En français

Was this virus sensitive to weather?

We ran two full exploratory and correlation studies to try to answer that question: one at country level, the other at US county level.

Study 1

Country-level data

Countries were not affected in the same way by COVID-19. Some countries reported close to zero deaths, while others already counted tens of thousands. Why such differences between countries?

For each country, we had collected the following data:

Here is the correlation matrix. As a reminder, correlation coefficients range from -1 to +1: +1 is a perfect positive correlation, -1 a perfect negative correlation, and 0 means no linear correlation.

Correlation matrix between demographic factors and COVID-19 mortality by country
Correlation matrix across countries

Four features were highly correlated with one another: wealth per inhabitant, average life span, average age, and quality of the health system. In a prosperous country, people tend to live longer, the average age is higher, and the health system tends to be better.

The declared death rate per million inhabitants was strongly correlated with those four features. That might sound contradictory, but it was a side effect: a better health care system meant a longer life span and an older population, which in turn meant a higher COVID-19 death toll.

Press freedom and the declared death rate were positively correlated: the less press freedom, the lower the declared death rate — suggesting the true toll was understated in some countries.

Temperature was negatively correlated with wealth per inhabitant (poorer countries tended to be hotter) and with the declared death rate. But an effect is correlated with its cause, and two effects will also be correlated with each other: did higher temperatures reduce deaths, or were there simply fewer deaths in hot countries because they tended to be poorer, with a younger population and a lower life expectancy? The correlation matrix alone could not answer that.

We took the analysis further by training a gradient boosting algorithm — widely considered one of the most reliable machine learning methods, and capable of picking up correlations even when they are not linear — on our dataset.

Feature importance according to the gradient boosting algorithm
Feature importance — gradient boosting
The most important variable was the average age of the population. Temperature had virtually no impact and was not retained by the algorithm.

Conclusion: temperature had no significant impact on how the pandemic evolved.

COVID-19 data: jhu.edu
Press freedom: rsf.org/fr/classement
Health system quality: april-international.com
Average life span: wikipedia.org — life expectancy
Average age, by country: wikipedia.org — median age
Wealth per inhabitant: wikipedia.org — GDP (PPP) per capita
Population density: wikipedia.org — population density
Population: wikipedia.org — population
Temperature, April 2020: tutiempo.net

Study 2

US county-level data

For each of the 3,242 US counties, we had collected data on:

We had trained our algorithms on this dataset, with the goal of identifying which features had an impact on how the epidemic evolved. We had been careful about potential biases.

census.gov — USA Counties
Google Sheets — dataset
timeanddate.com — historical weather
github.com/nytimes/covid-19-data

Map of US counties by temperature and epidemic progression
US counties — overview

There was no meaningful correlation between temperature and the speed of the epidemic's spread in the US. The country is vast, and in March 2020 it spanned a wide range of temperatures, from about 20°F to 90°F (-6°C to 32°C). The same held true for humidity and atmospheric pressure: no correlation emerged.

We built a correlation matrix heat map, and the coefficients between weather variables and the epidemic's evolution were consistently insignificant.

Correlation matrix heat map between weather and epidemic evolution by US county
Weather vs. epidemic evolution — correlation heat map
Conclusion: temperature and weather had no major direct impact on Covid. The first wave in Mexico, Texas, and Florida took place in the middle of summer, with temperatures around 95°F (35°C). What seemed to drive the spread of the virus was not the temperature, but people's behavior — in particular, time spent in confined spaces.