Approaches in malaria risk mapping continue to advance in scope with the advent of geostatistical techniques spanning both the spatial and temporal domains. A substantive review of the merits of the methods and covariates used to map malaria risk has not been undertaken. Therefore, this review aimed to systematically retrieve, summarise methods and examine covariates that have been used for mapping malaria risk in subSaharan Africa (SSA).
A systematic search of malaria risk mapping studies was conducted using PubMed, EBSCOhost, Web of Science and Scopus databases. The search was restricted to refereed studies published in English from January 1968 to April 2020. To ensure completeness, a manual search through the reference lists of selected studies was also undertaken. Two independent reviewers completed each of the review phases namely: identification of relevant studies based on the Preferred Reporting Items for Systematic Reviews and MetaAnalyses guidelines, data extraction and methodological quality assessment using a validated scoring criterion.
One hundred and seven studies met the inclusion criteria. The median quality score across studies was 12/16 (range: 7–16). Approximately half (44%) of the studies employed variable selection techniques prior to mapping with rainfall and temperature selected in over 50% of the studies. Malaria incidence (47%) and prevalence (35%) were the most commonly mapped outcomes, with Bayesian geostatistical models often (31%) the preferred approach to risk mapping. Additionally, 29% of the studies employed various spatial clustering methods to explore the geographical variation of malaria patterns, with Kulldorf scan statistic being the most common. Model validation was specified in 53 (50%) studies, with partitioning data into training and validation sets being the common approach.
Our review highlights the methodological diversity prominent in malaria risk mapping across SSA. To ensure reproducibility and quality science, best practices and transparent approaches should be adopted when selecting the statistical framework and covariates for malaria risk mapping. Findings underscore the need to periodically assess methods and covariates used in malaria risk mapping; to accommodate changes in data availability, data quality and innovation in statistical methodology.
The disproportionate decline of malaria risk overtime and between/within countries in subSaharan Africa attributed to biological, environmental, social and demographic factors has triggered a renewed interest in its finescale epidemiology.
Enhanced computational ability and availability of data of high quality and volume has enabled the quantification malaria risk burden in space and time leading to the proliferation of methods within a formal statistical framework.
The complexity of spatiotemporal models has increased, making inferential and predictive processes difficult to undertake.
The production of more granular estimates of malaria risk hinges on accessibility to and collection of timely data at finer resolutions.
Variable selection should be objectively developed to contribute to the maximum predictive accuracy of the spatiotemporal model.
Spatiotemporal approaches need to robustly quantify the subnational burden of malaria risk, as an epidemiological prerequisite to intervention strategies.
Investments in primary data collection at subnational scales, development and continuous application of robust modelling tools and approaches will be important for orienting malaria control and elimination efforts in the next decade.
As the malaria landscape diversifies, new tools will be required to not only highlight changes locally, but also to provide evidencebased insights into factors driving the change.
Global efforts to control and eliminate malaria are intrinsically linked to the Sustainable Development Goals.
The importance of malaria risk mapping in Africa can be traced back to the mid1950s when malaria epidemiology formed a critical prelude to the design of interventions aimed at eliminating malaria.
The science of malaria cartography has evolved from handdrawn risk maps to contemporary digital maps due to the demand for computational solutions and methodologies. These are needed to produce accurate estimates at a high spatial and temporal resolution to facilitate monitoring elimination progress within and between countries in SSA.
Previous reviews have been conducted to; identify environmental risk factors of malaria transmission,
The protocol guiding this review has been previously published.
All studies published between 1 January 1968 and 30 April 2020 were systematically searched through four electronic databases (PubMed, Web of Science, EBSCOhost and Scopus) using search terms defined in
Studies were screened independently by two authors JNO and CK for possible inclusion based on information provided in the title and abstract. Relevant studies based on the research questions were subsequently appraised on their eligibility for fulltext review. The fulltext review entailed the application of a more stringent inclusion/exclusion criteria for selecting studies to be included for data extraction. Additional papers were identified by examining the reference lists of retrieved studies and by contacting the authors where necessary. Emerging discrepancies were resolved by consensus and by an independent arbitrator (BS). A comprehensive and pilot tested form was used for data extraction.
Peerreviewed studies that employed spatial, temporal and spatiotemporal modelling techniques in malaria risk mapping in SSA were considered. A spatial model was defined as one that explicitly included a geographical index, while a temporal model included a time index. Studies using at least one visualisation or modelling technique (with or without covariates) for assessing the burden of malaria were included. Commentaries, expert reviews and/or reports that did not include original research were read, and only relevant studies cited included.
A standardised extraction form was used to independently extract the data by two reviewers (JNO and CK). The tool was first piloted and refined accordingly. Discordance between the reviewers with respect to the information extracted was resolved by consensus and by consulting with an independent arbitrator (BS). For each selected study, the following information was extracted (
Bibliographic information (Author, year, study setting and period, primary unit of analysis, spatial and temporal resolution).
Study objective(s)
Data aspects (data sources, malaria data, covariate type)
Analytical method (modelling approach(es), assumptions, cluster detection techniques, statistical tests, diagnostic/validation checks).
Results and discussions (key findings, modelling gaps, recommendation(s)).
A previously used 8point scoring criteria
A total of 7189 studies were retrieved from the various databases with 170 studies fully screened after the title and abstract review. Ultimately 107 studies were included for review and underwent quality assessment and synthesis (
Study flow from literature search to data extraction and analyses.
The distribution of studies by geographical scale and scope varied across SSA. Five (5%) studies were continental in scale, 48 (45%) studies were national and 52 (49%) studies were subnational. Kenya (10 studies) and Tanzania (9 studies) had the highest number of publications included in the review (
Geographical scale and scope of studies. Geographical scale (municipality, district, province/state, country) of studies is given in grey boxes. The studies covered 27 countries in subSaharan Africa with East Africa being the most represented subregion.
The longest study period spanned 115 years,
Bar—chart with a trend line (red) showing the total number of included studies.
The median score was 12 out of 16, with 16 representing the highest possible quality. The overall quality score of the reviewed studies ranged from 7 to 16. Two studies were of low quality, 22 studies were of medium quality, 42 studies were of high quality and 41 studies were of very high quality (
From the review, global, continental, national and subnational databases/repositories provided a rich source of both malaria data and covariates used for modelling. These sources comprised of geographically referenced surveys used by 34 (32%) studies, 20 (19%) studies used population databases and 10 (9%) studies used government records. Routinely collected data from the Health and Demographic Surveillance System were used in 16 (15%) studies. Sources of climatic and environmental covariates consisted of ground station observations used by 17 (16%) studies and remotely sensed satellite surrogates of climate, urbanisation and topography were employed by 49 (46%) studies (
Data sources
Type  Source  No  References 
Global/continental databases  Malaria Transmission Intensity and Mortality Burden across Africa  1 

Mapping Malaria Risk in Africa databases  9 
 
World Pop/Afripop  14 
 
Food and Agriculture OrganisationFood Security and Nutrition Analysis Unit  1 
 
Global Rural and Urban Mapping project  2 
 
WHO database on malaria drug resistance  1 
 
Global Lakes and Wetlands Database  3 
 
UN World Urbanisation prospects database  4 
 
National databases  Health and Demographic Surveillance System  16 

Census  6 
 
National statistical agencies  10 
 
Demographic Health Survey  7 
 
Malaria Indicator Survey  12 
 
Subnational databases  Crosssectional surveys  9 

Cohort studies  5 
 
Cluster surveys  1 
 
Entomological/parasitological surveys  5 
 
Remote sensing  Moderate Resolution Imaging Spectroradiometer  28 

Africa Data Dissemination Service  8 
 
United States Geological SurveyEarth Resources Observation and Science Centre  8 
 
Health Mapper  8 
 
Shuttle Radar Topographic Mission  5 
 
WorldClimGlobal Climate database  7 
 
Tropical Rainfall Measuring Mission  3 
 
Early Warning System  3 
 
Climate Research Unit  3 
 
National Oceanic and Atmospheric Administration  2 
 
Water Resources Institute  1 
 
World Wildlife Fund  1 
 
Africover  1 
 
Famine Early Warning Systems Network Land Data Assimilation System  1 
 
Ground station data  Meteorological data  17 

In this review, variable selection techniques were explicitly specified by 47 (44%) studies. These techniques varied substantially; with the frequentist approach used in 14 (13%) studies to assess the (uni and multi) variate association between malaria outcomes and its covariates being the most common. Significant covariates were included if their nominal p value was less than 0.001,
Analytical methods used in malaria risk mapping
Category  Method  No  References 
 Stepwise procedures  11 

Preliminary frequentist analysis  14 
 
Totalset analysis  6 
 
Principal component analysis  6 
 
Bayesian stochastic search  3 
 
LASSO penalty  2 
 
Literature review  2 
 
Spike and slab  2 
 
BMA  1 
 
 Rate map  63 

Dot map  25 
 
Case counts  16 
 
 Spatial scan statistic  15 

Global Moran’s/  6 
 
Getis Ord statistic  3 
 
Local Moran’s  7 
 
 Geostatistical models  27 

Bayesian CAR models  15 
 
Time series models  9 
 
Bayesian Kriging  5 
 
Conventional Poisson  7 
 
Conventional logistic  4 
 
GAM  2 
 
Negative binomial regression  1 
 
GWR  1 
 
ANN  1 
 
BRT  1 
 
 Data partitioning  24 

Deviance information criterion  19 
 
Akaike information criterion  8 
 
Root mean squared error  7 
 
Variogrambased algorithm  7 
 
Mean absolute prediction error  6 
 
Mean error  3 
 
Bayesian information criterion  2 

ANN, Artificial neural network; BMA, Bayesian model averaging; BRT, Boosted regression tree; CAR, Conditional autoregressive; GAM, General additive model; GWR, Geographically weighted regression; LASSO, Least absolute shrinkage and selection operator.
Data preprocessing procedures were employed in 37 (35%) studies. The verification of geographical coordinates by either paper maps
The type and number of covariates included in malaria models varied across studies. Different categories encompassing climatic and environmental, sociodemographic and malaria intervention covariates were identified. The most common covariates in the environmental domain were rainfall and temperature used in 61 (57%) and 59 (55%) studies, respectively; while the most common sociodemographic covariates used in 12 (11%) studies were population size and age. Malaria interventions (insecticidetreated bed nets, indoor residual spraying and artemisininbased combined therapy) used in 32 (30%) studies and transmission seasonality used in 28 (26%) studies were also common. Detailed variations and adaptations covariates are presented in
Covariates used in malaria risk mapping
Indicator  Metric  No  References 
Malaria Outcome  Malaria incidence/cases  50 

Malaria prevalence  37 
 
Malaria risk  12 
 
Malaria mortality/deaths  5 
 
EIR/Estimate/Mosquito density/ abundance  3 
 
Rainfall indices  Rainfall/precipitation  44 

Monthly rainfall  10 
 
Annual rainfall  5 
 
Weekly rainfall  2 
 
Temperature indices  TSI  10 

LST  19 
 
Mean/min/max temperature  28 
 
Weekly temperature  2 
 
Vegetation indices  NDVI  31 

EVI  17 
 
Annual EVI  2 
 
Monthly EVI  1 
 
Leaf area index  1 
 
GIS Derived  Distance to nearest water source  34 

Distance to main road  6 
 
Distance to health facility  4 
 
Distance to urban centre  2 
 
Distance to border  1 
 
Elevation  Altitude  10 

Elevation  11 
 
Land cover  Land cover  8 

Humidity  Relative humidity  8 

Weekly humidity  1 
 
Evapotranspiration  2 
 
Vapour pressure  3 
 
Evaporation  1 
 
Digital Elevation Models  DEM derivatives  Wetness index/CTI  2 

Slope  5 
 
TWI  1 
 
Aridity index  1 
 
Reflectivity  Stable lights  1 

Visibility  1 
 
Wind  Wind speed  3 

Demographic factors  SES  9 

Gender/Sex  6 
 
Age  12 
 
Population density/size  12 
 
Livestock ownership  2 
 
Urbanisation  8 
 
Development  1 
 
Wealth index/category  4 
 
Building/Housing material  4 
 
Time  Year/Month of survey  3 

Time period  1 
 
Transmission seasonality  28 
 
Malaria intervention  ITN/LLIN ownership/coverage/use  19 

IRS  8 
 
ACTs  5 
 
Treatment seeking rate  3 
 
Reporting and testing  1 
 
None  None  19 

ACTs, Artemisininbased combined therapy; CTI, Compound topographic index; DEM, Digital elevation models; EIR, Entomological inoculation rate; EVI, Enhanced vegetation index; GIS, geographical information system; IRS, Indoor residual spraying; ITN, Insecticidetreated bed nets; LLIN, Long lasting insecticidal nets; LST, Land surface temperature; NDVI, Normalised difference vegetation index; NDWI, Normalised difference water index; SES, Social economic status; TSI, Temperature suitability index; TWI, Topographic wetness index.
A variety of spatial, temporal and spatiotemporal methods were employed to visualise malaria risk patterns, explore spatial clusters and model risk across space and time in SSA. Measurement of malaria burden varied across studies with the type of outcome informing the modelling framework. The most common malaria metric used in models, was incidence used in 50 (47%) studies and prevalence used in 37 (35%) studies.
In settings of low malaria transmission, local and global spatial cluster detection methods were used in 31 (29%) studies to identify significant geographical variation in malaria risk patterns (
The Bayesian spatial only and spacetime kriging—a statistically unbiased and robust interpolation method appropriate for study settings with limited data; was used in five (6%) studies, to predict risk at unsampled locations,
Using pointreferenced data sourced from multiple independent surveys, 27 (25%) studies applied both the modelbased geostatistical (MBG) and Bayesian MBG methods to analyse, predict and map malaria risk. In this framework, the spatiotemporal dependency was modelled as a Gaussian process in fourteen (13%) studies.
Structure of the spatiotemporal models
ID  References  Year  Space  Time  Space time 
1  Abellana  2008  CAR  
2  Alegana  2016  Markov random field  –  Gaussian 
3  Alegana  2013  –  –  CAR 
4  Alemu  2013  –  Temporal trend – ARIMA  – 
5  Amek  2012  Gaussian  AR (1)  – 
6  Amratia  2019  Gaussian  –  – 
7  Appiah  2011  –  –  STOK 
8  Awine  2018  –  –  SARIMA 
9  Bejon  2010  Cluster analysis  Temporal trends  – 
10  Bejon  2014  Cluster analysis  –  – 
11  Belay  2017  –  Temporal trends  – 
12  Bennett  2013  –  –  Gaussian 
13  Bennett  2016  Gaussian  –  – 
14  Bennett  2014  CAR  CAR  CAR 
15  Bhatt  2015  Markov random field  AR (1)  Gaussian 
16  Bisanzio  2015  Markov random field  B – splines with RW (2)  – 
17  BM & OE  2007  CAR  –  – 
18  Bousema  2010  Hotspot analysis  –  – 
19  Ceccato  2007  Cluster analysis  –  – 
20  Chipeta  2019  –  –  Gaussian 
21  Chirombo  2020  Markov random field  Markov random field  Gaussian 
22  Cissoko  2020  Cluster analysis  Temporal trend  
23  Colborn  2018  –  –  Gaussian 
24  Coulibaly  2013  Cluster analysis  –  – 
25  DePina  2019  Cluster analysis  Temporal trend  _ 
26  Diboulo  2016  Gaussian  –  – 
27  Ferrão  2017a  –  Temporal trend  ARIMA  – 
28  Ferrão  2017b  –  Temporal trend  ARIMA  – 
29  Ferrari  2016  Cluster analysis  –  – 
30  Gaudart  2006  Cluster analysis  Temporal trend  ARIMA  – 
31  Gemperli  2006  Exponential correlation function  –  – 
32  Gething  2016  –  P – splines with RW (1)  – 
33  Giardina  2015  Gaussian  –  – 
34  Giardina  2012  Multivariate Normal  –  – 
35  Giardina  2014  Gaussian  –  – 
36  Giorgi  2018  –  –  Gaussian 
37  GómezBarroso  2017  Cluster analysis  –  – 
38  Gosoniu  2012  Gaussian  –  – 
39  Gosoniu  2010  Gaussian  –  – 
40  Gosoniu  2006  Gaussian  –  – 
41  Houngbedji  2016  Normal  –  – 
42  Ihantamalala  2018  Cluster analysis  –  – 
43  Ikeda  2017  –  –  SOM 
44  Ishengoma  2018  –  Temporal trends  – 
45  Kabaghe  2017  Gaussian  –  – 
46  Kabaria  2016  –  –  BRT 
47  Kamuliwo  2015  Cluster analysis  –  – 
48  Kang  2018  Gaussian  AR (1)  – 
49  Kangoye  2016  Cluster analysis  –  – 
50  Kanyangarara  2016  –  –  – 
51  Kazembe  2006  Gaussian  –  – 
52  Kifle  2019  Cluster analysis  Temporal trends  SARIMA  
53  Kigozi  2016  –  Temporal trend ARIMA  – 
54  Kleinschmidt  2000  Kriging  –  – 
55  Kleinschmidt  2001a  Kriging  –  – 
56  Kleinschmidt  2001b  Kriging  –  – 
57  Kleinschmidt  2002  Normal  Normal  – 
58  Mabaso  2005  CAR  –  AR (1) 
59  Mabaso  2006  CAR  AR (1)  – 
60  Macharia  2018  –  –  Gaussian 
61  Mfueni  2018  –  –  – 
62  Midekisa  2012  –  Temporal trend  SARIMA  – 
63  Millar  2018  –  –  – 
64  Mirghani  2010  Cluster analysis  –  – 
65  Mlacha  2017  Cluster analysis  –  – 
66  Mukonka  2014  –  Temporal trends  – 
67  Mukonka  2015  Cluster analysis  –  – 
68  Mwakalinga  2016  Cluster analysis  –  – 
69  Ndiath  2015  Cluster analysis  –  – 
70  Ndiath  2014  Cluster analysis  –  – 
71  Nguyen  2020  Gaussian  –  Gaussian 
72  Noor  2013a  Gaussian  –  GRF 
73  Noor  2008  Gaussian  –  – 
74  Noor  2012b  –  –  GRF 
75  Noor  2009  –  –  GRF 
76  Noor  2014  Gaussian  AR (2)  – 
77  Noor  2013b  –  –  GRF 
78  Noor  2012a  Gaussian  –  Stationary Gaussian 
79  Nyadanu  2019  Cluster analysis  –  – 
80  Okunola  2019  Cluster analysis  –  – 
81  Onyiri  2015  Gamma  –  – 
82  Ouedraogo  2018  –  Temporal trend ARIMA  – 
83  Ouédraogo  2020  CAR  AR (1) / Temporal trends  
84  Peterson  2009  Cluster analysis  –  – 
85  Pinchoff  2015  –  –  – 
86  Raso  2012  Multivariate Normal  –  – 
87  Rouamba  2020  CAR  CAR  Gaussian 
88  Rumisha  2014  Gaussian  AR (1)  – 
89  Selemani  2015  Cluster analysis  –  – 
90  Selemani  2016  CAR  AR (1)  – 
91  Sewe  2016  –  Natural cubic spline  – 
92  Seyoum  2017  Cluster analysis  –  – 
93  Shaffer  2020  Cluster analysis  Temporal trends  – 
94  Simon  2013  Cluster analysis  –  – 
95  Siraj  2015  CAR  –  – 
96  Snow  2017  CAR  CAR  – 
97  Snow  1998  –  –  – 
98  Solomon  2019  Cluster analysis  –  – 
99  Ssempiira  2018a  CAR  AR (1) / temporal trend  – 
100  Ssempiira  2018b  CAR  AR (1) / temporal trend  – 
101  Ssempiira  2017b  CAR  –  – 
102  Ssempiira  2017a  –  –  – 
103  Sturrock  2014  CAR  Temporal trend  – 
104  Yankson  2019  Gaussian  –  – 
105  Yeshiwondim  2009  –  –  – 
106  Zacarias and Andersson  2011  CAR  AR (1)  – 
107  Zacarias and Majlender  2011  CAR  RW (1)  – 
AR, autoregressive; ARIMA, autoregressive integrated moving average; BRT, boosted regression tree; CAR, conditional autoregressive; GRF, Gaussian random field; RW, random walk; SARIMA, seasonal autoregressive integrated moving average; SOM, selforganising maps; STOK, spacetime ordinary kriging.
Using observations aggregated over distinct geographical region/spatial partitions/adjacent units (eg, census tract, administrative boundaries); 15 (14%) studies used the Bayesian conditional autoregressive (CAR) models, to explore the spatial and spatiotemporal variation of malaria risk. To account for the temporal dependency between consecutive time points; seven studies (6%) used the first order autoregressive AR (1) prior process, whereas one study (1%) used the random walk of order one RW (1) prior process (
Generalised linear modelling framework, such as the Poisson, logistic regression, negative binomial and geographically weighted regression, was used in fifteen studies (14%). These models explored the association of malaria counts or rates and its correlates, using appropriate exponential distribution families. Machine learning techniques such as the artificial neural network and the boosted regression tree were used to analyse incidence patterns and to examine malaria prevalence, respectively (
A range of different validation techniques were used to assess model fitness and to select the optimal predictive models. The most commonly used approach entailed partitioning the data for model training and validation and was employed in 24 (22%) studies. The training set was then used to validate the predictive model fit, whereas the validation set was used for assessing the model predictive ability. The representative holdout datasets were selected using a spatially and temporal declustered algorithm,
The rapid expansion of methods and data informs the need to guide future spatial and spatiotemporal modelling of infectious diseases in SSA (
Schematic illustration of the spatiotemporal modelling framework for malaria risk in subSaharan Africa.
Scalable guidelines for rigorous and transparent statistical methodology are necessary for reproducible malaria risk estimation. This review offers a comprehensive appraisal and synthesis of methods and covariates used in malaria risk mapping in SSA in the last five decades.
Highresolution maps revealing the spatiotemporal variation of malaria endemicity are useful for estimating malaria burden, quantifying the effectiveness of control initiatives and assessing the progress towards its elimination nationally and subnationally. However, malaria risk mapping efforts in SSA are rarely based on routinely collected data. Instead, periodic and costly household survey’s data have traditionally been used in modelling malaria risk. To address this challenge and obtain robust estimates reflective of the subnational burden, WHO initiated the high burden to high impact approach in 2018, which underscored the need for reliable national data systems. This is considered central to the understanding of malaria burden in low transmission settings and in the most vulnerable populations.
The steady growth of satellite, remote sensing platforms and curated databases has made available a rich suite of both environmental and socioeconomic covariates at a finer level of detail useful for mapping malaria risk at high spatial and temporal resolution. Validating the quality of available satellite data prior to their inclusion in malaria studies remains central to achieving robust estimates.
While malaria incidence and prevalence metrics can be modelled from routine health information systems and sample surveys respectively, caution should be taken when interpreting estimates as both metrics are products of interacting factors such as interventions, sociodemographic and environmental factors that may contribute to the overall risk. A concise picture may be achieved by measuring malaria indicators at a finer spatial scale and exploring the nature and scope of the interaction. Data on malaria mortality as an outcome were sparse, and efforts must be made to increase data collection and improve the sensitivity and specificity of malaria mortality burden attribution in SSA.
The paucity of continuous, reliable data necessary to yield estimates with greater geographical and temporal richness is a growing concern in the era of evidencebased public health. A highquality, routinely collected data avail an alternative source of malaria metrics for continuous analysis over time.
Improving the precision of malaria risk estimates largely depends on limiting subjective decisions. These decisions may impact on the modelling process, even as more covariates becomes accessible at finer geographical and temporal resolutions. Studies have shown large variables to be desirable for prediction, whereas small sets of variables to be meaningful for inference.
Environmental and climatic factors influence mosquito vector abundance, distribution and longevity; at different time scales
Complex decisions involving key modelling components such as covariates to include, preliminary data preprocessing and diagnostics checks demands advanced statistical knowledge. Extensive computational algorithms and complex spatiotemporal data structures may limit the applicability of these modelling approaches to experts. Furthermore, complex models used to represent malaria heterogeneity may not necessarily represent the truth on the ground. Thus, the statistical uncertainties around model estimates should be carefully examined, and the varying quantities and quality malaria data, that informs modelling approaches accounted for.
The review highlights the prominence and flexibility of geostatistical methods in modelling spatial and spatiotemporal malaria patterns, at policyrelevant units and thresholds.
Bayesian hierarchical CAR models are useful for modelling spatially correlated areal data by smoothing noisy estimates and leveraging information from adjacent regions. However, choosing an appropriate prior specification for the parameters defining the spatial interaction is inevitable and sometimes challenging. Notably, the spatial dependence among neighbouring regions is accounted for by assuming a CAR process in the random effects. For example, in the Besag York and Mollie/convolution model, locationspecific spatial effects are assumed to follow a normal distribution with the mean equal to the average of its neighbours and the variance considered to be inversely proportional to the number of neighbours. In the Leroux
Overall, malaria risk mapping has increased dramatically over the last decades, with novel methods advanced to meet the quest for accurate estimates of malaria burden. Whereas most approaches are built on classical statistical methods, recent advances in computing, availability of geographically referenced data have ushered/propagated new techniques designed to address existing challenges. These approaches include ensemble modelling, neural networks, simulationbased methods and bootstrap models to better capture spacetime interactions.
As malaria landscape diversifies in the next decade, investments in primary data collection at subnational scales, development and continuous application of robust modelling tools will continue to be important priorities in malaria control and elimination efforts. In the era of open data policy and reproducible research, our review reiterates the importance of periodically reviewing, validating and updating malaria maps to accommodate new data sources, improved data quality, enhanced computing power and novel methodological approaches. Variable selection procedures should be data driven and objectively developed to the maximise the predictive accuracy of malaria risk mapping. The spatiotemporal modelling framework should incorporate practical challenges facing control and elimination of malaria in SSA. These challenges are: human migration within and among endemic zones, mapping asymptotic infection reservoirs and accounting for differential immunity within a population.
The review search strategy was exhaustive and transparent, in accordance with the current methodological guidelines and included studies have provided a fair depiction of malaria risk mapping efforts in SSA. The methodological approach of the included studies was diverse, making metaanalysis inappropriate. The review considered only studies published in English and relevant papers published in other languages might have been excluded.
Malaria risk mapping remains an important component for understanding the burden of malaria in SSA. The review has described modelling approaches and examined covariates used in mapping malaria risk in different epidemiological contexts. As malaria transmission continues to decline in SSA, the use of metrics that accurately describes changes in its transmission intensity across space and time will be important for the design and implementation of evidencebased control and elimination measures.
The authors thank the College of Health Sciences, systematic review services and library services at the University of KwaZuluNatal for providing training and resources at the initial phases of the review.
Alberto L GarciaBasteiro
@Pete_M_M
JNO, BS and RWS conceived and designed the systematic review. JNO, BS and CK conducted the literature search, study selection and data extraction. JNO wrote the first draft of the manuscript with assistance from CK and BS. PMM and RWS revised the draft critically for important intellectual content. All authors read and approved the final version of the manuscript.
JNO acknowledges support from the University of KwaZulu Natal, College of Health Sciences postgraduate scholarship scheme. RWS is supported as a Wellcome Trust Principal Fellow (#103602 and 212176) that also supported PMM. PMM acknowledges support for his PhD under the IDeALs Project part of the DELTAS Africa Initiative (DEL15003). The DELTAS Africa Initiative is an independent funding scheme of the African Academy of Sciences (AAS)'s Alliance for Accelerating Excellence in Science in Africa and supported by the New Partnership for Africa's Development Planning and Coordinating Agency with funding from the Wellcome Trust (107769) and the UK government. RWS and PMM are grateful to the support of the Wellcome Trust to the Kenya Major Overseas Programme (203077).
None declared.
Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Not required.
Not commissioned; externally peer reviewed.
All data relevant to the study are included in the article or uploaded as online supplemental information. The data supporting conclusions made in this review are available in the detailed reference list.