Maps, molecules, and multi-level modeling: Understanding the epidemiology of Escherichia coli O157 with clustered data

From 2000--2002, over 800 human cases of Escherichia coli O157 were reported in Alberta, Canada. Using spatial scan statistics yearly spatial, temporal, and space-time clusters were identified. In each year, spatial clusters were identified in the south of the province and temporal clusters were identified during the summer to early fall period. The locations of spatial clusters varied among years, but were fairly similar regardless of whether analyses included sporadic and outbreak cases or only sporadic cases alone. However, the spatial clusters identified for the entire study period were highly sensitive to the inclusion/exclusion of outbreak cases, and had the potential to lead to different hypotheses regarding the role of cattle farming in disease transmission. The space-time clusters identified as outbreaks using spatial scan statistics were validated qualitatively using epidemiological evidence and/or molecular data based on pulsed-field gel electrophoresis (PFGE). A randomization test was then applied and evaluated, using known outbreaks and analytical space-time clusters, and shown to be a reasonable tool to determine if isolates associated with a putative outbreak were more closely related, based on PFGE banding pattern, than expected by chance alone. Logistic regression models using various techniques to correct for auto-correlation revealed that outbreak cases tended to be younger and were more likely the result of person-person transmission than sporadic cases. Negative binomial and multi-level Poisson models revealed that population stability, the aboriginal composition of communities, and the economic links between communities and major urban centers were statistically significant risk factors associated with rates of disease among census subdivisions. The statistical significance of cattle density, recorded at a higher geographic level, depended on the method used to correct for over-dispersion, the number of levels included in the multi-level models, and the choice of using all reported cases or only sporadic cases. The impact of outbreak identification and limits to the spatial resolution of demographic and agricultural data are discussed as issues that limit and potentially bias the results of studies based on surveillance data.