malariaAtlas: an R interface to global malariometric data hosted by the Malaria Atlas Project

malariaAtlas currently enables users to download, visualize and manipulate three types of data: parasite rate (PR) survey data; administrative boundary shapefiles; and a large suite of rasters covering a range of modelled outputs related to malaria research (see Table 1). Georeferenced PR survey data is a core component of MAP’s data estate and a common measure of malaria endemicity [1, 28]. The PR survey points entered into MAP’s database are screened for robust sampling methods and geographic specificity to ensure they provide representative parasite species-specific information on the local prevalence of malaria infection. This database includes 73,326 survey points as of July 2018 (64,685 Plasmodium falciparum; 14,412 Plasmodium vivax), covering the period 1975–2017. In addition to georeferenced data on malaria endemicity, up-to-date and topologically correct shapefiles of a region’s administrative boundaries are fundamental to visualizing, interpreting and analysing spatial epidemiological data. As such, MAP maintains a collated set of subnational administrative boundary shapefiles assembled from various publicly available sources (see [29]). MAP also makes a large number of raster grids publicly available, representing the major outputs of MAP’s spatiotemporal epidemiological research. At the time of writing, 86 raster surfaces were available to download using malariaAtlas. These cover a variety of relevant metrics, such as predicted malaria parasite prevalence, clinical incidence and malaria-attributable fever [8, 30‐32]; prevalence of malaria-related human blood disorders [33‐35]; predicted risk of zoonotic Plasmodium knowlesi infection [36]; predicted mosquito vector distribution and relative abundance [37‐40]; coverage of insecticide-treated bed nets (ITNs), indoor residual spraying (IRS) and artemisinin-based combination therapy (ACT) [8]; and travel time to cities [41]. By providing an R-interface to MAP’s hosted survey data, shapefiles and rasters, malariaAtlas enhances direct and reproducible access to this data source.

Using malariaAtlas to download and visualize data from MAP in R is achieved through four main classes of functions as outlined in Table 2. These include: ‘list’ functions that allow the user to see how much data is available for a given data type; ‘get’ functions for data downloads; ‘autoplot’ methods that enable quick visualisation of downloaded data using functions from the ggplot2 package [42]; and a number of utility functions that enable common manipulations of downloaded data (see Table 2).

Within malariaAtlas, the functions listPoints, getPR and autoplot.pr.points provide a quick and simple way of downloading and visualising publicly available PR survey data hosted by MAP. listPoints returns a data.frame outlining the countries for which parasite rate survey data is available in MAPs database. getPR returns a data.frame of geo-located PR point data including: number of individuals examined; number of positive diagnoses by species; age range of the sample population; sampling date and location information; diagnostic method(s) used; and source citation. Arguments are included to enable queries based on location (Continent; Country Name; 3 letter ISO code; or spatial extent) and species (either P. falciparum or P. vivax). The returned data has the additional class ‘pr.points’ which enables quick visualization of downloaded points using autoplot. A subset of the PR survey points in MAP’s database remain confidential, in accordance with the respective data-use agreements under which they have been shared. For these confidential data points, MAP has either limited or no permission to share measured PR values and/or geo-location data, however citations to the original data source are provided for all downloaded points. Accordingly, data-sharing restrictions for any given point are provided in the ‘permissions_info’ column of a downloaded pr.points data.frame. Figure 1 illustrates the use of malariaAtlas to download and visualise PR survey points, including maps of (a) the full database of available P. falciparum PR points at the time of publication (Fig. 1a, b) all PR survey points hosted by MAP from Tanzania (Fig. 1b).

Analogous to the functions described above, listShp, getShp and autoplot.mapShp allow users to download and visualise the set of shapefiles collated by MAP (see Table 2). listShp returns a data.frame indicating all administrative regions covered by these shapefiles along with their administrative level and corresponding parent administrative unit. getShp returns either a SpatialPolygons object or mapShp object (as chosen by the user) containing polygons at either ADMIN0 (national) or ADMIN1 (state; province) levels for any given country; and down to ADMIN3 level for some malaria-endemic countries. Quick visualisation of mapShp objects is possible through an autoplot method.

Rasters are the final datatype available to download and visualise using malariaAtlas via the functions: listRaster, getRaster and autoplot.mapRaster. listRaster returns a data.frame that serves as a catalogue of rasters available to download using getRaster, mirroring the catalogue of rasters available on MAP’s online interactive explorer tool (map.ox.ac.uk/explorer). This data.frame includes columns that provide descriptive metadata including an abstract outlining raster content, a citation to the original publication associated with a given raster, and the time period covered for time-varying raster datasets. getRaster provides the means to download one or more raster layers at a time, queried by location (using either an input SpatialPolygon shapefile or a user-defined extent (xmin, xmax, ymin, ymax)), and year (for temporally dynamic raster datasets). The data is returned as a Raster* object: a RasterLayer for a single raster; a RasterBrick for two or more rasters of the same extent/resolution; or a list of Raster* objects for two or more rasters of differing extents/resolutions. Downloaded rasters represent the mean predicted value from various geostatistical models. For further information on specific modelling approaches and/or associated uncertainty of predicted values users are encouraged to consult the associated publication (citation information available via listRaster) or to contact MAP directly. The utility function as.mapRaster converts any object downloaded using getRaster into a mapRaster object (long-format data.frame with columns x, y, z (longitude, latitude, value) and raster_name) enabling tabular manipulation and ggplot-friendly visualisation. Quick visualization of mapRaster objects is provided via included autoplot methods. Figure 1c illustrates example code used to download and quickly visualise a raster for a given shapefile extent via malariaAtlas.

Three additional utility functions have been added to provide an easy means to perform common data manipulations. extractRaster allows users to download values from MAP rasters at specific point locations supplied in a user-specified set of coordinates (see malariaAtlas Vignette; [27]). This enables users to input a list of locations (latitude, longitude) and get back the associated raster value (e.g. malaria prevalence) for each location. as.mapShp and as.mapRaster provide a means of converting between Spatial* class objects (for polygon data) or Raster* class objects (for raster data) to the malariaAtlas data.frame-based object classes mapShp and mapRaster respectively. This permits tabular manipulation and ggplot-friendly plotting through provided autoplot methods. convertPrevalence is an additional utility function that provides a principled approach to age-standardization of malaria prevalence data [43], based on models defined by Smith et al. [28] for P. falciparum and Gething et al. [30] for P. vivax. Altogether, the above functions provide a simple means of downloading, visualising and manipulating spatial malariometric data. The flexibility of R as a statistical software platform and the wealth of existing R packages enable users to easily extend their analysis beyond these functions and integrate malariaAtlas into more complex analytical workflows.

To further aid the dissemination and use of these data, malariaAtlas modules were developed for the species distribution modelling software zoon [44]. Zoon provides a modular framework for species distribution modelling, allowing users to collect and model data in a simple pipeline. Species distribution modelling is a subfield of ecology in which the spatial distribution of an organism is estimated from known presence and absence (if available) locations. There are strong parallels between species distribution modelling and parasite rate mapping as both use binomial data to estimate a spatial probability surface; in species distribution modelling this surface is the probability of species occurrence while in parasite rate mapping the surface is probability of infection. Two zoon modules have been added (‘malariaAtlas_PR’ and ‘malariaAtlas_covariates’) allowing parasite rate surveys to be used as response data and raster data to be used as covariates within a zoon workflow. The parasite rate data offers a useful benchmark dataset for testing new methods. However, state-of-the-art models of malaria prevalence (e.g. [8]) are currently beyond the scope of zoon, and as such zoon is not expected to be directly used for risk mapping and/or policymaking.

The first mock analysis illustrates the use of malariaAtlas to download response and covariate data for use in spatial epidemiological analysis. P. vivax parasite rate (PvPR) survey points and covariate raster data were downloaded using malariaAtlas (see Box 1) and used to fit a Bayesian geostatistical model of malaria risk (see full example code in Additional file 1). For illustrative purposes, an arbitrary spatial extent was chosen for this analysis. All PvPR points in the study area were downloaded using getPR, and then subsetted to only publicly available data for analysis. convertPrevalence was used to standardize values to all-ages PvPR (see Fig. 2a; Box 1). The R-INLA package [45, 46] was used to fit a Bayesian geostatistical model with a binomial likelihood to these data. Covariate data included rasters of environmental factors (night-time land surface temperature [47]; log elevation [48]; rainfall [49]) and log travel time to the nearest city (downloaded using getRaster as in Box 1, hereafter referred to as ‘human accessibility’; [41]). These fixed effects were given minimally informative (INLA default) priors.

The spatial autocorrelation in the data was modelled using a continuous, spatial Gaussian random field with a Matern covariance function [45]. The hyperparameters of the random field were given Penalised Complexity (PC) priors, which by design prefer a simpler model with a smoother random field [50]. The hyperparameters of the random field are the range (the distance within which the correlation of the field is essentially zero) and the standard deviation (the amount the field can vary). For the current model, the priors on these values were parameterised by setting the probability that the range of the field was smaller than an extreme minimum value (2 decimal degrees) as 0.01 and the probability that the standard deviation of the field was greater than an extreme maximum value (2.7) as 0.01. A random field with a standard deviation of 2.7 would be able to explain all the residual variance from a previously fitted logistic regression. The above prior was thus defined such that this undesirable level of overfitting was unlikely.

The fitted model was used to predict PvPR across the spatial extent of the study area (see Fig. 2b). Within this model, night-time temperature and elevation were significant predictors of PvPR (estimated coefficients (95% CI) of − 0.98 (− 1.70 to − 0.30) and − 1.43 (− 2.69 to − 0.38) respectively), while human accessibility did not significantly predict PvPR (− 0.16 (− 0.44 to 0.16)). Overall interpretation of these results is limited due to its small sample size and arbitrary spatial extent. Nevertheless, this mock analysis illustrates the use of malariaAtlas to download spatial malariometric response and covariate data for incorporation into further analysis.

The second mock analysis demonstrates how malariaAtlas can be used to access malariometric data within a zoon workflow [44]. As an illustrative example, this analysis investigates whether including mosquito occurrence data can improve predictive models of PfPR, using data from a second arbitrary study area (bounded by latitudes of − 24 and − 15 and longitudes of 44 and 49). A simple spatial validation scheme was implemented, using PfPR data from north of latitude − 20 (28,921 individuals from 208 locations) as a hold-out validation data set. Logistic regression models were fitted to two datasets and their predictive performance was compared. The first data set was simply the PfPR data from 116 locations and 8546 individuals south of latitude − 20. The second dataset was comprised of the same PfPR data with the addition of known occurrence locations of Anopheles arabiensis and Anopheles gambiae collected from the Global Biodiversity Information Facility [51], treating each vector occurrence location as equivalent to a single positive case of P. falciparum (total 147 locations and 8592 individuals/mosquitoes; see Fig. 3a). For covariates, WorldClim layers 1, 4, 12 and 15 (mean and within-year variation of temperature and precipitation [49]) as well as human accessibility [41] were used. PfPR data and human accessibility rasters were downloaded using malariaAtlas zoon modules (see Box 2). Model performance was compared using the AUC (Area Under the Curve) model evaluation criterion which assessed the ability of each model to correctly assign an infected/non-infected status to individuals in the hold-out set.

Including mosquito occurrence data very marginally improved predictive performance. AUC was 0.577 without mosquito occurrence data and 0.578 with the addition of mosquito data. Maps created using both models are shown in Fig. 3b, c showing almost identical outcomes. It is worth noting that the difference in model performance has no practical relevance. However, this serves as an illustrative example of how malariaAtlas data can be used within zoon to test new methods. Larger scale comparisons, and a less naive approach to incorporating mosquito data, would be needed to truly examine whether this method has analytical merit.