 deaths were recorded in the State of New Mexico COVID-19 disease registry and public health database between March 1, 2020, and March 26, 2021. The total number of COVID-19 deaths, tests, and vaccinations are provided by NMDOH aggregated at the ZCTA level. A total of 202 ZCTAs (55%) had at least one reported COVID-19-related death as of March 26, 2021. The estimate provided by NMDOH may be low due to underrepresenting the number of cases and deaths in ZCTAs on Tribal lands, at correctional facilities, and at military bases.

This study was carried out in collaboration with the Division of Epidemiology and Response of the NMDOH with approval from the UNM Health Sciences Center Human Research Protections Office.

Sociodemographic Covariates

To account for the extent of social determinants of COVID-19 deaths, we incorporated data from the CDC-ATSDR Social Vulnerability Index (SVI) reporting census tract-level sociodemographic characteristics, household characteristics, racial composition, and housing type. More granular estimates of racial composition at the ZCTA level were derived from the 2020 American Community Survey (ACS). Rurality or urbanity of ZCTA units were derived from both the USDA RUCA and the Frontier and Remote Area (FAR) codes.

Data Harmonization

Because COVID-19-related deaths were aggregated at ZCTA level, we applied dasymetric mapping, which creates population-weighted coefficients to reapportion data aggregated at different geospatial configurations and scales (Eicher and Brewer 2001), such as SVI data, which are aggregated at the census tract level. Granular population distributions were obtained from the Oak Ridge National Laboratory LandScan estimates of 24-hour ambient population for every square kilometer (Oak Ridge National Laboratory 2020). RUCA and FAR data are both available at the ZCTA level. Summary statistics of input data aggregated by ZCTA are provided in supplementary Table S1.

Environmental Pollution Estimates

To estimate potential environmental pollution from metals, we applied a geospatial approach for predicting the release of metal contaminants from abandoned and active mining sites and fracking sites as described in Lin et al. (2020). The models here and those described in Lin et al. (2020) specifically characterize the theoretical distribution of fugitive mine waste through various environmental processes. We constructed additional GIS layers proximity to mines by type and oil/natural gas production and aerial radiometric scans.

The locations of mine sites and features were derived from the Mineral Resources Data System (MRDS), a compendium of records curated by the USGS containing the point locations and primary commodities of mines across the coterminous U.S. (Mason and Arndt 1996). All point locations were used in the present analysis because the dataset does not provide sufficient information to discriminate mine activity or reclamation status.

The point locations within New Mexico of all reported unconventional oil and natural gas production wells via hydraulic fracturing, or fracking, were obtained from the FracFocus Chemical Disclosure Registry (2023). Fracking sites are primarily grouped in the southeast and northwest portions of the state (Figure 1). We modeled the density of fracking wells per ZCTA as the number of wells divided by the area in square kilometers of every ZCTA plus a 50-km buffer.

The potential aeolian (windblown) distribution of mine and fracking waste in dust was modeled two ways. The wind index (WI) was derived to estimate downwind exposure from mine sites at every location based on prevailing wind direction using the following equation (Ryan et al. 2008; Lin et al. 2020):

$$WI=\frac{1-\text{cos}\theta -\beta }{2}$$ (1)

where $\theta$ is the straight-line distance from contaminant sources – in this case, active and abandoned mine and fracking sites – $\beta$ is the 30-year hourly average prevailing wind direction at any geographical location – in this case, every 30m grid cell in a digital elevation model (DEM) (U.S. Geological Survey 2022) – derived from the North American Regional Reanalysis (NARR) dataset (NOAA Physical Sciences Laboratory 2023). The output is a scaled value ranging from 0 to 1, where a WI of 1 indicates locations directly downwind from sources, 0.5 are locations directly perpendicular to sources, and 0 are locations directly upwind from sources.

Topographic wind exposure (TWE) expands on the wind index by considering terrain variability. For example, the leeward slope of a hill or mountain is less exposed to contaminant-bearing winds than the windward side. TWE calculates the angle of incidence between wind and surficial features and is also dependent on prevailing wind averages according to the following equation (Antonić and Legović 1999; Böhner and Antonić 2009; Lin et al. 2020):

$$\text{cos}\alpha =\text{cos}\mu \text{sin}\beta +\text{sin}\mu \text{cos}\beta \text{cos}\delta -\gamma$$ (2)

We adapted the methodology described in Lin et al. (2020) such that $\text{cos}\alpha$ is the angle of exposure, $\mu$ is the slope and $\gamma$ is the aspect of a land surface, and $\beta$ and $\delta$ are the horizontal angle and 30-year prevailing wind direction, respectively. TWE values range from −1 to 1 where values approaching −1 indicate low TWE and those approaching 1 are highest.

Next, following work detailed in Lin et al. (2020) we classified landforms using the topographic position index (TPI) as valleys, steep slopes, lower slopes, moderate slopes, flat land, and ridges. Finally, we computed the median value per ZCTA of estimated uranium-238 from aerial gamma-ray surveys conducted as part of the National Uranium Resource Evaluation (NURE) program (Kucks 2005), as well as median concentrations of arsenic, cadmium, copper, mercury, nickel, lead, uranium, and uranium in NURE soil and sediment samples. All environmental variables were normalized from 0–1 for statistical modeling.

Statistical Modeling

The common approach for determining the suitability of spatial models is to evaluate the effect of spatial nonstationarity in global linear models before proceeding to local ones (Comber et al. 2020; Chen et al. 2024). We first modeled the number of deaths due to COVID-19 using zero-inflated models to account for the large number of NM ZCTAs (145 of 396 as of March 31, 2021) reporting zero deaths.

When such a model is applied with the number of deaths as outcome, we expect the population size and the interaction of population size with other variables to be strong predictors. Therefore, the outcome of interest was the relative number of COVID-19 related deaths in a ZCTA rather than the absolute number of such deaths using:

$$logitr=\text{log}\left(\frac{r}{1-r}\right)$$ (3)

$$\text{where}r=\frac{(\text{\#}COVID-19deaths\text{)}}{(\text{\#}testsgiven\text{)}}\text{and}r\text{is}\text{in}(0,1).$$

For the 226 observations where $logitr$ was defined, we fit models to estimate 1) the association between exposures and $r$ when $N\mathit{\left(}COVID\text{-}\mathit{19}deaths\mathit{\right)}>\mathit{0}$; and 2) the association between exposures and the probability that $N\mathit{\left(}COVID\text{-}\mathit{19}deaths\mathit{\right)}>\mathit{0}$. The first question we sought to answer was thus how the rate of COVID-19-related mortality differed from the absolute number of reported deaths. To answer this question, we model $logitr$, using only observations for which $N\mathit{\left(}COVID\text{-}\mathit{19}deaths\mathit{\right)}>\mathit{0}$). However, we cannot rule out the possibility of instances where the reporting of COVID-19-related deaths per ZCTA is 0 may be inaccurate and that the number of COVID-19-related deaths in some ZCTAs is in fact at least 1. To address this uncertainty, we then defined a binary variable, Y, where:

$$Y=1ifN\mathit{\left(}COVID\text{-}\mathit{19}deaths\mathit{\right)}>\mathit{0};\mathit{0}ifN\mathit{\left(}COVID\text{-}\mathit{19}deaths\mathit{\right)}=\mathit{0}$$ (4)

and modeled $Y$ using all observations for which the predictor variables have measured values.

Predictor Variables

The variables we considered as potential predictors in all models included:

• The proportion of the population living in poverty obtained from the SVI, $proppov$

• The proportion of the population over 65 years in age obtained from the SVI, $prop65$

• The logarithm of ZCTA land area, $logarea$

• A variable denoting a combination of rurality (RUCA) and remoteness (FAR), $fmm$, adjusted for difference in location and scale in, where

  $$fmm=\text{max}(RUCA-1,2\times FAR)$$ (5)

• The logarithm of uranium in parts per million, $logeU.PPM$

• Proximity to mining features, $MineProx$

• Wind index, $WindIndex$

• Topographic wind exposure $TWE$

• The logarithm of the landforms variable, $logLandforms$,

• The proportion of land that is held in trust by the federal government for Native American reservations, $pc{t}_{TL}$, and

• Proportion of populations grouped by race as defined by the 2020 American Community Survey as: non-Hispanic white), b) Native American or Pacific Islander ($p_{NA}$) c) Hispanic ($p_{Hisp}$), or d) other ($p_{Other}$).

• Estimated total population derived from LandScan ($logpop$) was also considered as a potential predictor of the binary outcome defined in Equation 1.

The percentage of ZCTA land area that is Tribal reservation land and the proportion of population that is Native American were zero for several observations, and were transformed respectively with:

$$logpc{t}_{TL}=\text{log}\left(0.00001+pc{t}_{TL}\right)ifpc{t}_{TL}isdefined$$ (6)

$$log{p}_{NA}=\text{log}\left(0.00001+pc{t}_{NA}\right)ifpc{t}_{NA}isdefined$$ (7)

Due to the large number of covariates and predictors that we considered, and to avoid having predictor variables that were highly correlated, Principal Component Analysis (PCA) was used on both the sociodemographic covariates and soil metal concentrations. PCA is a technique for reducing the dimensionality of a dataset in order to ease interpretability while preserving as much statistical information as possible (Jolliffe and Cadima 2016). In PCA, an orthogonal linear transformation is applied to reframe the original potentially correlated data into a smaller number of uncorrelated variables called principal components (PCs). The first four PCs of sociodemographic covariates contained more than 90% of the total variation (Table 1). Two separate PCA were conducted for soil metal concentrations – one containing arsenic and another without due to significant missingness in arsenic measurements. PC loadings for soil concentrations are reported in supplemental table S2.

Table 1.

Loadings of the first four principal components from the PCA of sociodemographic covariates

	PC 1	PC 2	PC 3	PC 4
The proportion of the population living in poverty, $proppov$	0.4075	0.21	0.4148	0.771
The proportion of the population over age 65, $prop65$	−0.2634	−0.2664	0.8846	−0.2355
Log Proportion Tribal lands, $logpc{t}_{TL}$	0.5172	−0.1834	0.06024	−0.3472
Log Percent Native American, $logpc{t}_{AIAN}$	0.5126	−0.2566	0.05473	−0.2553
Percent non-Hispanic White $pc{t}_{anglo}$,	−0.4436	−0.4833	−0.08783	0.2774
Percent Hispanic, $pc{t}_{hisplat}$	−0.1935	0.7429	0.1766	−0.2958

These PCs will be labeled $PropRacCatComp$. $X$, where $X=1,2,3,4$.

Global Modeling

$Logitr$ is continuous and was modeled with ordinary linear regression. The $Y$ of Equation 4 is binary and was modeled using logistic regression.

The number of observations that can be used is further diminished if we attempt to use data on metals in soil/sediment. As stated above, if we attempt to include the effects of five metals, we would include as potential covariates the four principal components in Supplemental Table S2a; if we try to include the effects of all metals but arsenic we would include as potential covariates the three principal components in Supplemental Table S2b.

Model 1 tests the binary outcome, Y, model 2 tests the continuous outcome, $Logitr$. Each test took the following approach: first a full model was fitted, using all variables in the set of potential predictors. Then we applied backward stepwise model selection using the Akaike information criterion (AIC) to remove predictor variables that were not significantly associated with the outcome using a 95% confidence level (p>0.05). When starting with full models with metal-based variables (Table 3), the reduced models either included no metal-based PCs, or in one case (binary outcome, using PCs based on four metals, 191 observations), one metal PC appeared in the reduced model, but with high uncertainty (p>0.1). Therefore, we concluded that the models based on smaller numbers of observations with available soil and sediment metal concentration variables do not show results that are both different from the models with no metal data and also significant. The final reduced models are shown in Tables 3 and 4.

Table 3.

Models for binary outcome, $Y$. This model uses 333 observations, with no data for five metals

	Estimate	Std. Error	z-Value	$Pr(>|z|)$	FDR Test (j/m)* $\delta$:
					$\delta =0.05$	$\delta =0.01$	Reject $H_0$
(Intercept)	−20.520	10.440	−1.967	0.049	0.030	0.006	0
logpop	3.640	1.808	2.014	0.440	0.027	0.005	1
FMM	−0.197	0.064	−3.091	0.002	0.013	0.003	0
logarea	0.655	0.166	3.935	0.000	0.003	0.001	1
MineProx	12.570	11.190	1.123	0.262	0.043	0.009	0
WindIndex	−1.895	0.932	−2.033	0.042	0.023	0.005	0
logLandForms	−0.625	0.380	−1.646	0.100	0.033	0.007	0
logeU.PPM	−1.068	0.653	−1.636	0.102	0.037	0.007	0
densityFracking	−2.735	2.658	−1.029	0.304	0.050	0.010	0
PropRacCatComp.1	0.472	0.129	3.651	0.000	0.007	0.001	1
PropRacCatComp.2	2.586	1.096	−2.359	0.018	0.017	0.002	0
PropRacCatComp.3	0.769	0.243	3.170	0.002	0.010	0.003	1
logpop:MineProx	−2.628	1.967	−1.336	0.182	0.040	0.008	0
logpop:densityFracking	0.558	0.500	1.115	0.265	0.047	0.009	0
AIC	225.8
p-value of $x^2$ statistic	0
Tjur’s COD	0.6141

Table 4.

Models for logit r. This model uses 202 observations, with no data for five metals.

	a) Model after variable selection				b) Full model
	Est.	SE	t-Val	$Pr(>|t|)$	Est.	SE	t-Val	$Pr(>|t|)$
(Intercept)	−3.773	0.184	−20.480	0.000	−4.385	0.788	−5.562	0.000
FMM	0.054	0.018	3.018	0.003	0.045	0.021	2.082	0.039
WindIndex	−1.152	0.243	−4.750	0.000	−1.473	0.356	−4.139	0.000
TWE	−0.314	0.222	−1.414	0.159	−0.307	0.245	−1.253	0.212
PropRacCatComp.4	0.400	0.097	4.120	0.000	0.383	0.106	3.607	0.000
logarea	−	−	−	−	0.025	0.054	0.453	0.651
MineProx	−	−	−	−	0.204	0.412	0.496	0.620
Landforms	−	−	−	−	−0.159	0.144	−1.104	0.271
logeU.PPM	−	−	−	−	−0.061	0.231	−0.264	0.792
PropRacCatComp.1	−	−	−	−	−0.020	0.053	−0.384	0.702
PropRacCatComp.2	−	−	−	−	0.044	0.037	1.169	0.244
PropRacCatComp.3	−	−	−	−	−0.028	0.055	−0.504	0.615
Logit V	−	−	−	−	0.133	0.114	1.171	0.243
AIC	510.4				524.8
Adjusted R2	0.2757				0.2581
p-val. of F statistic	6.48E-14				1.05E-09

Once a model was fit, we evaluated how well the fitted values approximate the known outcomes. For all models, AIC was calculated at each step, and AIC is shown for all models. The models for $Logitr$ are linear models. For these, a standard measure of goodness of fit is the adjusted R².

The outcome defined in Equation 2 is binary. For a binary outcome, R², adjusted or not, is not an appropriate measure for goodness of fit. One measure that is suited to binary outcomes and is often used is Tjur’s coefficient of discrimination (COD). This uses the fact that observations can be divided into two groups: those for which the outcome $Y$ is 0, and those for which $Y$ is 1. When modeling a binary outcome, we are estimating the probability that the outcome is 1. That is, $Y_i$ is an estimate of $P\left(Y_i=1\right)$. When $Y=1$, $\widehat{Y}$ should be close to 1; when $Y=0$, $\widehat{Y}$ should be close to 0. Tjur’s COD is the difference of the means of fitted values for the two groups:

$$R_{Tjur}^2=\frac{1}{n_1}\sum \widehat{\pi }(y=1)-\frac{1}{n_0}\sum \widehat{\pi }(y=0)$$ (6)

Local modeling

Global regression models fail to account for spatial nonstationarity (Brunsdon, Fotheringham, and Charlton 1996; Siordia, Saenz, and Tom 2012; Kwan 2021), whereby place-based health relationships vary across geographic space. In the current study, we argue that exposure to environmental contaminants is largely influenced by spatially varying wind patterns, and that the relationship between windborne metals and population-level exposure is dependent on proximity to mining activities – the distribution of which is heterogeneous.

We tested four commonly used modeling approaches to account for spatial relationships between the predictor variables and the outcome, $Logitr$: two spatial autoregressive (SAR) models – a spatial lag model and spatial error model – a geographically weighted regression (GWR), and a multiscale GWR (MGWR) of the reduced models described in Table 5. Spatial lag models violate the assumption that error terms are independent and introduces a lag coefficient ($\rho$) to the model and spatial error models violate the assumptions that both error and observations are independent by introducing a spatial error term ($\lambda$) to account for spatial diffusion between variables. GWR is a form of linear regression that models relationships between all points within a local neighborhood for each observation in the full set (Fotheringham 2011). MGWR models differ from GWR by selecting a unique neighborhood size for each predictor variable – this is particularly useful when the spatial influence of predictor variables is nonstationary.

Table 5.

Local model results

Model	R2	AIC	Effective DF	Test for spatial dependence	Stat.	p
Spatial error	0.25	184.171	197	Likelihood ratio	3.302	0.069
Lag coefficient ($\lambda$)		0.173
Std. Error		0.085
z-value		2.04
Prob		0.041*
Spatial lag	0.25	185.686	196	Likelihood ratio	3.787	0.052
Lag coefficient ($\rho$)		0.145
Std. Error		0.075
z-value		1.937
Prob		0.052
GWR	0.423	166.221	176.838	Moran’s I	−0.002	0.92
				Anselin Local	See Figure 2
				Moran’s I
Bandwidth [fixed]		109
MGWR	0.88	656.125	63.534	Moran’s I	0.041	0.407
				Anselin Local	See Figure 2
				Moran’s I
Bandwidth [TWE]		201
Bandwidth [WindIndex]		30
Bandwidth [PropRacCatComp.4]		9
Bandwidth [FMM]		12

We calculated the spatial autocorrelation of model residuals using the Moran’s I global test of spatial autocorrelation, which is defined as the correlation of a variable with itself across two-dimensional space (Barber 1988). If model residuals are spatially autocorrelated, then we conclude that the global model is poorly specified in that it fails to account for spatial dependence between variables. Spatial dependence may stem from measurement and approximation error, or from phenomena where the spatial or directional (anisotropic) nature of a characteristic are important defining features of that characteristic.

Results

Global models

Logistic model for binary outcome $Y$

Table 3 lists the variables used to predict any reported deaths due to COVID-19 after variable selection. All covariates representing soil metal concentrations were automatically removed in this step. The model described in Table 3 seems quite complex. However, many of the covariates were not statistically significant (p > 0.05). Therefore, their actual effect on the outcome is uncertain. According to the model, $Logpop$ and $logarea$ have positive coefficients with low p-values. So, $P$ (COVID-19-deaths) increases with population and area of ZCTA. The coefficient of $WindIndex$ is negative, so, increasing exposure to environmental contamination via wind tends to decrease the probability that there will be COVID-19 deaths. The variable $fmm$ has a negative coefficient. So, zip codes that are more rural and remote tend to be more likely to have zero deaths reported as being due to COVID-19. Finally, principal components 1–3 of the PCA of Table 1 all appear in the model with small p-values. PCs 1 and 3 have positive coefficients, PC 2 has a negative coefficient. However, the interaction between PC 2 and $logpop$ has a positive coefficient. Additionally, we calculated the false discovery rate (FDR) to account for the expected proportion of false positives among the declared significant results. The FDR shows that the significant findings ($logarea$, $AceRacCatComp.1$, $AgeRacCatComp.3$, and $fmm$) are likely true effects rather than artifacts of multiple comparisons.

PC 1 is high when the proportion of Native Americans in the population and the proportion living in poverty are high, but the proportion of non-Hispanic Whites is low. PC 3 is dominated by the proportion over 65; the PC is high when this proportion is high, but to a lesser extent when proportion living in poverty is high. PC 2 is high when the proportion of the population that is Hispanic is high, and the proportions of non-Hispanic whites or Native Americans are low. So, the probability of having any deaths recorded as due to COVID-19 is higher for zip codes that have a large proportion of population that is Native American, a small proportion that is non-Hispanic white, and/or a large proportion of population living in poverty, a large proportion of population over 65, a small proportion that is non-Hispanic white, and/or a large proportion of population living in poverty, or a large proportion of population that is Hispanic, and small proportions of population that were non-Hispanic white or Native American. However, this is true only if the total population is fairly large. A zip code with a largely Hispanic population, but a small total population, is likely to show zero deaths due to COVID-19.

Linear model for logit r

Now consider the model for $logitr$ (Table 4). Similarly, all soil metal concentration covariates were removed following variable selection in this model, which only has four covariates: $fmm$, $fWindIndex$, $fTWE$, and the fourth component of the sociodemographic PCA described in Table 1. This fourth PC is dominated by the proportion in poverty. It is positive for the non-Hispanic white racial proportion, negative for other racial proportions. The coefficient of this PC in the model is positive. The model says that ZCTAs that tend to be downwind of environmental contamination tend to have lower rates of COVID-19 deaths per number of tests, ZCTAs that are more remote tend to have higher rates of COVID-19 deaths per number of tests, ZCTAs with higher proportions of the population living in poverty tend to have higher rates of COVID-19 deaths, and that the effect in item 3 is stronger for ZCTAs with higher proportions of non-Hispanic white, and not as strong for ZCTAs with higher proportions of Hispanics or Native Americans.

The numbers of vaccinated individuals per zip code were estimated from the numbers of vaccinated per county as of 3/3/2021. From this we can find the vaccination rate:

$$V=\frac{\text{\#}vaccinated}{estimatedtotalpopulaitonofZCTA}$$ (7)

and calculate $logitV$. There was one ZCTA for which the estimated number of vaccinated was larger than the estimated population. This is likely because of a centralized vaccination clinic serving a large rural Tribal population. The single dose vaccination rate for the county in which the ZCTA is located is 95%. To deal with this case, $V$ is assumed to be 0.95 for this ZCTA.

Models for $logitr$ were fit in which $logitV$ was included as a potential predictor. When model selection was applied using backwards stepwise regression, $logitV$ was always removed. This included interactions in models that allowed interactions. So, there did not seem to be a significant association between the rate of deaths attributed to COVID-19 ($r$) and the vaccination rate $V$, as of March 31, 2021.

Additionally, we considered the effect of confounders on the reduced linear model by including results from the full model in Table 4. The reduced model, after stepwise variable selection, appears to perform similarly to the full model in terms of explaining the variance in the outcome (as indicated by adjusted R²) but with fewer predictors and a lower AIC. This suggests that the reduced model provides a more parsimonious representation of the relationship between the predictors and the outcome. In terms of adjusting for confounders, both models seem to appropriately account for the significant predictors. The reduced model retains the significant variables identified in the full model, ensuring that key confounders are still adjusted for while eliminating unnecessary predictors.

Local models

Before conducting local regression models, we determine appropriateness by testing the spatial autocorrelation of global model residuals using the Moran’s $I$ test of spatial autocorrelation. If p-values from this test are significant (e.g., <0.05) we conclude that model residuals are spatially random, indicating no significant structure. On the other hand, if p-values are significant, we conclude that underlying spatial relationships are not adequately explained by the global modeling approach. Results indicate that the residuals of the binary model for $Y$ are spatially random (I = −0.018, z-score = −0.608, p = 0.543), indicating the global model was correctly specified. Conversely, the residuals of the linear model for $logitr$ were significantly spatially autocorrelated (I = 0.079, z-score = 2.599, p = 0.009). Additionally, we measured the degree to which residuals from the GWR and MGWR models cluster using the Anselin Local Moran’s I test of spatial association (Anselin 1995). If geographical features demonstrate significantly similar values to other features within a spatial neighborhood, they are clusters of either high or low values. Conversely, if significantly dissimilar values are either high or low outliers. Outliers or clusters of residuals suggest some structure in spatial relationships that are not fully accounted for by local models. Figure 2 shows that some ZCTAs demonstrate clustering and outlier relationships in both the GWR and MGWR models (p < 0.05) in some regions of the study area. These clusters and outliers must be interpreted alongside the spatial distribution of other significant model terms. Therefore, all local models test the outcome, $logitr$, using the same predictor variables described in Table 4: $fmm$, $WindIndex$, $TWE$, and $PropRacCatComp.4$ on 202 observations. A test for the variance inflation factor (VIF) in the linear model indicates that there is no significant multicollinearity between these model terms.

Figure 2.

Local clustering of model residuals in MGWR (left) and GWR (right).

We conducted four spatial regression models to account for spatial dependencies between observations in the linear model (Table 5). The first two spatial autoregressive (SAR) models add an additional term to account for spatial dependencies between terms and the outcome. First, a spatial error model was conducted in GeoDa (version 1.20.0.36). Results suggest that the global model explains only slightly more variability than the global linear model, but a likelihood ratio test for spatial dependence indicates that the spatial effects are accounted for. Second, a spatial lag model was also conducted using GeoDa. While spatial dependence in the model was likely the result of more than measurement error, the results did not indicate significantly improved model performance from the spatial lag model.

Third, because neither $\rho$ nor $\lambda$ were significant at p < 0.05 in either SAR model (Table 5) and because model strength was reduced compared to the global model, a GWR was used to account for spatial dependence between variables using the GWModel package (version2.3–1) in R (version 4.2.2). We defined neighborhood size by optimizing the AIC by the number of neighbors from each ZCTA. Neighbors are defined using edge and corner continuity between ZCTA polygons and spatial weights are assigned using a bisquare kernel. The results indicate significant model improvement in terms of variation explained and a lower overall AIC. Moreover, the Moran’s $I$ test demonstrates that residuals are randomly distributed ($I=-0.002$, p = 0.920), suggesting adequate model specification in terms of underlying spatial relationships.

Fourth, while the GWR appears to be well-specified in terms of spatial autocorrelation of residuals, R², and AIC (Table 5), the wind index coefficient demonstrates unexpected results in that it is negative for every observation. Suspecting that the spatial influence of each variable varies, we applied an MGWR model in R to account for differing degrees of influence within the model. Neighborhood sizes for each explanatory variable were determined by backfitting the changing values of the residual sum of squares over 600 iterations. Residuals in this model are similarly randomly distributed (I = 0.041, p = 0.407). The MGWR explains slightly more overall variance than the GWR, and despite a higher AIC the model effectively captures spatial heterogeneity of model coefficients. Table 6 and Figure 3 provide the distribution of model coefficients. Because the MGWR model calculates a linear relationship between the outcome and predictor variables at every observation, the significance of relationships is also spatially heterogeneous. We see that the coefficient for the wind index is negative in large portions of the study area, but positive and significant in regions heavily impacted by mining activity (Figure 4).

Table 6.

Summary statistics of MGWR model coefficients.

	Min.	1st Qu.	Median	3rd Qu.	Max.
Intercept	−3.21	−2.14	−1.82	−1.62	−0.09
TWE	1.08	1.10	1.11	1.11	1.12
WindIndex	−3.63	−0.91	−0.58	0.48	1.20
PropRacCatComp.4	0.39	0.44	0.70	0.71	0.77
FMM	−0.31	−0.11	−0.07	0.00	0.19

Figure 3.

Boxplots showing distribution of MGWR coefficients.

Figure 4.

Local wind index coefficients from MGWR model.

Discussion

The logistic model for the binary outcome, $Y$, does not provide information about the impact of COVID-19 but rather evaluates the completeness of reporting. This model identified four ZCTAs where the predicted number of COVID-19-related deaths is at least 1 but where the reported number of deaths is 0. Each of these ZCTAs is highly rural with a population of less than 600. We cannot rule out inaccurate reporting and conclude that these data have been suppressed. Therefore, the logistic model suggests that rural and remote portions of New Mexico are somewhat more likely to have reported 0 COVID-19-related deaths than more urbanized ones. However, we also see that the size of zip codes is positively associated with increased mortality. On the surface these findings may appear to be contradictory, but we do also see a positive relationship with sociodemographic PCs that are dominated by elderly Native American populations, who were among some of the hardest hit in the early stages of the pandemic (Hatcher et al. 2020; Rodriguez-Lonebear et al. 2020; Sequist 2020; Gorris et al. 2021).

Soil metal concentrations were automatically removed in both the logistic model and the linear model for $logitr$, the relative number COVID-19-related deaths to tests per ZCTA, after stepwise model selection. However, both models do include potential sociodemographic and environmental drivers of disease mortality. Despite this, we see negative relationships between potential windblown exposure to fugitive waste from resource extraction sites in both models. Beyond this negative association, the linear model for $logitr$ similarly indicates positive associations between COVID-19-related deaths and rural or remote zip codes, as well as poverty – this finding is consistent with the literature (Hirko et al. 2020; Sequist 2020; Gorris et al. 2021). Conversely, stepwise model selection consistently removed the estimated ZCTA-level vaccination rate, $logitV$, despite meta-analysis showing significant positive odds ratios of pooled vaccine effectiveness against incidence rate, hospitalization, and mortality after both first and second doses (Rahmani et al. 2022).

Overall, results from the MGWR model are different from the linear model for $logitr$. To wit, while sociodemographic characteristics captured by $PropRacCatComp.4$ are consistently positive, $FMM$ coefficients are generally negative. Environmental transport of dust captured by $TWE$ and $WindIndex$ both have positive coefficients in the MGWR but are negative in the global linear model. We are primarily interested in the distribution of $WindIndex$ coefficients, of which a large proportion are positive and negative. Furthermore, by calculating a t-statistic at each observation we can map which ones are statistically significant (Figure 4) and demonstrate that the model is significant for $WindIndex$ coefficients in both directions.

The spatial heterogeneity of $WindIndex$ coefficients is likely driven by the physical geography of New Mexico. The state is bisected by the Rio Grande (Figure 1), and the western portion has more variable mountainous terrain than the east. This mountainous terrain has been the setting for decades of hard rock mining that frequently occurred on lands largely inhabited by Native American populations (Hoover et al. 2020). The southeastern plains of New Mexico sit on the Permian Basin, a geologic formation with significant petroleum reserves (Frenzel et al. 1988). This topographic and industrial variability is highlighted by the $WindIndex$ variable, which explicitly considers terrain variability in chorus with proximity to and density of point sources of environmental exposures, be those hard rock mines or oil and natural gas wells.

We determined the suitability of the MGWR model by its ability to effectively capture spatial heterogeneity of environmental coefficients. MGWR is an ideal modeling approach when the scale of influence of predictor variables varies greatly, as is the case here. By not forcing a predetermined neighborhood on all variables a priori, as GWR does, MGWR consumes fewer degrees of freedom by allowing some modeled relationships to be global (Oshan, Smith, and Fotheringham 2020). This is the case with TWE in the present model, which has a neighborhood of 201 observations.

Conclusion

The present analysis demonstrates complex associations between socio-environmental predictors and large proportions of COVID-19-related deaths in New Mexico during the first year of the pandemic. These associations are spatially nonstationary, meaning that potential exposures vary across space and geographic context. Specifically, the wind index, which captures the redistribution of toxic metal-bearing fugitive dust from hard rock mines and fracking wells is both positively and negatively correlated with COVID-19-related mortality in different regions across New Mexico. Regression models that attend to spatial nonstationarity by examining local relationships within a larger study area have the potential to enhance health-focused research and are highly applicable in contexts with complex geographical properties.

We show that New Mexican populations living both in closer proximity to extractive industries and in complex mountainous terrain are more likely to have experienced increased incidences of COVID-19-related deaths in the first year of the pandemic. New Mexico is a highly rural state with large Tribal populations segregated into disproportionately underserved and mostly rural regions (Farrigan and Parker 2012; Yellow Horse, Deschine Parkhurst, and Huyser 2020). Disproportionate environmental and social burdens experienced by Indigenous communities across the state are deeply entrenched and rendered these populations particularly vulnerable to adverse effects of the pandemic (Yellow Horse, Deschine Parkhurst, and Huyser 2020).

Our models do not contradict numerous findings showing the relationship between social vulnerability and COVID-19-related hardships across the U.S. However, the local effect of environmental metals exposures was almost impossible to capture using traditional global modeling techniques. For example, using MGWR we demonstrate that the effect of potential windborne exposures to metal-bearing dusts is only significant and positive in pockets across the state – local variability in this relationship is so pronounced that the overall effect is negative in global models. Local spatial relationships between sociodemographic disparities and disproportionate environmental exposures are masked in global modeling approaches in what is a clear example of Simpson’s paradox (Sachdeva and Fotheringham 2023). Specifically, we see coefficients change direction primarily in the wind index variable when aggregated by ZCTA.

Limitations

There are several limitations of the current study. First, nearly all health data are protected by an array of laws and regulations at multiple scales of authority – including state, federal, and institutional oversight. One of the most common methods for protecting individual privacy is to group data into spatial regions – typically this occurs along census designations like tracts or – as in the current analysis – ZCTA. Because these spatial configurations vary in both scale and zoning patterns, particular and challenging methodological problems characterize their use. One of the most persistent is the modifiable areal units problem (MAUP), which refers to the sensitivity in model calibration introduced by aggregating data into areal units (Openshaw 1984; Fotheringham and Wong 1991). Failure to account for and understand the extent of impact from the MAUP on analyses can have profound implications on both environmental and health research and policy.

We also faced challenges surrounding data quality that cannot be fully resolved. For example, fracking locations are voluntarily disclosed by well operators. The Shale Gas Production Subcommittee of the Congressional Secretary of Energy Advisory Board (SEAB) recommends public disclosure of fracking activity but there is currently no law or regulation specifically requiring such disclosures. All disclosures in New Mexico were freely obtained from the FracFocus Chemical Disclosure Registry (FracFocus 2023) dating back to January 1, 2011. All wells, regardless of current production status, were included in the present analysis. Similarly, the point locations of all mine features derived from the MRDS is a compendium of mine records compiled over a multidecadal period. Each record has varying degrees of completeness, and it is difficult to ascertain mine size, production volume, or reclamation status.

Finally, measurements of metal concentrations in soil and sediment are provided by the NURE program, dating back to the 1960s. While the documentation of the laboratory methods is thorough and we are able to clean the data effectively, we remain uncertain about any biogeophysical and anthropogenic processes that have contributed to the resuspension or change in geographical extent of metals in the earth’s crust since samples were collected.

Table 2.

Numbers of observations used in different models, by outcomes and sets of covariates.

Outcome	$Logitr$	$Y$
Type	continuous	binary
Model type	Linear	logistic
	# Observations used
Basic covariates	202	333
+ PCs for four metals	104	191
+ PCs for five metals	83	139

Highlights.

• The association between COVID-19-related mortality and socio-environmental exposures is spatially heterogeneous

• Underserved communities and racial minorities tend to experience a higher burden of environmental exposure, rendering them more vulnerable to COVID-19-related mortality.

• Spatial relationships between COVID-19-related mortality and its environmental and sociodemographic drivers are masked by global modeling approaches.

Data Statement:

Data are from the NMDOH and are managed under a use agreement between NMDOH and UNM-HSC. Data will be made available to interested researchers pending approval for dissemination of sensitive data by both the University of New Mexico Sponsored Projects Office and the New Mexico Department of Health.