Acute lower respiratory infection data
Data on hospital outpatient visits for ALRI were retrieved from the Guangdong Second Provincial General Hospital located in the southwest of the city (Figure 1), which is one of the highest-level (tertiary) hospitals in Guangzhou. According to the International Classification of Diseases, Tenth Revision (ICD-10), hospital outpatient visits with the primary diagnoses of pneumonia (J12-J18), bronchiolitis (J20-J21), and asthma (J45-J46) were obtained between February 2013 and December 2019. We aggregated the three subtypes of ALRI to a series of daily time-series data[17-20].
Air pollution and meteorological data
Daily concentrations of air pollution during the study period were obtained from 11 air monitoring stations in Guangzhou (Figure 1), including inhalable particulate matter (PM10), coarse particulate (PMc), fine particulate matter (PM2.5), nitrogen dioxide (NO2), sulfur dioxide (SO2) and ozone (O3). Following previous study, the PMc concentrations were calculated by subtracting PM2.5 from PM10, because PM10 was consisted of PM2.5 and PMc. Air pollution measurement details have been previously described. Approximately 1% of observation days had missing data for the air pollution, the missing data were imputed using a linear interpolation approach (the “na.approx” function in “zoo” package in R).
Daily meteorological data (mean temperature and relative humidity) were obtained from the National Weather Data Sharing System (http://data.cma.cn/). Because there is potentially high correlation among different air pollutants and meteorological factors, we examined the pairwise Pearson correlation coefficients among these variables.[22, 23]
The ALRI data, daily air pollution concentrations and meteorological data were linked by date. Following prior similar epidemiology studies, the association between PM pollution and hospital outpatient visits for acute lower respiratory infections diseases was examined by an over-dispersed generalized additive Poisson model (GAM). In the model, public holidays (PH) and day of the week (DOW) were adjusted for as categorical variables. Seasonal patterns, long-term trends, temperature, and relative humidity were controlled for as smoothing splines. In accordance with the approaches used in previous studies[25, 26], we selected six degrees of freedom (df) per year for temporal trends, a df of 6 for moving average temperature of the current day and previous three days (Temp03) and relative humidity (RH).
Considering the delayed health effects of air pollutants, we examined the lag effects for different lag structures. We begin with the same day (lag0) up to five days lag (lag5) in the single-lag day models. We also considered the accumulated effects of multi-day lags (moving averages for the current day and the previous 1, 2 and 3 days [lag01, lag02, and lag03]).
In order to evaluate the potential effect modifiers of the PM pollution-ALRI associations, we conducted stratified analyses by gender (male vs. female), age group (age <5 vs. age 5-14), and season (warm vs. cold). The warm season was defined as from April to September, and the cold season was from October to March. The 95% confidence interval (CI) of the difference between group was calculated by the formula below:
where Q represents the estimated coefficient in each stratum, and SE is the corresponding standard error. The difference was considered as statistically significant if the 95% CI did not include unity.
To examine the robustness of the main models, we applied a series of sensitivity studies. The main findings were assessed by changing the df in the smooth functions for temporal trends and meteorological factors. Additionally, we adjusted for gaseous air pollutants (SO2, NO2, and O3) in two-pollutant models. The models were regarded as robust if there were no significant changes after df-changed or further adjustment for gaseous air pollutants.
Counterfactual analyses on the burden of ALRI attributable to air pollution
We estimated the burden of ALRI attributable to PM2.5, PMc, and PM10 by calculating the difference between the observed ALRI outpatient visits and the counterfactual visits predicted using well-recognized reference values of air pollution recommended by the World Health Organization (WHO) and our previously built over-dispersed generalized additive Poisson models. This difference between the observed and counterfactual ALRI outpatient visits represents the estimated burden of ALRI outpatient visits associated with particulate matter of different sizes. The counterfactual scenarios were set to be hypothetical values of PM2.5, PMc, and PM10 set by the WHO Air Quality Guidelines (24 hours mean: 25 μg/m3 for PM2.5, 25 μg/m3 for PMc, and 50 μg/m3 for PM10). The observed air pollution levels lower than the reference values were kept the same in the counterfactual scenario. The 95% CIs were constructed using 1,000 bootstrap replicates with replacement for each model.[28, 29]
In all statistical analyses, a P value ≤0.05 was considered statistically significant. All data cleaning, aggregation, and visualization, and statistical analyses were done using statistical computing environment R (version 4.0.5).