MATERIALS AND METHODS

Mortality data

Daily mortality data between 1995 and 1998 were obtained from the Census and Statistics Department. Mortality data covered all deaths reported in Hong Kong, and were coded according to the 9th revision of the international classification of diseases (ICD). Daily time series datasets were constructed for mortalities from “all diseases of the respiratory system” (ICD 461-519) and its subsets, chronic obstructive pulmonary diseases (COPD, ICD 490-496) and pneumonia and influenza (ICD 480-7). Daily datasets were also constructed for mortalities from “all diseases of the cardiovascular system” (ICD 390-459) and its subsets, ischaemic heart disease (IHD, ICD 410-414) and cerebrovascular disease (ICD 430-8).

Air quality and weather data

Air pollution data between 1995 and 1998 were obtained from the Environmental Protection Department. Hourly concentrations of four air pollutants: SO₂, NO₂, PM₁₀, and O₃ were monitored in eight monitoring stations interspersed in different districts of Hong Kong using pulsed fluorescence, gas phase chemiluminescence, tapered element oscillating microbalance, and ultraviolet absorption, respectively.²⁰ Twenty four hour mean concentrations of NO₂, PM₁₀, and SO₂ were calculated. As the formation of O₃ is dependent on sunlight, a daytime (0900–1700) 8 hour mean concentration of O₃ was used for analysis.²² In view of the low ambient concentration of CO in Hong Kong, monitoring of CO has been confined to only one station in recent years. We have therefore excluded CO from our study. Daily mean temperatures and relative humidity for the same period were obtained from the Hong Kong Observatory.

Statistical analysis

A Poisson regression model was constructed in accordance with the air pollution and health: the European approach (APHEA) protocol.²² The following terms were included to construct the core model: day of the time series (t), days of the week, trigonometric functions to control for seasonal variations (sin 2kπt/365 and cos 2kπt/365, where k=1, 2, 3, 4, and 6, represent cycles of 12, 6, 4, 3, and 2 months respectively), temperature and humidity.^21–23 To control for overdispersion, the covariance matrix was modified by multiplying the dispersion parameter ϕ, and the scaled deviance and log likelihoods used in likelihood ratio tests were divided by ϕ. The function obtained by dividing a log likelihood for the Poisson distribution by a dispersion parameter is an example of a quasi-likelihood function.^24,25 To control for autocorrelation, the autocorrelation functions plot of the residuals was examined and significant terms, up to lag day 7, were retained in the model.^23,26 Different pollutants may affect mortality with variable time lags. To test the influence of an individual pollutant on mortality, a single pollutant model was constructed for each of the four pollutants by adding its daily concentration to the core model. The best fitting lag period of each pollutant was found by testing its concentration on single lag days and on cumulative lag days (moving averages). Lag days from 0 (same day) to 5 (5 days before) were tested for O₃; lag days from 0 to 3 were tested for the other three pollutants. The lag period used to construct the model that gave the smallest Akaike's information criterion value was then selected for each pollutant.²⁷ The relative risk (RR) and its 95% confidence intervals (95% CIs) for a 10 μg/m³ increase in the concentration of each pollutant were then calculated. To study the combined effects of the pollutants, multipollutant models were constructed. Firstly, all four pollutants (as continuous variables) were entered into the core model irrespective of significance. Non-significant (p≥0.05) terms were then successively removed from the model by a process of backward elimination, until only the significant terms remained. Owing to the non-trivial correlation of some pollutants, a pairwise approach was adopted as in the APHEA protocol if more than one pollutant seemed to be associated with the outcome, the associations with one pollutant stratified by the level of the other pollutant was sought.²² With this approach, we explored the interaction between different pollutants by analysing each pair of pollutants using one as a continuous variable and the other as high and low concentrations dichotomised by the median, and their interaction term (between a pollutant and a high concentration of another pollutant). The influence of the concentration of a pollutant (expressed as a continuous variable) on a high and low concentration of another pollutant was then studied.^22,23 Interaction was considered to be present when the p value of the interaction term was less than 0.05. As an ambient temperature has a significant influence on mortality,¹⁶ we also explored the possible interaction of cold seasons (December to February, when the mean monthly temperature was below 20°C) with the effects of the air pollutants by introducing a dummy variable for season in the model.²² To investigate dose-response relations, we derived from the single pollutant models the respective RRs of daily deaths by pollutant concentrations categorised into deciles with the lowest decile as reference.²

Harvesting, a hypothesis that a high death toll on one day due to air pollution would deplete the pool of vulnerable people and result in fewer deaths on succeeding days, was investigated according to the procedure described by Spix et al.²⁸ In short, the coefficients of the interaction between the mean daily mortalities on the previous k days (where k ranged from 2 to 21) and each of the pollutants were checked. Harvesting was considered to be present if any of the coefficients was significantly negative.

RESULTS

During the 4 years, there were 128 229 deaths of which 58 347 (46%) were caused by respiratory and circulatory diseases. Table 1 shows the daily number of deaths by cause and the pollutant concentrations. The Pearson's correlation coefficients between air pollutants and meteorological variables are presented in table 2. After fitting the core model, residuals were plotted against the predicted values for diagnostic checking. No cyclical pattern could be discerned in the residual plot. Table 3 summarises the RRs of daily deaths for different disease categories. The lag days for the pollutants that fitted the model best for all respiratory diseases varied for different pollutants, ranging from 1 to 2 single lag days and 0 to 1 cumulative lag days. For all cardiovascular diseases, the “best lag” ranged from 0 to 2 single lag days and up to 0 to 2 cumulative lag days. The overdispersion parameter (ϕ) was 1.14 and 1.09 respectively for respiratory and cardiovascular diseases. The autocorrelation coefficients (r) of the models ranged from 0.005 to 0.052. For all four pollutants, there was a significant increase in mortality for respiratory diseases that ranged from 0.8% to 1.5% per 10 μg/m³ increase in concentration of pollutant. For COPD, the risks were higher and significant for all pollutants except SO₂. For pneumonia and influenza, the RRs were significant for NO₂ and SO₂ only. For all cardiovascular diseases and cerebrovascular diseases, the RRs for the best fitting lag days of all four pollutants were non-significant. A significant increase in mortality from IHD, ranging from 0.9% to 2.8%, was associated with a 10 μg/m³ increase in the concentration of all four pollutants.

In the multipollutant model for all respiratory mortalities, SO₂ (RR=1.015) and O₃ (RR=1.010) remained after eliminating the non-significant pollutants (table 4). In the three and four pollutant models, O₃ was the only significant pollutant. The RR of O₃ was stable in all three models, and similar in magnitude to that in the single pollutant model. For COPD, O₃ was the only significant pollutant in the two, three and four pollutant models, with RRs slightly lower than in the single pollutant model. For pneumonia and influenza, SO₂ was significant after eliminating non-significant terms from the model. For IHD mortalities, NO₂ was the only significant pollutant in the four, three, and two pollutant models, its RRs being smaller than that in the single pollutant model.

No significant interaction between any pair of pollutants was found in the pairwise analyses, or between any pollutant and the cold season. None of the coefficients of the interaction terms between the pollution concentration and the mean mortalities of the previous days from lag day 2 up to day 21 were significantly negative, suggesting the absence of a harvesting effect. When RRs of deciles of the pollutant concentrations were derived from each single pollutant model, with the lowest decile as reference, an upward trend in RRs was found for all pollutants from the lowest to the highest deciles (fig 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

A regression line is fitted to show the upward trend of the RRs. RRs were derived from the Poisson regression models in which the concentrations of the individual pollutants were replaced by their respective deciles. See statistical analysis in the methods section.

DISCUSSION

This study provides additional information for our previous study on hospital admissions,²¹ and the many time series studies on air pollution and mortality in temperate countries.^{1–11,13,15,17–19,28,29,33–35,38,39} Although hospital admissions are influenced by socioeconomic and personal factors, mortality is the ultimate and most serious of all health outcomes. As explained earlier, both the mortality and air pollutant datasets were of reasonable quality. Our statistical methods followed the APHEA protocol, which facilitated the comparison of results. Significant associations were found with respiratory mortalities for all four pollutants in the single pollutant models. For the different respiratory diseases, the best lags ranged from 1 to 2 single lag days and 0 to 4 cumulative lag days. These lag periods were statistically chosen by model fitness. The durations of the lag periods, however, seemed reasonable, if we assume a short interval (in terms of several days) between exposure and death, which varies with different diseases. It is also reasonable to assume that mortality might be associated with cumulative exposure over several days. The risk estimates for respiratory mortalities, at 0.8% to 1.5% per 10μg/m³ increase in pollutant concentration were similar in magnitude to results reported elsewhere.³⁰ These mortality risks were, however, lower than risks of hospital admissions we reported previously (1.3% to 2.2% per 10 μg/m³ increase).²¹ Of the respiratory diseases, the RRs for mortalities from COPD, the most sensitive mortality to air pollution, were higher for O₃ and NO₂ (at increases of 3.4% and 2.3% respectively). The RRs of cardiovascular mortalities were non-significant for all four pollutants, by contrast with our previous findings for hospital admissions for cardiovascular diseases.²¹ Mortalities for IHD were significantly associated with all four pollutants, suggesting that the associations of air pollution were disease specific. Of cardiovascular diseases, the RRs of deaths from IHD (at 1.028 and 1.024 respectively) were the highest for NO₂ and SO₂. However, a direct comparison of the magnitude of the RRs among the air pollutants can be misleading, because of differences in molecular weight. By contrast with our findings, a pooled analysis among European countries showed equivocal RRs of cardiovascular and respiratory deaths for NO₂ but significant RRs for black smoke, SO₂,and O₃.²⁹ The summary RRs for SO₂ and O₃ in western European cities are similar in magnitude to ours. In Hong Kong, the mean concentrations of O₃, NO₂, and PM₁₀ are generally higher, but the concentrations of SO₂ are much lower than in European cities. Our findings of a significant RR of IHD for SO₂ suggest that the threshold has not been reached at the current SO₂ concentrations.

Owing to the high correlation between pollutants, it is not clear whether the observed effect of one pollutant in the single pollutant model represents that of an underlying pollutant. Much research focus has been given to the respirable fraction of particulates, and more recently, fine particulates (PM _2.5).^31,32 In the multipollutant model for both respiratory diseases and IHD, no significant effect for PM₁₀ could be discerned. It should be noted that in some studies in the United States and Europe where the effects of particulates were significant, single pollutant models were used.^{2–6,12–15,17,29} These results agreed with that of our single pollutant model for PM₁₀. Our findings of significant associations of SO₂ and O₃ with respiratory deaths are in accord with those in many European cities and elsewhere.^{4,5,8,12,15,17,33–36} The problems of collinearity among pollutants in a multipollutant model are well recognised. In our data, the concentrations of NO₂ and PM₁₀ were highly correlated on the same days (table 2), and on the best lag days that were used in the multipollutant model (r=0.73). The correlation between all the other pairs of pollutants ranged from trivial to moderate (Pearson's r ranged from −0.07 to 0.45 for the best lag days). Hence, it is statistically valid to include SO₂, O₃, and either of NO₂ and PM₁₀ in the models. For all respiratory mortalities and COPD, the RR for O₃ was remarkably stable in all the models. A significant association with O₃ for respiratory mortality in our multipollutant model is biologically plausible. Firstly, O₃ is a highly toxic oxidant pollutant with known adverse effects on the respiratory system.³⁷ Secondly, the concentration of O₃ in Hong Kong is higher than in many cities in Europe and the United States, possibly related to the abundance of sunshine, even in winter months. In our earlier study on hospital admissions, O₃ was significantly associated with admissions for both respiratory and cardiovascular diseases.²¹

The high ambient concentration of NO₂ in Hong Kong is a possible explanation of our finding of a significant association with mortalities from IHD. Significant associations between NO₂ and all cardiovascular mortalities have been reported in some studies.^4,10,13 However, the relations between NO₂ and cardiovascular mortalities were inconsistent in a meta-analysis of pooled data from four European cities.²⁹ The reasons for the lack of association of PM₁₀ with any respiratory or cardiovascular mortalities in a multipollutant model are not clear. The effects of PM₁₀ might have been masked by those of NO₂ and O₃—the concentrations of both were fairly high. Neither could we explain why SO₂, rather than PM₁₀, was associated with deaths from pneumonia and influenza, especially when the mean concentration of SO₂ was relatively low and that of PM₁₀ was fairly high. The chemical composition of particulates might be related to the effects on health. In Hong Kong, diesel vehicle exhaust is a major source of PM₁₀, although crustal dust and marine aerosols are important sources as well.²⁰

No interaction was detected in this study between any two pollutants, whether analysed in pairs or when all four pollutants were considered together. Significant interactions between particulates and SO₂ have been reported in Athens and Lyon in the APHEA mortality studies.^15,33 The discrepancies might be due to differences in concentrations of pollutant between Hong Kong and these cities. The absence of interaction with the cold season could be related to the relatively mild winters in Hong Kong, where the mean monthly temperatures from December to February ranged from 15° to 19°C and the mean daily temperature rarely drops below 10°C. When analyzed by deciles of pollutant concentrations, a dose-response relation was found for all pollutants and respiratory mortalities, with no evidence of a threshold effect.

The role of CO on cardiovascular mortalities has been investigated elsewhere. Touloumi et al reported a positive association between ambient CO concentrations and daily mortality in Athens,³⁸ but this association was reduced considerably after including CO and black smoke or SO₂ in the same model. Carbon monoxide was not considered as a copollutant in other time series analyses.³⁹ Owing to the insufficiency of data, we did not evaluate the effect of CO, which might be relevant to cardiovascular mortalities. The ambient concentration of CO in Hong Kong is low. The mean 24 hour concentration was 800 μg/m³, less than one tenth of the local air quality objective of 10 000 μg/m³.²⁰

The absence of a harvesting effect of respiratory and cardiovascular mortalities within a 3 week period suggests that the shortening of life is likely to be substantial. Schwartz suggested that deaths from chronic obstructive airway diseases were mostly brought forward by a few months, whereas some (but not most) deaths from pneumonia might be brought forward by a few days.⁴⁰ We have not investigated the harvesting effect of specific causes of death on different time scales.

In conclusion, we found significant associations of all four air pollutants with mortalities from respiratory disease and IHD, the effect size being strongest for COPD. Although the multipollutant model selected SO₂ and O₃ as significant pollutants for respiratory deaths and NO₂ for deaths from ischaemic heart disease, other pollutants such as PM₁₀ might play an important part. The joint effects of the air pollutants are still poorly understood and comparative studies in cities with different pollutant profiles are warranted. The ecological design of the study precludes the inference of cause and effect. As the current ambient concentrations of these pollutants seem to be above the threshold, some health benefits should be noticeable if air pollution is reduced. A reduction in morbidity and mortality after the implementation of an intervention programme will add evidence to the hypothesis of a causal link between air pollution and ill health.