Background
Maps of disease distribution are an essential tool for optimizing the allocation of resources for malaria interventions [
1,
2]. There have been a number of attempts to develop malaria transmission maps at different geographic scales based on expert opinion [
3,
4]; deterministic biological models driven by the conceptual relationship between transmission and environmental covariates [
5]; and empirical transmission models based on entomological inoculation rates [
6,
7] or human infection prevalence data [
8‐
17]. These methods suffer several limitations: expert opinion maps are subjective; deterministic models ignore the secular effects of expanded coverage of interventions that supersede the influence of climate on the epidemiology of malaria and do not quantify uncertainty around model results. Where studies have used observational data to predict malaria distributions, most have used historical data collected opportunistically from secondary sources [
10,
15,
16] that did not involve random sampling and/or a sampling framework optimized for spatial analysis.
Arguably the greatest need for malaria maps is at the periphery of stable, endemic areas where decisions about the delivery of standard suites of interventions, such as those promoted by the Roll Back Malaria (RBM) initiative to support malaria control in high transmission areas, may become less appropriate or cost-efficient. In areas of perceived low malaria risk there is little empirical information on the risks and intensity of transmission. As such the semi-arid regions of the horn of Africa remain less well described epidemiologically compared to the rest of malaria endemic sub-Saharan Africa (SSA) and there are no contemporary national maps of the extents of malaria risk. The Malaria Atlas Project (MAP) while maintaining a global remit in its efforts to improve the cartography of malaria [
2] is equally committed to developing national mapping initiatives with country partners, where the data available can support rigorous cartography. Somalia represents the first such example.
A Plasmodium falciparum malaria prevalence map for Somalia is presented here using Bayesian geostatistical analysis of community-based parasite prevalence survey data. The data used in this analysis have several unique features that minimize some of the problems of using retrospectively assembled data: first the community data were derived from random sample surveys undertaken as part of national malaria or nutritional surveys; second all the data were collected using similar methodologies; and finally all the data represent contemporary infection prevalence between 2005 and 2007.
Discussion
There has been little historical description of the basic epidemiology of malaria transmission in Somalia. In 2002, an application was made to and successfully approved by the GFATM to support the funding of a suite of interventions and strategies managed by a consortium of non-government and governmental agencies across the three main zones of Somalia [
51]. This application, similar to other successful applications and RBM policies in neighbouring, higher-intensity transmission countries of Kenya [
52], Tanzania [
53] and Uganda [
54] involved a monitoring & evaluation component to investigate intervention coverage and
P. falciparum infection rates. In Somalia rapid, sample malaria intervention and parasitological surveys of communities have now become part of a routine component of rolling nutritional surveillance surveys across the country [
55]. Consequently, despite being a country without a functioning research capacity and a fragile health system, Somalia is now one of the 87
P. falciparum endemic countries worldwide with the largest series of infection prevalence data [
56,
57].
Simple summaries of the data suggest that large parts of the country, particularly in the north, have very low human infection prevalence (Table
1). These summaries, however, mask spatial heterogeneities in risk that are important for better targeting of interventions and maintaining aggressive surveillance. A Bayesian geostatistical approach to predict
Pf PR throughout Somalia is used here. In the north, the inclusion of the survey and environmental covariates appeared not to make a significant difference to model fit, while in the south they improved the model fit. Predictions of endemicity class membership made in the north were associated with lower prediction probabilities and generated generally lower AU-ROC values (Table
3 & Figure
3c). This greater prediction uncertainty in the north is due largely to the comparatively fewer empirical data points compared to the south (Figure
2). This disparity was essentially driven by the population distribution: approximately 65% of Somalia's population live in the South and communities in the North are more scattered in isolated settlements [
58].
Although the environmental covariates selected for inclusion in the Bayesian geostatistical model were significantly associated with
Pf PR when examined in the non-spatial multivariate model (Tables
2 &
3), none remained significant when spatial correlation was accounted for in the north and only precipitation and temperature remained significant in south. Overall the inclusion of these covariates accounted for a relatively small proportion of spatial variation suggesting that other unmeasured factors might be influencing the spatial distribution of malaria prevalence. These factors might include proximity to artificial breeding sites such as wells, dams, boreholes and seasonal streams and/or the use of interventions to prevent malaria at the household level. It has recently been demonstrated that in southern Somalia, the use of insecticide treated nets (ITN) reduced the prevalence of infection by as much as 54% [
30]. Mapping the household or community levels of ITN use at high spatial resolutions is not currently feasible at a national scale. Similarly, the mapping of fluctuating, localized vector breeding sites would require very detailed spatial reconnaissance and infection mapping at finer scales than is currently possible using public domain covariate data at national scales. Furthermore, communities where sample sizes were less than 40, most of which could not be geo-located, were excluded from the analysis and these might have resulted in information loss for some areas of Somalia. Although the difference, in terms of mean parasite prevalence, was minimal between the excluded and included surveys, future analysis should include all data regardless of sample sizes given the Bayesian analytical approach implicitly adjusts for sample size.
Despite the constraints described above, the use of Bayesian geostatistics to model
Pf PR does provide a valuable method to define sub-national spatial variation in prevalence, and a baseline against which future changes in prevalence can be quantified intervention coverage expands. Under such a scenario the value of the environmental covariates might be expected to wane further, particularly in areas of very low transmission intensity where the environment currently supports homogenously low transmission conditions. The similar levels of performance observed between the univariate and multivariate models for the north of Somalia may be evidence of this view. In addition, the relatively higher coverage of ITN among the communities closest to the two rivers in the south might explain the lower predicted prevalence in their immediate vicinity consistent with the observational data and reported effectiveness of ITN [
30].
Population density or a derived categorisation of urbanisation, with known influences on malaria transmission [
59,
60], would have been a worthy candidate covariate for testing in this study and in determining accurately the population at risk against varying malaria endemicity. However, the reliability of settlement and population data in Somalia is highly questionable. The last national census was undertaken in 1971 and the displacement and migration over the last 20 years of civil unrest has been substantial. Development agencies and non-governmental agencies working in Somalia continue to update a semi-quantitative database of settlement locations and population counts but its fidelity is unknown. The absence of an accurate national census also hampers the linkage of spatial malaria risk to populations-exposed to risk. Notwithstanding the precision and scale of calculating populations at risk, aggregated district-level estimates of population in 2004 across the 120 districts of Somalia have been used and assigned each district the dominant
Pf PR risk class. From these numbers it can be estimated that approximately 75% of Somalia's estimated 7.4 million people live in areas that support unstable or very low
Pf PR (0–5%) transmission and less than 0.1% live in areas classified as high, intense transmission (
Pf PR > 40%). Areas of low
Pf PR include many communities where infection prevalence was observed as zero (Table
1). In these locations it is assumed that these observations represent a statistical zero (i.e. resulting from a limited sample in areas of very low transmission) rather than implying a true absence of infection risk [
56]. This is important to highlight because routine sample surveys in such areas demand considerably larger samples [
45,
61] or the use of serological markers of parasite exposure [
62] to truly exclude the possibility of transmission.
In communities exposed to low
Pf PR, such as the majority of the population in Somalia, the risk of disease is low and spread across all age-groups. These are fundamentally different epidemiological conditions to areas of high transmission where functional immunity is developed early in life and a higher disease burden is experienced in young children and pregnant women [
63‐
66]. Tailoring the existing intervention recommendations in the Somalia National Malaria Strategy [
67] to the spatial transmission patterns shown in Figure
3 will be a challenge to the agencies providing malaria control services nationwide.
Conclusion
The use of routine, nationwide surveillance of infection prevalence is key to monitoring the changing epidemiology of malaria in all countries scaling up coverage of malaria preventative strategies. Including RDTs in on-going community-based health surveillance is a cost-effective means of assembling this information. The use of geostatistical methods can help focus surveillance efforts and define those areas where uncertainty exists, guiding future sampling [
49,
68]. Coupled with better estimates of where people live, these should provide the basis for informed estimates of disease burden [
63] and how these might change with changing infection-risk exposure. Somalia has a range of political and economic barriers that might limit the success of a strategic, epidemiologically driven malaria control programme. It has been possible to demonstrate, however, that the foci of greatest disease risk are predominantly concentrated in one area in the South and that infection risks are very low in the northern reaches of the country. Moreover, although the density of survey sites and hence the uncertainty of the modelled output varies spatially, also it has been demonstrated that, despite constant civil disturbance, routine survey data can be assembled to inform strategic decision making. Finally, areas where model uncertainties are greatest, predominantly in the north of the country, should be the focus of any future parasitological surveys to improve further the precision of the prevalence maps.
Acknowledgements
The authors are grateful to the WHO-MERLIN and FAO/FSAU survey team for their invaluable supervision and support during the field surveys and subsequent data entry and cleaning and Bruno Moonen of MERLIN specifically for helping with training for the parasite survey. We thank Priscilla Gikandi and Victor Alegana for additional data cleaning and geo-referencing. We are also grateful to Anand Patil and Andy Tatem for statistical and GIS advise and to Simon Brooker, Carlos Guerra and Emelda Okiro for their comments on the manuscript.
Funding for the WHO-MERLIN 2005 surveys were provided by the UN Trust Fund for Human Security and the GFATM. FAO/FSAU funded training of assessment teams, data collection, paid enumerators and data entry clerks for the 2007 surveys. The FAO/FSAU nutrition surveillance project is funded primarily by OFDA-USAID and receives support from UNICEF, SIDA and EC for conducting nutrition assessments in Somalia. RDTs and anti-malarial treatment were provided by UNICEF through GFATM funding (SOM-202-G01-M-00). AMN is supported by the Wellcome Trust as a Research Training Fellow (#081829). SIH is supported by the Wellcome Trust as Senior Research Fellow (#079091). RWS is supported by the Wellcome Trust as Principal Research Fellow (#079081). AMN, SIH and RWS acknowledge the support of the Kenyan Medical Research Institute. The funders did not have a role in study design, data collection and analysis, decision to publish, or preparation of manuscript. This work forms part of the output of the Malaria Atlas Project (MAP:
http://www.map.ox.ac.uk), principally funded by the Wellcome Trust, U.K.
Authors' contributions
AMN was responsible for data cleaning, analysis, interpretation and production of the final manuscript and revisions, ACC contributed to the data analysis, interpretation and production of final manuscript, PWG contributed to the data analysis, interpretation and production of final manuscript, GM was responsible for the study design, supervision of data collection, cleaning and contributed to the preparation of the final manuscript, MB provided the necessary interface with community leaders, and the Ministry of Health for approval and was responsible for the data collection, cleaning and contributed to the preparation of the final manuscript, TS assisted in the survey design, supported the field investigations, provided the interface with local ministry of health and helped in the preparation of the manuscript, SIH contributed to overall scientific direction interpretation and preparation of the final manuscript and revisions, RWS was responsible for overall scientific management, analysis, interpretation and preparation of the final manuscript and revisions.