A STUDY OF SHORT-TERM EFFECTS OF AMBIENT AIR POLLUTION ON PUBLIC HEALTH

Summary Statistics of Levels of Air Pollutants and Meteorological Variables

Tables 5 and 5a present the summary statistics of various measures of daily concentrations of individual pollutants and meteorological variables for the study period (1994 - 95, and the first half of 1996). The mean daily values are well below the Air Quality Objectives set by the Environmental Protection Department. Mean daily levels of air pollutants are generally higher in the first half-year of 1996 than in 1994 - 95, with an increase of 3% for RSP, 5% for NO2, 14% for SO2 and 36% for O3.

The three-monthly mean of the daily concentrations of air pollutants are shown in Tables 6 to 9. Levels are generally higher in colder seasons for NO2 and RSP. A distinct pattern is observed for O3, when levels in the last quarter of each year are about twice those of the other quarters. The levels of SO2, after reaching a peak in the third quarter of 1994, show a steady decline throughout 1995 but increase slightly in 1996.

* first half-year

* first half-year

* first half-year

first half-year

Time Trends of Air Pollutants

The time trends of individual air pollutants for 1994 - 95 and 1996 (first half-year) are shown in Figures 1 - 4 and 1a - 4a respectively. For NO2 and RSP, a distinct cyclical pattern with a long wavelength is observed, and higher levels in winter. For O3 and SO2, shorter wavelengths are seen, with peaks in early spring and late autumn for the former and mid-summer for the latter. There is no obvious increasing trend for all four pollutants during the 1994 - 94 period. Levels of O3 are generally higher in the first half of 1996 compared to 1994 - 95.

Summary Statistics of Hospital Admissions

Summary statistics of daily counts of hospital admissions for respiratory and circulatory diseases in 1994 and 1995, and in the first half of 1996 are shown in Tables 10 and 10a respectively. The daily mean numbers of admissions in 1996 (first half-year) was generally higher than those in 1994 to 1995 (by 32% for respiratory diseases, 20% for cardiovascular diseases and 27% for both combined).

* Total = Respiratory diseases + Cardiovascular diseases

** Acute myocardial infarction

* Total = Respiratory diseases + Cardiovascular diseases

The total numbers of admissions for 1994-95 by age and type are shown in Table 10b. A total of 172,157 hospital admissions for respiratory and cardiovascular (circulatory) diseases were recorded in the selected hospitals during the period 1994-95 and used in the Poisson regression model while 171,829 were used (i.e., excluding 328 admissions with no record of age) in the analysis by age group

Age group

Time Trends of Hospital Admissions

Time trends of hospital admissions for respiratory and cardiovascular diseases for the period 1994 - 1995 and the first half of 1996 are shown in Figures 5 - 7 and 5a - 7a respectively. An increasing trend throughout the overall study period is evident, with a cyclical pattern more obvious in 1994. This was contributed largely by respiratory diseases, with an approximately three-month cyclical pattern, while hospital admissions for cardiovascular diseases displayed a sinusoidal pattern with a longer period. For the first half of 1996, the sinusoidal pattern was less obvious (Fig. 5a - 7a).

Correlation between Individual Air Pollutants, and between Air Pollutants and Hospital Admissions

The correlations between individual air pollutants from 1994 to 1995 and for each season are shown in Table 11. Overall, the correlation between NO2 and RSP was good, while those between RSP and O3, NO2 and SO2, and NO2 and O3 were fair. In general, the correlation between pollutants was higher in summer, except between SO2, and NO2, where the correlation coefficient was the highest in winter. The high degree of collinearity between air pollutants, in particular, NO2 and RSP cautions against the use of a multiple pollutant model in the analysis.

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

The correlations between individual air pollutants and hospital admissions for respiratory and cardiovascular diseases from 1994 to 1995 are shown in Table 12. In general, the (unadjusted) correlation was poor. Of all four air pollutants, mean NO2 and RSP showed positive and significant correlations with total hospital admissions as well as for respiratory and cardiovascular diseases. Mean and minimum daily temperatures were negatively correlated with hospital admissions.

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

Parameter Estimates using Multiple Linear Regression Model

Using data on total hospital admissions (respiratory and cardiovascular diseases), a multiple linear regression model was fitted to estimate the partial regression coefficients of variables in the core model (without air pollutants) and the coefficient of determination using SAS Program Version 6.1 (SAS Institute Inc., 1996). R2 (coefficient of determination) was 0.6609 indicating that the confounding variables alone, accounted for about two third of the variations in daily hospital admissions (Table 13). When the daily hospital deaths due to respiratory and cardiovascular diseases combined were fitted into the model, the R2 was 0.3098, indicating a poorer fit compared with the hospital admission data.

Dependent Variable : Logarithm* of total (respiratory + cardiovascular) admissions

* Logarithmic transformation was used to "normalize" the data.
** Caption of variables:

# NS: Not statistically significant

Parameter Estimates using Poisson Regression Model

The logistic procedure of the SAS 6.1 Programme (SAS Institute Inc., 1996) was used to estimate the parameters and their risk ratios in the core model (Table 14). The variables which were statistically significant in the Poisson core model were almost identical to those in the multiple linear regression model.

Response Profile

Deviance and Pearson goodness-of-fit Statistics

Model fitting and testing global null hypothesis = 0

* Akaike Information Criterion
# Schwartz Criterion
$ -2 log Likelihood

The -2 log Likelihood has a chi-square distribution under the null hypothesis (that all the explanatory variables in the model are zero), and a p value is given for this statistic. The AIC and SC statistics give two different ways of adjusting the -2 log Likelihood statistic for the number of terms in the model and the number of observations used. Lower values of the statistic indicate a better fitting model.

^ This value, ø, is an estimate of overdispersion parameter.

Analysis of maximum likelihood estimates

* Caption of variables:
t: time trend
Year: year-effect indicator
I1 - I6: days of the week (Monday to Saturday)
S1-4: sine terms for seasonality
CS1-4: cosine terms for seasonality
Holiday 1: Public Holiday
Holiday 2: Day after Public Holiday

The fitted Core Model and the Residual Plot

Figures 8 - 9 show the predicted number of daily admissions based on the multiple linear regression core model (without air pollutants), and the residuals of the model. After fitting the core model, the residuals appear to be more randomly dispersed and no obvious trend can be discerned. Figures 10 - 17 show the deviance residuals, by levels of pollutants and time (days) when each of the four pollutants is fitted into the Poisson regression model (Williams' method). The residual plots by pollutant levels and by time appear dispersed with no recognizable pattern,

Risk Estimates using the Multiple Pollutant Model

Tables 17 to 19 summarizes significant main effects (pollutants) and interactions for total (respiratory and circulatory diseases) admissions, admissions for respiratory diseases, and admissions for circulatory diseases. In the multi-pollutant model, SO2 was not significant and therefore not included in the stepwise selection process.

For total (respiratory and circulatory) admissions, significant pollutants were RSP (lag 0-3), NO2 (lag 0-1) and O3 (lag 0-5). As there was significant interaction between NO2 and O3, their respective RRs were expressed at three arbitrary levels of each other, namely the 25th percentile, the median (50th percentile) and the 75th percentile of the levels of pollutants (Table 17). A synergistic effect between the two pollutants was observed (i.e., the RR of each pollutant was higher at a higher level of the other pollutant than at a lower level). Compared with the single pollutant model, the magnitude of the RRs for RSP, NO2 and O3 was generally smaller.

A. RSP

B. NO2 (at different levels of O3)

C. O3 (at different levels of NO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

For respiratory admissions, the following were significant pollutants and interaction terms:- RSP (lag 0-3), NO2 (lag 0-3), O3 (lag 0-3), RSP with O3, and NO2 with O3 (Table 18). As O3 interacted with both RSP and NO2, the RRs of ozone were expressed at three different levels of NO2 and of RSP. At every level of RSP, the RR of O3 (albeit insignificant in some cases) increased with higher levels of NO2. The marginally significant RR of 0.84 for NO2 at 25% of O3 was probably a spurious finding. The interaction of RSP with O3 was antagonistic, i.e., the RR for RSP decreased with increasing levels of O3, while the RR for O3 decreased with increasing levels of RSP at every level of NO2. When interpreting the RR of O3, it should be noted that, as RSP and NO2 were correlated, there were few days in the dataset with high levels (75th percentile and above) of one pollutant and low levels (below 25th percentile) of the other.

A. RSP (at different levels of O₃)

B. NO2 (at different levels of O3)

C. O3 (at different levels of RSP and NO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

For circulatory diseases admissions, the following pollutants and their interaction were significant: NO2 (lag 0-1), O3 (lag 0-5), NO2 with O3 (Table 19). As before, NO2 and O3 showed synergistic interactions, but reached significance only at higher pollutant levels.

A. NO2 (at different levels of O3)

B. O3 (at different levels of NO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

The multiple pollutant model was also applied to different age groups (Tables 20 to 25). For the age group 0-4, respiratory admissions, RSP (lag 0-3) and O3 (lag 0-3) were the only significant terms in the model. No significant interaction was found. As the number of admissions for circulatory diseases for this age group was comparatively small (420 vs. 30,263 for respiratory admissions), the RRs were not computed.

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

When analyzing total hospital admissions (respiratory and circulatory diseases combined) among the 5 to 64 years age group (Table 21), the following parameters were selected, NO2 (lag 0-1), O3 (lag 0-5) and their interaction. Similar to the findings in all ages combined, a synergistic effect between NO2 and O3 was evident.

A. NO2 (at different levels of O3)

B. O3 (at different levels of NO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

For respiratory admissions (Table 22), the following were included in the model, RSP (lag 0-3), SO2 (lag 0), O3 (lag 0-3), and the interaction between SO2 and O3. It should be noted that SO2 (lag 0) was included despite being insignificant by itself, because the interaction term was significant. The effect of O3 was enhanced at higher levels of SO2. For circulatory diseases admissions, no pollutants were selected in the model.

A. RSP

B. SO2 (at different levels of O3)

C. O3 (at different levels of SO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

Among those aged 65 years and above, for total (respiratory and circulatory diseases) admissions, the selected parameters were SO2 (lag 0), NO2 (lag 0-1), O3 (lag 0-5) and the interaction term NO2 with O3 (Table 23). The interaction was synergistic, as in the 5-64 age group. SO2, which was not selected in the multiple pollutant model in the younger age groups or in all ages combined, was a significant pollutant in this age group.

A. SO2

B. NO2 (at different levels of O3)

C. O3 (at different levels of NO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

For respiratory diseases admissions, significant terms were: pollutants - SO2 (lag 0), RSP (lag 0-3), NO2 (lag 0-3) and O3 (lag 0-3); interactions - RSP with O3 and NO2 with O3 (Table 24). As with total admissions, SO2 was a significant pollutant. Similar to findings for all ages combined, RSP interacted antagonistically with O3, and synergistically with NO2. However, RR for the latter was found to be (marginally) below unity at low levels of O3 (25%).

A. SO2

B. RSP (at different levels of O3)

C. NO2 (at different levels of O3)

D. O3 (at different levels of RSP and NO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

For circulatory diseases admissions, NO2, O3 and their interaction term were entered in the model (Table 25). (As the interaction term was significant, O3 was included in the model, despite the fact that it was not significant by itself). As before, the interaction between NO2 and O3 was synergistic.

A. NO2 (at different levels of O3)

B. O3 (at different levels of NO2)

* 0.05 p 0.01
** 0.01 p 0.001
*** p < 0.001

Validation of the Model with 1996 Dataset

Time series plots of the observed total (respiratory and cardiovascular) hospital admissions in 1996 and the predicted numbers based on the 1994 - 95 model were shown in Figures 18 to 21. On visual inspection of the graphs, the predicted values were found to generally coincide with the peaks and troughs of the observed values. A systematic downward bias in total (respiratory and cardiovascular) admissions was observed for all four air pollutants in the last four months of the first half-year (i.e., the "observed" values being higher than the "expected" values). However, the model was found to fit fairly well during the first two months of 1996.