This set of analyses is the most comprehensive to date, based on data accounting for 40% of the global case load. Simultaneously analysing data from various affected areas in Greater China together revealed consistencies and explained heterogeneities that would have otherwise remained unexplored. Our combined database is unique in that of the three regions, only mainland China is a member of the WHO whereas Hong Kong is an observer and Taiwan is a non-member. Such geo-political influences have posed an almost insurmountable challenge to official supranational agencies in pulling together the respective databases . Database construction during an outbreak is critical to underpin analyses that will inform policy decisions regarding both disease control and treatment of infected patients. Key to the utility of such databases is their scope, format, accuracy and timeliness. While we have largely succeeded on the first two counts, the fact that we are only reporting results seven years after SARS receded into its animal reservoir points to an urgent deficiency in the global public health research infrastructure. Nevertheless, this report underscores the positive steps forward in recent years.
Our findings confirm the widely reported heterogeneous CFRs in the three regions [1, 21, 22]. However, discounting the highly selected XTS Hospital sample, the Beijing CFR approximated that of Hong Kong. On the other hand, Taiwan had a 64% higher adjusted case fatality compared to Hong Kong. Clearly one of the most important explanations for the observed low case fatality ratio in our subset of patients from Beijing is the inherent selection bias. While we studied 36% of the reported probable SARS cases in Beijing, from three of the largest SARS hospitals, our subset is by no means representative. The majority (74%) of our patients were from the XTS Hospital, a facility which was built especially to house SARS patients and opened halfway through the epidemic. Only one of the 680 inpatients was directly admitted, the remainder were transferred from other hospitals in Beijing. Although some of the XTS patients experienced symptom onset as early as the end of March, admission to XTS would have required survival until at least May 1, when XTS opened. Further, it is less likely that patients in serious condition (for example requiring mechanical ventilation) would have been transferred. As further confirmation of the systematic selection of the XTS sample, we found that patients admitted to XTS Hospital were epidemiologically biased, as shown by having shorter onset-to-admission periods and longer onset-to-discharge periods (data not shown).
An interesting related finding is that the relatively higher CFR also applied consistently to each subgroup including low risk groups such as younger patients and HCWs, which suggests that the presence of some common factors not included in our models (e.g. treatment protocols, hospital setting) may have independently increased CFRs in Taiwan. However, the treatment of choice in Hong Kong and Taiwan was the combined use of ribavirin and corticosteroids [14, 23], and a large meta-analysis has shown that these treatments were likely ineffective . We also observed similar patterns of CFRs for patients with the four most common comorbidities (diabetes, cardiovascular disease, ischaemic heart disease and chronic obstructive airways disease; data not shown) between Hong Kong and Taiwan and we did not observe exceptionally high case-fatality ratios among the patients with comorbidities in Taiwan. Hence the difference is probably not attributable to anti-viral or immunologic regimens but could have been related to different levels of general supportive care. Overall, our analysis found that age and pre-existing comorbid conditions were two major determinants of fatality, which is consistent with existing studies [14, 25–30]. Male sex was significantly associated with increased risk of fatality in Hong Kong after adjustment for other important confounding factors, consistent with a previous study in Hong Kong that identified a sex effect in unadjusted analyses of aggregate data . The reasons for an increase in risk of death among males remain unclear. In addition we found that admission before symptom onset, most likely due to nosocomial acquisition, was also significantly associated with higher fatality rate after adjustment for comorbidities in Taiwan .
In terms of epidemiologic parameters, the incubation period had mean 4.4, 5.7 and 6.9 days and a 95th percentile of 12.4, 19.7 and 17.9 days in Hong Kong, Beijing and Taiwan respectively, with some degree of variation between regions . Together they can reliably inform the diagnostician the likelihood of SARS in a suspected case and advise the public health practitioner the appropriate period of quarantine. While most countries used a ten-day quarantine period, based on early estimates of the incubation period , our findings suggest that a longer period may be appropriate if SARS were to reemerge . However differences between regions in the incubation distribution, as well as in the onset to discharge/death distributions discussed below, may be due to different interpretations of disease onset between the regions.
As would be expected, our results show that onset-to-admission periods shortened throughout the epidemic (Figure 2) as public awareness improved and public health interventions were implemented [14, 34, 35]. However, in the three regions, there was no evidence that shorter onset-to-admission period resulted in lower fatality (see Additional file 3), possibly due to the lack of effective treatment and that viremia only peaks after seven to ten days .
Taiwan's shorter onset-to-death distributions, indicating quicker deterioration, coupled with the consistently higher fatality ratios for most subgroups raises the possibility that the management of these patients did not compare favorably with those in the other two regions, especially if we assume the viral agent and host genetic susceptibility were similar throughout. However without more detailed clinical data, for example levels of viral load or as a proxy lactate dehydrogenase at admission reflecting initial illness severity [14, 29], it is not possible to discern here whether management and treatment, or some other factors, were responsible for the higher fatality ratios in Taiwan. Alternatively, differences in environment may be responsible for some observed differences, given the known associations between for example smoking and respiratory disease [36, 37]. The relatively longer onset-to-discharge periods in Beijing could likely be explained by different clinical or official protocols .
Our description of the variability between regions in symptoms at presentation (Figure 3) is the first such comparison in the literature. The higher rates in Hong Kong of almost every symptom could be due to differences in reporting (or asking) behavior. This again points to the need for a universal information supply chain, from case and symptom definitions to guidelines in history taking and data coding, for newly emerging or resurging diseases of supranational interest. As previously noted all three areas analyzed here shared the identical viral strain, as well as the same ethnic gene pool, so it is less likely that differences are due to the infective agent or host genetic factors. Although there are differences in the absolute rates of the various symptoms, we note that a previously derived clinical prediction rule for SARS  has been validated on patients from Taiwan .
Finally, a few limitations bear mention. As in most previous studies [14, 29, 41, 42], our analysis was based on probable cases of SARS according to the WHO definition  rather than laboratory confirmed cases because the latter definition may be biased toward including more survivors, particularly among the earlier cases . Furthermore, rapid diagnostic SARS tests were not available until fairly late in the epidemic and had poor sensitivity for detecting the disease [43, 44]. The WHO case definitions may have been applied differently in the three regions, or a different percentage of probable cases might have been caused by other pathogens, which may have led to the observed differences. Furthermore, while SARS patients in Beijing were classified according to the WHO definition, but a more recent detailed case review has found that some reported 'probable' SARS cases may have been misclassified . We note that odds ratios should be interpreted with caution since they may be a poor approximation to relative risks with outcomes that have a high prevalence. Asymptomatic and subclinical infections of SARS were not considered in our analysis, although there is strong evidence that very few existed . A final limitation is that, as previously discussed, the Beijing patients in our database mostly were hospitalized in XTS Hospital and were found to be epidemiologically different from patients in other hospitals, which makes it difficult to generalize our results to all patients in Beijing.