We developed an ecological study using secondary data.
Data were drawn from the registry of cholera treatment centers (CTCs) and oral rehydration points (ORP) functioning during the cholera epidemic in Harare and Chitungwiza. Medecins Sans Frontieres, in collaboration with the Department of City Health of the Ministry of Health & Child Welfare, implemented and managed three CTCs, in the Budiriro Polyclinic, in the Beatrice Road Infectious Diseases Hospital and in the city of Chitungwiza. Ten ORPs were functioning in Harare city and one in Chitungwiza.
Data used in this study were the aggregated number of cholera cases by suburb of residence, topographic average elevation of each suburb in meters, and the distance in meters from suburbs’ geometric centre to Chitungwiza (the epicenter of the epidemic).
Population figures by suburb were calculated from the official census of Harare and Chitungwiza, completed in 2002. To estimate the populations’ figures at the time of the epidemic, we employed an average constant annual growth rate of 3%, as estimated by the Population Division of the Department of Economic and Social Affairs of the United Nations Secretariat
During the epidemic, a cholera case was defined as any patient presenting 3 or more liquid stools and/or vomiting for the last 24 hours
We described the elevation by suburbs and distance of each suburb to the epicenter (Chitungwiza) in meters. Geographical information was integrated into the GIS database as vector polygons. Using ArcGIS ‘zonal statistics’ module, we measured for each polygon average ground elevation and the distance between each suburb centroid (geometric center of the polygon) to Chitungwiza.
To describe the outbreak, we calculated attack rates for each suburb, including 95% exact confidence intervals (CIs). We assumed that the attack rates follow a Poisson distribution
[24, 25]. We then estimated rate ratios and their respective 95%CIs, taking the whole Harare rate as reference.
To describe the relation between the cholera cases and elevation, the Pearson’s correlation coefficient (R) between the cholera rates and the mean elevation of each suburb was calculated. Then, a scatter plot was built to show the relation between the observed rates and the average elevation by suburb.
A generalized linear mixed model (under the assumption of a Poisson distribution) was used to model the relation between the risk of cholera by suburb and the elevation in meters of Harare: The model assumed a random intercept, with the number of cholera cases as the dependent variable, the log of the population figures by suburb as the offset, and the suburb’s average elevation as the independent variable
We use prior knowledge about the dynamic and spread of the epidemic to allow for spatial correlation between neighboring suburbs based on the distance in meters to the epicenter (Chitungwiza)
]. Therefore, the random intercept of the mixed model was used to account for the dependency of neighboring suburbs, overdispersion and allowing for clustered variance estimation
]. In detail, we constructed a categorical cluster variable Q
, based on the quintiles (
) of the distance of each suburb to Chitungwiza. Hence, our final model looks as follows:
The derived risk ratio of the model was interpreted as the cholera risk per 100 meters of increase in the elevation, taking into account overdispersion and spatial correlation of the data.
Finally, the predicted rates of cholera by suburb were derived from the model using an Empirical Bayesian estimation. This Empirical Bayesian approach consists of computing a weighted average between the raw rate for each suburb and the regional average, with weights proportional to the underlying population at risk
. Estimated rates were illustrated in a figure in order to show the relation between the predicted rates and the elevation in meters. Because of the random intercept, represented by the quintile variable (Q1 to Q5), predicted rates by suburbs were clustered in groups following the specification of the model.
A sensitivity analysis was developed to confirm the consistency of our findings, consisting of an iterative random selection of suburbs to build the cluster variable for the estimation of the random effect. The aim of this process was to estimate the variability of the coefficient of the elevation, taking our first estimation as reference and proving its consistency
For the GIS component of the analysis (mean elevation and distance to the epicenter), a recent (2008) high resolution satellite image of Harare was used. In order to assure a good integration with field surveys, the satellite image was geometrically corrected to fit a set of control points surveyed with a GPS. This image was used as a spatial reference for all GIS analysis. Information from the Harare’s topographic map (Produced by the Surveyor General of Zimbabwe), was digitized on top of the satellite image using Arcview software. The elevation information was extracted from the SRTM database (a digital terrain model obtained from the Space Shuttle mission). In this database, ground elevation is provided for each 90x90 meter cell of the model. The extent of each urban district was integrated to the GIS database as vector polygons. Using ArcGIS ‘zonal statistics’ module, average ground elevation and the distance between each suburb centroid (geometric center of the polygon) to Chitungwiza were measured for each polygon.
The statistical software used for this study was Stata v.11.2 and the module ArcMAP of ArcGIS (ESRI, v. 9.2) software for spatial and geographical representation.