Abstract

1. Introduction

The recent outbreak of SARS-CoV-2 infections, which causes COVID-19 in humans, has accelerated the development of global health policies to manage and mitigate the risks of infectious diseases (Wilder-Smith et al., 2020). Concurrently, the medical scientific community has been mobilized to provide clinical support during the pandemic (Kupferschmidt and Cohen, 2020). This pandemic has also attracted the interest of non-medical researchers hoping to understand potential environmental drivers of SARS-CoV-2 prevalence and to predict future outbreaks via correlative methods such as species distribution models (SDMs).

Coro (2020) presented an implementation of a SDM to predict global areas of high SARS-CoV-2 infection rate from geophysical and social-related spatial covariates that he found to be correlated with high infection rates. While contributing to a broader understanding of the potential geographic scope of the pandemic is a commendable effort, the application of a correlative model to estimate SARS-CoV-2’s ecological niche given the epidemiology of transmission, as well as technical aspects of model implementation, highlights the drawbacks of using this approach to predict future outbreaks of SARS-CoV-2. We contend that such SDM implementations are prone to spurious results, meaning that they have a high potential for finding correlations between mechanistically unrelated variables, especially when model assumptions are violated and spatial scaling between covariates and mechanistic processes are mismatched. Therefore, extreme care should be taken when disseminating and overinterpreting correlative research without strong a priori mechanistic hypotheses, particularly in the midst of a pandemic, as incorrect conclusions about drivers of viral spread may have negative consequences if used as guidance by policy makers. Indeed, Coro's model did identify correlations between putative explanatory variables and high infection rates. However, below we argue that such correlations are unlikely to be biologically relevant both because the variables considered lack mechanistic plausibility in this setting as well as other flaws in model implementation.

Herein, we present our concerns with the foundational premise of Coro 2020 to model the ecological niche of SARS-CoV-2 as well as the implementation and interpretation of modeling results. We suggest that Coro's (2020) claim that “Generally, the model indicates a high infection rate in areas characterized by an annual moderate-high level of CO₂, moderate-low temperatures, and moderate precipitation” is problematic for two primary reasons: 1) SDMs do not accurately predict the dynamic geography of SARS-CoV-2 transmission because the underlying drivers of viral spread are dominated by human behavior (Carlson et al., 2020); and 2) the paper overlooks substantial predictive misclassifications, model assumptions and validations.

2. SDMs are not suitable for modeling emerging SARS-CoV-2 spread

SDMs typically use the locations where a species is recorded as being present and, if possible, where the species is recorded as being absent, to build a statistical model of the occurrence of the species with environmental covariates in order to predict its distribution. The correct application of an SDM requires that the species has a clear environmental niche, even if that niche is unknown to the researcher, and that its distributional data are reflective of that niche (Václavík and Meentemeyer, 2012; Gallien et al., 2012). The use of an SDM to predict viral spread does not meet these requirements because SARS-CoV-2 presumably originated from a single spillover event (Andersen et al., 2020) and further transmission has resulted from human-to-human contact and not through a vector with an identifiable ecological niche (Carlson et al., 2020). Moreover, the virus is continuing to spread across the globe into new environments, driven primarily by human social contact patterns (Liu et al., 2020), such that the distribution of occurrence at any one point in time likely does not accurately reflect any sort of theoretical ecological niche (Chipperfield et al. 2020). Although Coro's (2020) model includes human-related parameters of population density and CO₂ levels in addition to climatic variables, they inadequately describe other social-related factors that drive transmission beyond human abundance such as, but not limited to, social contact dynamics, behavior changes from perceived threat of infection, governmental policy and its timing of implementation, socioeconomic factors across different regions of a country and across countries, and the timing of infection peaks.

3. Flaws in model implementation and validation

Coro (2020) dichotomized continuous values of infection rate using as threshold the average number of confirmed cases per capita at the peak of the pandemic in Italy to derive a response variable with the categories "occurrence" and "absence" of "high infection rate". The dichotomization of continuous data has well-known negative consequences, such as overestimation of effect sizes and loss of measurement reliability (MacCallum et al., 2002), overestimation of the differences between groups (Altman and Royston, 2006), and distortion of the relationship between predictive and response variables (Selving, 1987).

Furthermore, Howard et al. (2014) demonstrate that using a proxy of abundance (such as per-capita infection rate) as input for a SDM lead to distribution estimates significantly better than those derived from presence-absence data because the signal abundance-suitability is lost when all presences are treated as equals. For example, treating above-average per-capita infection rates as equals in the presence category may lead to a high bias if “very high” infection rates are scarce within the dataset and “just above average” infection rates are environmentally clustered by chance due to the North-to-South climatic and infection gradients in Italy. In such a case, high suitability values yielded by MaxEnt will be centered around the presences with lower infection rates within the “presence” category. Such a problem may only worsen if there is not much difference between infection rates at both sides of the threshold.

Using such knowledge, Coro (2020) could have made a sound choice by using per-capita infection rates as a continuous response variable in a regression model, rather than obscuring potential relationships between the response variable and the predictors through a convenient (MaxEnt can only use a dichotomous variable as response) albeit arbitrary dichotomization. As a result, biases in Coro's analysis cannot be assessed, and remain masked behind the apparent reliability of a colorful map.

Moreover, Coro's paper does not acknowledge or address violations of model assumptions beyond the existence of a viral niche that significantly affect results and their interpretation such as scaling mismatch between species’ data (modeled here as locations with high viral incidence rate) and predictor variables, and equilibrium of the species with its environment (Guisan and Zimmerman, 2000). Additionally, the poor model validation described in the manuscript undermines the paper's primary conclusions. To wit, the conclusions that “climatic parameters such as air temperature and precipitation (or air humidity) play a critical role at defining locations that may be subject to a high infection rate” is unsupported based on the model diagnostics presented in the paper. We elaborate these details in the next three sections.

4. Conclusion

We found little convincing evidence that this global-scale ecological niche model predicts SARS-CoV-2infection rate due to the drawbacks that we identified in the model implementation, validation, and interpretation as well as in the premise of an identifiable ecological niche for the spread of this particular virus. All modelling exercises require careful consideration of modeling assumptions in relation to study design and data collection for making interpretations that are meaningful and repeatable. Given the limitations of SDMs in particular, we recommend that future efforts to forecast SARS-CoV-2 outbreaks and/or to predict the spatial occurrence of other viruses with a similar ecology to SARS-CoV-2 be avoided (Carlson et al., 2020). Fundamentally, these outbreaks represent a non-equilibrium process with an environmental niche constrained almost entirely by that of its host organism, in this case humans, and with range dynamics that SDMs have a poor track record of predicting well. There may be cases for limited application of well-considered SDMs for directly transmitted diseases in appropriate scenarios such as the modelling of the occurrence of reservoir hosts of infectious disease transmissible from animals to humans (Zhu and Peterson, 2014; Carlson et al., 2016) and/or to make assessments of wildlife ecology and conservation strategies (Higgins et al., 2012). However, the direct application of an SDM for the occurrence of a directly-transmitted disease with a host species that has affected an estimated 95% of the earth's land surface is never likely to be informative.

Declaration of Competing Interest

The authors whose names are listed immediately below certify that they have NO conflict of interests in the subject matter or materials discussed in this manuscript.

Acknowledgements

References