The study was approved by the Ethics Committee of Guangzhou Eighth People’s Hospital. The hospitalization date, diagnosis, CD4 cell count, age, and sex were extracted from the records from January 1, 2014, through December 31, 2019. Informed consent was exempted because individual identifiers were not used. The diagnosis of PTB in HIV-infected patients was based on medical history, symptoms, physical examinations, and laboratory tests, e.g., isolation of mycobacterium tuberculosis from sputum or bronchoalveolar lavage, and confirmed by at least two physicians. The patients who resided outside Guangzhou city were excluded. In case of repeated hospitalizations, only the first admission was included.
Meteorological and air pollutants data
The daily data of meteorological factors including temperature, humidity, wind speed, and pressure, and air pollutants including PM2.5, SO2, carbon monoxide (CO), nitrogen dioxide (NO2), and ozone (O3) in Guangzhou were collected from Weather Underground, IBM (https://www.wunderground.com) and National Air Quality Study Platform (https://www.aqistudy.cn), respectively, as described in our previous study (24). The monthly mean concentrations of these variables were calculated for further analyses.
Spearman’s rank correlation and scatter plot were used to explore the relationships between PTB cases and incidence, meteorological factors, and air pollutants. In generalized linear model (GLM), we included meteorological factors, humidity, and wind speed, which correlated with PTB cases with a P value less than 0.4 in order to minimize the loss of information. Other meteorological factors, temperature and pressure, were excluded because they had little correlation with PTB cases (P > 0.4) but strongly correlated with various air pollutants (|r| > 0.7), indicating collinearities between them(17, 25). We adopted a quasiPoisson regression model to combine a GLM:
Yt ~ quasiPoisson (µt)
Log (µt) = α + β1 (air pollutant) + ns (humidity, df1) + ns (wind speed, df2) + β2 (month)
Yt represents the number of monthly PTB cases or incidence, and µt is the expected value of Yt. The meteorological factors of humidity and wind speed were controlled by a natural cubic spline function (ns) with three degrees of freedom (df1 and df2) in accordance with previous reports (26, 27). The variable “month” was used to control the impact of months. Air pollutants with a P value less than 0.15 in single-pollutant model were entered into multi-pollutant analyses.
We also established distributed lag non-linear models (DLNM) (28) to explore the associations between air pollutants and PTB cases. The variables in DLNM were similar to those in GLM except that a cross-basis function for air pollutants, as well as a “time” variable for controlling long-term trends, were used. In cross-basis functions, “ns” and “poly” functions were applied to fit the exposure-response and lag-response relationship (28). Given that the incubation period of tuberculosis is typically no more than two years (median, 15.38 months) (29–31), the maximum lag was set to 18 months via exploratory analysis (31). To evaluate the effect of air pollutant exposure on PTB cases, the median concentration of each pollutant was set as a reference (32), and the relative risk (RR) and cumulative RR for a 10-unit increase in the concentration of each air pollutant were calculated. The DLNM model is as follows:
Yt ~ quasiPoisson (µt)
Log (µt) = α + β1Tt, l + ns (humidity, df1) + ns (wind speed, df2) + β2 (month) + β3 (time)
Tt, l represents the cross-basis function for air pollutants.
Sensitivity analyses were conducted to confirm the robustness of our results: (1) change the degrees of freedom (1, 2, and 4–8 df) in the “ns” function of humidity and wind speed variables in GLM; (2) change the maximum lag of 18 months to 12 or 6 months in DLNM. The analyses were performed with “dlnm” packages in R software (version 4.1.1). All P values were two-sides and a P value less than 0.05 was considered statistically significant.