Summary of Main Findings

In accordance with the research hypothesis, we have found statistically significant, positive associations between daily hospital admissions and deaths due to respiratory and circulatory diseases and the measured levels of various parameters of air pollutants. In the single pollutant model, O3 and NO2 had the highest relative risks for hospital admissions for respiratory diseases in general and asthma in particular. The associations between hospital admissions and all four air pollutants, in terms of magnitude of relative risk, were largest for asthma, followed by respiratory diseases, total (respiratory and cardiovascular diseases) admissions and circulatory diseases (where RR was marginally significant for RSP). RSP and O3, however, were negatively associated with acute myocardial infarction, though with marginal significance. Risks of death from respiratory diseases were highest for O3, NO2, and RSP, in decreasing order of magnitude, whereas risks of cardiovascular mortalities were significantly (but only marginally) increased for RSP only.

When all four pollutants were analyzed in the multiple pollutant model, RRs were generally smaller in magnitude compared to the single pollutant model. SO2 was insignificant for all ages combined, but became significant in those aged 65 and above for respiratory and combined admissions. For respiratory diseases admissions and total (respiratory and cardiovascular diseases) admissions, synergistic effect between O3 and NO2 was observed. By contrast, the interaction of RSP with O3 was antagonistic for respiratory diseases admissions.

Validity of Model

When validated against the 1996 (first half-year) dataset, the model developed using the 1994-95 dataset was found to be satisfactory, as the "predicted" values coincided broadly with the peaks and troughs of the "observed" data in the first half-year of 1996. The model fitted especially well with data for the first two months of 1996, but under-estimated the observed values in the remaining four months for all pollutants. This might have been due to the absence of a "year-effect" indicator for the 1996 (first half-year) data. The long term upward trend of hospital admissions could not therefore be adjusted for in the 1996 data when fitted to the 1994-95 model, leading to an underestimate of the number of admissions in 1996.

Comparison with Other Studies

Many epidemiological studies have been reported documenting health effects of major air pollutants at concentrations below existing guidelines and standards (Brunekreef et al., 1995). Earlier studies, mostly in the United States, were conducted independently. More recent work in Europe was designed according to a standardized protocol (APHEA protocol). This approach provides a basis for comparison between different countries. However, it should be cautioned that a direct comparison of numerical risk estimates is not appropriate, due to the following: differences in the primary datasets (types and levels of air pollutants, numbers of hospital admissions, sub-sets of disease codes used), and specific modelling assumptions and procedures. The following summarizes key findings in time series studies of acute health effects of various air pollutants.

In the United States, the role of particulates has been the primary focus of research in recent years. Significant adverse effects of particulates on health outcomes, measured by daily mortality and hospital admissions for respiratory and cardiovascular diseases have been reported in ecological studies (Dockery et al., 1994; Samet et al., 1996; Schwartz, 1996; Schwartz et al., 1992; Schwartz et al., 1995). Independent analyses using the same source of data (Philadelphia data) over a longer time period have led to different conclusions (Moolgavkar, Luebeck, Thomas & Anderson, 1995). This highlights the dependence of the study results on the nature and characteristics of the dataset used.

In Europe, within the APHEA study, positive and significant associations between particulates and respiratory mortality or hospital admissions due to respiratory diseases have been reported in France, Greece, Italy and Spain (Zmirou et al., 1996; Dab et al., 1996; Touloumi, Samoli & Katsouyanni, 1996; Vigotti, Rossi, Bisanti, Zanobetti & Schwartz, 1996; Sunyer, Castellsague, Saez, Tobias & Anto, 1996), but not in Britain, Slovak Republic, Finland and The Netherlands (Ponce de Leon, Anderson, Bland, Strachan & Bower, 1996; Bacharova et al., 1996; Ponka & Virtanen, 1996a; Schouten, Vonk & de Graaf, 1996).

The magnitude and statistical significance of the associations between air pollutants other than particulates and measures of health outcome also varied between studies. For ozone, significant positive associations with hospital admissions for respiratory diseases were reported in Britain (Ponce de Leon et al., 1996) and with hospital admissions for asthma and ischaemic heart diseases in Finland. (Ponka et al., 1996b; Ponka et al., 1996a) For SO2, significant associations with total mortality were reported in Germany (Spix et al., 1996), with mortality and hospital admissions for respiratory diseases in Italy (Vigotti et al., 1996), with respiratory and/or cardiovascular mortality in France (Zmirou et al., 1996; Dab et al., 1996), with total and cardiovascular mortality in Poland (Wojtyniak & Piekarski, 1996), and with total, respiratory and cardiovascular mortality in Spain (Sunyer et al., 1996).

For NO2, most participants of the APHEA project reported no significant associations with health outcomes. Positive associations with elderly mortality and cardiovascular mortality in Spain and with asthma in France (Paris) were the only exceptions (Sunyer et al., 1996; Dab et al., 1996). Indoor exposure to NO2 has not been shown to be associated with respiratory illnesses in children in a cohort study by Samet (1993).

No significant health effects were found for SO2 and TSP in the Slovak Republic in Eastern Europe (Bacharova et al., 1996). Inconsistent and conflicting results were reported in Polish cities (Wojtyniak et al., 1996). Significant but negative associations have also been reported, e.g., SO2 and NO2 in Amsterdam where levels of air pollutants were relatively low (Schouten, Vonk & de Graaf, 1996).

In this study, we have demonstrated consistent and significant positive association between concentrations of all four air pollutants (NO2, SO2, RSP and O3) with hospital admissions for respiratory diseases in general and asthma in particular. All air pollutants except SO2 were also associated with respiratory mortalities. In the Interim Report, a simple Poisson regression model was used, without correction for serial correlation and overdispersion. Using the logistic regression model (Williams' Method) in this Report, the results were broadly similar. Compared to the simple Poisson model, slightly higher RR for the respective pollutants were found, but the confidence intervals of the RR were broader. Samet et al. (1995) used the "iteratively weighted and filtered least-squares model" by Zeger to correct for serial correlation and overdispersion. When comparing this with the simple Poisson model, he found that the estimated regression coefficients were insensitive to such corrections. The standard error derived from the simple Poisson model was under-estimated by 10% - 30% (and hence the statistical significance was over-estimated) when no serial correlation and overdispersion was assumed. Our estimates using Williams' model, compared with the simple Poisson model (adopted in the Interim Report), concurred with these findings by Samet et al. (1995). This means that, even accounting for overdispersion, our findings of significant associations between air pollutants and health effects are likely to be true.

One possible reason for our demonstration of highly significant findings is that the daily mean numbers of hospital admissions for respiratory and cardiovascular diseases were much higher than in other cities (about 7 times that in Milan, almost twice that in Paris, 30% higher than in London). However, compared to many studies, our time-series period was relatively short.

In respect of the magnitude of the relative risks, our findings (RR=1.39) for O3 were much higher than that for The Netherlands (RR for O38 hr max lag2 = 1.043 in summer) (Schouten et al., 1996). For SO2, our findings (RR = 1.13 and 1.25 for respiratory diseases admissions and asthma respectively) were also higher than in Italy, (RR=1.04, 1.05 for age groups 64 and 15-64 respectively for respiratory diseases admissions) and in France, (RR = 1.042 and 1.175 for respiratory diseases admissions and asthma respectively) (Vigotti et al., 1996; Dab et al., 1996). For RSP and NO2, our results were also higher than risk estimates presented in the APHEA and United States studies. One possible reason for our findings of higher risks for all pollutants is that, contrary to most Western cities, a significant proportion of the population in Hong Kong live in close proximity to the sources of emission of air pollutants, such as along motor highways, busy roads and near industrial premises. As mentioned earlier, one must caution against a direct, numerical comparison of risk estimates between individual studies due to variations in the study design, data quality, modelling, the range of parameter levels (e.g., air pollutant levels and meteorological variables) and the choice of lag in the model.

As with studies elsewhere, inconsistencies were found in our case, where a marginally significant reduction in risk of acute myocardial infarction (AMI) was observed for RSP and O3. While there is no pathophysiologic basis for this finding, the mean number of daily hospital admissions for AMI was only four, compared to 101 daily admissions for cardiovascular diseases. Hence, statistical instability cannot be ruled out.

Single Pollutant versus Multiple Pollutant Models

The issue of single pollutant versus multiple pollutant models poses one of the most difficult problems for environmental epidemiologists. In a multiple pollutant model, one can theoretically assess the effect of one pollutant given the presence (or adjusting the effect) of the other pollutants. However, collinearity between air pollutants is considerable, as exemplified in the high correlation coefficient (0.79) between NO2 and RSP in our 1994 - 1995 data. This implies that the use of a multiple pollutant model may render the results statistically unstable. Schwartz (1996) commented that the inclusion of several collinear pollutants in the same model risked letting the "noise" choose which pollutants were significant. To study the independent effects of the pollutants, he chose a location (Tucson, Arizona) where the PM10 levels were poorly correlated with O3 and SO2 (Schwartz, 1997). Earlier air pollution research, which focused on single pollutants, were "largely driven by the regulatory imperative" (Moolgavkar et al, 1997). However, there are shortcomings in the use of the single pollutant model. First, it is not possible to identify a particular pollutant as the principal cause of the measured effect. Second, one cannot examine the combined effects of more than one pollutant, given that they individually do exert some health effect in animal and human studies. Lastly, the partial b are usually overestimated compared to the multiple pollutant model. Moolgavkar emphasized that "the focus (of research) must shift from individual pollutants to the consideration of air pollution as a complex mixture" (Moolgavkar et al, 1997). The approach recommended in the APHEA protocol, is to fit the regression model into different strata (e.g., centiles) of levels of air pollutants. Another method is to use factor analysis, a statistical technique in which air pollutants can be grouped into a number of factors according to their correlations with one another.

In this study, we applied Ridge regression (a method which dealt with the collinearity problem) but found the results were comparable to those without using this method. This comparability of results indicates the stability of our multiple pollutant model. Another indication of model stability is the finding that the RRs were broadly in the same direction (but smaller in magnitude) as those in the single pollutant model. One important advantage of the multiple pollutant model is the ability to study interactions between pollutants.

The synergistic effect between O3 and NO2 might be explained by the fact that both were oxidizing pollutants and thus potentiated the effect of each other. By contrast, the antagonistic interaction of RSP with O3 could be due to the chemically reducing properties of RSP which partially neutralised the damaging effect of the highly reactive and oxidizing agent O3.

Age group-specific analysis showed that infants and children below the age of five and those aged 65 and above were at higher risks compared to the other (5-64 years) age group. While SO2 appeared to affect the elderlies selectively (they were probably more susceptible), the RR was not significant among infants and children, who might be less exposed to this predominantly outdoor pollutant.