Finding malaria hot-spots in northern Angola: The role of individual, household and environmental factors within a meso-endemic area

Background: Identifying and targeting hyper-endemic communities within meso-endemic areas constitutes an important challenge in malaria control in endemic countries such like Angola. Recent national and global predictive maps of malaria allow the identification and quantification of the population at risk of malaria infection in Angola, but their small-scale accuracy is surrounded by large uncertainties. To observe the need to develop higher resolution malaria endemicity maps a predictive risk map of malaria infection for the municipality of Dande (a malaria endemic area in Northern Angola) was developed and compared to existing national and global maps, the role of individual, household and environmental risk factors for malaria endemicity was quantified and the spatial variation in the number of children at-risk of malaria was estimated. Methods. Bayesian geostatistical models were developed to predict small-scale spatial variation using data collected during a parasitological survey conducted from May to August 2010. Maps of the posterior distributions of predicted prevalence were constructed in a geographical information system. Results: Malaria infection was significantly associated with maternal malaria awareness, households with canvas roofing, distance to health care centre and distance to rivers. The predictive map showed remarkable spatial heterogeneity in malaria risk across the Dande municipality in contrast to previous national and global spatial risk models; large high-risk areas of malaria infection (prevalence >50%) were found in the northern and most eastern areas of the municipality, in line with the observed prevalence. Conclusions: There is remarkable spatial heterogeneity of malaria burden which previous national and global spatial modelling studies failed to identify suggesting that the identification of malaria hot-spots within seemingly mesoendemic areas may require the generation of high resolution malaria maps. Individual, household and hydrological factors play an important role in the small-scale geographical variation of malaria risk in northern Angola. The results presented in this study can be used by provincial malaria control programme managers to help target the delivery of malaria control resources to priority areas in the Dande municipality. © 2012 Magalhães et al.; licensee BioMed Central Ltd. The study protocol was approved by the Angolan Ministry of Health Ethics Committee. One day before the survey, a field worker visited each household to provide explanations on the study and obtain individual informed consent, signed or marked with a fingerprint from the selected mother or caregiver. All households that agreed to participate were enrolled in the study after handing in the signed consent form. Participants found to be positive for P. falciparum infection were treated with ACT. The Demographic Surveillance System (DSS) established in 2009 within the CISA project (Centre for Health Research in Angola, translated) in the Dande municipality, aims to collect longitudinal data on the population’s structure, dynamics and geographical location [6,27]. Dande municipality, Bengo Province, north-western Angola, is a largely rural area 60 km north of the capital Luanda. According to the census conducted between September 2009 and April 2010 this area of about 4,700 km2, has 59,844 registered inhabitants in 15,604 households distributed in 69 hamlets. The study area of the CISA DSS includes three of five communes (Caxito, Mabubas and Úcua) within Dande municipality. A community-based cross-sectional epidemiological survey in the CISA DSS area was conducted between May and August 2010 [6]. Detailed information about the sampling design and survey management are detailed in Sousa-Figueiredo et al.[6]. Briefly, a total of 972 households were selected and included in the study, distributed in 36 hamlets, including a total of 960 mothers (mean age 33.3 years, range 16 to 80 years) and 2,379 children aged ≤15 years (mean age 5.9 years, range 6 months to 15 years) in Caxito (794 children), Mabubas (904 children), and Úcua (681 children) [6]. Prior to the survey, field workers were trained to interview caregivers using a questionnaire, which collected information relating to the caregiver and her children. Questions included: demographic information (age, sex), occupation, access to healthcare and history of previous treatment, and questions related to malaria (e.g. knowledge and bed-net ownership and utilization). Malaria parasitaemia was detected by preparing blood smears in the field which were stained with 10% Giemsa. Malaria parasites were screened by two independent microscope technicians using a double-blind approach [28]. In the field participants were tested using a rapid diagnostic test – Paracheck-Pf (Orchid, India) allowing for treatment of positive cases. The geographic information available in the DSS data warehouse included the coordinates (longitude and latitude) of the majority of the households which were extracted on-site using Garmin® GPSMAP® 60Cx handheld global positioning system (GPS) receivers, and household coordinates were collected with a precision of two to six metres. The codes of houses included in the survey were linked to the CISA DSS data warehouse to obtain household information, which included type of flooring and ceiling. Because 30% of households included in the survey did not have geographic coordinates the geographical centre of the sector (the lower subdivision of the hamlet) was used as the geographical unit for the analysis. Centroids were estimated based on a digitalized map of the hamlets, using the spatial analyst tool in the GIS. Due to their large size, six hamlets are divided into sectors (subdivision of the hamlet) and in these cases the centroid of the sector was used as the geographical location for the analysis. This enabled 95% of individuals included in the survey to be successfully georeferenced to a total of 43 locations. Figure ​Figure11 shows the geographical location of the centroids of the 43 locations where the prevalence survey was conducted. Geographical distribution of malaria in children aged <15 years in the Dande municipality in Angola. A shapefile of large perennial inland water bodies was obtained from the Food and Agriculture Organization of the United Nations [29]. The shapefile of large perennial inland water bodies was updated by adding the location of rivers, lagoons and irrigation canals in the study area. The location of rivers, lagoons and irrigation canals was obtained by recording the path of these features on the GPS with 10-metre intervals. Layers for these features were created by uploading the GPS information on the geographical information system (GIS) ArcGIS 9.3.1 (ESRI, Redlands, CA) and using the vectorization procedure in the GIS. The distance to rivers, lagoons, large perennial water bodies and irrigation canals was extracted for each sector in the GIS ArcGIS version 10.0 (ESRI, Redlands, CA). Electronic data for land surface temperature and rainfall for a 1 km × 1 km grid cell resolution were also obtained from the WorldClim data warehouse. A 5km resolution population surface derived from the Global Rural–urban Mapping Project (GRUMP) beta product was obtained from the Center for International Earth Science Information Network (CIESIN) of the Earth Institute at Columbia University [30]. The geographical centre of the sector was linked to the environmental data to obtain values of the environmental variables for each sector using the spatial overlay procedure in the GIS. For the purpose of the analysis the presence of malaria parasites in blood identified by the Giemsa method was used as the outcome variable of the analyses and thus all children were categorized into infected and not infected based on the presence of at least one parasite. In order to explain the spatial variation in malaria risk, the individual level variables age, sex, maternal knowledge of malaria, bed net usage and previous malaria treatment and five household variables: type of household walls, type of household roof, number of rooms in household, type of water source and type of sanitation facility were considered in the initial variable screening stage. These household variables were considered important as they could represent proxies for the economic status of the family as well as a measure of household exposure to malaria vectors. Environmental data such as land surface temperature, rainfall, distance to rivers, distance to irrigation canals, distance to lagoons and distance to health care centers were also considered in the variable screening stage. Univariate logistic regression models for a Bernoulli-distributed outcome and cluster correction by hamlet using robust standard errors were built to screen variables for inclusion in the final multivariable model, using the statistical software Stata version 10.1 (Stata corporation, College Station, TX). During this stage, variables with P>0.2 were excluded from further analysis; using that criterion, bed net usage, previous malaria treatment, the household variables type of household walls, number of rooms in household and type of water supply, and the variable distance to irrigation canals were subsequently excluded. Spatial dependence in the residuals of the final non-spatial multivariable model was investigated using a residual semivariogram which is a graphical representation of the spatial variation left unexplained by the covariates included in the model. The residuals of the final non-spatial multivariable model were extracted and analysed for spatial dependence in the statistical software R using the geoR package version 2.14.1 (The R foundation for statistical computing). A Bayesian geostatistical model of malaria prevalence was developed in WinBUGS 1.4 (Medical Research Council, Cambridge, UK and Imperial College London, UK) (see Additional file 1 for detailed model specification). The model included an intercept, the individual level variables age, sex and maternal knowledge of malaria, the household variables type of household roof and type of sanitation facility, the environmental variables land surface temperature, rainfall, distance to rivers, distance to lagoons and distance to health care centers, and a geostatistical random effect. The covariate effects were summarized by using the mean and 95% credible intervals (representing the range of values that contains the true value with a probability of 95%). The geostatistical random effect modelled spatial correlation as a function of the separating distance between pairs of hamlets. Model predictions were used to generate a malaria risk map for boys of older ages (the subgroup with the highest risk of malaria) for the entire study area in ArcGIS version 10.0. The prediction model included the individual level variables age and sex and the variables of the physical environment temperature, rainfall, distance to lagoons, distance to health care centre and distance to rivers. While using age and sex allowed predicting to the subgroup most at risk to malaria in the study area (i.e. boys or older age), the use of the variables of the physical environment allowed prediction across a continuous landscape. To determine the discriminatory performance of the model predictions relative to observed prevalence thresholds (10% and 50%), the area under the curve (AUC) of the receiver operating characteristic was used. An AUC value of >0.7 was taken to indicate acceptable predictive performance [31]. A map of predicted standard deviation was also generated to depict the uncertainty around the mean predicted malaria endemicity. The global malaria map for 2011 [10] was included in the GIS and the shapefile of the administrative boundaries of the Dande municipality was used as a mask to extract the spatial distribution of malaria for the study area (see Additional file 1). The mean malaria endemicity in the malaria risk map and in the malaria map for the study area from the global malaria map for 2011, was estimated based on the summary statistics provided by the raster properties in the GIS. The national malaria map for Angola using MIS from 2007 [26] was not available to us and comparisons between this map and malaria risk map developed in this study were made visually. The comparison between spatial distributions of predicted malaria endemicity for the different studies was made visually. To estimate the number of children at-risk of malaria in the six communes of Dande municipality for 2011, complete and up-to-date numbers of children aged ≤15 years was retrieved from the CISA DSS database. However this information was not available for the communes of Barra do Dande and Quicabo; for these communes the predictive malaria risk map was used and multiplied it by a population map of ≤15 yrs for 2011. This population map was generated by multiplying a population map for 2009 (see methods for details) by the population growth factor for Angola for the period 2010–2015 and by the average proportion of ≤15yrs in Angola for the period 2010–2015, available from the UN World Population Prospects website [32].