Reliability of Hospital Data

With the exception of Caritas Medical Centre (CMC) in Shamshuipo (which coded its 1994 records in another system incompatible with the rest), all "relevant" hospitals (i.e., those that admit patients with acute respiratory and cardiovascular illnesses through A & E or 24-hour Outpatient Department) under the Hospital Authority were captured in this study. In 1995, the total number of hospital admissions (all causes) into CMC was about 8% of the total number of admissions into the 12 hospitals. Hospital admissions for respiratory and cardiovascular diseases into CMC (7379 in number) was also about 8% of the total admissions into the 12 hospitals (93,276). Hence, the exclusion of this hospital should not affect the study results substantially. As explained under Section 4.2.1, the exclusion of private hospitals and specialist hospitals has an even smaller effect. Because a large proportion (over 90%) of the total number of hospital beds in Hong Kong was provided by the Hospital Authority, the external validity of the sample is good. However, problems do exist with this dataset. There were substantial duplicate entries in two hospitals which were subsequently deleted. Other coding errors included mistakes or the lack of age data in some records, but the numbers were relatively small. The numbers of hospital admissions in some hospitals were much higher in 1995 than in 1994. This could have been due to a combination of several reasons. First, the migration from the IPAS (where coding of diagnosis for all admissions was incomplete) to the MRAS system (where coding rate was nearly 100%) might have artifactually boosted the number of hospital admissions. Second, an increase in the availability of beds in some hospitals in 1995 might affect the decisions by the doctors at the A & E Departments to admit more patients, and the phased operation of A & E service in Yan Chai Hospital probably had similar effects. While these "artifacts" could not be fully accounted for in the analysis, they were adjusted for by introducing the linear and quadratic time trend variables and a year-effect indicator.

Changes in treatment or diagnostic practices may theoretically affect the numbers of hospital admissions and coding of cases. However, we are not aware of any major changes in the treatment of cardiovascular and respiratory diseases which may affect the need for hospitalization. The use of the ICD 9 classification ensures uniformity in coding practice.

It should be noted that the number of daily mortalities in the selected hospitals did not reflect the total mortality picture in Hong Kong, and the analysis (Table 16) only provides supplementary information on air pollution risk. Territory-wide data on daily mortality was available only since 1995.

Finally, it must be recognized that in studies using routinely collected data, the primary dataset will contain various types of errors (miscoding, clerical etc.), which may not be easily detected and/or rectified.

Reliability of Air Pollutant Data

The validity of air pollutant data has been commented on in Section 4.2.2. Incompleteness of the data was the major problem, especially in respect of ozone (where data were available for only two out of seven monitoring stations). For the same reason, analyses could not be performed for carbon monoxide and for Yuen Long station. As mentioned in the 'Methods' section (4.2.2), missing values were imputed for the time series analysis according to the APHEA protocol to ensure comparability of study methods.

Choice of Model

The core model was constructed according to the APHEA protocol. Even within this standardized framework, modelling problems can arise. In Finland, problems were encountered in the control of cyclicity and autocorrelation (Ponka et al., 1996b). In our dataset, serial correlation and overdispersion were evident. While a number of techniques have been advocated to adjust for these effects, we have chosen Williams' Method because of its relative simplicity. Compared with the uncorrected model, the parameter estimates were generally higher but the confidence intervals were wider. Applying the multiple pollutant model (which was capable of estimating the effects of individual air pollutants "independently", i.e., adjusting the effects of the others), we detected significant interactions of some pollutants by comparing the relative risks of one at different levels of the other pollutant.

With only two years of data, it was not possible to correct for confounding variables like influenza epidemics which may occur every two years. Yet, seasonal and epidemics corrections have been recognized as the most important modelling steps in Germany (Spix et al., 1996).

A comparison of the datasets shows that both hospital admission data and air pollutant data for the year 1994 were less comprehensive than for 1995 and the first half-year of 1996. The problems with the 1994 dataset have been commented before (See 4.2.1 and 4.2.2). It is highly recommended that a further time series study should be performed using datasets for the two-year period of 1995 - 1996. This should enable the construction of a statistical model which is less prone to unrecognized biases due to poor data quality.

Ecological Fallacy

This study, as with all time series studies, is inherently subject to ecological fallacy, which refers to the inability to relate individual outcomes to personal exposure to the measured pollutants when analyzing group data from a heterogeneous population. Therefore, it is not valid to make inferences about exposure-disease relationship at the individual level. In contrast to individual-based studies, ecological studies are more prone to systematic errors and unrecognized confounding factors. These are often hard to detect and, even if known, their effects are difficult to adjust for statistically. However, confounding factors like cigarette smoking (which are relevant in individual based studies) need not be addressed in a time series design, because the prevalence of smoking does not change over short time periods and therefore does not influence the results.

It has been observed that independent non-differential mis-classification of an exposure indicator will usually result in a bias away from the "no effect hypothesis" in ecological studies. "Whenever feasible, ecological studies using aggregated data should be supplemented by individual-level studies in a hybrid epidemiologic analysis" (Briggs, et al., 1996).