Characterizing local-scale heterogeneity of malaria risk: a case study in Bunkpurugu-Yunyoo district in northern Ghana

Bunkpurugu-Yunyoo is located in Northern Region, Ghana (10.6° N 0.0° W). The district exhibits a gradual slope from the rocky Gambaga escarpment in the north/northwest to riverine plains in the south/southeast, dropping from 518 to 128 m above sea level within a relatively small area (30 km by 40 km). The district lies in the Guinea Savannah zone, characterized by a unimodal rainy season from May to October, which peaks in August–September while the remainder of the year is typically dry. The mean annual rainfall is between 100 mm and 115 mm, with annual temperature ranging between 30 °C and 40 °C. The predominant malaria vector species are Anopheles gambiae sensu stricto and Anopheles funestus [21]. In 2010, the estimated total district population was 122,591, with an estimated under-five population of 21,373. The study district is predominantly rural, with the exception of two settlements that exceed a population threshold of 5000, and are thereby designated as ‘urban’ according to the Ghana Statistical Service, namely Nakpanduri (population 6179) and Bunkpurugu (population 11,106) [22]. Most of the population live in mud-walled compounds and engage in rain-fed, small-scale farming as well as small-scale trading.

Malaria transmission in BYD is characterized by strong seasonal variations that closely follow rainfall patterns, with a peak lasting up to 3–4 months between August and November [6]. The study area is defined as the set of communities that were eligible for spraying by the PMI-funded IRS programme in 2011–2013 (Fig. 1). The study boundaries do not perfectly align with the official administrative district boundary because the sampling frame for the study consisted of the list of communities that fell under the public health jurisdiction of the BYD office of Ghana Health Service (GHS). As a result, this list leaves out communities handled by the East Mamprusi office of GHS along the extreme northeastern border and the southeastern corner of BYD.

The district had benefitted from two recent mass distributions of long-lasting insecticide-treated nets (LLINs). In 2010, the LLINs programme covered children under 5 years and pregnant women and in 2012 they covered all other members of the household, resulting in approximately > 75% coverage across BYD [2, 23]. The other major malaria prevention and control intervention in the region had been IRS, which began in 2011 using alphacypermethrin 0.4% WP (ICON^®10CS, Syngenta, Basel, Switzerland) and shifted to organophosphates in 2013 due to declining susceptibility of Anopheles spp. to the pyrethroid insecticides [21].

Surveys were conducted twice annually at periods of expected peak (i.e., the end of the rainy season, October–November) and trough (i.e., the end of the dry season, March–April) levels of malaria parasitaemia. Each survey was collected over a 2-week period.

Children under 60 months of age and their caregivers were randomly selected for each survey, using a multi-stage randomized cluster sampling design. Probability proportional to population was used to randomly sample approximately 72 communities per survey across the study area from a GHS roster of 238 communities. Within each community, survey teams visited 15–17 households with children under 5 years old, selected randomly from an inventory of such households that had been conducted within 6 months prior to each survey. Teams were instructed to invite any child under 5 years old who was available on the day of the survey and had slept in the household the night before for participation. The resulting sample (average 24 children per community) covered approximately 20% of the under-five population in each survey, based on the 2010 census data and community lists developed by the GHS [3]. Thin and thick blood films were prepared in the field and later read by two qualified microscopists at the Noguchi Memorial Institute for Medical Research laboratory in Accra to determine microscopic parasitaemia. In cases of discordant readings, a senior microscopist determined the final result. Prior to analysis, children aged under 6 months were removed from the dataset, based on previous studies suggesting that young infants have enhanced malaria protection due to maternal antibodies [24, 25].

Using a hand-held Etrex© GPS device (Garmin), field technicians obtained GPS coordinates for a readily identifiable central point in each community, such as a church, school, or chief’s palace. GPS coordinates were not obtained at the household level because of feasibility and ethical considerations. The geo-coding of communities permitted the dataset to be enhanced with remotely sensed and geographic variables.

Remote sensed variables Malaria transmission has been shown to be strongly related to satellite-derived environmental and socio-demographic factors [26]. Weiss et al. [27] carried out a comprehensive assessment of spatially gridded covariates that are likely to be associated with the malaria transmission cycle. Based on their results a suite of satellite and geographically derived covariates were assembled for modelling purposes (Table 1). The selection of environmental covariates was partly based on availability of raster data that closely matched the survey times. Elevation is widely used in malaria mapping due to its association with precipitation and temperature [27] and it was extracted from the 90-m resolution Shuttle Radar Topography Mission Digital Elevation Model (SRTM-DEM). The normalized difference vegetation index (NDVI), a proxy for vegetation cover, was obtained from Moderate Resolution Imaging Spectroradiometer (MODIS) products using the 16-day composite. Based on this product, the maximum NDVI was calculated within a 32-day period prior to the start date of each survey. Vegetation cover is a useful proxy for characterizing vector habitat for Anopheles spp. commonly found across Africa [28]. The land surface temperature (LST) for day and night was also obtained from MODIS (MOD11) products using the 8-day composites, which were then used to calculate a monthly average LST 32 days prior to survey start date. Temperature has been widely accepted to be an important component in malaria transmission, largely based on its influence on mosquito survival, development, breeding, and biting rates [8, 29, 30], whereas precipitation is a proxy for available stagnant water puddles that are ideal habitat for mosquito larvae. Previous studies have shown that a 9 to 12-week time lag exists between rainfall and onset of malaria transmission [31]. Therefore, the variable of cumulative rainfall for a 3-month period 30 days prior to survey start date was created, based on the Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) data while long-term precipitation patterns at finer resolution than CHIRPS (1 km) were extracted from WorldClim datasets [32]. Long-term precipitation was included because of its higher spatial resolution compared to cumulative rainfall. Night-time lights for 2011 were obtained from National Oceanic and Atmospheric Administration (NOAA) satellite products at 1-km resolution [33], being a useful proxy for poverty and infrastructure (i.e., electrified housing are more likely to be walled and be less suitable habitat for vector development) [34]. Because densely populated areas are often poor habitat for Anopheles to breed [35], Population density at 100-m resolution was extracted from WorldPop [36] and validated this product using the census data for the two main urban towns of Bunkpurugu and Nakpanduri. Finally, land use products by MODIS were expected to be an important factor [27] for vector habitat, however it was not included because there was little variation across the study district [37].

The use of too many covariates can lead to over-fitting of the model, as well as multicollinearity [53, 54]. To address this issue, a two-step procedure, commonly used in mapping exercises [9, 55], was used to select the best predictors from Table 1. First, correlation between covariates was calculated, and in cases of high correlation (i.e., \( R \) > 0.7), a single representative covariate was selected. The choice of which covariate to retain was based on the strength of association with malaria prevalence, spatial resolution, and its relevance to malaria epidemiology [27]. Second, a bi-directional, step-wise, regression analysis was run on a full multivariate model using the remaining covariates. The covariates associated with the model with the lowest Akaike information criterion (AIC) were used for subsequent analysis.

A Bayesian hierarchical geostatistical model was fitted independently to data from each survey. Let the malaria status \( M_{ijt} \) for child \( i \) in village location \( j \) in survey t be a binary variable, equal to 1 for a positive microscopy result and 0 otherwise. Using a probit regression framework, assume that \( M_{ijt} = 1 \) if the latent variable \( z_{ijt} \) is greater than 0 and \( M_{ijt} = 0 \) otherwise. The latent variable \( z_{ijt} \) is modelled as:where \( \varvec{x}_{{\varvec{ijt}}}^{\varvec{T}} \) is the design vector containing the age of each child, \( \varvec{\beta}_{t} \) contains the corresponding regression coefficients, and \( \alpha_{jt} \) is the village-survey level random-effect. Assume that:where \( \varvec{W}_{t} \) is the design matrix and \( \varvec{\gamma}_{t} \) is a vector of the corresponding coefficients. Furthermore, \( {\varvec{\Sigma}} \) is the spatial correlation matrix, where the correlation between villages \( k \) and \( l \) is given by the exponential parametric function \( exp\left( { - \frac{{d_{kl} }}{{\rho_{t} }}} \right) \). In this expression, \( d_{kl} \) is the Euclidean distance between villages k and l and \( \rho_{t} \) is the correlation decay parameter.

Finally, the priors are given by: where \( {\mathbf{\rm T}} \) is a diagonal matrix with diagonal elements equal to [10,1,…,1]. A customized Gibbs Sampler programmed in R [56] was used to fit this model separately for each survey.

The geospatial model performance was assessed based on its out-of-sample predictive ability using a tenfold cross-validation approach for each survey. Villages in the data were randomly divided into 10 sub-sets. The model was trained on 9 of the 10 sub-sets, and then used the estimated parameters to predict the expected prevalence for the remaining withheld sub-set. This was repeated using each testing sub-set. The out-of-sample predictive skill of the geospatial Bayesian model was compared to that of a standard generalized linear regression model (GLM), which is a commonly used statistical framework for modelling this type of response variable. A temporal validation was also done for the Bayesian model, where the model was trained on one survey and used to predict prevalence for future surveys. These temporal predictions were stratified by season (i.e., rainy season surveys were used to predict future rainy season surveys but not dry season surveys). Two statistical metrics were used for all validations: the log-likelihood and the mean absolute error (MAE) in relation to malaria prevalence in each village.

Once the geospatial model was fitted, the results were extrapolated to the neighbouring district: East Mamprusi. This district was chosen because it has similar topography and environment to Bunkpurugu–Yunyoo. Furthermore, a list of health facilities and urban centres for East Mamprusi was obtained from the GHS [57] and corroborated using expert information and the 2010 population census, respectively [22], enabling the determination of the distance to health facilities and urban centres. For an independent evaluation of the accuracy of these spatial extrapolation, the MAE was calculated between the predicted malaria prevalence rates in the 2011 rainy season and the 2011 Multiple Indicator Cluster Survey (MICS) prevalence estimates for locations that fell within this district [58].

The study included 10,518 children aged 6–59 months of age with a complete microscopy diagnostic test result for malaria parasitaemia collected in 438 communities across 3 rainy seasons and 3 dry seasons. The average age of participants was 31 months. The average community level prevalence was 44.46% (95% CI 43.83–45.74) during the entire study but prevalence varied considerably by survey, with a strong seasonal effect (Table 2).

As mentioned in Methods, an initial set of 14 spatial covariates were chosen based on the malaria mapping literature [27] (Table 1). NDVI, rainfall, long-term precipitation, and daytime LST were highly correlated (R > 0.7). Because NDVI had the strongest association with microscopy and this covariate had the finest spatial resolution, NDVI was retained and the other three highly correlated covariates were excluded from the step-wise selection. Because accessibility was highly correlated with distance to health facility (R = 0.71), given that distance to health facility was based on ground-truth locations of active health facilities at the time of each survey, it was retained instead of accessibility. A step-wise, model-selection approach using AIC revealed that the most important covariates were distance to urban centre, distance to health facility, elevation, NDVI, distance to road, distance to water, LST at night, and night-time lights. These parameters were used for geostatistical predictions (parameter estimates can be found in Additional file 2).

The Bayesian geostatistical model revealed a strong rural–urban gradient with prevalence generally increasing northeast to southwest (top and middle panels in Fig. 2) ranging from 19 to 90%. These results were particularly surprising given that the malaria prevalence maps for Ghana developed by the MAP suggest a much smaller range in prevalence (68–83%), without a clear geographical trend (bottom panels in Fig. 2). Malaria remained low in and around the urban centres throughout the study but this pattern was more distinct during rainy seasons versus dry seasons. The only notable temporal trend was a small reduction in prevalence in the southern portion of BYD in the final dry season.

The regression results indicate that increasing distance to urban centre was significantly associated with high malaria prevalence estimates during rainy seasons. Furthermore, higher elevation was linked to reduced parasite prevalence of malaria during the rainy seasons in 2010 and 2011, and the dry season in 2012. Finally, living further away from a health facility tended to increase the risk of high malaria prevalence, but this effect was only significant for dry season 2013. All other covariates showed no statistically significant association. Posterior estimates of the parameters used for predictive modelling are given in Additional file 2.

The model validation results using the average log-likelihood across all surveys (Fig. 3a) revealed that the Bayesian model generally had better out-of-sample predictive performance when compared to the GLM model, particularly during the dry seasons. Similar results also arose in relation to the mean absolute errors (Fig. 3b).

The temporal validation results reveal an average MAE of approximately 12.7%, without much difference in error for temporal predictions of 1 versus 2 years (Table 3). These MAE values are similar to those obtained when performing the tenfold cross-validation, suggesting that temporal predictions within a 1- or 2-year time interval are as accurate as spatial predictions.

Predicted malaria prevalence in East Mamprusi was substantially higher during the rainy season when compared to the dry season (Fig. 4). As expected, there is a distinct trend of lower transmission close to urban areas and the 2012–2013 dry seasons showed a significantly lower predicted prevalence. Finally, as expected, there is distinctly higher uncertainty across East Mamprusi compared to Bunkpurugu-Yunyoo for all surveys. The results for survey three (rainy season 2011) and MICS 2011 survey data were used to determine the reliability of these spatial extrapolations. MICS 2011 contained 10 clusters that fell in Bunkpurugu-Yunyoo and East Mamprusi district. Of these, six were in Bunkpurugu-Yunyoo and four in East Mamprusi. No location in survey data matched the MICS locations; so unfortunately, direct ground-truthing was not possible. The MAE was equal to 19.5 and 11.2% for Bunkpurugu-Yunyoo and East Mamprusi, respectively. The overall MAE for all 10 locations in both districts was 16%. While this may seem large, it is important to note that this margin of error is not substantially larger than those obtained with the tenfold cross-validation and temporal prediction exercises in Bunkpurugu-Yunyoo. Furthermore, this margin of error is enough to adequately distinguish the main spatial prevalence trends in the region, something that current national prevalence maps are not able to detect. For comparison purposes, the MEA for these 10 locations with the MAP 2011 estimates was 37.6% (graphical representation can be found in Additional file 2: Figure S1).