This study is the most detailed assessment to date of the risks of SARS acquisition associated with HCW involvement in medical procedures, their infection control practices and the demographic and clinical characteristics of SARS patients. 
Amongst this group of HCWs caring for SARS patients immediately prior to and during intubation, the strongest predictor of SARS transmission from patient to HCW was whether or not the patient under care was a “superspreader”. Twenty-three of the 26 HCWs who became infected with SARS were infected by four of 45 patients. Lack of adherence to infection control procedures was also associated with transmission of SARS-CoV; however, individual patient characteristics beyond superspreader status were not.
The substantial heterogeneity in the number of secondary infections created by each case for many sexually transmitted and vector-borne diseases has led to the general rule that 20% of patients are associated with 80% of new cases. Although heterogeneity in individual patient transmission in diseases spread by direct contact and respiratory droplets is less well recognized, it has been clearly described for measles 
, rubella 
, Staphylococcus aureus 
, and tuberculosis 
, as well as for SARS-CoV in previous publications. 
Understanding this heterogeneity is important, because modeling studies demonstrate that the models incorporating variability in transmission differ substantially from standard outbreak models, and, in these models, individually–targeted interventions are much more effective than untargeted interventions. 
“Superspreading” appears to be a normal feature of disease transmission, and one that must be understood if we are to effectively prevent the spread of respiratory infection. The presence of heterogeneity in transmission also makes the interpretation of observational cohort data about risk factors for transmission difficult: our analysis illustrates the substantial potential impact of confounding in such cohorts. Observational cohort data may have very limited value for assessing transmission of influenza unless patient transmission heterogeneity can be taken into account.
Our findings with respect to HCW activity risk factors are similar to those of other studies assessing HCW risks unadjusted for patient factors, but illustrate the complexity of analyses of cohort studies in these settings. 
The factors associated with SARS in other analyses are somewhat different, but essentially all are related to procedures that bring workers into proximity with a patient's airway for prolonged periods of time, or with unprotected faces. In keeping with data from Teleman et al. 
our highest estimated HCW risk in GEE models was eye or mucous membrane exposure to body fluids (OR
7.3), while in CART analysis, the primary HCW related risk factor was whether or not eye protection was worn. This should not be interpreted as meaning that conjunctival contact in particular is a primary mode of spread of SARS CoV: when exposure to droplet spray occurs, is it generally not possible to distinguish exposure to eyes versus other mucous membranes. Absence of eye protection results in exposure of facial skin, and transmission could subsequently be from facial skin to hand to other mucous membrane. It is also possible that absence of eye protection is a marker for reduced adherence to other precautionary measures for which adherence is not adequately captured by self-report.
The range of different types of healthcare providers infected emphasizes that healthcare worker safety is not an issue limited to one profession, or to those workers with less education or control over their workplace situation. The finding in our study and those of others that a relatively small amount of education was associated with significant increases in adherence to precautions and reductions in infection also highlights the fact that, at least in some situations, education alone is enough to provide significant safety benefits. 
Where possible, hospital planners should consider building plans for “just-in-time” training into pandemic and outbreak responses.
CART and GEE logistic regression were complementary techniques for identifying HCW and patient characteristics most associated with transmission of SARS-CoV. Advantages of CART include the ability to identify interactions between variables by identifying specific combinations of variables which place HCWs at higher risk of acquiring SARS, modeling of nonlinear relationships between the dependent and independent variables, ability to handle numerical data that is highly skewed and categorical data with either ordinal or non-ordinal structures and its ease of interpretation. While logistic regression models are not as flexible in handling this variety of data, they yield odds ratios and p values, which are useful for quantifying risk and measuring the statistical significance of relationships between variables. The CART analyses were comparable with logistic regression GEE models with respect to sensitivity, specificity, and positive and negative predictive values.
There are several limitations to this study. Although we used patient charts to enhance recall, and validated our questionnaire 
, HCW recall of various exposures may have been imperfect due to the stress of caring for SARS patients and the time from exposure to interview, may have been biased by HCW outcomes, or may themselves have introduced biases if some were more accurate or complete than others. While our assumption that patients who were known to have infected other HCWs also infected the three HCWs whose source of infection was unclear had the potential to overestimate the superspreader phenomenon, secondary analyses confirmed our findings. While we might speculate that presence in the room during an ECG identified as a risk factor because other variables incompletely adjusted for duration of time in the room when a patient is deteriorating rapidly, we do not have a satisfactory explanation for why this variable is associated with SARS-CoV transmission. Finally, since the greatest risk of transmission of SARS-CoV occurred in the cohort of HCWs caring for the patient with the lowest PaO2
(P/F) ratio, the superspreader effect was confounded with the P/F ratio effect, and it is not possible to conclude that low patient P/F ratio is associated with SARS-CoV transmission.
Although some authors have assumed that all viral respiratory infections have the same relative modes of transmission, such that identified risk factors and/or interventions that prevent transmission for one can be assumed to be true for others 
, it is not clear that knowledge about risk factors for SARS coronavirus infection can be directly applied to other diseases such as influenza. It is clear, however, that, during the SARS outbreak, HCW exposures to body fluids occurred frequently and adherence to recommended precautions was often incomplete, putting HCWs at significant risk of infection. Thus, research into the incidence of and risk factors for influenza transmission in acute care hospital settings, and into interventions effective in minimizing transmission, is urgently needed.