Introduction Maps are essential for all aspects of the coordination of malaria control . In an international policy environment where the malaria control community has been challenged to rethink the plausibility of malaria elimination [2–4], malaria cartography will become an increasingly important tool for planning, implementing, and measuring the impact of malaria interventions worldwide. The last global map of P. falciparum endemicity was published in 1968 . In common with all previous maps of the global distribution of malaria [6–10], and to a large extent those that followed [11–16], the map (i) suffered from an incomplete description of the input data used; (ii) defined contours of “risk” using subjective and poorly explained expert-opinion rules; and (iii) provided no quantification of the uncertainty around predictions. Here we describe the generation of a new global map of malaria endemicity that overcomes these major deficiencies. Geographic Scope of the Modelling The global spatial limits of P. falciparum malaria transmission have been mapped recently by triangulating nationally reported case incidence data, other medical intelligence, and biological rules of transmission exclusion, derived from temperature and aridity limits to the bionomics of locally dominant Anopheles vectors [17,18]. The results of this exercise stratified the world into three classes: the spatial representation of no risk, unstable risk (P. falciparum annual parasite incidence [PfAPI] 5% to 100 km2] and small [>25 km2] polygons ; removing those surveys that could not be, or were only geo-positioned imprecisely; and removing those that could not be temporally disaggregated into independent surveys or for which the date was unknown), 7,991 PfPR surveys remained (Figure S1.2 in Protocol S1). All PfPR data were then age-standardized to the 2- to 10-y age range before mapping using an algorithm based on catalytic conversion models first adapted for malaria by Pull and Grab . This algorithm was found to perform best out of a set of candidate standardization procedures and is described in detail elsewhere (Protocol S1.3) . The final dataset was stratified into three major global regions (Figure 1): the Americas; Africa, Yemen, and Saudi Arabia (Africa+); and Central and South and East Asia (CSE Asia) (Protocol S1.4). This division allowed these biogeographically, entomologically, and epidemiologically distinct regions [8,16] to be considered separately, whilst retaining sufficient data in each region for meaningful analysis. These global divisions were further supported by observing the distinct spatial structure of the PfPR2−10 data in each region, illustrated by their semi-variograms (Figure S1.1 in Protocol S1). Malaria transmission-specific approaches to mapping urban, peri-urban, and rural extents were developed, the rationale for which is described in detail elsewhere (Protocol S2) . In brief, all urban extents (UEs) defined by the Global Rural Urban Mapping Project (GRUMP) alpha version UE mask (GRUMP UE) [60,61] were identified at 1 × 1 km spatial resolution (Protocol S2.1) . Within these extents, those areas containing population densities greater than 1,000 people per km2 according to the Gridded Population of the World version 3 population density surface [60,61] were then mapped . All surveys were then assigned as either urban (Gridded Population of the World version 3 ≥ 1,000 km2 within GRUMP UE), peri-urban (Gridded Population of the World version 3 5% to 5% to 5 to 5% to 5 to 5 to < 40%) and 0.345 billion under conditions of high risk (PfPR2−10 ≥ 40%) (Figure 7; Table 4). In the areas of intermediate risk, mathematical modelling suggests that by taking ITNs to scale, the interruption of P. falciparum malaria transmission might be achieved, whereas in the high transmission areas, malaria transmission will be more intractable and require aggressive control with suites of additional and complementary interventions [19,55]. Statistical Implementation and Model Validation The modelling procedure presented here represents a large scale implementation of modern Bayesian geostatistical techniques and incorporates a number of novel components. The incorporation of an age-standardization model has allowed the coherent assimilation of survey data obtained across a wide variety of surveyed age ranges whilst acknowledging the uncertainty introduced by this additional source of variation. Likewise, the use of a fully spatiotemporal random field has allowed surveys from as early as 1985 to be incorporated in the prediction of contemporary P. falciparum endemicity in a statistically and epidemiologically plausible framework. MBG techniques are exceptionally computationally demanding even for small prediction problems. To our knowledge this is the first time these procedures have been applied to any disease at the global scale. This computational burden has also imposed a number of restrictions on the modelling procedure that may have improved predictive capability. In particular, the current model adopts a single mean and covariance function within each global region, representing an assumption of second-order stationarity within each. Approximations to nonstationary random fields adopted in smaller scale studies [32,76] represent possible refinements to the current model, but were considered computationally infeasible globally. Assessment of the various validation statistics revealed that the model performed satisfactorily for each of the three performance aspects: predicting PfPR2−10 point values and endemicity class, and providing realistic measures of prediction uncertainty. Given the highly variable nature of P. falciparum endemicity over even short distances, an overall correlation of 0.82 between the model predictions and validation data, and an average absolute error magnitude of 9.75% PfPR2−10 represents an unexpected level of precision. Certain aspects of the uncertainty measures output by the model are suboptimal: in particular, the tendency to underestimate slightly the probability of PfPR2−10 taking very low values. Nevertheless, given the multitude of sources of uncertainty that are captured and propagated though the modelling framework, the resulting uncertainty predictions represent a rich source of information in the generation of output products for decision makers. The model was fitted using MCMC [77,78]. MCMC is an extremely powerful algorithm, and is the only general-purpose, computationally tractable algorithm available for many Bayesian problems. However, it is an approximate algorithm. No fail-proof method for estimating its error is available, but using a heuristic method (Protocol S1.3) we estimated that our “Monte Carlo error” is unimportant relative to the uncertainty in our actual posterior distributions. The information contained in the maps presented here and the associated uncertainty varies across a range of geographical scales. The large-scale variation in endemicity described between regions and countries is unambiguous, robustly quantified, and of direct use to global planners. As progressively finer scales are considered, however, the utility of these maps for local malaria control managers diminishes although this is heavily dependent on the local availability and density of survey points. The appropriate threshold and metric of uncertainty will vary enormously for different end users and applications of the maps. As a rule-of-thumb, however, it is suggested that the differentiation in endemicity between areas smaller than the first administrative level would be inappropriate for most countries. Examination of the frequency distributions for all-year and 2007 input PfPR2−10 data, and for the predicted PfPR2−10 surface, revealed a number of important features. Firstly, 2007 data from all three regions displayed substantially smaller median and maximum values and were more positively skewed than data from all years considered together (compare Figure 8A and 8B). Secondly, there were marked differences in all regions between the distribution of 2007 data values and the distribution of values from the predicted PfPR2−10 surface (compare Figure 8B and 8C). Specifically, the latter distributions had larger medians, were less positively skewed, and for the Americas and Africa+ had substantially smaller maximum values. The overall shift towards higher PfPR2−10 in the predicted surfaces can be attributed to the spatial clustering of the survey locations. It must always be remembered that the set of surveys collated represents an opportunistic sample driven by the motivations and constraints of a multitude of individuals, organizations, and governments. Visual examination of this set reveals a considerably larger proportion located in lower endemicity regions than would be the case in a spatially random sample and, as such, summary statistics of these raw data display a substantial bias. By predicting endemicity over a continuous surface, the MBG process compensated implicitly for this clustering in the output maps and the resulting frequency distribution was not biased in the same way. The MBG process makes predictions at unsampled locations using linear combinations of survey data. For this reason, the resulting surfaces are inevitably smoother than the raw data from which they are predicted. One feature of this smoothing process is that the range of extreme high and low values in the predicted surface is likely to be smaller than that displayed by the input data. This explains why the frequency distributions for the predicted PfPR2−10 surface cover substantially smaller ranges of values than those of the input data. An important implication of this smoothing effect is that the predicted surface provides a more robust prediction of endemicity at larger scales but is less able to represent faithfully the short-scale variations occurring over very short distances. Using Environmental Covariates to Make Continuous Maps The extreme limiting effects of climate covariates have been incorporated comprehensively in the definition of the stable and unstable limits of P. falciparum malaria transmission described above . There is an illusory attraction in the further use of environmental covariates to increase complexity and improve predictive accuracy in MBG endemicity mapping. This is because such analyses are based on the assumption that the contemporary distribution and endemicity of malaria approximates its fundamental niche [79,80]. This assumption is unfounded because the global distribution of malaria has contracted substantially  since its hypothesised maximum distribution circa 1900 . Moreover, it is not known to what extent the environmental determinants of the remaining distribution reflect this fundamental niche, how these relationships might vary spatially, and therefore, what artefacts might be introduced by their inclusion in the analyses. In addition, it is not trivial to obtain “adequate” environmental covariates at a global level with the required spatial and temporal fidelity [63,81]. Finally, the degree to which these relations would be further obscured by ongoing and spatially variable intervention efforts is also unquantified. An increasing body of evidence points to these intervention effects being substantial, to have accelerated in the post 2000 period, and to represent a spatial mosaic of influence that would act to confound substantially any modelled relationships [82–90]. Unsurprisingly, no statistical support was found for the inclusion of a range of climate  and remotely sensed  environmental covariates (Protocol S1.7). In eschewing the use of environmental covariates in this analysis framework, the output maps are determined only by the input survey data and the assumptions of the modelling. This choice ensures a maximally parsimonious baseline, against which future changes may be audited. Potential Geostatistical Improvements In embracing the MBG approach, the rationale for excluding surveys with a sample size below 50 is diminished, as the uncertainty in relation to the population sampled is explicitly modelled by the technique (Protocol S3). This exclusion rule was devised at a time before MBG could be applied at a global scale and will be revised in future iterations of the map. The spatial resolution with which these MBG techniques could be reasonably implemented on a computer cluster was on a 5 × 5 km grid. The entire process took an average of one month at this spatial resolution and has been estimated to take one year to run on a 1 × 1 km spatial grid. There are no plans to increase the spatial resolution of the output maps at the global scale because they are robust for the regional planning purposes for which they are intended. For smaller areas, such as PfPR data rich countries where higher spatial resolution maps may be desirable to support national control plans, however, MBG outputs to 1 × 1 km grids can be considered . Moreover, at these national scales, the fidelity of the geo-positioning of the input PfPR survey data may have an important influence on the uncertainty of the predictions, so procedures that can help incorporate these effects into the modelling may also need to be investigated [91–93]. In this study, the uncertainty likely to be contributed by geo-positioning errors was thought to be trivial in relation to the scales of spatial variation in observed endemicity and given the global scale of model outputs. We were not able to improve the age-correction model's predictive performance by modelling the age-dependent sensitivities of microscopy and rapid diagnostic tests separately or by modelling diagnostic specificity. The accuracy in the determination of PfPR by microscopy or rapid diagnostic tests were assumed to be equivalent in these analyses, but the sensitivity of the diagnostic technique [94–98] could be included into a future iterations of this MBG framework. No solution could be found to applying these MBG techniques across large tracts of ocean (for example in the Caribbean, Madagascar, and the Indonesian archipelago), given the global distribution of the PfPR data and the lack of data in some regions (Figure 1). Potential biogeographical influences on malaria transmission on islands are ignored by these analyses. Future map iterations would ideally have sufficient data to treat islands separately or sufficient information on the distribution of Anopheles vectors to help inform the predictions . We have incorporated the ability for the analyses to be cognisant of secular trends in the PfPR data and of annual variations in transmission. This map does not provide a full description of seasonal malaria dynamics [99–101], however, and further information on the global variation of malaria seasonality might inform future map iterations. The Road Ahead: Public Domain and Dynamic Maps These mapped surfaces are made available in the public domain with the publication of this article. The underlying data used in their predictions are due for public release in 2009 , and the online infrastructure to host this service is under development. The MAP team anticipate providing annual updates of this P. falciparum global malaria endemicity map and the accompanying PfPR database. Annual updates will also be required to reflect the changing spatial limits of stable and unstable P. falciparum malaria transmission  in order to define accurately the limits within which endemicity predictions need to be made. If the international community is successful in rolling back malaria, informed decisions will need to be made about the temporal discontinuity between the spatial limits of P. falciparum malaria transmission (defined, where possible, by the average PfAPI in the three most recently recorded years ) and the endemicity data (PfPR collected since 1985). It is obvious that the predicted map represents a snapshot of the year 2007 from a malaria endemicity that changes through time. No degree of statistical sophistication can circumvent the fact that additional data will increase the fidelity of the map, by either increasing the spatial resolution of the malariometric surveys or updating an existing survey location with more recent information. The methods have been devised specifically so that these surfaces can be updated rapidly. The predominantly univariate approach adopted also means changes in future maps' iterations can be attributed reliably to finding more data in areas of high uncertainty (changes in space) or to changes brought about through intervention success or disease recession (changes in time), rather than any temporal and spatial mix of the relationship of the PfPR2−10 data and the environmental covariates. We encourage the submission of additional existing data to improve the map in areas where we have least spatial accuracy, and new data to sustain future production of updated contemporary maps. Current areas of highest uncertainty are indicated to a good approximation by the inverse of the class prediction probability (Figure 5), although future work is aimed at refining this information. Therefore, an immediate priority is to generate regional maps showing the optimal location of new surveys that would need to be implemented to maximally reduce the variance in the existing endemicity surface for the minimum cost. These solutions are substantially more involved than the list of areas with highest variance provided here because (i) each new survey will change the structure of the spatial variance and affect the optimal location of the next survey; (ii) both the number and spatial distribution of surveys will affect the outcome and require multiple simulations to converge on optimal solutions; and (iii) potential survey locations will need to be weighted appropriately by the distribution of the human population. Immediate MAP Goals The initial focus of the MAP has been P. falciparum  due to its global epidemiological significance  and its better prospects for control and local elimination . We have not yet addressed the significant problem of P. vivax burden  despite its increasingly recognised clinical importance [104–106], but have archived over 2,500 P. vivax parasite rate surveys with which to start this process. Another immediate goal is in refining global burden of disease estimates for P. falciparum (both morbidity  and mortality [48,107,108]) to support global estimation of antimalarial intervention and commodity needs. The statistical methods used in this analysis will allow the next iteration of burden estimates to represent more holistically and robustly the uncertainty in predictions. In the medium term, combinations of these global endemicity maps with forthcoming maps of the distribution of the dominant Anopheles vectors of human malaria  should empower malaria control managers to make more informed decisions regarding interventions appropriate to the bionomics of their local suite of vectors. In the long term we hope to not only monitor and evaluate progress with these maps, but to increase our ability to model future malaria endemicity and support objective assessment of where in the world it might be possible to eliminate malaria. Conclusions The state of the P. falciparum malaria world in 2007 represents an enormous opportunity for the international community to act [109,110], but these actions remain considerably under-resourced . Regardless of whether nations champion sustained, intensive control or reach for the higher ambition of malaria elimination [2–4,74,112–114], the intermediate intervention paths are similar . This cartographic resource will help countries determine their needs and serve as a baseline to monitor and evaluate progress towards interventional goals. We wish to continue to work alongside individuals, countries, and regions to improve future iterations of this map and document hopefully these intervention successes. Supporting Information Alternative Language Text S1 Translation of the Article into French by Frédéric Piel and Stéphanie Loute (1.04 MB DOC) Click here for additional data file. Alternative Language Text S2 Translation of the Article into Chinese by Robert Li (438 KB DOC) Click here for additional data file. Alternative Language Text S3 Translation of the Article into Indonesian by Iqbal R.F. Elyazar and Siti Nurlela (1.08 MB DOC) Click here for additional data file. Alternative Language Text S4 Translation of the Article into Vietnamese by Bui H. Manh (549 KB DOC) Click here for additional data file. Alternative Language Text S5 Translation of the Article into Spanish by Carlos A. Guerra (796 KB DOC) Click here for additional data file. Protocol S1 The PfPR Malariometric Survey Database S1.1 Summary of Data Search and Data Abstraction Procedures S1.2 Data Exclusion Rules S1.3 Age-Standardisation S1.4 Semi-Variograms of PfPR2−10 Data by Region S1.5 Geostatistical Filter for the Detection of Extreme Outliers S1.6 Malariometric Survey Data Summary and Descriptive Statistics S1.7 Relationships with Environmental Covariates (3.4 MB DOC) Click here for additional data file. Protocol S2 Demographic Databases and Procedures S2.1 Parasite Rate Survey Urban/Peri-Urban/Rural Classification Rules S2.2 Urban/Peri-Urban/Rural Status and Prevalence S2.3 GRUMP alpha Human Population Surface S2.4 PAR Derivation (2.5 MB DOC) Click here for additional data file. Protocol S3 Model Based Geostatistical Procedures S3.1 Overview of the Statistical Model S3.2 Prior Specification S3.3 Age-Standardization S3.4 Implementation Details S3.5 Overview of Map Generation (23 MB DOC) Click here for additional data file. Protocol S4 Model Validation Procedures S4.1 Creation of the Validation Sets S4.2 Procedures for Testing Model Performance S4.3 Additional Results (26 MB DOC) Click here for additional data file.