Methods
Case reports in NSW, Australia from the beginning of the epidemic in January to the end of May 2020 were accessed.1 Those whose infection source was determined to be locally-acquired, and whose postcode of residence was reported, were included. A daily time-series of cases was created, from which separate series preceding and following the epidemic peak (31 March) and for individual NSW public health units (PHUs, see Figure 1A) were created. Based on the reported postcode the closest weather observation station was identified.2Daily observations of the following factors were downloaded: rainfall (mm) and temperature (°C), relative humidity (%) and wind speed (m/s) recorded at 9am and at 3pm.2 The mean values for each day were estimated to create time-series of weather data. Additional series of daily differences between 9am and 3pm temperature, relative humidity and wind speed were created. Thus, 10 predictor time-series were created for modelling.
The data was analysed based on the exponential and descending phases of the epidemic overall, and for 6 PHUs (those PHUs reporting <100 cases were excluded), to determine the effect of epidemic phases and locations on the association between climatic factors and case reports. Thus, 14 separate time-series analyses were performed.
A PHU-average Spearman correlation coefficient matrix was first calculated to avoid multicollinearity among variables. Then, a univariate generalized additional model (GAM) was fit and variables withP value <0.1 in univariate analysis in all the PHUs were selected for multivariate analysis. A standard two-stage approach was then applied to evaluate the PHU-specific and NSW-average associations between short-term exposure to climate factors and cases. In the first stage, a quasi-Poisson GAM was used to estimate the association between PHU-specific climate factors and the daily count of cases. A 14-day exponential moving average (EMA) was used to represent the effects of climate factors. Natural splines of time were included to control short-term temporal trend; its optimal degrees of freedom (df ) was chosen based on Quasi AIC (QAIC). In the second stage, a meta regression model with random effects were used to obtain NSW-average risk estimate of meteorological factors on cases. To estimate the overall relationship, exposure-response curves were plotted using the GAM with natural spline’s knot setting at its median (df = 2). A sensitivity analysis was performed by modifying the EMA (14 days to 13 or 15 days), and changing df for natural splines of time (3 to 2 or 4). R4.0.1 software (R Foundation for Statistical Computing, Vienna, Austria) was used to perform all analyses.