Preliminary evidence that higher temperatures are associated with lower incidence of COVID-19, for cases reported globally up to 29th February 2020

Melanie Bannister-Tyrrell; Anne Meyer; Celine Faverjon; Angus Cameron

As of 29 th February 2020, provinces in China (excluding Hubei) reported between 0 and 68 imported cases; ADM1 outside China reported between 0 and 26 imported cases (noting that as case travel history was not always reported, classification of imported versus local case was not always possible). The number of presumed locally-transmitted cases ranged from 0 to 1,220 in provinces in China (excluding Hubei), and 0 to 552 in ADM1 outside China (Table 1) . Average temperatures at ADM1 centroids in China (excluding Hubei) ranged from -16.8°C to 20.4°C, and from -18.7°C to 31.9°C outside China (Table 1) . Average temperature (as a linear and quadratic term combination) was strongly associated with count of local COVID-19 cases (likelihood ratio test = 19.4, df = 2, p=0.00006). Although one observation (Daegu province in South Korea) was identified as an outlier and had high leverage, the model results did not change when this observation was removed from the dataset. The model results indicate that there was a negative correlation in the predicted number of cases with temperature from 1°C and above ( Figure 2 ). For example, at mean values for the other variables, an increase in average temperature from 1°C to 9°C was associated with a decrease in predicted cases at ADM1 level from 24 cases to 19 cases, respectively. Similarly, an increase in average temperature from 10°C to 19°C was associated with a decrease from 18 to 7 predicted cases at ADM1 level, respectively. The pseudo R-squared values for the final model with and without temperature were 0.44 and 0.39, respectively, indicating that the inclusion of the temperature effect only provided a relatively modest improvement in model fit. There are several important limitations to this study. These results remain preliminary, as they only include confirmed cases as of February 29 th 2020, at which point reported local transmission outside China was relatively limited. There was no data available on many characteristics that affect rate of spread within a region, especially the interventions initiated in response to the detection of imported or locally transmitted cases. Furthermore, the model could not be fitted with a random intercept for country to account for clustering of ADM1 units within countries, as the uneven distribution of the number of affected provinces by country led to model instability. However, the early detection score likely captures part of the country-level variance. Is it also important to note that the classification of cases as local or imported was based on available information, and in some ADM1 (in particular in China, South Korea and Italy) the first imported case could not be identified. These results need to be confirmed by repeating the analysis as the pandemic progresses, and including data on implemented interventions to contain or mitigate COVID-19 as it becomes available.