Trade, uneven development and people in motion: Used territories and the initial spread of COVID-19 in Mesoamerica and the Caribbean.

Mesoamerica and the Caribbean form a region comprised by middle- and low-income countries affected by the COVID-19 pandemic differently. Here, we ask whether the spread of COVID-19, measured using early epidemic growth rates (r), reproduction numbers (R t ), accumulated cases, and deaths, is influenced by how the 'used territories' across the regions have been differently shaped by uneven development, human movement and trade differences. Using an econometric approach, we found that trade openness increased cases and deaths, while the number of international cities connected at main airports increased r, cases and deaths. Similarly, increases in concentration of imports, a sign of uneven development, coincided with increases in early epidemic growth and deaths. These results suggest that countries whose used territory was defined by a less uneven development were less likely to show exacerbated COVID-19 patterns of transmission. Health outcomes were worst in more trade-dependent countries, even after controlling for the impact of transmission prevention and mitigation policies, highlighting how structural effects of economic integration in used territories were associated with the initial COVID-19 spread in Mesoamerica and the Caribbean.

Introduction

COVID-19 is a pandemic disease caused by SARS-CoV-2, a novel coronavirus of zoonotic origin [1]. In September 2020, half a year into the pandemic, as many COVID-19 cases had been reported in the Western Hemisphere as in the rest of the planet [2], with the United States reporting the largest number of deaths globally, around 200,000 [3]. Patterns of high transmission were similar to those previously observed in Europe and East Asia after the first global wave of transmission in early 2020 [3,4]. For COVID-19, a previous analysis, restricted to Costa Rica, Panamá and Uruguay, suggested that early epidemic growth may plateau faster in more equitable societies [5]. Thus, transmission patterns may vary across space, associated with a region's uneven development in interrelated social, economic, environmental, and health realms. This uneven development is partially reflected in differences in socio-economic indicators among the countries and territories of the region. Indeed, income inequality has been widely recognized as a major driving force of adverse health outcomes [6]. Associated mechanisms of action include the reduced funding of public services [7], such as health care and education. This, in turn, can lead to the erosion of social organizations (which can offer mutual benefit via social cooperation) and of “social capital” [7]. High levels of social cooperation and solidarity can lead to the participation of diverse, often excluded, groups in inclusive policy and decision making [8]; more vulnerable groups are more likely to actively engage in health and social campaigns for health promotion when their views are considered [9] and those views are more likely to be noticed when groups are politically organized [10].

However, beyond income inequality and the erosion of social cooperation, solidarity and political engagement, one of the largest problems that societies face is lack of equitable access to essential resources for the human development of those societies or for their social, cultural and biological well-being [11], despite the abundance of resources that could be leveraged toward such ends. To help understand these contradictions, many researchers have pointed to trade between the global ‘core’ and the ‘periphery’, as explored in the theory of uneven development offered by economist Samir Amin [12] among others [13]. In his research, Amin proposed that global capitalism develops and divides the territories of the world into a ‘core’, composed of highly developed and wealthy countries, and a ‘periphery’, composed of developing (or “underdeveloped” in Amin's own words) countries. The periphery adjusts structurally to satisfy resource demands by the core. ‘Unequal exchange’ arises between core and periphery (contra to mainstream accounts of trade, in which trade is theorized to be both equal and mutually beneficial [14,15]) with the end result being that needs of the core countries are met, but those of the periphery are not [12]. Developing, i.e., periphery, countries are periodically coerced and their policies distorted not only by core countries and transnational corporations but also by supranational institutions, such as the World Trade Organization, the World Bank, the International Monetary Fund, and the global financial system generally, among others [16]. These dynamics lead to the transfer of wealth that could be put to greater social good from the periphery to the center, and a bias toward policies that suit the economic and political interests of the center at the cost of undermining human development in the periphery [12,16]. There are positive feedback loops, akin to a Sisyphean punishment, where no matter how much effort is made to reach higher levels of development, developing countries are hindered in their struggle to reach satisfactory levels of human development in an increasingly globalized world where territorial boundaries no longer separate the reach of what, post-independence, were intended to be sovereign countries [16]. In the end, pervasive power imbalances drive global crises, both human and more broadly environmental threatening life on our planet [16].

Both a cause and consequence of the unequal exchange between the center and periphery, human movement has been recognized as connected to disease spread through history [17,18]. Recent research trends in disease ecology, including analyses about COVID-19, have suggested that early transmission and epidemic growth of diseases might be modulated by human movement. Efforts to measure human movement patterns across spatial scales are generally based on individualized data collected from cellphones and social network usage or travel statistics through air, land or maritime terminals [19]. However, these efforts have tended to alienate the phenomenon of human movement from deeper regularities driven by structural causes, thus reducing our capacity to fully understand the underpinnings of novel disease emergence and spread.

In response to such static treatments of territories as merely stages for human movement, the geographer Milton Santos [20] developed the concept of “used territory,” which sees territories as dynamic entities formed by historical processes that serve as material and social bases for human actions. “Used territories” widen the scope of variables and parameters that can be evaluated to study and understand the emergence of novel diseases and can serve as the functional basis for the analysis of pandemic disease spread.

Building on concepts from the social sciences, including notions of “human development,” “uneven development” and “used territory” as articulated within and across economics, sociology, and geography, we evaluate the impact that development, trade and travel indicators have in driving COVID-19 epidemiological parameters in each of the studied countries and territories. Our underlying hypothesis is that variables that measure concepts from theories developed outside the natural and medical sciences influence pandemic disease spread. We expect that this influence can be examined using quantitative methods, yielding inferences that improve our understanding of disease spread. This innovative approach seeks to understand the emergence and spread of diseases as the expression of coupled social and natural systems. The epidemiological parameters include the cumulative number of cases and deaths 50, 75 and 100 days after the introduction of the disease in each of the studied territories as well as from standard mathematical epidemiological models used to describe early epidemic growth [21] and the time-varying reproduction number [22], i.e., the number of secondary infections expected from a primary infection [23]. We focus our analysis in Mesoamerica and the Caribbean by virtue of it being a region that is horizontally integrated, i.e., a region composed by neighboring territories [20], yet diverse in terms of economic and social development.

Materials and methods

Data

We analyzed epidemiological data and data from variables associated with human development, uneven development, and used territory, considering up to 16 epidemiological parameters and 28 parameters measuring aspects of human development, uneven development and used territory data for the 22 territories analyzed. In our analysis, we included the following territories: Haiti, Antigua and Barbuda, Bahamas, Barbados, Dominica, Grenada, Jamaica, St. Kitts and Nevis, St. Lucia, St. Vincent and the Grenadines, Trinidad and Tobago, Cuba, Puerto Rico and the Dominican Republic, located in the Caribbean; and Belize, Costa Rica, El Salvador, Guatemala, Honduras, Nicaragua and Panamá located in Mesoamerica. We also included data from Uruguay to serve as an outgroup, as defined in evolutionary biology, in the different analyses we performed. In evolutionary biology, an outgroup is an element that is distantly related to the group being studied [24]. In that sense, Uruguay is spatially distant from Mesoamerica and the Caribbean. However, Uruguay shares some similarities with the region with respect to population size, area, and development. We excluded other territories, mainly Caribbean islands that are overseas territories of European countries or the United States and Caribbean states that are part of continental South America.

The epidemiological and social science theory parameters are, respectively, described in the following subsections: “epidemiological data” as well as “human development, uneven development, and used territory data.”

All raw data employed in this study are available in supplementary online material Supplement S1, while disease count data sources are described in Table S1, with additional data sources in Table S2. For all non-COVID-19 variables, we used estimates for 2019 or the most recent year when the estimates were available; when there was one missing observation, we imputed the missing value using the regional average for either Mesoamerica or the Caribbean. Table 1 shows all the variables and parameters collected and abbreviations used throughout the rest of this article.

Table 1

Abbreviations for parameters and the corresponding variables in development, travel and trade used in this study. Parameters and variables are for the studied territories.

Abbreviation	Parameter or variable
r	Exponential growth rate
m	Exponential growth deceleration
A	Intercept for epidemic growth
Rt[day]	Reproductive number n days after the detection of the first case
C[day]	COVID-19 Cases n days after the detection of the first case
D[day]	COVID-19 Deaths n days after the detection of the first case
ASK	Area in square kilometers
POP	Territory population size estimate for 2020
PSK	Population per square kilometer
PP0	No. of prevention policies before the first case detection
PP25	No. of prevention policies 25 days after the first case detection
MP0	No. of mitigation and suppression policies before the first case detection
MP25	No. of mitigation and suppression policies 25 days after the first case detection
HDI	Human development index
IHDI	Inequality adjusted human development index
%HD	Percentage reduction in HD as function of inequality
UHC	WHO Universal Health Care index score
GDP	Gross domestic product per capita in US$
GDP2	Gross domestic product per capita in purchasing power parity
AUS	Aid flow from the United States of America in US$
AUK	Aid flow from the United Kingdom in US$
CN	Number of International territories connected through the main airport
CC	Number of International cities connected through the main airport
PA	Total number of passengers traveling through the largest airport
TO	Trade openness overall (in both goods and services)
TM	Trade openness (in merchandise alone)
FI	Inward flow of foreign direct investment
FO	Outward flow of foreign direct investment
SI	Inward stock of foreign direct investment
SO	Outward stock of foreign direct investment
EX	Number of export types
IM	Number of import types
CE	Concentration of exports index
CI	Concentration of imports index

Epidemiological data

Daily data of COVID-19 cases confirmed by molecular diagnostics were collected from various sources that included each territory's ministries or secretaries of health (Table S1). We compiled a 100-day-long time series for each territory starting the day the first case was detected in each (Fig. 1 ).

Fig. 1

Cumulative COVID-19 cases in countries (A) Mesoamerica and Uruguay, and (B) the Caribbean. The Y axis is on a log scale. The inset legends show the name of the countries and their respective ISO 3166-1 alpha-2 two letter code.

We used the disease data to estimate disease transmission parameters related to early epidemic growth and the time-varying reproduction number (R ), as described below.

Early epidemic growth parameter estimation using a general growth curve

We compared early COVID-19 epidemic growth across Mesoamerica and the Caribbean using a generalized-growth model to characterize outbreaks [21] based on the following ordinary differential equation:

C(t) is the cumulative number of cases at time t after the first case detection, r is a parameter for case growth rate, and assumed to be positive, i.e., r > 0, and p is a non-negative parameter related to the case growth rate. When p = 0, cases increase at constant value; when p = 1 case growth is a perfect exponential growth curve; and when 0 < p < 1, the model has sub-exponential growth. The equation presented in (1) has the following solution [21]:where A depends on the initial condition C(0) and m > 1 is related to p as follows:

The parameter m is expected to decrease as COVID-19 cases decreased in response to transmission reduction policies, while parameters r and A likely reflect structural differences across health systems in Mesoamerica and the Caribbean. We fitted the model presented in (2) by maximum likelihood estimation. We used the mle2 command from the R package bbmle [25]. We fitted the model using a Poisson maximum likelihood function, where the initial parameter search was done using a Genetic Algorithm [26]. Then a local optimization was performed using the Nelder-Mead algorithm [27]. We estimated confidence intervals by profiling the likelihood for each parameter [25]. Estimated parameters for each country are presented in supplementary online Table S3.

Rt estimation based on the serial interval

We examined the time-varying reproduction number (R ), which, similar to R , measures the expected number of new cases generated by a given case, but does not assume that the whole population is susceptible to the disease. To estimate R given the above, the only additional information necessary is the serial interval of the pathogen under study, which is the time distribution between symptom onset in an initial case and symptom onset in secondary cases linked by contact tracing to that initial case [22]. We based our serial interval on an estimate from China [28], which was based on the largest sample size available for our period of study and assumes a 10% pre-symptomatic transmission rate before the originating infected person shows any symptoms [28]. To account for potential uncertainties, we took full advantage of the parameter estimates presented by Du et al. [28].These estimates allowed us to set the 95% credible intervals from their estimates as the minimum and maximum values for the mean and SD of the serial interval, while also setting a standard deviation of 0.2 [5] and employing the estimation method developed by Thompson et al. [22]. We reduced potential biases in the estimation of R by considering imported cases as potential sources of new local cases, but not as result of local transmission [22]. We estimated R at 25, 50, 75 and 100 days after detecting the first case, using the EpiEstim R package [22]. As noted, we chose this method given that its estimation is just based on knowing the serial interval of the pathogen [22]. Our approach thus differs from methods that have more assumptions and require additional parameters for the estimation of R [29,30]. R estimates are presented in the online Supplementary Table S4.

Data on COVID-19 prevention and mitigation policies

We also collected data about policies directed to prevent, mitigate and suppress SARS-CoV-2 transmission during the study period, as these might have created heterogeneities in transmission patterns. For this, we used the coronavirus government response tracker developed by the University of Oxford (Table S2) to count the number of policies in place before the detection of the first COVID-19 case and 25 days after the first case was detected in each territory. We specifically included the following prevention policies: the presence of public information campaigns, a defined testing protocol and a contact tracing protocol. To account for mitigation and suppression, we included the following: school closing, workplace closing, cancelation of public events, restrictions on gatherings, public transport closing, stay-at-home requirements, restrictions on internal movement and international travel controls. Fig. 2 shows the timeline for the different policies and the detection of the first case across the studied territories.

Fig. 2

COVID-19 timeline in Mesoamerica the Caribbean and Uruguay. In the plot the date for the first case and the implementation of the different prevention, mitigation and suppression policies is indicated by symbols along time (x axis). Supplementary online Fig. S1 shows a magnified version of the timeline during March 2020.

Human development, uneven development, and used territory data

Ancillary geographical data relevant to all tested social science theories

We included population size estimates to account for heterogeneities that might emerge from different territories having differences in population size and density per area unit. Territory surface areas (in km2) were calculated from Natural Earth 1:50 m scale Admin 0 geometries (Table S2) and used as denominator to estimate population density by km2 for each territory.

Human development data

To measure human development, we employed the 2018 Human Development Index (HDI) and the Inequality-adjusted Human Development Index (IHDI) from the United Nations Development Programme (UNDP, Table S2). The HDI is a “summary measure of average achievement in key dimensions of human development: a long and healthy life, being knowledgeable and have a decent standard of living” (Table S2). The HDI is then corrected by discounting a percent representing inequality in the components (%HD) of human development, thus producing the IHDI (Table S2). We also included the Universal Health Coverage (UHC) service index, a metric developed by the World Health Organization (WHO) to measure access to “health services (including prevention, promotion, treatment, rehabilitation and palliation) of sufficient quality to be effective while also ensuring that the use of these services does not expose the user to financial hardship” (Table S2). UHC values range between 0 and 100 (Table S2). These variables quantify the role of human development in the initial spread and growth of COVID-19 across Mesoamerica and the Caribbean.

Uneven development data

We compiled a series of economic indicators developed by the World Bank including the gross domestic product (GDP) per capita, i.e., the average net economic value added by each resident in a given territory, to measure the role of “uneven development” in the initial spread of COVID-19 across Mesoamerica and the Caribbean. The economic indicators were measured in US$ and corrected by purchasing power parity, a normalized value to account for differences in price levels across countries (Table S2). We also included data on net bilateral aid flows from the United States of America and the United Kingdom (Table S2) into the studied territories from Mesoamerica and the Caribbean. The net bilateral aid flows measure the difference between government grants and loans and repayments of those loans between those two pairs of countries, measured in millions of current US$. We consider these aid flow variables could also provide a measure of uneven development since countries of the periphery are more likely to receive aid from wealthier countries [12,31], and to that end, selected the United States and the UK, the two main donors in the study area (Table S2).

Used territory data

To quantify the “used territory” impact in the initial spread of COVID-19 across Mesoamerica and the Caribbean, we included data on the number of connected countries, connected international cities and passengers traveling (last annual estimate available) through the main airport for each of the studied territories. These details and data sources are presented in Table S5. Other variables were related to international trade. We included merchandise trade openness, i.e., the ratio of exports plus imports in merchandise goods to that of the territory's GDP (Table S2) as well as overall trade openness, the sum of exports and imports of goods including services divided by a country's GDP from the World Bank (Table S2). We also included additional variables on the composition and diversity of trade from the United Nations Conference on Trade and Development (UNCTAD, Table S2). From the “Merchandise: Product concentration and diversification indices of exports and imports, annual,” table, we included the number of products imported and exported at the three-digit SITC, Rev. 3 level. We also included countries' trade concentration indices (normalized Herfindahl-Hirschman indices). For the latter, values range between 0 and 1, where high values, i.e., close to 1, indicate that imports and/or exports are concentrated in a few products (Table S2). Finally, we included UNCTAD variables related to foreign direct investments (from the table “Foreign direct investment: Inward and outward flows and stock, annual,” Table S2), which are standardized as percentages of GDP. These measure the investment of enterprises from a different country within a target country (in the context of our study, the territories in Mesoamerica and the Caribbean) and can be measured as flows or stocks, which, respectively, measure transactions and accumulated values held at the end of an annual period. For the analysis, we employed the inward metric as it measures the impact of foreign investments within the territories of Mesoamerica and the Caribbean (Table S2). Trade and foreign direct investment data are key to understanding the role of the “used territory” as defined by Santos [20], since these variables relate to the structural underpinnings for the movement of human beings [16]. We hypothesize these variables might play a key role beyond the one directly attributable to human movement on the spread of COVID-19. We note that trade relations are understood within several schools of thought to play integral roles within processes of uneven and under-development and the constitution of the used territory. Here, we group them with the latter for analytical convenience but discuss their intrinsic connections in the discussion.

Cluster analysis

To characterize differences and similarities among the studied countries, we performed a cluster analysis. In the context of this study, cluster analysis finds potential groups of countries with similar transmission patterns or social science parameters. Briefly, a cluster analysis is a technique that looks at the similarities in the characteristics of objects to separate them in groups, with objects being more similar to members of their own cluster than to members of a different cluster. We specifically did an agglomerative nesting hierarchical cluster analysis [32] of both epidemiological variables and variables associated with uneven and human development, as well as the used territory. In this type of clustering, objects (here, countries) are systematically grouped in a bottom-up manner where more similar objects are put within the same cluster [32].

The cluster for epidemiological variables included parameters r and A estimated from the early epidemic growth for all the territories considered in this study, Cases, deaths and R at day 25, 50, 75 and 100 after detecting the first infection in each of the studied territories. The clustering for variables dealing with the used territory, human development, and uneven development included: HDI, %HD, IHDI, UHC, GDP, GDP2, AUS, AUK, CN, CC, TO, TM, IM, EX, CE and CI (for explanations of these parameters, see Table 1). We did not include PA, FI, FO, SI and SO due to these having two or more missing values. Before performing the cluster analysis, we normalized all variables by subtracting the mean and dividing by the standard deviation so distances between territories were not dominated by the variables with the largest numerical scale [32]. In the cluster analysis for disease parameters, we set unreliable R estimates (Table S4) to zero as wide confidence limits emerged given that no new cases were reported for periods equal to or longer than the serial interval for disease transmission (Fig. 1). Setting these values to zero is an appropriate choice as disease transmission was not observed in the days before the dates (t = 50, t = 75, t = 100) for which R was estimated [22]. The zero imputation was unlikely to have created artifacts that affected cluster estimation, as R values were likely near zero [33]. To represent the results from the cluster analysis, we made dendrograms. We also present the clusters in a bi-dimensional projection, based on a principal component analysis (PCA) of the variables considered in the cluster analysis [34]. To ease the interpretation of clusters we also performed a segmentation analysis of the clusters, where mean values for each parameter and each cluster were estimated [35,36]. We then made parallel coordinates plots with the standardized values, i.e., removing the mean and dividing by the standard deviation of each variable considered in the cluster analysis [37], of the estimated means for each cluster and parameter [38].

Econometric modeling

We quantified, using econometric linear models, the impact of variables measuring used territories, human development, and uneven development, and policies to control SARS-CoV-2 transmission on early transmission COVID-19 epidemiological parameters. Prior to fitting the linear models, we started with an exploratory analysis of our full dataset. We first examined the Pearson correlation [39] among the epidemiological variables to select which ones to analyze (Fig. 3 A). That preliminary analysis showed that cumulative cases and deaths at the chosen dates were highly correlated across the territories studied. Thus, we focused our analysis on cases and deaths at t = 100 days. Meanwhile, R at t = 25 was positively correlated with death and cases at all times and was the only time at which we were able to reliably estimate this parameter for all the studied countries (Table S4).

Fig. 3

Correlation plots for COVID-19 in Mesoamerica and the Caribbean (A) Cumulative Cases C, Cumulative deaths D and estimated model parameters, including R, r, m and A. For C, D and R, t = 25, 50 or 100 indicates the time of the observations (B) Used territory and human and uneven development covariates considered in the analysis. A full description of the epidemiological parameters and covariates are described in the methods sections and the abbreviations are explained in Table 1. The scale at the bottom is for the Pearson correlation coefficient, square size is likewise proportional to the correlation value.

In contrast, we did not analyze R with econometric tools. The R estimates at t = 50, t = 75 and t = 100 had between 6 and 8 unreliable estimates each (Table S4) and the correlation with other epidemiological variables suggested that their patterns were either similar to those of cases and deaths or parameters from the early epidemic growth model. We chose r as it was not associated with cases or deaths on any selected dates when examining the exponential growth parameters. Moreover, r was estimated for all territories, something not possible with m, which could not be estimated for Dominica, Dominican Republic, St. Kitts and Nevis and Uruguay. Meanwhile, the covariates (Fig. 3B) had a clustering pattern where human development variables, GDP measurements and foreign direct investment stock (inward and outward) were highly correlated.

Similarly, variables related to travel through the main airport, CC, CN, and PA, among the studied territories were highly correlated (Fig. 3B). Given that SI and SO were correlated with human development variables and that two or more observations were missing, we did not consider these variables in our analysis. Similarly, PA, FI and FO were not further considered as those variables also had two or more missing observations. We then selected which variables to consider in the initial “full models” based on the correlation of the covariates with r, Ct and Dt at t = 100.

For r, we used linear regression as the variables are assumed to have errors, denoted by , with an equal, independent and normal distribution [40]. For r, the initial model was described by the following general equation:

The model presented in (4) accounted for heterogeneities shaped by different population sizes by including weights proportional to population size in the error matrix [40]. Since cases, C, and deaths, D, at t = 100 were count variables, we employed generalized linear models for count variables, using a Poisson distribution. For C and D at t = 100, the initial model was described by the following general equation:

For models described by equations (4), (5), is an intercept; for Dev[j], we alternatively tested HD, IHD, GDP or GDP2, as all these variables were highly correlated (Fig. 3B). In these equations, travel was either CC or CN, import[j] either IM or CI and export[j] either EX, if import[j] was IM, or CE otherwise. In models for C and D, described by equation (5), the link() function is a natural logarithm. Models had population size included as an offset, indicated by offset(log(POP)), to account for heterogeneities in disease counts that emerge from the fact that cases in a given territory are constrained by the population size of that territory [41]. The resulting 16 models described by equations (4), (5) were simplified through backward elimination [42]. Backward elimination is an algorithm by which nested models (i.e., models with one parameter less than the “full model” within which they are nested) are compared through the Akaike Information Criterion (AIC) [42]. The model minimizing AIC is selected at each step until a “best model” is found which is not significantly different in its explanatory power from the “full model” but becomes significantly different from the full model if further simplified, through the removal of additional covariates. The resulting best models were then compared using the AIC, a metric that trades off model fit and the number of parameters in a regression [42]. The resulting models were then subject to model diagnostics using the variance inflation factor, VIF, and factors above a threshold value of ten (VIF>10) were removed [43]; we also removed variables that were not significant.

For each model, we inspected diagnostic plots to assess homoscedasticity (examining the residuals as a function of fitted values), normality (using a Q-Q plot) and assessing influential points on parameter estimates by plotting residuals as a function of the leverage [40].

To further ensure we did not overfit parameters, we also performed a jackknife [44], a method for parameter inference based on refitting the model leaving one observation out at a time (loo). The parameters fit at each iterative step of the jackknife are used to estimate average model parameters from all the, i.e., n, loo models. With the jackknife, bias-corrected standard errors for each parameter can be estimated using the deviation of parameter estimates from all the loo models according to the following formula equation [44]:which can then be used for inference against a normal distribution [45].

We also performed a bootstrap [40], given this is considered one of the best methods for linear model cross-validation [46]. For the linear regression presented in equation (4), we performed a nonparametric bootstrap. A nonparametric bootstrap consists of resampling residuals with replacement from a regression model, which are added to the fitted values, refitting the model and building confidence intervals based on the distribution of estimated parameters from the bootstrap replications [34]. We performed a parametric bootstrap for the count regression models represented by equation (5) [47]. The fitted values were simulated as mean values from a count distribution, Poisson or negative binomial. These simulated values were used to refit the model and build confidence intervals based on the distribution of estimated parameters. We specifically used this method for the count regressions to more accurately represent the original distribution of the count data [47], as models assumed that both coronavirus cases and deaths (equation (5)) were proportional to population size. All bootstraps were set to 10,000 replications for each one of the best models.

Finally, we also evaluated the goodness of fit of the best models by estimating a predictive coefficient of determination, PR [48] which is an R based on the reduction of the variance in model predictions with data not used for model fitting, measured as the mean squared error (MSE, the mean of the square difference between observed and predicted values), in relation to the variance (var) of the response variable, i.e., r, C and D, according to the following equation [49]:

This model performance metric is used to assess prediction, in addition to fit [48].

Results

The cluster analysis results are presented in Fig. 4 . At least one cluster, formed by Panamá and the Dominican Republic, was observed when examining the epidemiological parameters (Fig. 4A and B). The rest of the clusters had some overlapping member territories as indicated by the color-coding. Epidemiological parameters (Fig. 4A) allowed us to define three clear clusters of transmission patterns. Uruguay, as expected, served as an outgroup by not clustering with the rest of the studied territories. In the cluster analysis, Panamá and the Dominican Republic, the two countries with the highest total cases, as shown in Fig. 1, formed a cluster. Reflected by the cluster branch's height, the other two clusters showed more similarities. One of the clusters included Nicaragua plus territories that have been relatively unaffected by the coronavirus (Fig. 1), including Belize, Antigua and Barbuda, St. Lucia, St. Kitts and Nevis, St. Vincent and the Grenadines, Barbados, and Trinidad and Tobago. The third cluster included territories affected by the coronavirus reached a transmission plateau by day 100 after detecting the first COVID-19 case. This cluster included Cuba, Costa Rica, El Salvador, Haiti, Puerto Rico, Guatemala, Honduras, Bahamas, Grenada and Jamaica. Here it is key to note that these patterns seem to have arisen independently of the COVID-19 related health policies (Fig. 2) and are not an artifact of the epidemiological parameters having different numeric magnitudes, given that all variables were standardized. The goodness of fit of this cluster analysis was high, as indicated by an agglomerative coefficient of 0.79.

Fig. 4

Cluster analysis results. Dendrogram representation of the agglomerative clusters (A) Epidemiological parameters (B) Used territory and human development and uneven development parameters. Parallel coordinates plots for cluster segmentation (C) Epidemiological parameters (D) Used territory and human development and uneven development parameters. In panels A & B branches are colored to indicate clusters. In panel B Haiti, Cuba, Puerto Rico and Jamaica did not belong to any cluster but have the same color to indicate this property. In panels C & D clusters are indicated, respectively, using the same colors used to represent the clusters in panels A & B, and different type lines are used to represent Cuba, Haiti, Jamaica and Puerto Rico in panel D, which are indicated in the inset legend. The x-axis of panels C & D indicate the different parameters we used in the cluster analysis. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

The same clustering patterns were observed when the epidemiological parameter clusters were projected over a PCA-based bi-dimensional projection of the studied epidemiological variables (Fig. S2A). PCA loadings (Table S6) suggest that cases and deaths were larger in Panama and the Dominican Republic. In contrast, the other two clusters were mainly separated by differences in R and epidemic growth parameters as we discussed above. This pattern can be more easily observed in a parallel coordinates plot (Fig. 4C) which clearly shows that cluster segmentation was driven by the Dominican Republic and Panamá having a larger burden of cases and deaths through time. In contrast, Uruguay and the third cluster (including most of Mesoamerica as well as Cuba and most of the Greater Antilles territories) differed on R and the early epidemic growth parameters, which in these two clusters were larger than in the second cluster (including Belize and Nicaragua and several Caribbean island countries). The mean values for each epidemiological variable and cluster are presented in Table S7.

Meanwhile, the cluster of variables associated with human development, uneven development and used territory (Fig. 4B) showed patterns that revealed similarities and differences in compared epidemiological patterns. Four territories did not cluster with any other: Haiti (the most different from all other territories as revealed by the height of its divergence from the others), Cuba, Jamaica, and Puerto Rico. The lack of clustering partially explains the agglomerative coefficient of 0.59, which is lower than observed for the epidemiological parameters. The rest of the territories belonged to four clusters. One of these clusters, related with Puerto Rico, included Panamá and the Dominican Republic, just as observed for the epidemiological parameters. A second group was formed by territories that can be seen as more developed since it included Uruguay (one of the most developed nations in Latin America), Costa Rica, Barbados, Trinidad and Tobago, and The Bahamas. This second cluster was related to a third cluster containing the remaining Caribbean territories that we studied. A fourth cluster included the rest of the Mesoamerican territories. Inspection of the clusters over a PCA-based bi-dimensional projection of the studied variables (Fig. S2B) confirmed this pattern. Based on the PCA loadings (Table S8), the major factor separating the clusters was human development, which was high in the cluster containing Uruguay, but also for Cuba and Puerto Rico, and was smaller for the cluster of Mesoamerican territories, where human development was, nevertheless, higher than for Haiti. This pattern can be observed in the parallel coordinates plot of Fig. 4D. The lower connectivity with other places separated the Caribbean territories from the other clusters. The connectivity also appeared to be a dominant factor in separating Panama and the Dominican Republic into a different cluster (Fig. 4D). The connectivity was measured by the number of connected nations (CN) and connected international cities (CC) at the main airport of each territory. Overall, the cluster analysis is suggestive that clusters based on epidemiological patterns are similar to clusters based on variables related to development and the used territory, even if the match was not perfect. The mean values for the variable and clusters of Fig. 4D are presented in Table S9.

A different way to look at the relationship between disease and social science theory parameters is by performing an econometric analysis. We explored the sixteen potential models explaining the early epidemic growth rate r. We found all candidates for the best model (Table S10) considered concentrations of imports (CI), the number of international cities connected at the main airport (CC) and the number of prevention (PP0) and mitigation (MP0) policies already in place when the first case was detected locally. All VIF diagnostics suggested that variables were unidentifiable, implying there was no reason to worry about problems with parameter estimation due to high correlation in the predictors. The best model was not significantly different from the full model F 9,8 = 1.101, p > 0. 451, meaning fewer parameters explained as much of the variability in r as a model with more parameters. The best model (Table 2 ) showed that the early epidemic growth increased with all the variables considered (CC, CI, PP0 and MP0). Overall, this model successfully explained the sources of variability, having a coefficient of determination above 60%. Specifically, the model had an 2 = 0.61, which was reduced by one-third when using the model to predict observations not considered in the fit ( 2 = 0.40). All parameters in this model had VIF below 10 (Table S10). The inspection of model diagnostic plots suggests that there were neither outliers, significant deviations from normality, nor influential points that affected parameter estimates (Fig. S3). As expected, parameter estimates from the jackknife were similar to the estimates obtained by ordinary least squares (Table S11). Confidence intervals from the nonparametric bootstrap also suggest none of the parameters was overfitted (Table S11).

Table 2

Parameter estimates for the best linear regression model explaining the COVID-19 epidemic growth rate (r) across nations from Mesoamerica and Caribbean.

Parameter	Estimate	Std. Error	t value	Pr(>|t|)
Intercept	−1.753	0.483	−3.63	0.00206*
No. of Connected cities	0.013	0.004	3.20	0.0053*
Concentration of imports index	6.980	1.622	4.30	0.00048*
No. of prevention policies before the first case detection	0.448	0.102	4.38	0.00041*
No. of mitigation policies before the first case detection	0.108	0.037	2.89	0.0101*

*Statistically Significant (P < 0.05).

Meanwhile, the best model (Table 3 ) explaining the number of cases, C, at t = 100 days after detecting the first local case dropped a large number of variables after backward elimination (Table S12) and diagnostic checks (Fig. S4). The final model had a negative binomial distribution instead of a Poisson distribution, which precluded comparing the full and best model, provided models were not nested because their overdispersion parameters were different, and thus not comparable [47]. Thebest model's approximate 2 indicated it explained up to 79.45% of the data variability, with a predictive capacity ( 2) of 60.96%. This model showed that cases increased with the number of types of goods exports (EX), connected cities at the main airport (CC) and overall trade openness in both goods and services (TO). By contrast, cases decreased with the number of coronavirus transmission prevention policies in place (PP25), and model diagnostics showed that assumptions were met (Fig. S4), so were VIF diagnostics (Table S12). The jackknife for this model also found all significant and similar parameters to those found with the negative binomial generalized linear model (Table S13). The parametric bootstrap did not confirm the significance of ASK and TO, a result that might be an artifact of the bootstrap asymptotically approximating the distribution of parameters at the cost of increasing type II error, i.e., failing to reject the null hypothesis when false [40,44,47].

Table 3

Parameter estimates for the best negative binomial generalized linear model explaining total COVID-19 Cases 100 days after the detection of the first case across nations from Mesoamerica and Caribbean. In the model estimates for population size in 2020 for each country were used as offset to account for differences in the size of the population at risk.

Parameter	Estimate	Std. Error	z value	Pr(>|z|)
Intercept	−9.081	0.717	−12.666	<2e-16*
No. of export types	0.00640	0.00176	3.642	0.00027*
No. International cities connected through the main airport	0.0259	0.0058	4.502	6.74E-06*
Trade Openness (both merchandise/goods and services)	0.0194	0.0048	4.017	5.89E-05*
No. of prevention policies 25 days after the first case detection	−0.650	0.204	−3.189	0.00143*
Overdispersion	4.09	1.23

*Statistically Significant (P < 0.05).

The best model (Table 4 ) explaining the number of deaths, D, at t = 100 days after the first locally detected case likewise dropped a large number of variables after backward elimination and diagnostic checks (Table S14), with the final model also having a negative binomial distribution instead of a Poisson distribution, thus again precluding a full model comparison. This model had an 2 = 0.64 and a 2 = 0.41. Deaths increased with the numbers of international cities connected through the main airport (CC), the concentration of imports (CI), overall trade openness in both goods and services (TO), and aid flow from the US (AUS). Deaths decreased with the concentration of exports (CE). Like in all previous models, assumptions were met (Fig. S5), so were VIF diagnostics (Table S12). The jackknife for the deaths model confirmed the significance of CC and provided similar estimates for all other parameters (Table S15). The parametric bootstrap, meanwhile, suggested that both CC and AUS were significant while the lower 95% CI of TO was nearly significant, a result also suggested by the jackknife (P < 0.08) (Table S15). However, further simplifications to the model resulted in a significant loss of model fit and different overdispersion parameters, a problem suggesting that all parameters presented in Table 4 were not overfitted [34]. These results highlight the limitation in negative binomial model selection when overdispersion is not fixed. Overdispersion is a factor that makes reliable model comparison impossible [42,47]. In addition, the satisfactory VIF (Table S14) and model diagnostics (Fig. S5) results suggest that the model was not overfitted.

Table 4

Parameter estimates for the best negative binomial generalized linear model explaining total COVID-19 Deaths 100 days after the detection of the first case across nations from Mesoamerica and Caribbean. In the model estimates for population size in 2020 for each country were used as offset to account for differences in the size of the population at risk.

Parameter	Estimate	Std. Error	z value	Pr(>|z|)
Intercept	−12.372	0.681	−18.161	<2e-16*
No. International cities connected through the main airport	0.024	0.006	4.165	0.00003*
Concentration of imports index	4.339	2.155	2.013	0.044*
Concentration of exports index	−4.091	1.295	−3.158	0.0016*
Trade openness (both merchandise/goods and services)	0.011	0.005	2.064	0.039*
Aid flow from the United States of America	0.0043	0.0014	3.001	0.002*
Overdispersion	4.11	1.93

*Statistically Significant (P < 0.05).

Discussion

Relevance of social science theory to describing disease spread

The main common feature of our results is that variables related to social science theories about used territory, human development, and uneven development were shown to be highly instrumental in explaining the initial spread of COVID-19 in Mesoamerica and the Caribbean. As summarized in a cartographically explicit way in Fig. 5 , COVID-19 spread patterns and variables that measure development and used territories have a degree of overlap. This highlights how connections between territories of the region we studied and the structures of the territories those connections are associated with are a driving force for the COVID-19 pandemic just as are proximal biological phenomena related to human movement [50] and pathogen spillover [1]. The overlap between functional cluster types in Fig. 5 is most evident for the Dominican Republic and Panamá, the two worst-hit countries by COVID-19 in the region during the first 100 days of transmission analyzed in this study. The Dominican Republic and Panamá clustered together when examining both their COVID-19 data and different indicators of human and uneven development as well as of used territory. A similar overlap pattern was observed for Guatemala, El Salvador and Honduras in Mesoamerica as a cluster in both epidemiological and social science parameters, as were states from the Lesser Antilles in the Caribbean of Antigua and Barbuda, St Kitts and Nevis, Dominica, St. Lucia and St. Vincent and the Grenadines. On the other hand, Caribbean states from the Greater Antilles did not cluster with other territories when looking at the indicators from the social science theories. Nevertheless, they had similar patterns of COVID-19 spread as observed for Haiti, Cuba, Puerto Rico and Jamaica.

Fig. 5

Map of Mesoamerica and the Caribbean displaying relations between disease and used territory, human development and uneven development parameters. Weavings are used to depict the two patterns of clustering. In the map we also included Uruguay, since data for that country was used as an outgroup for the underlying analyses.

The feasibility of used territories driving disease spread

These patterns can be expected under Santos' conception of the used territory as “made up of objects and actions, and … a synonym for human space, inhabited space,” unlike the traditional definition of territory as the land of a state [20]. Santos insightfully argued spatial patterns can go beyond “regional” by considering “horizontalities” as territorial links that emerge from absolute space contiguity. Thus, horizontalities are akin to the rich observations of Tobler on spatiality [51,52], which have been condensed and reinterpreted by others as Tobler's First Law of Geography, the principle stating that ‘nearby’ phenomena are likely more interrelated [53,54]. In that sense, we can expect the overlapping clusters of territories in Mesoamerica, the Lesser and Greater Antilles, that we described across a spatially contiguous geography as illustrated in Fig. 5. More generally, Uruguay illustrates horizontalities, an outgroup, not clustering with Mesoamerica and the Caribbean when examining its disease patterns. However, “verticalities”, defined by Santos as forces connecting what is not contiguous in absolute space, also shape used territories [20]. Verticalities can explain clusters like the one observed for the Dominican Republic and Panamá, which, although not contiguous in absolute space, are connected by having similar social, political and economic relations. Similarly, when looking at indicators related to people in motion, economic structure and trade, the presence of Uruguay in a cluster with countries of Mesoamerica and the Caribbean can be partially explained by human development metrics based on the ‘capabilities approach’ developed by Amartya Sen [11,55]. This clustering pattern can also be articulated by Amin's theory of uneven development [12].

Higher human development does not prevent early epidemic growth

Following Sen [11,55], we would expect that a higher level of human development would have implied substantial protection against the spread of COVID-19 in Mesoamerica and the Caribbean. Yet we did not find any human development metrics associated with the epidemiological parameters considered a response in our econometric models. We found that r grew with the concentration of imports, a variable positively correlated with human development indices (Fig. 3B). Indeed, the raw correlations indicated a positive association of r with both the human development index (correlation = 0.21) and its inequality-adjusted version (correlation = 0.14). This positive association might reflect that countries with higher human development were better able to diagnose SARS-CoV-2 infections. In the studied region this might be the case for countries like Nicaragua, with low human development, that have taken no action to contain, or even diagnose, COVID-19 [56]. This, for instance, contrasts with the laudable epidemic surveillance that Panamá has had in the region, being the country with the best surveillance system in the studied region, the country most successful in diagnosing the virus at the beginning of the pandemic, and the only one that was using molecular information to trace SARS-CoV-2 patterns of spread and biological evolution [5,57].

On the other hand, this counterintuitive association, where more human development was associated with larger epidemics, might reflect biases inherent to the human development index construction. Given its components, human development is strongly associated with the gross domestic product per capita, as shown in Fig. 3B. Although able to account for individual wellbeing at a particular scale, the human development index was unable to reflect important societal-scale vulnerabilities (even in an ‘inequality-adjusted’ form). The data shows that more human development implied less resistance to a pandemic infectious disease such as COVID-19's spread across the studied territories. This raises concern about the formulation of development metrics. It is desirable to reduce the gap in material wellbeing across and within societies and this measure may correlate with some dimensions of human possibility, as originally intended by Sen [11,55]. Still, linear measures of development may be misleading in complex systems.

Indeed, complex systems are prone to showing non-linear associations that change through time [58], as we observed with the prevention policies. Those measures went from having a positive association with transmission as observed in early epidemic growth rates to having a negative association with the number of cases 100 days after the detection of the first case. And these association patterns are unlikely to be spurious, given the robustness of our cross-validated results, which come from models explaining at least 60% of the data variability, a value well above what is commonly reported in many areas of the social sciences in which models explaining 35% of the variance may be exceptional [40,42,59].

Uneven development articulated by used territories

Given the limitations of measurements for “human development,” and after controlling for the impact of COVID-19 prevention and mitigation policies, we can try to understand the verticalities defining used territories by resorting to the principles of uneven development developed by Samir Amin [12,60]. The framework developed by Amin and others [13] has been widely used to study other systems under the influence of uneven development. Successful examples include examinations of ecologically unequal exchange and deforestation [61] but also applications to the study of health problems [62] such as maternal and infant mortality [[63], [64], [65]] and obesity [66,67], as well as infectious diseases like HIV-AIDS [31], malaria [68], tuberculosis [69] and Ebola [70]. According to uneven development and related theories [12,13], peripheral states have a general tendency to have an economic profile and global entanglements that negatively influence their abilities to mobilize resources necessary to avoid or mitigate social problems [71], such as the health crisis that has emerged with the COVID-19 pandemic. Constraints on state capacity and the erosion of social organization in peripheral countries can emerge from dependency, arising from the types and conditions of external aid from core economies and inward flows of foreign direct investment. Foreign direct investment flow is often viewed positively, and is the objective of certain political strategies, but can also be a modern measure of the selective plundering of peripheral societies by transnational actors based in core states. Likewise, trade openness measures a country's successful integration into the global economy and a measure of a country's exposure to, and likely structural adaptation to playing particular roles within the global geographical, political economy. All provide relevant measurements for the uneven development that channels and can constrain the ability of peripheral states to mobilize resources to fight diseases like COVID-19 and that were considered in this analysis.

Our analysis found that aid received from the United States was positively and robustly associated with increases in deaths by COVID-19 in Mesoamerica and the Caribbean. This positive association suggests that uneven development played a role in the spread of COVID-19 in Mesoamerica and the Caribbean. A clearer pattern was observed for increases in trade openness, positively associated with increased cases and deaths during our study period.

Similarly, other aspects of unequal trade were associated with increased transmission and burden of COVID-19 in Mesoamerica and the Caribbean. We observed that the concentration of imports increased early epidemic growth and deaths during the study period, likely, at least in part, reflecting emergent structures of trade between core economies and regional blocks within a global political economy of vast differences in power. These dependency mechanisms might explain observed patterns in Caribbean countries like Dominica, whose major imports come from Japan, the major market for its exports [72], but beyond bilateral patterns of trade; they might reflect realities observed in Mesoamerica and the Dominican Republic following the Central American Free Trade Agreement (CAFTA), where the elimination of tax barriers has eased importations from the United States in the countries covered by that treaty [73]; and similarly, with Panamá, which has an independent agreement with similar terms [74]. This promotion of economic specialization [75] has destroyed the ability of peripheral states to satisfy needs for essential goods [76], including those essential to deal with pandemics of infectious diseases through domestic production, as observed for COVID-19 and the shortage of ventilators and other medical supplies in Latin America [77,78].

Meanwhile, exports were associated with a higher number of cases; but when the concentration of exports appeared in tandem with the concentration of imports, it reduced the number of deaths. An explanation for this changing influence may be rooted in the role that exports play in the economy and the shape of the relationship between export concentration and the wealth of nations, as measured by the GDP per capita, which is roughly “U” shaped [79], and the fact that diversification does not always lead to an economy autonomous from market shocks [80]. The non-monotonic relationship between concentration of exports and GDP is, indeed, quite interesting as it might explain the positive effect in the regression analysis if the following are captured: an increased susceptibility in countries whose exports are concentrated by way of primarily extracting natural resources [81] and limitations to diversification in states with small populations or territories [82]. But at the same time, the exports effect on deaths might become negative when it accounts for a national specialization in producing export goods that reflects not the typical core-led development pattern observed but a successful societal mobilization around alternative development strategies that Amin described as ‘delinking’. Delinking is not a matter of a country pursuing autarky or rejecting the outside, but the country moderating the terms of its external interactions as part of a reorientation of its social and economic relations to serve the needs of its peoples [83]. A complete investigation of these issues is beyond the goals of this study given numerous unresolved debates around how exports and their trade concentration relate to the growth of wealth [84]. However, when not paired with the concentration of imports, concentrations of exports strongly support the explanation that elements of uneven development are verticalities constructing used territories that accelerated the spread of COVID-19 in Mesoamerica and the Caribbean.

The relevance of used territories for the development of spatial theories based on relational geographies

As discussed, the used territory offers unique advantages to study phenomena that couple natural and human systems. However, its origin within the discipline of geography enables it to go one step beyond by framing those connections within the spatiality of landscapes. These are landscapes whose attributes, length and time scales, and even appropriate spatial coordinate systems for analysis coevolve with coupled dynamical systems [85]. In that sense, beyond what is possible in human development theory, and as complementary to human and uneven development theories, the used territory is a theoretical tool that eases the understanding of relational geographies [[86], [87], [88]]. It contextualizes the placement of people in motion. As measured by the number of international cities connected through the main airport, the motion factor was positively associated with increases in the early exponential growth rate of cases, in cases and in deaths by COVID-19 in Mesoamerica and the Caribbean. Thus, our results illustrate the power of the used territory and of attention to relational geographies generally, as conceptual tools to pose problems within large enough frameworks to find meaningful solutions [89].

Conclusion

Our cross-national and relational geographic approach to studying the COVID-19 pandemic in Mesoamerica and the Caribbean showed that global political-economic roles of countries are also extremely important to understand how diseases spread within societies. This approach went beyond standard epidemiological analyses focused on public health measures to prevent and mitigate disease transmission, which estimate transmission parameters and alienate human movement from structural factors shaping space as defined by used territories. Some proxies for spatial connections examined in this study were directly and positively associated with SARS-CoV-2 transmission (measured in terms of epidemic growth, cases and deaths), such as the size of the network of cities connected across major airports (parameter CC in Table 1). However, our results also support the epidemiological relevance of social theories that stress a broader spectrum of power-laden global connections. We found that variables key to the formation of space in used territories, e.g., concentration of imports, trade openness and international aid, were positively associated with increases in disease transmission and deaths. These variables, conceived within social science theory, might remain important over much longer timescales than the unfolding of the COVID-19 pandemic and deserve further study in developing mathematical models dealing with spatial and spatio-temporal disease spread.

Author contributions

LFC and LRB Conceptualization, LFC, DOS and LRB Formal analysis, LFC, MDF, LAH and RMR Investigation, LFC, LAH, RMR and LRB Data curation, LFC, LAH & LRB Funding acquisition, LFC Project administration, LFC, MDF, DOS and LRB Methodology, LFC, DOS and LRB Software, LFC Supervision, LFC Validation, LFC, DOS and LRB Visualization, LFC and LRB Writing - original draft, LFC, MDF, LAH, RMR, DOS and LRB Writing - review & editing & approving final version.