Meteorological factors and domestic new cases of coronavirus disease (COVID-19) in nine Asian cities: A time-series analysis

Zonglin He; Yiqiao Chin; Jian Huang; Yi He; Babatunde O. Akinwunmi; Shinning Yu; Casper J.P. Zhang; Wai-kit Ming

We evaluated the normality of the daily new cases and meteorological data by examining their skewness and kurtosis. We also estimated the Pearson correlation and covariances between the daily COVID-19 new cases and daily meteorological factors using STATA 14.0. Generalized additive models (GAMs) with a Poisson family and logarithm link function were used to estimate the associations of daily COVID-19 new cases with average temperature and relative humidity. GAMs are useful for identifying exposure-response relationships from various types of data, particularly in exploring nonparametric relationships (32) . The GAM analysis was performed in R software (version 3.6.0) using the package "mgcv". We first established a basic temporal model for COVID-19 cases without including meteorological variables. To adjust for long-term trends and seasonality, we included penalized spline functions of time in the model. The degree of freedom (df) for time was optimized by minimizing of the absolute values of the partial autocorrelation function (PACF) of residuals for lags up to 30 days (33) (34) (35) (36) . Additionally, the selection of an optimal model was based on the lowest Akaike's Information Criterion (AIC). Secondly, we built meteorological models based on the temporal models to account for the lagged effect of meteorological variables and the incubation period of COVID-19 (7, 37) . Specifically, we examined the effect of meteorological variables with different time lags including one-day lag (Lag 1d), three-day lag (Lag 3d), five-day lag (Lag 5d), single-week lag (Lag 7d), and two-week lag (Lag 14) to capture immediate effects and lagged effects, respectively. Automated penalized splines were used to fit the association between the daily new cases and each of the meteorological variables. The date when the accumulated cases exceeded 30 in each city or country was selected as the inception of the date incorporated in the model to equalize the starting speed of outbreak and thus avoid miss-interpretation and overfitting.