Introduction The international agenda shaping malaria control financing, research, and implementation is increasingly defined around the goal of regional elimination –. This ambition ostensibly extends to all human malarias, but whilst recent years have seen a surge in research attention for Plasmodium falciparum, the knowledge-base for the other major human malaria, Plasmodium vivax, is far less developed in almost every aspect –. During 2006–2009 just 3.1% of expenditures on malaria research and development were committed to P. vivax . The notion that control approaches developed primarily for P. falciparum in holoendemic Africa can be transferred successfully to P. vivax is, however, increasingly acknowledged as inadequate –. Previous eradication campaigns have demonstrated that P. vivax frequently remains entrenched long after P. falciparum has been eliminated . The prominence of P. vivax on the global health agenda has risen further as evidence accumulates of its capacity in some settings to cause severe disease and death –, and of the very large numbers of people living at risk . Amongst the many information gaps preventing rational strategies for P. vivax control and elimination, the absence of robust geographical assessments of risk has been identified as particularly conspicuous , . The endemic level of the disease determines its burden on children, adults, and pregnant women; the likely impact of different control measures; and the relative difficulty of elimination goals. Despite the conspicuous importance of these issues, there has been no systematic global assessment of endemicity. The Malaria Atlas Project was initiated in 2005 with an initial focus on P. falciparum that has led to global maps – for this parasite being integrated into policy planning at regional to international levels , –. Here we present the outcome of an equivalent project to generate a comprehensive evidence-base on P. vivax infections worldwide, and to generate global risk maps for this hitherto neglected disease. We build on earlier work  defining the global range of the disease and broad classifications of populations at risk to now assess the levels of endemicity under which these several billion people live. This detailed depiction of geographically varying risk is intended to contribute to a much-needed paradigm shift towards geographically stratified and evidence-based planning for P. vivax control and elimination. Numerous biological and epidemiological characteristics of P. vivax present unique challenges to defining and mapping metrics of risk. Unlike P. falciparum, infections include a dormant hypnozoite liver stage that can cause clinical relapse episodes , . These periodic events manifest as a blood-stage infection clinically indistinguishable from a primary infection and constitute a substantial, but geographically varying, proportion of total patent infection prevalence and disease burden within different populations , –. The parasitemia of P. vivax typically occurs at much lower densities compared to those of falciparum malaria, and successful detection by any given means of survey is much less likely. Another major driver of the global P. vivax landscape is the influence of the Duffy negativity phenotype . This inherited blood condition confers a high degree of protection against P. vivax infection and is present at very high frequencies in the majority of African populations, although is rare elsewhere . These factors, amongst others, mean that the methodological framework for mapping P. vivax endemicity, and the interpretation of the resulting maps, are distinct from those already established for P. falciparum , . The effort described here strives to accommodate these important distinctions in developing a global distribution of endemic vivax malaria. Methods The modelling framework is displayed schematically in Figure 1. In brief, this involved (i) updating of the geographical limits of stable P. vivax transmission based on routine reporting data and biological masks; (ii) assembly of all available P. vivax parasite rate data globally; (iii) development of a Bayesian model-based geostatistical model to map P. vivax endemicity within the limits of stable transmission; and (iv) a model validation procedure. Details on each of these stages are provided below with more extensive descriptions included as Protocols S1, S2, S3, and S4. 10.1371/journal.pntd.0001814.g001 Figure 1 Schematic overview of the mapping procedures and methods for Plasmodium vivax endemicity. Blue boxes describe input data. Orange boxes denote models and experimental procedures; green boxes indicate output data (dashed lines represent intermediate outputs and solid lines final outputs). U/R = urban/rural; UNPP = United Nations Population Prospects. Labels S1-4 denote supplementrary information in Protocols S1, S2, S3, and S4. Updating Estimates of the Geographical Limits of Endemic Plasmodium vivax in 2010 The first effort to systematically estimate the global extent of P. vivax transmission and define populations at risk was completed in 2009 . As a first step in the current study, we have updated this work with a new round of data collection for the year 2010. The updated data assemblies and methods are described in full in Protocol S1. In brief, this work first involved the identification of 95 countries as endemic for P. vivax in 2010. From these, P. vivax annual parasite incidence (PvAPI) routine case reports were assembled from 17,893 administrative units . These PvAPI and other medical intelligence data were combined with remote sensing surfaces and biological models  that identified areas where extreme aridity or temperature regimes would limit or preclude transmission (see Protocol S1). These components were combined to classify the world into areas likely to experience zero, unstable (PvAPI 0.9) to have a PvPR1–99 less than 1% were classed as unstable. Areas in which Duffy negativity gene frequency is predicted to exceed 90%  are shown in hatching for additional context. Modelling Plasmodium vivax Endemicity within Regions of Stable Transmission We adopt model-based geostatistics (MBG) ,  as a robust and flexible modelling framework for generating continuous surfaces of malaria endemicity based on retrospectively assembled parasite rate survey data , , . MBG models are a special class of generalised linear mixed models, with endemicity values at each target pixel predicted as a function of a geographically-varying mean and a weighted average of proximal data points. The mean can be defined as a multivariate function of environmental correlates of disease risk. A covariance function is used to characterise the spatial or space-time heterogeneity in the observed data, which in turn is used to define appropriate weights assigned to each data point when predicting at each pixel. This framework allows the uncertainty in predicted endemicity values to vary between pixels, depending on the observed variation, density and sample size of surveys in different locations and the predictive utility of the covariate suite. Parts of the map where survey data are dense, recent, and relatively homogenous will be predicted with least uncertainty, whilst regions with sparse or mainly old surveys, or where measured parasite rates are extremely variable, will have greater uncertainty. When MBG models are fitted using Bayesian inference and a Markov chain Monte Carlo (MCMC) algorithm, uncertainty in the final predictions as well as all model parameters can be represented in the form of predictive posterior distributions . We developed for this study a modified version of the MBG framework used previously to model P. falciparum endemicity , , with some core aspects of the model structure remaining unchanged and others altered to capture unique aspects of P. vivax biology and epidemiology. The model is presented in full in Protocol S3. As in earlier work , , , we adopt a space-time approach to allow surveys from a wide time period to inform predictions of contemporary risk. This includes the use of a spatiotemporal covariance function which is parameterised to downweight older data appropriately. We also retain a seasonal component in the covariance function, although we note that seasonality in transmission is often only weakly represented in PvPR in part because of the confounding effect of relapses occurring outside peak transmission seasons . A minimal set of covariates were included to inform prediction of the mean function, based on a priori expectations of the major environmental factors modulating endemicity. These were (i) an indicator variable defining areas as urban or rural based on the Global Rural Urban Mapping Project (GRUMP) urban extent product , ; (ii) a long-term average vegetation index product as an indicator of overall moisture availability for vector oviposition and survival , ; and (iii) a P. vivax specific index of temperature suitability derived from the same model used to delineate suitable areas on the basis of vector survival and sporogony . Age Standardisation Our assembly of PvPR surveys was collected across a variety of age ranges and, since P. vivax infection status can vary systematically in different age groups within a defined community, it was necessary to standardise for this source of variability to allow all surveys to be used in the same model. We adopted the same model form as has been described  and used previously for P. falciparum , , whereby population infection prevalence is expected to rise rapidly in early infancy and plateau during childhood before declining in early adolescence and adulthood. The timing and relative magnitude of these age profile features are likely distinct between the two parasites in different endemic settings , , and so the model was parameterised using an assembly of 67 finely age-stratified PvPR surveys (Protocol S2), with estimation carried out in a Bayesian model using MCMC. The parameterised model was then used to convert all observed survey prevalences to a standardised age-independent value for use in modelling, and then further allowed the output prevalence predictions to be generated for any arbitrary age range. We chose to generate maps of all-age infection prevalence, defined as individuals of age one to 99 years (thus PvPR1–99). We excluded infection in those less than one year of age from the standardisation because of the confounding effect of maternal antibodies, and because parasite rate surveys very rarely sample young infants. We deviated from the two-to-ten age range used for mapping P. falciparum ,  because the relatively lower prevalences has meant that surveys are far more commonly carried out across all age ranges. Incorporating Duffy Negativity Since Duffy negative individuals are largely refractory to P. vivax infection , high population frequencies of this phenotype have a dramatic suppressing effect on endemicity, even where conditions are otherwise well suited for transmission . The predominance of Duffy negativity in Africa has led to a historical perception that P. vivax is absent from much of the continent, and a dearth of surveys or routine diagnoses testing for the parasite have served to entrench this mantra . However, evidence exists of autochthonous P. vivax transmission across the continent , and therefore we did not preclude any areas at risk a priori. Instead, we used a recent map of estimated Duffy negativity phenotypic frequency  and incorporated the potential influence of this blood group directly in the MBG modelling framework. The mapped Duffy-negative population fraction at each location was excluded from the denominator in PvPR survey data, such that any P. vivax positive individuals were considered to have arisen from the Duffy positive population subset. Thus in a location with 90% Duffy negativity, five positive individuals in a survey of 100 would give an assumed prevalence of 50% amongst Duffy positives. Correspondingly, prediction of PvPR was then restricted to the Duffy positive proportion at each pixel, with the final prevalence estimate re-converted to relate to the total population. This approach has two key advantages. First, predicted PvPR at each location could never exceed the Duffy positive proportion, therefore ensuring biological consistency between the P. vivax and Duffy negativity maps. Second, where PvPR survey data were sparse across much of Africa, the predictions could effectively borrow strength from the Duffy negativity map because predictions of PvPR were restricted to a much narrower range of possible values. Model Implementation and Map Generation The P. vivax endemic world was divided into four contiguous regions with broadly distinct biogeographical, entomological and epidemiological characteristics: the Americas and Africa formed separate regions, whilst Asia was subdivided into Central and South East sub-regions with a boundary at the Thailand-Malaysia border (see Protocol S2). This regionalisation was implemented in part to retain computational feasibility given the large number of data points, but also to allow model parameterisations to vary and better capture regional endemicity characteristics. Within each region, a separate MBG model was fitted using a bespoke MCMC algorithm  to generate predictions of PvPR1–99 for every 5×5 km pixel within the limits of stable transmission. The prediction year was set to 2010 and model outputs represent an annualised average across the 12 months of that year. Model output consisted of a predicted posterior distribution of PvPR1–99 for every pixel. A continuous endemicity map was generated using the mean of each posterior distribution as a point estimate. The uncertainty associated with predictions was summarised by maps showing the ratio of the posterior distribution inter-quartile range (IQR) to its mean. The IQR is a simple measure of the precision with which each PvPR value was predicted, and standardisation by the mean produced an uncertainty index less affected by underlying prevalence levels and more illustrative of relative model performance driven by data densities in different locations. This index was then also weighted by the underlying population density to produce a second map indicative of those areas where uncertainty is likely to be most operationally important. Refining Limits Definition and Population at Risk Estimates In some regions within the estimated limits of stable transmission, PvPR1–99 was predicted to be extremely low, either because of a dense abundance of survey data reporting zero infections or, in Africa, because of very high coincident Duffy negativity phenotype frequencies. Such areas are not appropriately described as being at risk of stable transmission and so we defined a decision rule whereby pixels predicted with high certainty (probability >0.9) of being less than 1% PvPR1–99 were assigned to the unstable class, thereby modifying the original transmission limits. These augmented mapped limits were combined with a 2010 population surface derived from the GRUMP beta version ,  to estimate the number of people living at unstable or stable risk within each country and region. The fraction of the population estimated to be Duffy negative  within each pixel was considered at no risk and therefore excluded from these totals. Model Validation A model validation procedure was implemented whereby 10% of the survey points in each model region were selected using a spatially declustered random sampling procedure. These subsets were held out and the model re-fitted in full using the remaining 90%. Model predictions were then compared to the hold-out data points and a number of different aspects of model performance were assessed using validation statistics described previously , . The validation procedure is detailed in full in Protocol S4. Results Model Validation Full validation results are presented in Protocol S4. In brief, examination of the mean error in the generation of the P. vivax malaria endemicity point-estimate surface revealed minimal overall bias in predicted PvPR with a global mean error of −0.41 (Americas −1.38, Africa 0.03, Central Asia −0.43, South East Asia −0.43), with values in units of PvPR on a percentage scale (see Protocol S4). The global value thus represents an overall tendency to underestimate prevalence by just under half of one percent. The mean absolute error, which measures the average magnitude of prediction errors, was 2.48 (Americas 5.05, Africa 0.53, Central Asia 1.52, South East Asia 3.37), again in units of PvPR (see Protocol S4). Global Plasmodium vivax Endemicity and Populations at Risk in 2010 The limits of stable and unstable P. vivax transmission, as defined using PvAPI, biological exclusion masks and medical intelligence data are shown in Figure 2A. The continuous surface of P. vivax endemicity predicted within those limits is shown in Figure 2B. The uncertainty map (posterior IQR:mean ratio) is shown in Figure 3A and the population-weighted version in Figure 3B. 10.1371/journal.pntd.0001814.g003 Figure 3 Uncertainty associated with predictions of Plasmodium vivax endemicity. Panel A shows the ratio of the posterior inter-quartile range to the posterior mean prediction at each pixel. Large values indicate greater uncertainty: the model predicts a relatively wide range of PvPR1–99 as being equally plausible given the surrounding data. Conversely, smaller values indicate a tighter range of values have been predicted and, thus, a higher degree of certainty in the prediction. Panel B shows the same index multiplied by the underlying population density and rescaled to 0–1 to correspond to Panel A. Higher values indicate areas with high uncertainty and large populations. We estimate that P. vivax was endemic across some 44 million square kilometres, approximately a third of the Earth's land surface. Around half of this area was located in Africa (51%) and a quarter each in the Americas (22%) and Asia (27%) (Table 1). However, the uneven distribution of global populations, coupled with the protective influence of Duffy negativity in Africa, meant that the distribution of populations at risk was very different. An estimated 2.48 billion people lived at any risk of P. vivax in 2010 (Table 1), of which a large majority lived in Central Asia (82%) with much smaller fractions in South East Asia (9%), the Americas (6%), and Africa (3%). Of these, 1.52 billion lived in areas of unstable transmission where risk is very low and case incidence is unlikely to exceed one per 10,000 per annum. The remaining 964 million people at risk lived in areas of stable transmission, representing a wide diversity of endemic levels. The global distribution of populations in each risk class was similar to the total at risk, such that over 80% of people in both classes lived in Central Asia (Table 1). 10.1371/journal.pntd.0001814.t001 Table 1 Area and populations at risk of Plasmodium vivax malaria in 2010. Region Area (million km2) Population (millions) Unstable Stable Any risk Unstable Stable Any risk America 1.38 8.08 9.46 87.66 49.79 137.45 Africa+ 20.60 1.86 22.46 48.72 37.66 86.38 C Asia 5.60 3.63 9.24 1,236.92 812.55 2,049.47 SE Asia 0.96 1.78 2.74 150.17 64.90 215.07 World 28.55 15.35 43.90 1,523.47 964.90 2,488.37 Risk is stratified into unstable risk (PvAPI 6,000/uL among inpatients classified as having not serious, serious, and fatal illness with a diagnosis of P. vivax compared to P. falciparum . Further, the majority of case reports describing severe and fatal illness with a diagnosis of vivax malaria typically show parasitemia >5,000/uL. In contrast, the World Health Organization threshold for severe illness attributable to hyperparasitemia with P. falciparum is >200,000/uL . In brief, the relationship between prevalence and risk of disease and transmission for P. vivax is distinct from that for P. falciparum, and it is weighted more heavily towards substantial risks at much lower parasite densities and levels of prevalence of microscopically patent parasitemia. The capacity of P. vivax hypnozoites to induce relapsing infections has a number of important implications. First, because dormant liver stage infections are not detectable in routine parasite rate surveys, our maps do not capture the potentially very large reservoir of asymptomatic infections sequestered in each population. Evidence is emerging that this hidden reservoir may be substantially larger than previously thought, with long-latency P. vivax phenotypes both prevalent and geographically widespread . Whilst not contributing to clinical disease until activated, these dormant hypnozoites ultimately play a vital role in sustaining transmission since they are refractory to blood-stage antimalarial chemotherapy and interventions to reduce transmission. Hypnozoites also ensure an ability of P. vivax to survive in climatic conditions that cannot sustain P. falciparum transmission. Second, the P. vivax parasite rates observed in population surveys detect both new and relapsing infections, although the two are almost never distinguishable. This confounds the relationship between observed infection prevalence and measures of transmission intensity such as force of infection or the entomological inoculation rate. This, in turn, has implications for the use of transmission models seeking to evaluate or optimise control options for P. vivax , , , . The current unavailability of any diagnostic method for detecting hypnozoites  and our resulting ignorance about the size and geographic distribution of this reservoir therefore remain critical knowledge gaps limiting the feasibility of regional elimination . It is also worth noting that conventional parasite rate data do not measure multiplicity of infection which is an additional potential confounding effect between observed infection prevalence and transmission intensity. P. vivax in Africa and Duffy Polymorphism Our map of P. vivax endemicity and estimates of populations at risk in Africa are heavily influenced by a single assumption: that the fraction of the population estimated to be negative for the Duffy antigen  is refractory to infection with P. vivax. A body of empirical evidence is growing, however that P. vivax can infect and cause disease in Duffy negative individuals, as reported in Madagascar  and mainland sub-Saharan Africa – as well as outside Africa , . Whether the invasion of erythrocytes via Duffy antigen-independent pathways is a newly evolved mechanism, or whether this capacity has been overlooked by the misdiagnosis of P. vivax in Africa as P. ovale remains unresolved , , . Whilst this accumulated evidence stands contrary to our simplifying assumption of complete protection in Duffy negative individuals, there is currently no evidence to suggest that such infections are anything but rare and thus are unlikely to have any substantive influence on the epidemiology or infection prevalence of P. vivax at the population scale throughout most of Africa. We also make no provision in our model for a protective effect in Duffy-negative heterozygotes, although such protection has been observed in some settings –. The movement and mixing within Africa of human populations from diverse ethnographic backgrounds complicates contemporary patterns of Duffy negativity and, in principle, could yield local populations with substantially reduced protection from P. vivax infection in the future. Indeed, the implications for our map of population movement go beyond the effect of Duffy negativity: the carriage of parasites from high to low endemic regions, for example by migratory workers, may play an important role in sustaining transmission in some regions and further research is required to investigate such processes. Mapping to Guide Control There exists for P. falciparum a history of control strategies linked explicitly to defined strata of endemicity, starting with the first Global Malaria Eradication Programme , ,  and undergoing a series of refinements that now feature in contemporary control and elimination efforts. Most recently, stratification has been supported by insights gained from mathematical models linking endemic levels to optimum intervention suites, control options, and timelines for elimination planning , –. In stark contrast, control options for P. vivax are rarely differentiated by endemicity, and there is little consensus around how this may be done. In part, the absence of agreed control-oriented strata of P. vivax endemicity stems from the biological complexities and knowledge gaps that prevent direct interpretation of infection prevalence as a metric for guiding control. It is also to some extent inevitable that the dogma of unstratified control becomes self-propagating: risk maps are not created because control is not differentiated by endemicity, but that differentiation cannot proceed without reliable maps. As well as providing a basis for stratified control and treatment, the endemicity maps presented here have a number of potential applications in combination with other related maps. First, there is an urgent need to better identify regions where high P. vivax endemicity is coincident with significant population prevalence of glucose-6-phosphate dehydrogenase deficiency (G6PDd). This inherited blood disorder plays a key role in chemotherapy policy for P. vivax because primaquine, the only registered drug active against the hypnozoite liver stage is contra-indicated in G6PDd individuals in whom it can cause severe and potentially fatal haemolytic reactions , . A new global map of G6PDd prevalence is now available (Howes et al, submitted) which can be combined with the endemicity maps presented here to provide a rational basis for estimating adverse outcomes and setting appropriate testing and treatment protocols. Moreover, in practice most clinical infections are managed without differentiating the causative parasite species: combining the endemicity maps for P. vivax and P. falciparum may therefore inform unified strategies for malaria control programs and policy . It has been proposed, for example, that artemesinin-based combination therapy (ACT) be adopted for all presumptively diagnosed malaria in areas coendemic for both species, as opposed to a separate ACT/chloroquine treatment strategy . Further, in some regions more than 50% of patients diagnosed with falciparum malaria go on to experience an attack of vivax malaria in the absence of risk of reinfection . This high prevalence of hypnozoites may also justify presumptive therapy with primaquine against relapse with any diagnosis of malaria where the two species occur at relatively high frequencies. Such geographically specific cross-parasite treatment considerations hinge on robust risk maps for both species. Future Challenges in P. vivax Cartography Numerous research and operational challenges remain unaddressed that would provide vital insights into the geographical distribution of P. vivax and its impacts on populations. Perhaps the highest priority is to improve understanding of the link between infection prevalence and clinical burden in both P. vivax mono-endemic settings and where it is coendemic with P. falciparum. Official estimates of national and regional disease burdens for P. vivax remain reliant on routine case reporting of unknown fidelity and are only crudely distinguished from P. falciparum . It is illuminating that only 53 of the 95 P. vivax endemic countries were able to provide vivax-specific routine case reporting data, and there is a clear mandate for strengthening the routine diagnosis and reporting of P. vivax cases. Cartographic approaches to estimating P. vivax burden can therefore play a crucial role in triangulating with these estimates to provide insight into the distribution of the disease independent of health system surveillance and its attendant biases , –. There is also a particular need to define burden and clinical outcomes associated with P. vivax in pregnancy ,  and other clinically vulnerable groups, most notably young children. Linking infection prevalence to clinical burden implies the need to better understand the contribution of relapsing infections to disease. Whilst the magnitude of this contribution is known to be highly heterogeneous, its geographical pattern is poorly measured and causal factors only partially understood , . Further challenges lie in understanding how P. falciparum and P. vivax interact within human hosts and how these interactions manifest at population levels. Comparison of the maps for each species reveals a complete spectrum from areas endemic for only one parasite through to others where both species are present at broadly equal levels. Whilst identifying these patterns of coendemicity is an important first step, the implications in terms of risks of coinfection and clinical outcomes, antagonistic mechanisms leading to elevated severe disease risk, or cross-protective mechanisms of acquired immunity remain disputed , –. Conclusions To meet international targets for reduced malaria illness and death, and to progress the cause of regional elimination, the malaria research and control communities can no longer afford to neglect the impact of P. vivax. Its unique biology and global ubiquity present challenges to its elimination that greatly surpass those of its higher-profile cousin, P. falciparum. Making serious gains against the disease will require substantive strengthening of the evidence base on almost every aspect of its biology, epidemiology, control and treatment. The maps presented here are intended to contribute to this effort. They are all made freely available from the MAP website  along with regional and individual maps for every malaria-endemic country. Users can access individual map images or download the global surfaces for use in a geographical information system, allowing them to integrate this work within their own analyses or produce bespoke data overlays and displays. We will also make available, where permissions have been obtained, all underlying P. vivax parasite rate surveys used in this work. Supporting Information Protocol S1 Updating the global spatial limits of Plasmodium vivax malaria transmission for 2010. S1.1 Overview. S1.2 Identifying Countries Considered P. vivax Malaria Endemic. S1.3 Updating National Risk Extents with P. vivax Annual Parasite Incidence Data. S1.4 Biological Masks of Transmission Exclusion. S1.5 Risk Modulation Based on Medical Intelligence. S1.6 Assembling the P. vivax Spatial Limits Map. S1.7 Refining Regions of Unstable Transmission after MBG Modelling. S1.8 Predicting Populations at Risk of P. vivax in 2010. (DOC) Click here for additional data file. Protocol S2 The Malaria Atlas Project Plasmodium vivax parasite prevalence database. S2.1 Assembling the PvPR Data. S2.2 Database Fidelity Checks. S2.3 Data Exclusions. S2.4 The PvPR Input Data Set. S2.5 Age-Standardisation. S2.6 Regionalisation. (DOC) Click here for additional data file. Protocol S3 Bayesian model-based geostatistical framework for predicting Pv PR1–99. S3.1 Bayesian Inference. S3.2 Model Overview. S3.3 Formal Presentation of Model. (DOC) Click here for additional data file. Protocol S4 Model validation procedures and additional results. S4.1 Creation of Validation Sets. S4.2 Procedures for Testing Model Performance. S4.3 Validation Results. (DOC) Click here for additional data file.