A basic requirement for mapping malaria risk across a country is an understanding of the distribution of its population. Modelling techniques for the spatial reallocation of populations within census units have been developed in an attempt to (i) disaggregate population count data to a finer spatial detail and (ii) convert population count data from irregular administrative units to regular raster layers (Linard et al., 2011; 2012).





 Population census size estimates and the boundaries of the corresponding census enumeration unit were acquired at the highest spatial resolution from the most recently publically available census (2007). Typical regional per-land cover class population densities were estimated from African countries for which very fine resolution population data were available, following approaches previously outlined (Linard et al., 2012). These typical population densities were then applied as weightings to redistribute census counts according to the land cover and to map human population distributions at a finer spatial resolution using dasymetric modelling techniques (Mennis, 2009). 


The modelling method distinguishes urban and rural populations in the redistribution of populations. The population map is based on the 2007 census data resolved to the district level. The population maps could be improved if census data at smaller geographic units were available. Each blue grid represents a geographic space at one of three time points. The red dots represent positions and time for which P. falciparum parasite prevalence data are available.


The black arrows indicate that the data points surrounding (in time and space) the square of interest are used to predict the likely parasite prevalence in the orange square. The procedures used to assemble, geocode, archive, model and validate the transformation of empirical P. falciparum parasite prevalence data to continuous predictions of age-corrected mean prevalence in children aged 2-10 years (PfPR2-10) are provided by Noor (2014), Snow et all (2015) and Snow and Noor (2015).

 In brief, we used information from available age-corrected survey data (sample size and numbers positive) at known locations (longitude and latitude) and times (year) with a minimal set of conservative, long-term covariates traditionally used in vector-borne disease mapping. These were brought together in a Bayesian hierarchical spacetime model, implemented through an adapted Stochastic Partial Differential Equations (SPDE) approach using Integrated Nested Laplace Approximations (INLA) for inference (www.rinla.org;


A total of 1,012 malaria prevalence surveys undertaken between 1982 and 2013 were identified in time and space through the data search process. Four surveys were excluded because data could not be disaggregated below the regional level. A total of 73 surveys were excluded because their sample sizes were less than 10 individuals. The remaining 939 surveys are shown by year in Figure 6.3. The data volumes to make reliable spatial predictions are temporally sparse between 1982 and 2002. We have therefore elected to only use data from the most data-rich period, 2002-2013 (n = 902).

A complete excel database of all geocoded surveys was provided for the NMCP alongside this report.

Full details of the data assembly, geocoding methods and classifications of species according to their role in malaria transmission are provided elsewhere (Snow et al., 2015). The database has been arranged as a site-specific, referenced 36 inventory to capture details of species identification recorded since the earliest surveys in 1900 through to the latest records in 2014. The full digital PDF library, database and bibliography accompanies this report. From each identified report, data extraction included whether a species was identified at a given site, methods used to capture adults or larvae and the methods used to speciate each anopheline collection. 


“Y” was recorded if a species was identified and “N” was only recorded when the absence of the species was reported. The database is therefore one of species presence, not absence or proportional presence of the different vectors. We have not assembled geocoded information related to vector resistance. These data have been carefully curated, validated and mapped by the IRBase initiative (Knox et al., 2014; Typically, national household surveys are designed to be precise at national and regional levels and rarely at lower levels, such as districts. Simply aggregating survey data to provide district level estimates of an outcome of interest will lead to values of low precision. Small Area Estimation (SAE) methods handle the problem of making reliable estimates of a variable at these units under conditions where the information available for the variable, on its own is not sufficient to make valid estimates (Rao, 2003; BIAS, 2007). 


We used hierarchical Bayesian spatial and temporal SAE techniques using a geoadditive regression approach (Banerjee et al., 2004; Best et al., 2003) to estimate the proportion of the population in each health district sleeping under an insecticide treated net (ITN) the night before the survey. This was done by health district for the years 2005, 2008-9 and 2010-11 and 2012-13. This method uses survey data from a health district and neighbourhood information from adjacent districts to smooth values at the health district level. Covariates were not used in this approach. However, if information on the distributions of ITNs by month were to become available for each health district, this would improve the precision of the estimates.