Aheto Malaria Journal (2022) 21:384 https://doi.org/10.1186/s12936-022-04409-x Malaria Journal RESEARCH Open Access Mapping under-five child malaria risk that accounts for environmental and climatic factors to aid malaria preventive and control efforts in Ghana: Bayesian geospatial and interactive web-based mapping methods Justice Moses K. Aheto1,2,3* Abstract Background: Under-five child malaria is one of the leading causes of morbidity and mortality globally, especially among sub-Saharan African countries like Ghana. In Ghana, malaria is responsible for about 20,000 deaths in chil- dren annually of which 25% are those aged < 5 years. To provide opportunities for efficient malaria surveillance and targeted control efforts amidst limited public health resources, the study produced high resolution interac- tive web-based spatial maps that characterized geographical differences in malaria risk and identified high burden communities. Methods: This modelling and web-based mapping study utilized data from the 2019 Malaria Indicators Survey (MIS) of the Demographic and Health Survey Program. A novel and advanced Bayesian geospatial modelling and map- ping approaches were utilized to examine predictors and geographical differences in under-five malaria. The model was validated via a cross-validation approach. The study produced an interactive web-based visualization map of the malaria risk by mapping the predicted malaria prevalence at both sampled and unsampled locations. Results: In 2019, 718 (25%) of 2867 under-five children surveyed had malaria. Substantial geographical differences in under-five malaria risk were observed. ITN coverage (log-odds 4.5643, 95% credible interval = 2.4086–6.8874), travel time (log-odds 0.0057, 95% credible interval = 0.0017–0.0099) and aridity (log-odds = 0.0600, credible inter- val = 0.0079–0.1167) were predictive of under-five malaria in the spatial model. The overall predicted national malaria prevalence was 16.3% (standard error (SE) 8.9%) with a range of 0.7% to 51.4% in the spatial model with covariates and prevalence of 28.0% (SE 13.9%) with a range of 2.4 to 67.2% in the spatial model without covariates. Residing in parts of Central and Bono East regions was associated with the highest risk of under-five malaria after adjusting for the selected covariates. Conclusion: The high-resolution interactive web-based predictive maps can be used as an effective tool in the iden- tification of communities that require urgent and targeted interventions by programme managers and implementers. *Correspondence: justiceaheto@yahoo.com; jmkaheto@ug.edu.gh 1 Department of Biostatistics, School of Public Health, College of Health Sciences, University of Ghana, Accra, Ghana Full list of author information is available at the end of the article © The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://c reat iveco mmons.o rg/ licen ses/ by/4. 0/. The Creative Commons Public Domain Dedication waiver (http:// creat iveco mmons.o rg/ publi cdomai n/ zero/1. 0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Aheto M alaria Journal (2022) 21:384 Page 2 of 15 This is key as part of an overall strategy in reducing the under-five malaria burden and its associated morbidity and mortality in a country with limited public health resources where universal intervention is practically impossible. Keywords: Malaria, Under-five malaria, Mapping malaria risk, Bayesian methods, Geospatial methods, Geostatistical methods, Interactive web-based mapping, Predictors, Sub-Saharan Africa Background about 20,000 children die annually of which 25% are Malaria is a deadly disease and remains one of the severe those aged < 5 years [7]. In Ghana, malaria is considered global public health and development challenge par- endemic in all the regions with national prevalence of ticularly in sub-Saharan Africa (SSA) where under-five 14% in 2019 against the previous prevalence of 21% and malaria infection is the leading cause of under-five mor- 27% in 2016 and 2014, respectively, among under-five tality (U5M) due to their vulnerability. It is the leading children. However, marked regional geographic dispar- cause of illness and deaths in most of the malaria affected ities exist in the under-five malaria prevalence in Ghana countries where young children and pregnant women with the highest prevalence recorded in the Western are the most affected groups. Notably, malaria is con- (27%) and lowest recorded in the Greater Accra (2%) sidered entrenched global health problem and Ghana is regions. Thus, the under-five malaria prevalence across of no exception due to its significant number of deaths the country varied [11], demonstrating the need for associated with the disease in the country [1, 2]. Malaria examining more localized spatial trends in malaria. is also a driver of low productivity and poverty on the Unfortunately, information on localized spatial distri- individuals and exerts financial burden on families and butions and predictors of under-five malaria supported the economy [3]. An estimated number of 405,000 per- with web-based mapping, which are critical for effec- sons died due to malaria infections out of which major- tive design of intervention strategies that will enhance ity were young children from SSA in 2018 as against the the survival of under-five children amidst available lim- 2017 figure of 416,000 deaths and 585.000 deaths in 2010. ited public health resources are not readily available. In 2016, there were an estimated number of 216 million At the global level, the United States President’s clinical episodes caused by malaria, an increase of 5 mil- Malaria Initiative (PMI) launched in 2005 led to an lion over the previous year [2, 4, 5]. Despite the global increased availability of insecticide-treated nets (ITNs), rapid malaria control efforts that led to malaria mortality anti-malarial treatment and rapid diagnostic tests and reduction by 25% from the year 2010 to 2016, the malaria indoor residual spraying, which led to a significant prevalence and mortality rates remain high in SSA coun- reduction in under-five mortality in SSA [12]. The suc- tries where 14 out of 15 countries in SSA accounted for cess of the “for a malaria-free world 2008–2015 ini- 80% of the world malaria burden with national and sub- tiative”, the Roll Back Malaria Partnership outlined national differences [2, 4, 6]. an action plan dubbed, “Action and Investment to In 2018, the most vulnerable group hardest hit by Defeat Malaria (AIM) 2016–2030” [13]. In May 2015, malaria are children under-five who accounted for 67% the Global technical strategy for malaria (GTSM) 2016– of all malaria deaths globally [5]. SSA had the highest 2030  which sets the target of reducing global malaria burden of malaria where about 90% of all malaria deaths incidence and mortality rates by at least 90% by 2030 occur with children under-five accounting for about 78% was adopted by the World Health Assembly. The strat- of these deaths [7]. This link between malaria and under- egy was updated in the year 2021 to reflect the lessons five deaths also poses a great danger to achieving the Sus- learned in the global malaria response between 2016 to tainable Development Goals (SDG) 3 target 2.1 because 2020. It provides a comprehensive framework to guide the U5M rates are among the health indicators of utmost nations in their efforts to fast-track progress towards importance globally. It is the goal 3 target 2.1 of the SDGs elimination of malaria by emphasizing the need for that is expected to be reduced to internationally agreed universal coverage of core malaria interventions for targets of at least 25 per 1,000 livebirths by 2030 [8], but all populations at risk. At the heart of the strategy is several countries especially those in SSA like Ghana are the utmost need to use high-quality surveillance data struggling to meet this target [9, 10]. Thus, addressing for decision-making [14]. The alignment of the time- the problem of under-five malaria will be beneficial to the frame of the vision of AIM and GTSM to that of the global fight against U5M. SDG underscores the need to address the problem of Ghana was among the 10 highest burden countries under-five malaria to ensure the realization of SDG in Africa in 2018 that reported the highest increase in goal 3. Nonetheless, under-five malaria continues to be malaria cases compared to the previous year [5], where a significant cause of childhood deaths in SSA which A heto M alaria Journal (2022) 21:384 Page 3 of 15 militates against the progress towards the achievement questionnaire to record the results of the anaemia and of the sustainable development Goal 3 target 2.1. malaria testing of the children aged 6–59  months. This In Ghana, despite several national policies and inter- study used data on under-five children from the bio- ventions (e.g. Community-based Health Planning and marker dataset which has malaria RDT results on 2867 Services (CHPS), Child Health Policy 2007–2015 and under-five children residing in 192 geographical locations National Health Insurance, 2014–2020 Ghana Strategic (clusters). Detailed description of the survey methods Plan for Malaria Control) [1, 10, 15] rollout to improve employed in the 2019 GMIS is available elsewhere [11]. and promote health of children, the under-five malaria The Ghana Malaria Indicator Survey (GMIS) is based and its resultant under-five mortality rates remain high in on a two-stage sampling design. The sampling was based the country. The focus of the 2014–2020 Ghana Strategic on ten administrative regions. Each region was divided Plan for Malaria Control is to scale up preventive inter- into urban and rural areas, resulting in twenty sampling ventions to reduce the malaria morbidity and mortality strata. Enumeration areas (EAs) were sampled from each burden by 75% by the year 2020 [16]. stratum. In the first stage, 200 EAs (97 in urban areas and There is limited knowledge on localized geospatial 103 in rural areas) were selected with probability propor- distribution of under-five malaria risk and how cer- tional to EA size. In the second stage of selection, approx- tain environmental predictors could help explain geo- imately 30 households were selected from each cluster to graphical differences in under-five malaria risk in Ghana. make up a total sample size of 6,002 households of which Also, malaria burden is a continuous phenomenon that 5388 were occupied at the time of field work. A total of requires high-quality surveillance data and constant sur- 5799 household were interviewed among the occupied veillance to inform malaria control strategies. households, resulting in 99.4% response rate. Of the 5246 This study, therefore, attempts to fill these gaps. The eligible women, about 5181 women aged 15–49  years study aims to estimate, predict, and map localized under- (representing 98.8% response rate) who were either per- five malaria risk using novel Bayesian geospatial model- manent residents of the selected households or visitors ling approaches while adjusting for critical environmental who stayed in the household the night before the survey factors, with the goal of identifying communities at high- were interviewed. All children aged 6–59  months from risk of malaria burden where control efforts, interven- the interviewed households were eligible for malaria test- tions, and further research can be targeted to address the ing upon parental or guardian consent [11]. problem of under-five malaria and its associated morbid- ity and mortality. Outcome variable The outcome variable of interest is the number of under- Methods five children with positive test on rapid diagnostic test Setting, design and sample (RDT) kit in each sampled cluster. The RDT malaria test Ghana is in West Africa and covers a total area of 238,538 was conducted by taking a drop of blood with the SD km2. It lies between latitude 4  and 12 N and longitudes BIOLINE Malaria Ag Pf RDT and tests for one antigen, 4 W and 2 E. It is bordered in the south by the Gulf of histidine-rich protein II (HRP-II), specific to Plasmodium Guinea, Côte d’Ivoire to the west, Togo to the east, and falciparum, the major cause of malaria in Ghana. The Burkina Faso to the north. Presently, Ghana has 16 RDT kit produces result in 15 min [11]. administrative regions. Data from the 2019 GMIS of the DHS program was Covariates used in this study [11]. The 2019 GMIS is the second Though the main goal of the study is to predict and map round of the survey with the first round conducted in under-five malaria risk, the study adjusted for selected 2016 which provides a population-based estimates of environmental factors to allow for examination of how malaria indicators as a supplement to the routine admin- these factors help explain some of the spatial variability istrative data collected in the country that are used to in under-five malaria risk across Ghana. These factors inform strategic planning and evaluation of the Ghana include insecticide-treated nets (ITNs) coverage (i.e., Malaria Control Programme [11]. Computer-assisted proportion of the population protected by ITNs), travel personal interviewing (CAPI) was employed to collect time (time required to reach a high-density urban centre), the data. In the survey, information on malaria preven- aridity (ranging from most arid to most wet), enhanced tion, treatment, and prevalence is collected. The data vegetation (ranging from least vegetation to most vegeta- is freely available online at DHS MEASURE Program tion), annual temperature (mean temperature), and pre- website [17]. Parents or guardians consent were sort for cipitation (average precipitation–per month). A detailed children aged 6–59  months who were tested for anae- description of the methods and procedures employed mia and malaria infection. The study used a biomarker to generate these geospatial covariates and their sources Aheto M alaria Journal (2022) 21:384 Page 4 of 15 are published elsewhere [18]. The consideration of these is v − 1 times mean-square differentiable and the s√cale environmental and climatic covariates was based on the parameter κ > 0 is related to the practical range ρ = 8v , k available literature on predictors of malaria and other the distance at which the spatial correlation approaches health outcomes [19–22]. Following recommended strat- 0.1 or is negligible, κv(.) is the modified Bessel function of egy [20, 23], the study accounted for the displacement of second kind and orderv > 0. the GPS coordinates of the sampled cluster locations by The model was implemented under the Integrated creating 2 km buffers for urban and 5 km buffer for rural Nested Laplace Approximation (INLA) approach [27] settings to ensure that the correct cluster centroids were with Stochastic Partial Differential Equation (SPDE) captured in the analysis. strategy [28]. Based on a previous study [20], a mesh for inference and prediction was created for the SPDE strat- Geospatial analysis egy because the data (i.e., geostatistical data) points in this study do not have explicit neighbours required by the SPDE strategy unlike areal data. The description of Model formulation the mesh creation is provided in Additional file  1. The Following a previous modelling approach [20], we detailed procedures for mess creation are published else- employed a Bayesian Geospatial model [20, 24] to study where [20, 29]. spatial risk in under-five malaria while adjusting for envi- In this study, nine (9) models were set up: two (2) ronmental predictors. Consider Yi to be the number of non-spatial models with different set of covariates under-five children with positive RDT test out of the total included, one spatial model without covariates, and five N under-five children sampled per geographical cluster. (5) spatial models with different set of covariates. The i Given the true malaria risk P(zi) at location z , the num- Watanabe-Akaike information criterion (WAIC) was i ber of under-five children with positive RDT test out of employed to investigate how well each of these nine (9) the total number of under-five children sampled follows a models fits the data, and to select the model that rela- binomial distribution formulated as: tively fits the data well among the competing models. The level of uncertainty in the fitted model estimates Yi|P(zi) ∼ Binomial(Ni,P(zi)), were quantified by estimating the 95% credible intervals and the standard errors and map these uncertainties ′ = + + continuously across the whole of Ghana. Furthermore, logit(P(zi)) β0 d(xi) β S(zi). the study compares the predictive maps for the spatial where β0 is the intercept parameter which by default is model with covariate and spatial model without covari- assigned Gaussian prior with mean and precision to be ates to examine if the included covariates in the spatial zero (0), d(.) is a vector of observed environmental pre- model explained some differences in malaria preva- dictors of the outcome variable Y , β is a vector of spatial lence predictive maps. The study investigated how well regression coefficients for the covariates which by default the predictive model performs in the presence of new was assigned Gaussian prior with mean zero (0) and pre- data via cross-validation procedure by splitting the data cision 0.001, and S(.) is a spatially structured random into training and validation sets, a common and gen- effect and follows a zero-mean Gaussian process with erally accepted model validation approach in this area variance σ 2 and a given correlation function [30]. The R-INLA package [29, 31] was used for all the analyses. { } ρ(u) = corr S(zi), S(zj) Model validation where u is the Euclidean distance between locations zi It is critical to examine how well the predictive model and zj . There are various parametric families for ρ(u) as performs, especially in the presence of new data. This outlined by Diggle (2007) [25]. In the current analysis, study employed cross-validation approach to assess the the study use the Matérn class of covariance function[26] predictive performance of the model under out of sam- given by ple procedure. First, the data was split into training and σ 2 validation sets, and set a seed of 123 to make the parti-( ( )) ( ) ( ) Cov S(zi), S zj = − k||zi − zj|| v Kv k||zv 1 i − zj|| .v tion reproducible. The model was trained on 75% of the 2 Ŵ( ) samples and tested on 25% of the samples. The study Here, ||. || denotes Euclidean distance, σ 2 represents assessed the model predictive performance by plotting the spatial variance, v is the shape parameter which the observed and the predicted malaria prevalence and determines the smoothness ofS(z) , in the sense that S(z) estimated the resultant correlation. A heto M alaria Journal (2022) 21:384 Page 5 of 15 Interactive web‑based mapping of the predicted malaria Results prevalence The study analyzed data on 2867 children aged below To support policymakers with readily available qual- 5 years residing in 192 clusters (communities). A total of ity data for targeted policy and intervention strate- 718 (25%) children under-five had malaria in the study in gies, especially for malaria surveillance amidst limited 2019. public health resources in these settings, the study produced interactive web-based maps for the pre- Model selection results dicted malaria prevalence to improve visualization and To select a good model among the competing models identification of higher risk communities for urgent we fitted to predict the malaria prevalence, this study intervention and further research in this setting where employed the Watanabe-Akaike information crite- universal intervention is practically impossible due to rion (WAIC). The model with the smallest WAIC value limited public health resources. The spatsurv, rgdal, is preferred. In all, nine (9) competing models were fit- leaflet, and sp packages in R version 4.2.0 and RStudio ted and Model 6 which contained aridity, ITN cov- [32, 33] were used to support the development of the erage, and travel time had the smallest WAIC value interactive web-based predicted malaria prevalence (WAIC = 689.79) compared to all other models fitted, maps. an indication of a better model fit for the study (Table 1). Thus, the study present and discuss the results based on Ethical consideration the full spatial model (i.e., Model 6, Table 1) presented in Permission was granted by DHS MEASURE Program to Table 2. use the 2019 GMIS data for the study. The data is freely available after a simple, registration-access request at the Predictors of malaria prevalence from the Bayesian spatial link  https:// dhspr ogram. com/ data/ datas et_a dmin/ index. models cfm. The protocol for the 2019 GMIS was approved by The study presents the results based on the full spatial the Ghana Health Service Ethical Review Committee and model (Model 6) in Table  2 and Fig.  1. ITN coverage ICF’s Institutional Review Board [11]. (log-odds 4.5643, 95% credible interval = 2.4086–6.8874), travel time (log-odds 0.0057, 95% credible inter- val = 0.0017–0.0099) and aridity (log-odds = 0.0600, The role of the funding source credible interval = 0.0079–0.1167) were found to be pre- The present study did not receive any support from any dictive of under-five malaria risk. The estimated spatial funding source. Also, the funders of the original survey variance ( σ 2 ) is 0.8772 (95% credible interval = 0.5061– played no role in the design, data collection, analysis, 1.2915) and the estimated range is 0.2917 (95% credible interpretation, writing of the manuscript, and the deci- interval = 0.1250–0.4886) while the kappa ( κ ) is 10.8059 sion to submit this manuscript. The author confirm that (95% credible interval = 4.6372–18.1236) (Table 2). he has full access to all the data in this study and accept Figure 1 shows the posterior (marginal) distributions of responsibility to submit for publication. the fixed and random (hyper) parameters of the Bayesian Table 1 Model selection for the fitted Bayesian Geospatial models Parameters WAIC Full non-spatial model Model 1: Aridity, ITN coverage, Travel times, Precipitation, Vegetation, Temperature 903.38 Model 2: Aridity, ITN coverage, Travel times 905.18 Spatial models Model 3: Null spatial model 698.39 Model 4: Aridity 696.63 Model 5: Aridity, ITN coverage 692.34 Model 6: Aridity, ITN coverage, Travel times 689.79 Model 7: Aridity, ITN coverage, Travel times, Precipitation 692.00 Model 8: Aridity, ITN coverage, Travel times, Precipitation, Vegetation 690.64 Model 9: Aridity, ITN coverage, Travel times, Precipitation, Vegetation, Temperature 691.91 WAIC Watanabe-Akaike information criterion Lower values of the WAIC indicate better model fit Aheto M alaria Journal (2022) 21:384 Page 6 of 15 Table 2 Predictors of malaria prevalence in the non-spatial and Geospatial model (i.e., full spatial model—Model 6) pre- spatial Bayesian models sented in Table 2, which provides a fuller understanding Parameter Mean log odds (95% Credible of the posterior distributions of the model parameters intervals) and the appropriate quantification of uncertainty around the estimates unlike the frequentist approaches. Full non-spatial model Intercept − 4.9129 (− 5.7024, − 4.1390) Geospatial analysis and interactive web‑based mapping ITN Coverage 4.0385 (3.1535, 4.9407) This study analyzed data on children residing in 192 Travel time to health facility 0.0037 (0.0020, 0.0054) communities which are geographically indexed. In Fig. 2, Aridity 0.0476 (0.0271, 0.0681) we showed the location (centroid of clusters) of commu- Spatial model nities used in this study and their respective empirical Null spatial model (observed) malaria prevalence (coloured) at the sam- Intercept − 1.1943 (− 1.4532, − 0.9415) pled locations. Communities with red highlighted circles σ 2(spatial variance) 1.3218 (0.8211, 1.8869) had observed malaria prevalence of between 75 to 100% Range nominal 0.2738 (0.1365, 0.4271) while those with blue highlighted had observed 0–25% κ(kappa) 11.1711 (5.6585, 17.6613) prevalence. Full spatial model Intercept − 2.9184 (− 4.0083, − 1.9530) Model validation results ITN Coverage 4.5643 (2.4086, 6.8874) Presented in Fig.  3 is the model validation results to Travel time to health facility 0.0057 (0.0017, 0.0099) determine the predictive ability of the final model, espe- Aridity 0.0600 (0.0079, 0.1167) cially in the presence of new data. Given the high corre- σ 2(spatial variance) 0.8772 (0.5061, 1.2915) lation of 0.95, the fitted Bayesian geospatial prediction Range nominal 0.2917 (0.1250, 0.4886) model is very good for predicting malaria prevalence κ(kappa) 10.8059 (4.6372, 18.1236) spatially. Fig. 1 Posterior distribution of the effect of the predictors of malaria prevalence and the hyper parameters in the Bayesian spatial model in 2019 among under-five children in Ghana. *Denotes significant covariates A heto M alaria Journal (2022) 21:384 Page 7 of 15 Fig. 2 Empirical (observed) malaria prevalence in study locations in Ghana, 2019. Each circle represents a study location Aheto M alaria Journal (2022) 21:384 Page 8 of 15 in Fig. 4, and the interactive web-based version of Fig. 6 can be found in Additional file  2: Fig. S3. The width of the 95% credible interval ranges from 2.2 to 74.3% with the highest observed in parts of Bono East (59.8 to 74.3%) and the lowest in parts of Greater Accra (2.2 to 16.7%) regions. Comparing spatial model with covariate and spatial model without covariates To permit better comparison and understanding between the malaria predictive maps, we fixed the scale for both the predictive maps for the spatial model with covariate and spatial model without covariates. The results showed that the inclusion of these covariates helped explain some of the differences in malaria prevalence across the whole of Ghana, especially in the Central, Bono East, Oti, and Bono regions (Fig. 7). Fig. 3 Model validation for the final Bayesian geospatial model for We presented the SEs for the estimates presented in predicting malaria prevalence among children under-five in 2019 in Ghana Fig. 7 to examine the level of precision of the estimates for the two models. We observed a lower level of uncer- tainty (i.e., better precision) for the estimates in the spa- tial model with covariates compared to the spatial model The study found significant geographical differences without covariates (Fig. 8). The mean SE associated with in the predicted malaria prevalence in the country. The the spatial model with covariates was 8.9% compared to overall predicted malaria prevalence was 16.3% (stand- the SE of 13.9% for the spatial model without covariates. ard error (SE) 8.9%) with a range of 0.7% to 51.4% in the spatial model with covariates and overall prevalence Discussion of 28.0% (SE 13.9%) with a range of 2.4 to 67.2% in the This study utilized novel and advanced Bayesian Geospa- spatial model without covariates. Thus, inclusion of the tial models which is often the preferred approach to dis- covariates contributed to explaining some of the geo- ease mapping [30] to characterize under-five malaria risk graphical differences found in the predicted malaria spatially in this study. The need to link health outcomes prevalence. Here, the focus is on the interpretation of the like malaria to residential location of people is of utmost results from the spatial model that included the covari- importance globally and is increasingly being recognized ates. Residing in parts of Central (> 41.3 to 51.4%), Bono by the international health community and develop- East (> 41.3 to 51.4%) and Upper East (> 31.1 to 41.3%) ment partners for disease surveillance, monitoring, and regions was associated with highest risk of under-five control efforts [6, 20, 30, 34–36]. Under-five malaria is malaria after adjusting for the selected covariates. Other among the leading causes of under-five mortality in sub- relatively high-risk regions include Upper East, Oti, Saharan Africa. Malaria monitoring and control pro- Bono, Ahafo and Western North that recoded a preva- grammes could heavily benefit from timely, relevant, and lence of > 31.1 to 41.3% whereas parts of Greater Accra, accurate high resolution predictive malaria prevalence Eastern, Northern, Volta, Upper West, Savannah, and maps at a more localized levels supported with interac- Ashanti regions showed some of the lowest prevalence of tive web-based mapping tools that identify communities 0.7 to 10.9% (Fig.  4). The interactive web-based version with highest burden of malaria risk to inform optimal of Fig.  4 can be found in Additional file  2: Fig. S1. The preventive and targeted control efforts aimed at reducing standard errors (SEs) were presented in Fig. 5 to quantify malaria related morbidity and mortality, especially in set- the uncertainty associated with our estimates presented tings where universal intervention is practically impossi- in Fig. 4. The interactive web-based version of Fig. 5 can ble due to limited public health resources. be found in Additional file 2: Fig. S2. The estimated mean Of particular interest in this study is quantification of SEs is 8.9% with a range of 0.67 to 20.3%, suggesting geographical differences in under-five malaria risk con- low level of uncertainty for the estimates, hence reliable tinuously over the whole of Ghana as indicated in the estimates. predicted spatial maps. The 5 × 5  km high resolution Presented in Fig.  6 is the width of the 95% credible predictive maps showed substantial geographical differ- interval for the predicted malaria prevalence presented ences in the predicted malaria prevalence and identified A heto M alaria Journal (2022) 21:384 Page 9 of 15 Fig. 4 Predicted malaria prevalence in 2019 among under-five children in Ghana. The interactive web-based version of this map can be found online Aheto M alaria Journal (2022) 21:384 Page 10 of 15 Fig. 5 SEs of predicted malaria prevalence in 2019 among under-five children in Ghana. The interactive web-based version of this map can be found online Aheto M alaria Journal (2022) 21:384 Page 11 of 15 Fig. 6 Predicted width of the 95% credible intervals of malaria prevalence in 2019 among under-five children in Ghana. The interactive web-based version of this map can be found online Aheto M alaria Journal (2022) 21:384 Page 12 of 15 Fig. 7 Comparing the predictive maps for spatial model with covariate (left panel) and spatial model without covariates (right panel) from the Bayesian Geospatial models for malaria prevalence in 2019 among under-five children in Ghana specific communities/towns with highest concentration without covariates, this study found that inclusion of the of malaria risk, supporting previous studies that observed covariates helped explain some of the geographical differ- that health outcomes like malaria, malnutrition, mortal- ences in malaria risk, and with better accuracy compared ity and other related health outcomes exhibit spatial pat- to the spatial model without covariates, suggesting the terns and that the identification of these geographical need for researchers in this field to account for environ- patterns are of outmost importance urgent and targeted mental and climatic factors that might help explain the public health policy and intervention, especially in pre- malaria risk in this population of children for targeted vention and control efforts with the goal of improving preventive and control efforts. health outcomes in populations at sub-national, national To improve visualization, understanding and target- and global levels [6, 7, 20, 35, 37–39]. The overall pre- ing of scarce available resources to those communities dicted national malaria prevalence is 16.3% (SE = 8.9%), who needed it most (i.e., children highest malaria burden characterized by substantial localized geographical dif- areas), it is recommended that the programme manag- ferences with the highest observed in parts of Central ers and readers use the interactive web-based versions of and Bono East regions (41.3–51.4%) and lowest in parts the predicted maps published online (see figure titles for of Greater Accra, Eastern, Northern, Volta, Upper West, URL) where they can zoom-in or zoom out on specific Savannah, and Ashanti regions (0.665–10.9%). The find- towns or communities where the predicted malaria risk ings provide critical information to malaria control is highest or lowest. Generally, the level of uncertainties programme managers and other stakeholders in public associated with our estimates are low, suggestive of rea- health for urgent and targeted malaria preventive and sonably accurate estimates. Cross-validation was per- control efforts, where universal intervention is practically formed to examine how well our model performs on a impossible amidst limited public health resources. new data. The results show a very high correlation of 95%, Comparing the predictive maps of the estimates of suggesting that the model is good for correctly predicting our spatial model which included covariates and the one malaria risk spatially in this population of children. A heto M alaria Journal (2022) 21:384 Page 13 of 15 Fig. 8 Comparing the level of uncertainty between maps for spatial model with covariate (left panel) and spatial model without covariates (right panel) from the Bayesian Geospatial models for malaria prevalence in 2019 among under-five children in Ghana The study found environmental and climatic factors countries participating in the DHS program. Also, the like ITN coverage, travel times and aridity to be positively findings are relevant to the wider population of Ghana- predictive of under-five malaria prevalence. Increase ian children and similar populations elsewhere due to the in aridity index ranging from most arid to most wet nationwide coverage and representativeness of the survey increases the risk of malaria infection [40], while increase at the national level. Just like any other study, this study is in travel times to reach a high-density urban centre was subject to some limitations so the results should be inter- associated with increased risk of malaria infection, and preted with caution: data on spatially referenced malaria both findings are in the expected direction. Unexpect- data on policy and interventions, distance to the nearest edly, increase in ITN coverage, which is increase in pro- water bodies, type of housing which might explain some portion of population protected by ITNs was associated of the geographical differences in malaria risk observed with increased risks of malaria infection. This could be were unavailable to be included in the models. due to effect of suppressor variable and/or undetected The findings from the present study provide a critical multicollinearity [41]. tool for malaria surveillance and monitoring and assess- The key strength is the ability of the modelling ing progress in the fight against malaria and served as approach to borrow information from sampled locations an evidenced base for malaria control programme man- to create predictions and interactive web-based spatial agers and other stakeholders in public health to direct maps for both the sampled and the unsampled locations their resources to communities at utmost need, especially in the study over the whole of Ghana, while simultane- in countries like Ghana where available public health ously adjusting for environmental and climatic factors. resources are very limited, making it practically impos- This study accounted for the displacement of the cluster sible to rollout a universal intervention. Unlike national, locations, which is typical of the DHS data which ensures regional or district level estimates that masked real local- that similar approach can be applied accurately on other ized differences in risk levels (i.e., ecological fallacy), the Aheto M alaria Journal (2022) 21:384 Page 14 of 15 modelling and mapping approach enabled more localized Sustainable Development Goals; U5M: Under-five mortality; WAIC: Watanabe-Akaike information criterion. evaluation of malaria risk as a continuous phenomenon on finer scales (both at sampled and unsampled loca- Supplementary Information tions) over the whole of Ghana, allowing for effective and The online version contains supplementary material available at https:// doi. efficient allocation of the limited available public health org/ 10. 1186/s 12936- 022- 04409-x. resources dedicated to malaria prevention and controls efforts to communities at greatest need. Thus, this study Additional file 1. Building the mesh, SPDE and projector matrices for contributes to better understanding of the issue of under- estimation and prediction five malaria burden in Ghana. Additional file 2. Figures for supplementary material for online interactive web-based maps for Figs. 4, 5, 6 Conclusion Acknowledgements The study investigates, model, predict, and presents pre- Thank you to the MEASURE DHS Program for granting access and making the dictive maps of geographical differences in under-five data freely available for the study. malaria risk over Ghana. The Bayesian Geospatial mod- Author contributions elling of the environmental, and climatic predictors of JMKA developed the concept, secured, and analysed the data and wrote the malaria prevalence and interactive web-based spatial first draft manuscript. JMKA wrote and reviewed the various sections of the predictive maps provided in this study could be benefi- manuscript. JMKA reviewed the final version of the manuscript before submis-sion. JMKA read and approved the final manuscript. cial as an effective tool for the Ghana Health Service and her partners in the development of frameworks to miti- Funding gate malaria burden. This study identified communities Funding is not applicable to this paper. at highest risk of malaria that may require urgent and tar- Availability of data and materials geted interventions and further research amidst limited The datasets generated and/or analysed during the current study are available public health resources in this and other similar settings from the Measure DHS Program website http:// dhspr ogram.c om/ data/a vail able- datase ts. cfm. by public health officers, program managers and imple- menters, especially where it practically impossible to Declarations rollout a universal intervention. The modelling and spa- tial mapping approaches are critical as part of an overall Ethics approval and consent to participate strategy in reducing the malaria burden amidst limited The protocol for the 2019 Ghana Malaria Indicator Survey was approved by the Ghana Health Service Ethical Review Committee and ICF’s Institutional public health resources available in the country because Review Board. All data and other information collected were confidential. they can promote effective and sustainable malaria public Respondents’ names and identification numbers were removed from the health programs among under-five children in the coun- electronic database during analysis. The risk and benefits of participation in the survey were explained to respondents, including informed consent for try and other similar countries. To answer as-yet unan- the interview or blood collection, and informed consent was sought from all swered questions about why children residing in certain respondents. parts of Central, Bono East and Upper East regions were Consent for publication at highest risk while their counterparts in parts of Greater Not applicable. Accra, Eastern, Northern, Volta, Upper West, Savannah, and Ashanti regions were at lower risk, further research Competing interestsThe author declares that he has no competing interests. in the form of qualitative studies in addition to considera- tion and examination of further potential predictors left Author details 1 out in this study is warranted. The study further recom- Department of Biostatistics, School of Public Health, College of Health Sci-ences, University of Ghana, Accra, Ghana. 2 WorldPop, School of Geography mend that the DHS survey program managers and imple- and Environmental Science, University of Southampton, Southampton SO17 menters consider increasing the number of clusters to 1BJ, UK. 3 College of Public Health, University of South Florida, Tampa, FL, USA. be used in future surveys, and to include all districts in Received: 27 September 2022 Accepted: 7 December 2022 Ghana to improve the level of precision of the model esti- mates and the spatial predictions. References Abbreviations 1. Ghana Statistical Service (GSS), Ghana Health Service (GHS), ICF Interna- CHPS: Community-based Health Planning and Services; DHS: Demographic tional. Ghana Demographic and Health Survey 2014. Accra, Ghana, and and Health Surveys; EAs: Enumeration areas; GMIS: Ghana Malaria Indica- Rockville, USA, 2015. tor Survey; GSS: Ghana Statistical Service; ITNs: Insecticide-treated nets; 2. Centers for Disease Control and Prevention. Malaria: malaria’s impact INLA: Integrated Nested Laplace Approximation; MICS: Multiple Indicator worldwide. In.: CDC. https:// www.c dc. gov/m alar ia/ malar ia_ worldw ide/ Cluster Surveys; PMI: President’s Malaria Initiative; RDT: Rapid diagnostic test; impact.h tml. Accessed 13 Oct 2020. SPDE: Stochastic Partial Differential Equation; SSA: Sub-Saharan Africa; SDG: A heto M alaria Journal (2022) 21:384 Page 15 of 15 3. Thuilliez J, d’Albis H, Niangaly H, Doumbo O. Malaria and education: 28. Lindgren F, Rue H, Lindström J. An explicit link between Gaussian fields evidence from Mali. J Afr Econom. 2017;26:443–69. and Gaussian Markov random fields: the stochastic partial differential 4. WHO. World malaria report 2017. Geneva: World Health Organization; equation approach. J R Stat Soc. 2011;73:423–98. 2017. 29. Lindgren F, Rue H. Bayesian spatial modelling with R-INLA. J Stat Softw. 5. WHO. World malaria report 2019. Geneva: World Health Organization; 2015;63:1–25. 2019. 30. Odhiambo JN, Kalinda C, Macharia PM, Snow RW, Sartorius B. Spatial and 6. Weiss DJ, Lucas TCD, Nguyen M, Nandi AK, Bisanzio D, Battle KE, et al. spatio-temporal methods for mapping malaria risk: a systematic review. Mapping the global prevalence, incidence, and mortality of Plasmodium BMJ Glob Health. 2020;5: e002919. falciparum, 2000–17: a spatial and temporal modelling study. Lancet. 31. Rue H, Martino S, Lindgren F, Simpson D, Riebler A, Krainski E. INLA: func- 2019;394:322–31. tions which allow to perform a full Bayesian analysis of structured addi- 7. Yankson R, Anto EA, Chipeta MG. Geostatistical analysis and mapping of tive models using integrated nested Laplace approximaxion. R package malaria risk in children under 5 using point-referenced prevalence data in version 00-1404466487, 2014, http:// www.R- INLA. org.Accessed 20 Aug Ghana. Malar J. 2019;18:67. 2022. 8. GBD 2017 SDG Collaborators. Measuring progress from 1990 to 2017 and 32. R Core Team. R: A language and environment for statistical computing. projecting attainment to 2030 of the health-related Sustainable Develop- Vienna: R Foundation for Statistical Computing; 2022. ment Goals for 195 countries and territories: a systematic analysis for the 33. Taylor BM, Rowlingson BS. spatsurv: an R package for Bayesian inference Global Burden of Disease Study 2017. Lancet. 2018;392:2091–38 with spatial survival models. J Stat Softw. 2017;77:1–32. 9. GBD 2017 Mortality Collaborators. Global, regional, and national age-sex- 34. Golding N, Burstein R, Longbottom J, Browne AJ, Fullman N, Osgood- specific mortality and life expectancy, 1950–2017: a systematic analysis Zimmerman A, et al. Mapping under-5 and neonatal mortality in Africa, for the Global Burden of Disease Study 2017. Lancet. 2018;392:1684–735 2000–15: a baseline analysis for the Sustainable Development Goals. 10. Aheto JMK. Predictive model and determinants of under-five child mor- Lancet. 2017;390:2171–82. tality: evidence from the 2014 Ghana demographic and health survey. 35. Utazi CE, Thorley J, Alegana VA, Ferrari MJ, Takahashi S, Metcalf CJE, et al. BMC Public Health. 2019;19:64. Mapping vaccination coverage to explore the effects of delivery mecha- 11. Ghana Statistical Service GSS, National Malaria Control Programme, nisms and inform vaccination strategies. Nat Commun. 2019;10:1633. National Public Health Reference Laboratory, ICF. Ghana Malaria Indicator 36. Burstein R, Henry NJ, Collison ML, Marczak LB, Sligar A, Watson S, et al. Survey 2019. Accra, Ghana, and Rockville, USA; 2020. Mapping 123 million neonatal, infant and child deaths between 2000 12. Jakubowski A, Stearns SC, Kruk ME, Angeles G, Thirumurthy H. The US and 2017. Nature. 2019;574:353–8. President’s Malaria Initiative and under-5 child mortality in sub-Saharan 37. Aheto JMK, Utuama OA, Dagne GA. Geospatial analysis, web-based map- Africa: a difference-in-differences analysis. PLoS Med. 2017;14: e1002319. ping and determinants of prostate cancer incidence in Georgia counties: 13. Partnership RBM, WHO. Action and investment to defeat malaria evidence from the 2012–2016 SEER data. BMC Cancer. 2021;21:508. 2016–2030. Geneva: World Health Organization; 2015. 38. Aheto JMK, Yankson R, Chipeta MG. Geostatistical analysis and mapping: 14. WHO. Global technical strategy for malaria 2016–2030. Geneva: World social and environmental determinants of under-five child mortality, Health Organization; 2015. evidence from the 2014 Ghana demographic and health survey. BMC 15. Ministry of Health. Under five’s child health policy: 2007–2015. Accra, Public Health. 2020;20:1428. Ghana: Ministry of Health; 2007. 39. Fletcher IK, Grillet ME, Moreno JE, Drakeley C, Hernández-Villena J, Jones 16. National Malaria Control Programme (NMCP). Strategic plan for malaria KE, et al. Synergies between environmental degradation and climate vari- control in Ghana 2014–2020. Accra, Ghana: National Malaria Control ation on malaria re-emergence in southern Venezuela: a spatiotemporal Programme (NMCP); 2014. modelling study. Lancet Planet Health. 2022;6:e739–48. 17. MEASURE DHS. The DHS Program: Demographic and Health Surveys. 40. Okunlola OA, Oyeyemi OT. Spatio-temporal analysis of association Rockville: ICF Macro. https:// dhspro gram.c om/ data/a vaila ble- datase ts. between incidence of malaria and environmental predictors of malaria cfm. Accessed 21 Jul 2022. transmission in Nigeria. Sci Rep. 2019;9:17500. 18. Mayala B, Donohue R. The DHS program geospatial covariate datasets 41. Ludlow L, Klein K. Suppressor variables: the difference between ‘is’ versus manual. 3rd ed. Rockville, USA: ICF International; 2022. ‘Acting As.’ J Stat Educ. 2014;22:2. 19. Larsen DA, Hutchinson P, Bennett A, Yukich J, Anglewicz P, Keating J, et al. Community coverage with insecticide-treated mosquito nets and observed associations with all-cause child mortality and malaria parasite Publisher’s Note infections. Am J Trop Med Hyg. 2014;91:950–8. Springer Nature remains neutral with regard to jurisdictional claims in pub- 20. Aheto JMK, Dagne GA. Geostatistical analysis, web-based mapping, lished maps and institutional affiliations. and environmental determinants of under-5 stunting: evidence from the 2014 Ghana Demographic and Health Survey. Lancet Planet Health. 2021;5:e347–55. 21. Toh KB, Millar J, Psychas P, Abuaku B, Ahorlu C, Oppong S, et al. Guiding placement of health facilities using multiple malaria criteria and an inter- active tool. Malar J. 2021;20:455. 22. Aheto JMK, Pannell O, Dotse-Gborgbortsi W, Trimner MK, Tatem AJ, Rhoda DA, et al. Multilevel analysis of predictors of multiple indicators of childhood vaccination in Nigeria. PLoS ONE. 2022;17: e0269066. Ready to submit your research ? Choose BMC and benefit from: 23. Perez-Haydrich C, Warren JL, Burgert CR, Emch ME. Guidelines on the use of DHS GPS data. DHS Spatial Analysis Reports No.8. Calverton, USA: ICF • fast, convenient online submission International; 2013.https:// dhspr ogram. com/p ublic atio ns/ publi cation- SAR8-S pati al-A naly sis- Report s.c fm. Accessed 18 Apr 2021. • thorough peer review by experienced researchers in your field 24. Diggle P, Moyeed R, Tawn J. Model-based geostatistics (with discussion). J • rapid publication on acceptance R Stat Soc. 1998;47:299–350. • support for research data, including large and complex data types 25. Diggle P, Ribeiro PJ. Model-based geostatistics. New York, NY: Springer; 2007. • gold Open Access which fosters wider collaboration and increased citations 26. Matérn B. Spatial Variation. 2nd ed. Berlin: Springer-Verlag; 1986. • maximum visibility for your research: over 100M website views per year 27. Rue H, Martino S, Chopin N. Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations. J R At BMC, research is always in progress. Stat Soc. 2009;71:319–92. Learn more biomedcentral.com/submissions