How air pollution altered the association of meteorological exposures and the incidence of dengue fever

This study was carried out in Guangzhou, the capital city of Guangdong Province. It is the most affected city over the country in recent years, with 41 941 cases reported during 2004–2018, accounting for 73% burden in Guangdong and approximate 60% burden in China (www.phsciencedata.cn/Share/index.jsp). Detailed information on each dengue case including the date of disease onset, case category (clinical diagnosis or laboratory confirmation) and basic demographics is compulsively reported to the China National Notifiable Disease Surveillance System within the first 24 h after clinical diagnosis or laboratory confirmation by hospitals. We collected case data for the period between January 2015 and December 2019 from the Guangzhou Center for Disease Control and Prevention, and computed the daily number of cases as the primary outcome of the current study.

Meteorological exposures including daily average temperature (°C), cumulative precipitation (mm), average relative humidity (%), and average wind speed (m s⁻¹) were obtained from the China Meteorological Administration (http://data.cma.cn) for the period from 1/11/2014 through 31/12/2019. We also collected the data of daily concentration (μg m⁻³) of five main air pollutions from the China National Environmental Monitoring Centre (https://air.cnemc.cn:18007), including PM₁₀, PM_2.5, SO₂, NO₂, and O₃ for the same period. The daily observations were averaged across the monitoring sites in the study area as described in figure S1.

Mosquitoes favor dense vegetation (e.g. trees and bushes) to rest and breed since the vegetation could provide microclimate with less evaporation, higher humidity and lower wind speed [21–24]. Therefore, normalized difference vegetation index (NDVI), an established indicator of vegetation, was the most commonly used to represent the coverage of mosquito-favored habitats [25, 26]. The NDVI ranges from 0 to 1, with a greater value representing a higher vegetation coverage while a smaller value representing a higher water coverage. Usually, a value of 0.4–0.6 represents a median vegetation coverage [27]. We downloaded the Terra Moderate Resolution Imaging Spectroradiometer vegetation index—NDVI data (MOD13Q1, version 6, generated every 16 d at 250 m spatial resolution) from the Land Processing Distributed Active Archive Center (https://lpdaac.usgs.gov) for the same period as the atmospheric data. The average of the values across the cells covered by Guangzhou was calculated as the regional mean by using raster package in R software. The annual population of Guangzhou was collected from the Guangzhou Statistical Yearbook (http://tjj.gz.gov.cn) to adjust for the potential impact of background population changes.

Descriptive analysis was used for the grouped description of dengue cases, meteorological exposures, air pollution and NDVI. Besides the non-case days, we also defined low-incidence days (⩽4 cases) and high-incidence days (>4 cases) where 4 was the median number of cases on days with case reported. Means and standard errors for the whole study period and each group were acquired, ANOVA or Kruskal–Wallis H test were selected for the intergroup difference test according to the assumptions of normality and homogeneity of variance.

Generalized additive models (GAMs) with Poisson distribution were used to evaluate the meteorology-dengue association. The Poisson rather than quasi-Poisson distribution was used since we did not detect an over-dispersion issue with our data using the performance package in R software. We regressed daily counts of dengue against meteorological exposures while adjusting for main air pollution and NDVI. The long-term trend and seasonality were also fitted using a natural cubic spline (ns) of time with seven degrees of freedom (df) per year as commonly used in previous studies [28–30]. All the four meteorological factors were included in models. Besides, we also used the diurnal temperature range (DTR), maximum and minimum temperature instead of the average temperature to capture a more comprehensive picture of the impact of temperature. However, we only included three main air pollutions (i.e. PM_2.5, SO₂, O₃) without strong correlations with each other (Pearson's rho < |0.8|, figure S2) to avoid collinearity. The logarithm of the annual population was included as an offset. The model was specified as follows:

$$\begin{align}{{{Y}}_{{t}}}\;{{\sim Poisson}}\left( {{{{\mu }}_{{t}}}} \right)\end{align}$$

$$\begin{align}{\text{log}}\left( {{\mu _{\text{t}}}} \right) & = \alpha {{ + }}s\left( {{\text{TEMP}}} \right){{ + }}s\left( {{\text{PRE}}} \right){{ + }}s\left( {{\text{HUMD}}} \right)\nonumber\\ & \quad + s\left( {{\text{WIND}}} \right){{ + }}{\beta _{\text{1}}}{\text{P}}{{\text{M}}_{{\text{2}}{\text{.5}}}}{{ + }}{\beta _{\text{2}}}{\text{S}}{{\text{O}}_{\text{2}}}{{ + }}{\beta _{\text{3}}}{{\text{O}}_{\text{3}}}\nonumber\\ & \quad + {\beta _{\text{4}}}{\text{NDVI + ns }}\left( {{\text{time}},{\text{df = 7}} \times {\text{5}}} \right)\nonumber\\ & \quad {\text{ + offset}}\left( {{\text{log}}\left( {{\text{POP}}} \right)} \right)\end{align}$$

where ${\mu _t}$ refers to the expected daily counts of dengue cases on day t; $\alpha$ is the intercept; $s\left( . \right)$ is the smoothing function for meteorological exposures. We used the cubic spline with three df as was commonly used in literature [31–33]; ${\beta _i}$ indicates the coefficient of the corresponding term.

Since existing evidence mainly focused on the adverse effect of air pollution on the behaviors of adult mosquitoes, we selected 28 d as the maximum lag days to focus on the stage after the infection of mosquitoes, covering the extrinsic incubation period (∼15 d), the possible transmission period (0–7 d), and the intrinsic incubation period (∼6 d) of dengue virus [13, 34–36]. Specifically, we considered five scenarios of the cumulative effect: lag 0, lag 0–7, lag 0–14, lag 0–21 and lag 0–28. In each scenario, we computed the corresponding moving average of meteorological exposures, the concentration of air pollution, and NDVI for modeling. The scenario which minimized the Akaike information criterion was used to assess the relationships and for further analyses. If the relationship between a meteorological variable and dengue incidence showed a nonlinear trend, we performed a piecewise linear fitting according to the inflection point (the values of meteorological variables at which the effect reached a maximum or minimum). We reported the risk ratio (RR) (i.e. ${\text{RR}} = \exp \left( {{\beta _i}} \right)$) for each unit increase in an exposure.

To evaluate whether the associations of meteorological exposures with dengue incidence were modified by air pollution, we included an interaction term of each meteorological exposure and dichotomous air pollution stratum indicators (high level: concentration higher than the median concentration over the entire study period; low level: concentration lower than the median) into models. The following medians were used for strata definition: PM_2.5: 33.10 μg m⁻³; SO₂: 10.42 μg m⁻³; O₃: 89.92 μg m⁻³. Furthermore, we integrated the three dichotomous air pollution variables into an eight-level categorical indicator (table S1) to evaluate the combined effect of these factors as well as its interaction with the meteorological exposures. P for interaction was also computed.

To confirm that our results were robust, we performed sensitivity analyses by (a) changing the degrees of freedom for the spline functions from 3 to 4 and 5 respectively; (b) including each of the three pollutions in the primary models separately rather than simultaneously. We further applied a meta-regression to compare the results between sensitivity models and the primary models. All statistical analyses were conducted using R software (version 4.0.5).

Table 1 presents the characteristics of dengue cases and atmospheric factors between 2015 and 2019 (1826 d). A total of 3712 dengue cases were reported in Guangzhou during this period. The time series displays a long-term trend and seasonality with more cases reported in warmer months and less cases in other months (figure S3). The averages for meteorological exposures over the previous week were 22.37 °C (average temperature), 6.30 mm (cumulative precipitation), 81.29% (average relative humidity), and 2.08 m s⁻¹ (average wind speed), respectively. The average temperature over the previous week of non-case days was lower than that for days with case reported (P< 0.001), with means of 20.90 °C (non-case days) vs 25.60 °C (days with ⩽ 4 cases) and 25.57 °C (days with > 4 cases). Days with more cases (> 4 cases) tended to have a lower precipitation over the previous week than less-case (⩽ 4 cases) days and non-case days (4.24 mm vs 6.19 mm and 6.77 mm, P < 0.001). The average concentrations of PM_2.5, PM₁₀, NO₂, SO₂, O₃ over the previous week were 33.21 μg m⁻³, 52.79 μg m⁻³, 41.30 μg m⁻³, 10.23 μg m⁻³ and 90.84 μg m⁻³, respectively. Days with cases reported tended to have lower concentrations of PM_2.5, PM₁₀, NO₂ and SO₂ than non-case days (P < 0.001), but have higher concentrations of O₃ than non-case days (P < 0.001). As for NDVI, the average during our study period was 0.49, days with cases reported had a greater average of NDVI compared with non-case days (P < 0.001).

We chose the best cumulative scenario of 0–7 lag days for our models (table S2). Figure 1 illustrates the exposure-response relationship between meteorological exposures and daily dengue cases. The effects of average temperature (figure 1(A), P < 0.001) and cumulative precipitation (figure 1(B), P < 0.001) on dengue were non-linear (approximate logarithm-shaped and approximate U-shaped shaped) with cut-offs of 27.1 °C and 20.3 mm, respectively. However, the effects of both average relative humidity (figure 1(C), P = 0.908) and average wind speed (figure 1(D), P = 0.770) was not significant. Besides, the effects of minimum and maximum temperature had a similar pattern with the effect of average temperature (figures S4(A) and (B)), but the effect of DTR was statistically insignificant (figure S4(C)). According to figure 1, the effects of average temperature and cumulative precipitation could then be divided into two sets of linear associations.

Zoom In Zoom Out Reset image size

As shown in table 2, each degree increase in average temperature was associated with a 7.3% (95% CI: 3.5%–11.2%) increase in the daily number of dengue cases when the temperature was below 27.1 °C. Additionally, the effect of higher temperature (above 27.1 °C) showed a negative trend, but was not significant (P = 0.232). As for cumulative precipitation, each 1 mm increase in precipitation was related to a 2.7% (95% CI: 1.7%–3.8%) decrease of cases when the precipitation was below the threshold (20.3 mm). On the contrary, when the precipitation was above the threshold, each millimeter rise in precipitation was positively related to dengue, that is a 4.7% (95% CI: 1.2%–8.3%) increase in cases.

As depicted in figure 2, effect modification by PM_2.5 and SO₂ on the associations of meteorological exposures with dengue was observed, but the effect modification by O₃ was not significant. We found that each degree increase in the average temperature below 27.1 °C was associated with a greater RR of 1.11 (95%CI: 1.06–1.17) in high SO₂ conditions compared to low SO₂ conditions (figure 2(A)). However, when the average temperature was above 27.1 °C, we observed an inverse effect modification by SO₂, that is 1 °C rise in temperature was associated with a smaller RR of 0.78 (95%CI: 0.90–1.15) in high SO₂ conditions (figure 2(B)). The effect modification by PM_2.5 on precipitation-dengue association was similar to that by SO₂ on temperature-dengue association. Each millimeter rise in the cumulative precipitation below 20.3 mm was related to a greater RR of 1.00 (95%CI: 0.97–1.02) in high PM_2.5 conditions (figure 2(C)), but 1 mm increase in precipitation above 20.3 mm was related to a smaller RR of 0.93 (95%CI: 0.83–1.05) in high PM_2.5 conditions (figure 2(D)). As for average wind speed, each 1 m s⁻¹ rise was associated with a greater RR of 1.26 (95%CI: 1.07–1.49) in high SO₂ conditions (figure 2(F)). None of the three air pollutions modified the effects of relative humidity on dengue. The estimated RRs for humidity were consistent (P for interaction > 0.05) between the low- and high-level air pollution strata (figure 2(E)). For the modification effect on the impacts of three other temperature-related factors, only the RR of maximum temperature (<24.9 °C) was greater on high PM_2.5 d (figure S5(C), P for interaction = 0.025).

Zoom In Zoom Out Reset image size

As shown in figure 3, we observed similar effect modification by the multi-pollution indicator with that by single-pollution indicators in general. Compared with good air quality days (low PM_2.5 & low SO₂ & low O₃), the associations between temperature and dengue (positive effect of low-level temperature & negative effect of high-level temperature) were strengthened on some bad air quality days; whereas, the associations between precipitation and dengue (negative effect of low-level precipitation & positive effect of high-level precipitation) were offset in some bad air quality scenarios. As for humidity and wind speed, although some of the interaction terms were significant, the risk effects of these exposures were still not significant.

Zoom In Zoom Out Reset image size

Figure S6 is a summary of the results of sensitivity analyses. The associations of meteorological exposures with dengue incidence were almost robust after a series of adjustments for the primary model. Cutoffs derived based on GAM results in different models were similar with a maximum difference of 0.7 °C and 1.1 mm, respectively (table S3). Although the effect of wind speed on dengue in the single O₃ model (RR: 1.076, 95% CI: 0.955–1.214) was a bit higher than that in primary model (figure S6(E)), the difference was not significant (P = 0.534).

We found an approximately logarithm relationship between average temperature and dengue, which peaked at 27.1 °C. We also found that increased cumulative precipitation below 20.3 mm was negatively associated with dengue. These findings are in accordance with existing epidemiological and biological evidence [11, 37–40].

As for the cumulative precipitation above 20.3 mm, it was positively associated with dengue incidence. However, this result has not previously been described. On the one hand, although mosquitoes can fly on rainy days, persistent heavy rainfall threats the survival of mosquitoes in the open air [41, 42]. On the other hand, high rainfall can reduce people's outdoor activities [43, 44]. Therefore, humans and mosquitoes tend to get into houses on days with high rainfall thus leading to a possible increase in the density of vectors (some may have been infected already by dengue virus) in a limit space, which facilitates biting activities of these vectors and hence the transmission of dengue virus.

In contrast to earlier findings, we found another inconsistency about relative humidity-dengue association. Relative humidity is recognized as a key factor affecting vectors' oviposition and egg hatching [16, 45]. An appropriate relative humidity can increase the population density of vectors, facilitating the transmission of dengue [11, 31]. However, no significant association between relative humidity and dengue was detected in our study. This discrepancy could be attributed to the difference settings of lag days in different studies. A potential interpretation for that is atmospheric factors tend to affect the entire chain of dengue transmission, including the development (egg, larva, pupa, adult), survival and behavior of mosquitoes, as well as the extrinsic and intrinsic incubation period of dengue virus. Xiao et al reported an approximate logarithm relationship between relative humidity and dengue incidence at a lag of 4 months [31]. Whereas, we found that the relationship between cumulative exposures and dengue was best fitted with a 0–7 d lag in our study, it most likely only covered the adult stage of vectors and the intrinsic incubation period of dengue virus, which were less affected by relative humidity.

Regarding the interactive analysis, we found PM_2.5 and SO₂ could modify the effects of meteorological exposures in different patterns. Firstly, in high temperature condition (above 27.1 °C), we observed a significant negative association between temperature and dengue when SO₂ concentration was high. Synergistic effect of high temperature and high SO₂ concentration may lead to this variation compared to low SO₂ concentration situation. It may shorten the lifespan of vectors and reduce their population density. A meta-analysis on experiment data revealed that high temperature (above 25 °C) could significantly shorten the lifespan of adult mosquitoes [38]. Although no direct evidence confirmed the adverse effect of SO₂ on mosquitoes, Tan et al conducted an experiment and thought acidic SO₂ might cause tissue damage in the larvae of another kind of insect—butterfly [46]. Moreover, several studies suggested that other air pollution were harmful to mosquitoes, such as the repellency effect of O₃ and the adherence effect of fine particles [19, 20, 47], which may also be a potential possibility underlying our findings.

However, the risk effect of temperature under the threshold was strengthened when SO₂ concentration was high. Similarly, in low-precipitation conditions, the negative effect of increased precipitation was not significant anymore when SO₂ or PM_2.5 concentration was high. The reasons may also relate to human and mosquito behaviors. Chen et al and An et al suggested that a terrible air quality may reduce outdoor leisure activities [48, 49]. In addition, Wilder-Smith and colleagues thought mosquitoes may be driven indoors because of the adverse effect of pollution [50]. These two reasons could result in prolonged stay indoors of humans and create more chances of exposure to infected mosquitoes. According to above information, we can update current dengue-control strategies, that is, people may strengthen environment management and individual protection to reduce contact with mosquitoes on days with bad air quality.

We also found that wind speed was positively associated with dengue in higher SO₂ condition. A possible explanation for this result may be the dilution effect of pollution by wind [51], which alleviates the adverse effect of pollution on mosquitoes. However, the effect of wind on regional air pollution concentration is complex. Several key factors, including the locations of pollutant emission source, wind direction and speed, play vital roles in this process [52–54], which may be the possible reason why only the effect modification by SO₂ was detected in the study.

Besides above findings, we also observed that in high-precipitation condition, the positive association between precipitation-increase and dengue was not significant anymore when the concentration of air pollution was high (all three pollutions, but only the interaction term of PM_2.5 was significant). However, the modification effect by O₃ was not detected on any meteorological exposures. As for the modification by multi-pollution indicator, we found a similar trend as single-pollution indicator in general, except for the modification on the effect of wind speed. Analyses of these effect modification can help us identify the environmental conditions in which dengue transmission is more easily, and can help us establish more specific early warning systems.

The limitation of this study should be acknowledged. Firstly, we only included data from Guangzhou, which may result in an increased difficulty of conclusion extrapolation. However, Guangzhou is the largest city in South China with a population of 19 million. More importantly, it reported with the majority of dengue cases in the country in recent years. It is of great public health significance to investigate the modification effect of air pollution on meteorology-dengue associations in this area. Second, similar to other ecological studies, our study may suffer from the ecological fallacy and could not confirm the causal relationship. However, our study still provides a new insight on the potential impact of atmospheric factors underlying the dengue pandemic, which may offer ideas for future experimental researches.