As new cases of SARS CoV-2 (aka 2019-nCoV) Coronavirus are confirmed throughout the world and millions of people are being put into quarantine,  it is important to investigate the mix of factors affecting the spread of the epidemics. It is still early days, but, using existing data on positive cases of SARD-CoV-2 Coronavirus, Babak Naimi and I developed a large ensemble of 10 machine learning SDMs that project monthly variation in climate suitability of Coronavirus throughout a typical climatological year.

We were able to:

1) provide a preliminary characterisation of the realised ecological niche of SARS-CoV-2 Coronavirus and use it to predict for climate suitability for the Coronavirus around the world;

2) project the ecological niche of SARS-CoV-2 throughout the year following seasonal changes in temperature and precipitation.

A simple visualisation of the results (data from the 10th March) can be found below. It shows how modelled climate suitability for the Coronavirus changes with seasonal changes in climate suitability (map updates provided here).

Our analysis are currently under review but given the importance of these projections, we decided to deposit the manuscript in the preprint server medrxiv. We are updating the manuscript on a regular basis with new data and analyses, some of which address concerns expressed by peers (see technical comments, below).

A few examples of the inferences arising from models include:

  1. The pandemic, while global in extent, is unlikely to affect everywhere with the same intensity at the same time. Of course this is contingent not just on climate but also on the societal’s response to the crisis. The maps just reflect the modelled climate suitability for the Coronavirus and don’t take into account other relevant factors. This expected asynchronous pattern would allow public services to prepare with anticipation and economic actors to reduce uncertainty thus minimising risks for their activities. For example, while southern Europe is now badly hit and will likely continue to be through March and April, models anticipate reduced climate suitability for dissemination of the Coronavirus towards the end of May up until October.
  2. While patterns of out-door transmission of the coronavirus will likely be seasonal, the most important pulses are not between northern and southern hemisphere, as one would normally expect, but between warm temperate and cold temperate regions. That is, while the virus season in much of the warm temperate regions, like Hubei, Italy, Spain, Portugal, France, UK, Germany, the USA will cover parts of Autumn, Winter, and Spring, much of the cold temperate region, including Canada, Russia, and the Scandinavian countries will likely be exposed to climate conditions suitable for epidemic propagation from late spring up until the end of the summer. However, cold temperate regions rank amongst some of the ones with the lowest human population density and this pattern might contribute to neutralise the the effects of increased climate suitability for the virus during the summer.
  3. Arid regions are moderately at risk given climate conditions, compared to the temperate regions. The winter conditions is more prone to epidemic behaviour but still less climatically suitable than warm temperate regions, meaning that, all other things being equal, containment measures would be more effective implemented in Arid regions, like Iran, than in temperate regions. Of course, climate is one important variable but it does not act alone. The density, connectedness, and behaviour of hosts is critical.
  4. Comparatively with other climate zones, the wet tropics have reduced climate suitability for the Coronavirus. This does not imply zero risk as infected people can still reach these areas and infect other people locally (indoors, air conditioned spaces, are ideal for the propagation of the virus and according to a Chinese study made up to 80% infections in Wuhan). But it means the out-door ecosystem, as a whole, will likely be less favourable climatically for the virus potentially helping to constrain its capacity to reach epidemic status.
  5. The coarse patterns of climate determination inferred for SARS-CoV-2 are consistent with what one would expected given the finer resolution data obtained with the predecessor SARS-CoV (similar data to be still obtained for SARS-CoV-2).  

Technical Notes:
We have received a few critical comments in social media, relative to this post. We are strong believers of scrutiny by peers and thank all constructive criticism (and the less constructive criticism too) as they will allow us to refine both the models and the communication approach. As time goes by, we will summarise some of the recurrent criticisms below providing some responses to them.

  1. These results are not peer reviewed and in the middle of a crisis we should have waited for the peer review process to be completed to disseminate results: Our response: This is a harsh, and slightly unfair comment. We weighted carefully the pros and cons of early release. It is exactly because we are living a crisis moment that speedy communication is important and one of the consequences is crowd peer review, which enables a degree of scrutiny way higher than just procedural peer review. Procedural peer review can take months to be completed, sometimes more. Releasing insight that the virus is likely to display seasonal pattern is important now, not towards the end of the year (for example, the WHO still shares doubts that epidemic properties of the virus are seasonal). We understand the risks of early release of results, but we felt a moral obligation to share them. If new evidence is put forward that challenges our models and projections, fine, we are not in love with our models. Science moves when scientists make predictions and then test them. Unlike most model projections we have made in the past these ones will be quick to test.
  2. Given uncertainties associated with models, would it not be best to not disseminate results until we are more certain of them? Our response. George Box became famous for his quote “all models are wrong but some are useful”. The question is how do we define usefulness? One way is that models should be better than expected by chance. If a model describes a pattern that is qualitatively correct, even if incomplete or quantitatively inaccurate, it might still be useful. We are convinced that given the background (and references) provided in Point 6, there is good reason to expect a degree of seasonality in the incidence of outbreaks. That is the most parsimonious expectation, unless SARS-CoV-2 was fundamentally different from its relative SARS-CoV. It is not just our models: there are multiple lines of evidence in support of the link between the infectiousness of SARS-CoV and climate (existence of multiple lines of evidences is one of the criterion for the ‘gold standard’ in the validation section of Araújo et al. 2019 Science Advances). So, unless proven wrong by new data and models, the most reasonable expectation seems to be the one we propose: seasonality and differences in risk of epidemic infectiousness under different climate conditions. Another related question is: what is the cost of being wrong? Crisis management in a pandemic situation is reassessed on a weekly or even daily basis. Decision makers make decisions based on incomplete information in a compressed period of time. The role of forecasts is to help manage uncertainty by defining scenarios given data or prior knowledge. Our scenarios balance two types of errors: type I error (over-predicting risks locally) and type II error (under-predicting risks locally). For example, if we under-predict climate-induced risks in the topics, that is a Type II error. If we over-predict risk in the cold temperate regions, that is Type I errors. The alternative to our model assumes flat seasonal risk exclusively driven by policy and societal responses to the crisis. This scenario minimizes Type II errors, while inflating Type I errors (by discarding external environmental factors affecting risk).
  3. The virus is still spreading, thus not at equilibrium with climate. Surely this leads our model to overestimate seasonal changes. Our response: It is a possibility, as with any other empirical models, that non-equilibrium distributions truncate response curves thus leading to partly incorrect projections. This is a point we have made countless times in the literature (e.g., Thuiller et al. 2004 Ecography; Araújo & Pearson 2005 Ecography; Pearson et al. 2006 Journal of Biogeography; Araújo et al. 2011 Ecology Letters; Araújo & Peterson 2012 Ecology). However, a pandemic virus (globally distributed) is more likely to quickly reach equilibrium with climate than any other organism we (and others) have modelled so far. However, as we say above, it is still early days and there might be portions of the ecological niche that have yet not been occupied. This is an issue of concern for us and we are, consequently, updating the models (see here).
  4. Because of detection biases it is possible that the virus is passing unnoticed in the subsaharan African countries, thus biasing our estimates of low climate suitability in tropical regions. Our response: If such biases in the data exist they will surely affect the models (e.g., Thuiller et al. 2004 Ecography; Segurado et al. 2006 Journal of Applied Ecology). But we don’t know whether they exist or, how much, they affect the signal in the data. There is a fundamental difference having a few positive cases passing unnoticed to authorities, and having an epidemic passing unnoticed. We are modelling cases that have been confirmed by WHO that are likely to involve local transmissions.
  5. Spread of the virus is mainly a function of the connectedness of people and societies so it is better modelled through spatially-explicit approaches. Our response: We do not model spread of the virus but how climate suitability might contribute to the spread. Nevertheless, the virus has now officially reached a pandemic state. Connectedness is particularly important when modelling the direction and rate of spread early days of the outbreak (e.g. Wells et al. PNAS). Once reaching pandemic state, other factors like how the crisis is managed and how the environment may favour (or not) the spread of the virus can be important. Indeed, epidemiological models such as GLEaM, currently being used by a team at the University of Oxford to generate scenarios epidemic activity, use all of these factors combined.
  6. The Coronavirus lives at high temperatures in the human body, how would climate affect its ecological niche? Our response: Obviously, we are not modelling the ecological niche of the virus in the human body. We are modelling the viability of the virus outside the human body given environmental temperatures and precipitation (a surrogate for humidity, which is the relevant variable). Both variables had demonstrable effects in the incidence of SARS-CoV positive cases (e.g., Tan et al. 2005. Evidence Based Public Health Policy and Practice), and its survival rates, once expelled from the human body through sneezing or coughing (Chan et al. 2011. Advances in Virology). The infectiousness of the SARS-CoV-2 Coronavirus is thus expected to be related to its survival rate outside our body (unless it was structurally different from SARS-CoV, which seems unlikely), hence the importance of relating it to climate. Can SDMs capture those fine grain patterns? Time will tell, but the mechanistic link is there. If the density of positive cases is related to the infectiousness of the virus, related to its viability outside the human body, then SDMs using temperature and humidity-related variables can well capture meaningful patterns despite uncertainties coming from data quality, noise associated within quantities of indoor transmissions, and differences in the effectiveness of political responses to the epidemic.

Meanwhile, our post here caught the attention of:
Overpopulation Atlas

Three independent studies reaching similar conclusions to ours:
Temperature dependence of COVID-19 Transmission
Climate affects global patterns of COVID-19 early outbreak dynamics
Temperature, Humidity and Latitude Analysis to Predict Potential Spread and Seasonality for COVID-19

A recent review on seasonality of respiratory viral infections

Other useful reads: