Dr CM Wong
(Data analysis and report writing)

Mr Stefan Ma
(Computation and statistical advice)

Materials and methods

Findings

Discussion

Conclusions

For hospital admissions

Air pollution was found to have an effect on circulatory and respiratory diseases combined and separately (relative risk, RR=1.03-1.10; p<0.084) for all ages; on circulatory admissions (RR=1.05-1.10; p<0.001) for the 65 or above age group; on respiratory admissions, the effects of which appeared to be j-shaped from the younger to the older age groups; on asthma (RR=1.10-1.16; p<0.018, except SO₂); on chronic obstructive pulmonary disease (RR=1.08-1.14; p<0.0001); and on ischaemic heart disease (RR=1.04-1.09; p<0.051, except SO₂ and RSP).

For hospital deaths

Both NO₂ and O₃ were positively associated with circulatory and respiratory diseases combined and separately (for NO₂: RR= 1.10-1.14, p<0.038; and for O₃: RR=1.07-1.22, p<0.010).

Validation and composite score

The above estimates were consistent with and in between those obtained from similar studies using the European (APHEA) approach overseas and in the CUHK. But in addition, a composite score was derived from the four pollutants and was found to provide consistent estimates for all the health outcomes under study in all ages (RR=1.04-1.11, p<0.098, except hospital deaths due to circulatory diseases).

Several epidemiological studies have now shown an association between particulate air pollution and exacerbations of illness in individuals with respiratory disease and also increases in the numbers of deaths from cardiovascular and respiratory disease, particularly in the elderly. New hypotheses have been advanced to postulate mechanisms underlying these observed effects (Seaton et al 1995).¹

Respirable particulates (RSP) with an aerodynamic diameter of <10 um(or particles measured as black smoke by the smoke stain method) comprise the principal pollutant associated with these findings. In addition RSP sulphate and sulphur dioxide concentrations, are reported to be associated with all causes mortality and respiratory mortality in recent studies from the USA (Pope et al 1995)² and the UK (Anderson et al 1996)³ respectively. A recent review in the United Kingdom concludes that the associations between daily concentrations of particles and acute health effects principally reflect a causal relationship (Committee on the Medical Effects of Air Pollutants 1995).⁴ After a lengthy scientific review, the USEPA determined that new standard should be added for particulates less than 2.5 ? of aerodynamic in size and the welfare-base standards were also revised by making them identical to the health-based standards.⁵

In Hong Kong we have found that SO2, RSP and SO4 concentrations are associated with excess risks for symptoms of cough, phlegm and wheeze and also bronchial hyper-responsiveness (by histamine challenge test) in primary school children (Hedley et al 1993; Peters et al 1996; Tam et al 1994; Wong et al 1998).^6,7,8,9 However in the London study the strongest association with daily mortality was for ozone. The effects of ozone and black smoke were independent of the effects of other pollutants.

Evidence on the risks associated with other pollutants is variable and less consistent. In the recent London study (Anderson et al 1996)³ the NO2 (1 hour maximum) was associated with all causes mortality and cardiovascular mortality; however a negative effect was seen for respiratory mortality. A significant positive effect on mortality was seen for SO2 and all cause mortality in the warm season period.

Several authors point to the complex between-season covariation of several pollutants, which is sometimes negative and at other times positive.

Study design

It was an ecological study utilizing routinely collected hospital admission data, air pollutant concentration data and weather data by the Hospital Authority, Environmental Protection Department and the Observatory respectively. Variations in the daily number of hospital admissions due to circulatory and respiratory diseases were studied, and their relationships with each of the pollutants were modelled to assess the effects of air pollution on health after adjustment for time trends, seasonality, weather conditions and some other factors including days of week, holidays and days after holidays.

This study follows a previous one performed by the Chinese University of Hong Kong (CUHK)¹¹ which followed the general approach of the protocol of the APHEA (a European approach using epidemiological time series data), developed within the frame of the EC Environment 1991-94 Programme. However data for some disease categories (ICD Rubrics) included in the APHEA protocol were not analysed in the CUHK study. The data sets and disease categories used for this new study are shown in Table 1 below. Those categories which were excluded or missing from the CUHK study are indicated in the table.

* first half year
** data for ICD9 390-392, 393-398, 446-448, 451-459, 470 and 510-519 were not included for analysis in the CUHK Final Report
*** data for ICD9 401-405 and 500-508 were missing in the CUHK data files and were not analysed in the CUHK Final Report
( ) CUHK data in brackets
- not included for analysis in the CUHK study

In order not to exclude categories which might show effects from air pollution (e.g. hospital admissions for cardiovascular diseases under ICD9 390 - 429 which were found to be related to respirable suspended particulates and carbon monoxide by Schwartz (1997)¹³ but not all included in the CUHK study) and to ensure comparability to the results of those studies which follow the APHEA protocol (Bacharova 1996; Schouten 1996),^14,15 we decided to follow strictly the categories recommend by the APHEA protocol for this study.

Data for 1995-1996 (generated and cleaned by the Department of Community Medicine, the University of Hong Kong in this study) were used in establishing the statistical models and in estimating the effects; and data for first half year of 1997 were used for validation of the established models. It was also decided that data for the year 1994, which were used by the CUHK group, should not be used in this study because the data quality is not consistent with that of the 1995-96 data set to be used in this study. In 1994 only 3 of the 12 hospitals under study had already adopted the MRAS database; but the number increased to 7 in 1995. Besides, the percentages of valid daily data for air pollutant concentrations from the monitoring stations were lower in 1994 than those in the other years.

Databases

Hospital admission data: Hospitals included for generation of hospital admissions were the publicly funded hospitals (accounting for 90% of hospital beds in Hong Kong) which either had an accident and emergency department, or was a referral base from the accident and emergency department of another nearby hospital (9) or had a 24-hour outpatient department (2). One other hospital which was the only hospital in the most polluted district in Hong Kong, was also included. All hospitals should have a computerized system for inputting and retrieval of patient data. The hospitals included in the study, together with the type of information system they were using and an indication as to whether they had an A & E department, were listed in Table 2 below:

* IPAS - Integrated Patient Administrative System
MRAS - Medical Records Abstracting System

Patients admitted to hospitals between 1.1.1995 and 30.6.1997 with data on: dates of admission and discharge, socio-demographic information (age, gender, marital status, ethnic group, district of residence, pseudo identifier of patient), admission source, discharge diagnosis in ICD9 codes and discharge status were retrieved from the databases for each of the hospitals under study. In order to validate the completeness of the retrieved data, the total number of inpatients for the period 1.4.1995 to 31.3.1996, for each hospital, were compared with those reported in the Hospital Authority Statistical Report 1995/96. When there were big differences between the two, the Hospital Authority Information Technology Department was ask for an explanation and we then revised the databases if necessary until they were reasonably close to each other. This process took almost a half year to complete. The data are shown in Table 3 below:

# total hospital patient discharges
@ The excess 5409 cases in the HKU data set were day cases which could not be excluded when the Information Technology Department generated the data set due to missing of the identifier.

The total data sets were then extracted for circulatory (ICD9 390-459) and respiratory (ICD9 460-519) diseases. The numbers of hospital admissions by disease groups in the three years were as shown in the previous Table 1 and subset of specific disease categories are shown in Table 4 below:

* first half year

Acute myocardial infarction was not analysed, as had been done by the CUHK, because diagnosis for the disease has been changing and subject to misclassification over the past years.

Pollutant concentration data: Pollutant concentration data in CD-ROM were made available by the Air Services Group of the Environmental Protection Department with hourly data from all monitoring stations in Hong Kong. The following stations in various urban, suburban and industrial areas were included in the study:

Two stations, one in Yuen Long (in suburban area) due to the extent of missing data and one in Mong Kok (in urban area) which provided concentrations measured on ground level, were excluded from the study. The pollutants included in this study are in Table 6 below:

* TEOM - Tapered Element Oscillating Microbalance, an instrument for the continuous measurement of particulates matter in air

In order to maintain consistency and similar quality standards in the data, missing data were defined and replaced in accordance with the APHEA recommendations. The guidelines were slightly modified to suit local situations and these are described in Table 7 below:

Note: RSP measured by TEOM.

# A minimum of 67% non-missing daily data was used as criteria for inclusion of a pollutant in the analysis instead of 75%. This was set in the computer programme at the beginning of the study when the HKU was trying to adopt a similar procedure as that used by the CUHK. However this might not be necessary for data after 1994 as the data were more complete.

The data in monthly averages were comparable to those in the EPD 1995 and 1996 statistical reports. The percentage of data valid after replacement are shown in the following Table 8:

* the data are available only for first half year of 1997
- not available
# Abbreviations referred to Table 5.

A daily mean concentration representing the pollution level in all Hong Kong areas were obtained from valid data of all the stations for each of the pollutants under study.

Meteorological data: Daily means for temperature and relative humidity were obtained from the Environmental Protection Department which derived the data from the Observatory. The monthly means for the years 1995 and 1996 were comparable with those reported in the Hong Kong Annual Reports.

Statistical modelling

The statistical modelling methods followed the guidelines recommended by the APHEA protocol and are outlined as follows:

Validation of the model for 1995-96 was achieved internally by comparing the predicted with the observed number of health outcomes in the data set as well as by comparing the observed on the first half year of 1997 with the predicted from the model developed using 1995-96 data. Externally the effects per 50 ug/m3 concentration estimated by other studies following the APHEA protocol and those obtained from the CUHK study were compared with those in this study.

Back to Contents

Descriptive statistics

Summary statistics of daily hospital admissions for circulatory and respiratory diseases, both combined and separately by 3 month periods are shown in Basic Table A5.1 to A5.3 (in Appendix). Summary statistics of all the health outcomes used in this study are shown in Tables 9-12 below:

* Combined circulatory and respiratory diseases
COPD - Chronic obstructive pulmonary disease
IHD - Ischaemic heart disease

* Combined circulatory and respiratory diseases
# first half year
COPD - Chronic obstructive pulmonary disease
IHD - Ischaemic heart disease

* Combined circulatory and respiratory diseases

* Combined circulatory and respiratory diseases
# first half year

Summary statistics of daily mean temperature, relative humidity, and mean concentration of each of the pollutants are shown in the Basic Tables B1-2 and C4-7.

The correlations among the four pollutants, and between each pollutant and the health outcomes are shown in Table 13.

SO₂ and NO₂ were correlated with both circulatory and respiratory admissions, combined and separately. For RSP it correlated with circulatory admissions and circulatory and respiratory admissions combined. In studying whether these pollutants have an effect on early health outcome, it is important to adjust for confoundings which might explain the correlation.

The correlations among the pollutant concentrations, temperature and humidity are shown in the following Table 14.

Statistical modelling

Core models for hospital admissions: The following Tables 15-17 show the results of core models which explain the hospital admissions due to circulatory and respiratory disease categories in terms of daily linear and quadratic time trends, year effect, seasonality (using sine and cosine functions), day of week (compared with Sunday), holidays and days after the holidays, temperature and relative humidity. (Air pollutant concentrations are not included at this stage.)

For the two disease groups, both combined and separately, there were positive daily linear and quadratic trends over the two year period; and weekdays (Monday to Friday relative to Sunday), day after holidays, higher temperature and lower humidity were positively associated with hospital admissions (p<0.012, except for Friday effects in respiratory admission where p=0.708). Each model explained 64-79% of the variation in the health outcomes.

Similar effects were also found in individual age groups (younger than 15 years, 15-64 and 65 years and older age groups), for both disease categories separately, except in those younger than 15 years for respiratory diseases, (most p<0.10) (Basic Tables E 7-12). The above models explained 39-77% of the variations in the health outcomes.

For admissions due to asthma, chronic obstructive pulmonary and ischaemic heart diseases the effects due to the whole set of meteorological, seasonality and trend were similar except that numbers of hospital admissions were associated with lower temperature (p<0.044) (Basic Tables E 4-6). The models explained 31% to 49% of the variations in the health outcome.

Multiple R-Square = 0.7518 (from multiple regression on logarithmic transformed counts of health outcomes)

*Notes:

Multiple R-Square = 0.7869 (from multiple regression on logarithmic transformed counts of health outcomes)

* For explanation of variables, refer to Table 15.

Multiple R-Square = 0.6406 (from multiple regression on logarithmic transformed counts of health outcomes)

* For explanation of variables, refer to Table 15.

The daily number of hospital admissions in all ages due to circulatory and respiratory diseases combined and separate, due to asthma, chronic obstructive pulmonary and ischaemic heart diseases, as observed and predicted by the core models are depicted in Figures 5 (a1 - f1). The numbers of hospital admissions predicted by the models with pollutants for each of the corresponding health outcomes are depicted in Figures 5 (a2 - f2). For specific age groups and for circulatory and respiratory diseases combined and separately, the numbers of admissions are depicted in Figures 7 (a1 - b3). From the graphs there were strong seasonal variations which are predicted closely by the core models. However the observed numbers were varying to a greater extent than that of the predicted. After adding in the pollutant concentration to the core model, some of the excess variations were explained. In general both the observed and predicted numbers were more crowded in the upper bound than in the lower bound.

Core models for hospital deaths: For hospital deaths in all ages due to circulatory and respiratory diseases combined together and separately, the models explained 24% to 39% of the variations (Basic Tables E 13-15). The daily number of observed deaths and those predicted by the core models in 1995-96 are depicted in Figures 6 (a1 - c1). The numbers predicted by models with pollutants for each of the health outcomes are depicted in Figures 6 (a2 - c2). Because of the small observed numbers, the observed varied from the expected to a much greater extent for hospital deaths than for hospital admissions.

Models with pollutants for hospital admissions: The relative risks (RR) with 95% confidence intervals (95% CI) for hospital admissions in all ages due to each of the disease categories are shown in Table 18 below, with comparison to similar estimates from the CUHK reports. As the CUHK study used data of 1994-95 (instead of 1995-96) and included different ICD9 codes for circulatory and respiratory diseases, the results will not be directly comparable to those from the HKU study. All relative risks in the tests and in Tables 18-21 below were referred to 50 ug/m3 changes in concentration of the pollutants. However relative risks for 100 ug/m3 changes were also presented in Tables 18a-21a (Appendix B) for the pollutants correspondingly.

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

# Notes: Lag0-n denotes the mean of cumulative effects of lag from day 0 (same day) up to previous n days using up to 5 days (Lag0-5) for ozone and up to 3 days (Lag0-3) for other pollutants

The results will be discussed in 5.2-5.4.

Models of pollutants and hospital deaths: The relative risks (RR) with 95% confidence intervals (95% CI) for hospital deaths due to each disease category are shown in Table 19 below, with comparison to similar estimates from the CUHK reports:

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

The results will be discussed in 5.2-5.4.

Models with interactions between pollutants: The relative risks of hospital admissions for circulatory and/or respiratory disease in models with interaction terms between pollutants are as shown in Table 20 below:

(a) Effect of NO₂ on circulatory and respiratory diseases

(b) Effect of RSP on circulatory and respiratory diseases

(c) Effect of NO₂ on respiratory disease

(d) Effect of NO₂ on circulatory disease

*** p < 0.001

NO₂ interacted with SO₂ for circulatory and/or respiratory diseases, with the estimated RR ranging from 1.11 to 1.13 at lower levels of SO₂; and ranged from 1.05 to 1.07 at higher levels of SO₂. RSP has an effect on circulatory and respiratory diseases combined at lower levels of O₃ with RR 1.04 (95% CI: 1.02, 1.07) but no significant effect was seen at higher level of O₃ . The effects for a pollutant were stronger at a lower level of the other pollutant interacting with it. The other pollutants for circulatory and respiratory diseases combined and separately were not significant.

Models with interaction between pollutant and seasons: In studying whether the pollutant effects on hospital admissions for circulatory and respiratory diseases combined and separately, varied between seasons (i.e. spring: March - May; summer: June - August; autumn: September - November; winter: December - February), only the relative risk associated with O₃ were found to differ in the spring, being higher than in the other seasons.

The relative risk per 50 ug/m³ increase was 1.05 (95% CI 1.02, 1.08) in spring and 1.01 (0.99, 1.04) in other seasons for circulatory and respiratory diseases combined; and was 1.06 (1.03, 1.10) and 1.02 (0.99, 1.05) for respiratory disease alone.

Effect of composite score of pollutants: A composite score of the concentrations of the four pollutants was derived (for method, see footnote for Table 21), which extracts the maximum correlation among the pollutants and explains 68% of the variations among them. Using this composite score as the independent variable instead of each of the pollutants, the results with all those significant health outcomes are shown in Table 21 below.

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

# Composite score was generated from four pollutants (NO₂, SO₂, TEOM and O₃) by principal components analysis and the first principal component was used which explained 68% of the variance with loadings of 0.491, 0.105, 0.726 and 0.469 respectively.

(a) All ages
(b) 0-14 years old
(c) 15-64 years old
(d) 65 years or older

The results will be discussed in 5.2-5.4.

Adequacy of 1995-96 models (hospital admissions): The residues for hospital admissions due to circulatory and respiratory diseases combined and separately, after fitting the model for each of the pollutants are shown in Figures 9(a) - 9(c). Patterns indicating unexplained variations in the models are observed. There were no discernible patterns in the residuals for circulatory admissions. There were some unexplained patterns in the residual plots for respiratory and respiratory and circulatory admissions combined.

There may also be autocorrelations (i.e. correlation between consecutive observations) in the data. This was examined by the autocorrelation function plots in Figures 10(a) - 10(c). There was no evidence of autocorrelation in the circulatory admission residuals. However slight autocorrelation of less than 0.5 was present in the respiratory and respiratory and circulatory combined admission residuals.

Predictive validity of 1995-96 models (hospital admissions): After obtaining the models from the 1995-96 data, the predicted number of hospital admissions due to circulatory and respiratory diseases combined and separately in the first half year of 1997 were obtained. The observed numbers in the first half year of 1997 and the predicted numbers with model using each of the pollutants are depicted in Figure 8 (a-c). Overall the models for prediction of circulatory and for prediction of circulatory and respiratory admissions were better (i.e. closer to the observed) than those for the prediction of respiratory admissions. In all the three health outcomes observed there were rising trends in the later part of the first half of 1997. These phenomena were particularly prominent in respiratory admissions indicating that the long term effects of other determinants of the disease might have changed. The models developed according to two yearsÕ data were not able to pick up the changes. But it would not invalidate the models to be used for estimation of air pollution effect on health.

Validity and reliability of the models

The validity and reliability of the models were ensured and assessed in the following ways:

Summary of findings

All relative risks referred to an increase of 50 ug/m³ in the concentration of the pollutant.

Hospital admissions: Time trends, seasonality and weather conditions explained 31% to 79% of the variations in all the health outcomes under study. (In the CUHK report only those for admissions and hospital deaths due to circulatory and respiratory diseases combined were reported, which were 66% and 31% respectively.)

Irrespective of the amount of variations unexplained after fitting the core model, all the pollutants under study showed significant effect on daily hospital admissions for circulatory and respiratory diseases combined and separately with relative risk (RR) estimates ranged from 1.03 to 1.10 in all ages overall. (The CUHK results were 1.03-1.18 in the RR estimates.)

In analysis by age groups, apparently the effects of pollutants on circulatory diseases were stronger for the older age groups with significant RR of 1.05 to 1.10 found in 65 or older age group; and for respiratory diseases the effects by age groups were apparently j-shaped with RR 0.93-1.08 in the 0-14 years age group, 0.90-1.06 in the 15-64 age group, and 1.06-1.19 for the 65 or older age group. (In the CUHK report there were no comparable results.)

For specific disease categories, NO₂ , RSP and O₃ were positively associated with admissions for asthma, RR 1.10-1.16; all pollutants were positively associated with chronic obstructive pulmonary diseases, RR 1.08-1.14; NO₂ and RSP were significantly associated with ischaemic heart disease, RR 1.04-1.09. (In the CUHK results, only for asthma were available with RR (1.12-1.27.)

For all ages and the 65 and older age group, NO₂ and RSP exhibited the most consistent effects with estimates for RR 1.07-1.14 for NO₂ and 1.03-1.10 for RSP, significant for all health outcomes; O₃ exhibited the strongest effects with RR 1.04-1.19 all were significant except for ischaemic heart disease; SO₂ exhibited the least effects with RR 0.91-1.12, all were significant except for asthma and ischaemic heart diseases. However as the pollutants are highly correlated, a composite score which summarises the four pollutants, was constructed and was found to provide the most consistent (i.e. closely estimated) for all the health outcomes under study with RR 1.06-1.11, all statistically significant except for asthma and ischaemic heart disease. (No corresponding results were from the CUHK study.)

Hospital deaths: NO₂ was positively associated with hospital deaths due to circulatory and respiratory diseases combined and separately with RR 1.10-1.14 (same RR 1.10-1.14 for CUHK results); and O3 was similarly associated with these health outcomes with RR 1.07-1.22 (CUHK 1.13-1.27). Similar to that for hospital admissions, the composite score for pollutant concentrations was also found to provide consistent estimates for effects of air pollution on hospital deaths with RR 1.05-1.09 (no corresponding results for CUHK).

Comparison with APHEA studies

The estimates of relative risks in the validation study showed stronger and significant effects when compared with pooled results of Western European cities on hospital admissions due to respiratory diseases, asthma and chronic obstructive pulmonary diseases with our significant RR of 1.02-1.06 versus theirs of 1.01-1.03 for the 15-64 age group and 1.06-1.19 versus 1.02-1.04 for the 65 and older age group.

Effects on hospital deaths were also more significant and higher in the HKU study than those in the pooled up estimates for total mortality from Western European studies with significant RR 1.10-1.15 versus 1.01-1.04 for circulatory diseases and RR 1.14-1.22 versus 1.04-1.05 for respiratory diseases.

Comparison with CUHK study

However the estimates from the validation study, when compared with those from the CUHK study, were smaller with RR 1.04-1.07 versus 1.07-1.14 for hospital admissions due to circulatory and respiratory diseases combined for all ages; about the same for admissions due to circulatory diseases (RR 1.05-1.10 versus 1.04-1.12 for 65 or older); and smaller for respiratory disease (RR 1.06-1.19 versus 1.13-1.22 for the 65 or older age group).

In the analysis for interaction effects between pollutants, the validation study followed the APHEA protocol in principle (i.e. defining the interaction term by multiplying a continuous pollutant concentration variable with the other pollutant dichotomized into high and low level in order to avoid multicollinearity¹⁰) and this is a more conservative approach. The CUHK study took a more aggressive approach (in that several co-pollutants with interaction terms were put in the same model and subject to model selection by means of a stepwise procedure). Relatively smaller numbers of interactions were found in the HKU study compared to the findings reported in the CUHK report.

Limitations of the Hong Kong study

Only two years of data: Unlike most studies in other places, the Hong Kong study only had two yearsÕ data for the analysis. It would not be realistic to confidently predict the health outcome in the third year of the study according to models developed in the previous two years, with reasonable accuracy. However the main objectives of this kind of studies were to assess the health effects of air pollution, but not to obtain prediction for health outcomes.

Non-linearity of the effects: Although a moving average method was used to take account of the lag effect of air pollution, all the other covariates were modelled for their linear effects. Some other smoothing functions would be useful and have to be used such as the generalised additive modelling, for modelling the non-linear relationship between the health outcome and covariates.

Refinement of models: There are discernible patterns in the residuals of some models. We have used different cycles (obtained from spectral analysis) in fitting the models and found that there were no changes in the fitness and estimates of the models. However plots of the residuals against pollution levels suggested that the unexplained variations are not related to air pollution. For the sake of uniformity we keep the same covariates of the original models for our results.

Discrepancies between hospital deaths and total deaths: The hospital deaths included in this study represented only about half of all deaths occurred in the period. Also, they were analysed with the air pollutant concentration on the date of admission to hospital rather than on the actual date of death. These will produce discrepancies in the estimates compared to those when all deaths are included and analyses were based on date of death.

New direction: There are new issues raised in the newly developed APHEA II¹⁶ protocol, such as in harvesting effects of air pollution on pre-mature deaths, and regional differences in the effect estimates. These issues were pointed out in a separate operation manual but were not addressed in this study. They should be taken into account in any future study.

Critique of CUHK study