This study is conducted in Chongqing, which covers an area of 82,400 km2 in Southwestern China and has approximately 33 million registered residents in 2019 (Figure 1). It has a subtropical humid monsoon climate, a long and hot summer, and a short and warm winter.
Influenza Surveillance Data
The influenza-like-illness (ILI) was defined as patient who has acute respiratory infection with fever and at least one respiratory symptom (cough and/or sore throat). ILIs and influenza virus positive rates were obtained from sentinel influenza surveillance network in Chongqing, which have been stated in a previous study 6. Briefly, seven sentinel hospitals were selected based on higher accessibility to patients, higher qualifications of medical staff, adequate specimen storage capacity, and the desire of the physicians and nurses to participate voluntarily in the surveillance program. At each sentinel hospital, trained nurses and clinicians collected data on the counts of visits and the total number of ILI to outpatient and/or emergency departments, and collected nasopharyngeal swabs specimens then tested influenza virus by reverse transcription-polymerase chain reaction.
In this study, the activity of influenza virus was represented by weekly confirmed influenza cases every ten thousand outpatient visits, which was calculated by multiplying the weekly positive rates of influenza by the weekly ILI counts, on the scale of every 10,000 of the outpatient visits, similar to a previous study 7.
We obtained simultaneous weekly meteorological data, including mean temperature (°C), relative humidity (%), atmospheric pressure (hPa), wind speed (m/s), sunshine (hours), as well as aggregate rainfall (mm), observed in 12 weather monitoring stations in Chongqing from the China Meteorological Data Sharing Service System (http://data.cma.cn/). The weather monitoring stations distribution in Chongqing was shown in Figure 1.
The relationships between meteorological factors and the activity of influenza virus in the population are nonlinear and always lasting well beyond the exposure period. The weekly influenza activity, which was presented by weekly confirmed influenza cases every ten thousand outpatient visits, was modeled using a quasi-Poisson regression model combined with a distributed lag non-linear models (DLNM). The DLNM proposed by Gasparrini DLNMis a flexible model that estimated the nonlinearity and distributed lag effects of exposure-response relationships simultaneously 8.
In this study, we used DLNMs to explore the potential exposure-lag-response associations between multiple meteorological factors and influenza activity. We used the variance inflation factor (VIF) to assess the co-linearity. A VIF greater than 5 indicates multicollinearity 9. The results showed that the VIFs of the mean temperature and atmospheric pressure are 8.39 and 5.70, which suggested that the two variables should not be included in the model simultaneously.
A Poisson regression with a quasi-Poisson function was established to test the over-dispersion in influenza activity, which was presented by weekly confirmed influenza cases every ten thousand outpatient visits. All models were adjusted with other explanatory variables of meteorological factors, seasonality and long-term trend, and school holiday etc. The model structure is stated as following:
log[E(Yt)] = α + cb(climate variables, lag, df) +∑ns(Xj, df) + ns(time, df*8) + factor (holiday)
Where E(Yt) is the expected weekly confirmed influenza cases every ten thousand outpatient visits on week t; α is the intercept; cb( ) represents the cross-basis matrix of climate factors, including mean temperature, relative humidity, aggregate rainfall, wind speed and sunshine; df is the degree of freedom; Xj is the other explanatory variables of meteorological factors; time refers to duration of seasonality and long-term trend; holiday is an indicator variable which equals to 1 if week t is in school holidays and 0 otherwise. We used Akaike Information criterion (AIC) to choose the df, which was supported by other references 10-13. The weeks of lag structure in the models were determined by incubation period and infectious period of influenza virus 14, the Akaike Information criteria and other references 10, 11. We provided all AICs in the appendix (Supplementary Tab. S1).
We calculated the relative risk (RR) with corresponding 95% confidence interval (CI), relative to the reference levels. The reference levels were defined as the median values of mean temperature, relative humidity, wind speed, sunshine and aggregate rainfall. Moreover, we estimated the extreme effects by comparing the 97.5th percentiles and 2.5th percentiles to the median values. The RRs brought by high temperature (hot effect), high relative humidity (wet effect), high aggregate rainfall (rainy effect), high wind speed (windy effect), long sunshine (long-sunshine effect) were calculated by comparing the 97.5th percentiles to the median values. The RRs brought by low temperature (cold effect), low relative humidity (dry effect), low aggregate rainfall (rainless effect), low wind speed (windless effect), short sunshine (short-sunshine effect) were calculated comparing by the 2.5th percentiles to the median values.
To test the robustness of our results, sensitivity analyses were performed by changing (2-5) for climate variables and maximum lag weeks (2-4 weeks) in the model.
All statistical tests were two-sided, the p-value ＜ 0.05 was considered statistically significant. We performed all data analyses using R software version 3.4.2and used the “dlnm” package for the DLNMs.