Background
Malaria remains one of the most pressing public health and poverty-related issue in the developing world, particularly in sub-Saharan Africa [
1]. Each year, malaria might claim the lives of >1 million individuals. There are >500 million episodes of clinical
Plasmodium falciparum malaria and the global burden might exceed 40 million disability-adjusted life years (DALYs) [
2‐
4]. Mortality, morbidity and economic losses due to malaria could be reduced significantly if effective measures, such as sleeping under long-lasting insecticidal nets (LLINs) and access to prompt diagnosis and effective treatment using artemisinin-based combination therapy (ACT) were made available to all those in need [
5]. Interventions aiming at the control and local elimination of malaria require reliable risk maps in order to enhance the efficacy and cost-effectiveness of control measures. Since parasitaemia is correlated with clinical manifestations of malaria [
6], parasitaemia risk maps are a useful tool for the spatial targeting of control interventions. Ongoing blood sampling at the household level on a broad scale is expensive and not practical for surveillance purposes. District-level planning and targeting would be greatly facilitated by rapid and non-invasive identification of high-risk zones.
Over the past decade, geographical information system (GIS) and remote sensing technologies have been widely used for mapping malaria [
1,
7,
8]. However, purely GIS and remote sensing approaches have a number of shortcomings, due to their inability to quantify the relation between environmental factors and malaria risk and, consequently, infer predictions from statistical models [
9]. Furthermore, classical statistical models have been widely employed to evaluate the relationship between disease risk and demographic, environmental and socioeconomic factors, assuming independence of spatially-explicit data [
10‐
12]. Since disease data cluster in space, the assumption of independence is violated, and hence the statistical significance of the model covariates often overestimated [
13]. It follows that predictive risk models lack accuracy.
In recent work by the authors, Bayesian non-stationary geostatistical models were employed for spatial risk profiling of malaria [
9,
14,
15]. The strengths of these models are their accountancy for spatial dependence in the data, and the assumption of non-stationary spatial processes. The use of non-stationary models is further justified on the ground that local characteristics related to human behaviour and environment, including vector ecology, depend on location. Consequently, assuming stationarity may provide unreliable results when analyzing spatially-explicit disease data.
Here, risk factors and spatial patterns of
P. falciparum parasitaemia among school-aged children in a high endemicity setting of western Côte d'Ivoire are elucidated. An integrated approach, using GIS and remotely-sensed environmental data, questionnaire and parasitological survey data and Bayesian geostatistical models was employed. Finally, the use of non-stationary models for risk profiling of
P. falciparum parasitaemia at a regional scale was explored. The identified risk factors can help district health planners to implement malaria control interventions in a spatially-explicit manner, followed by monitoring and surveillance so that control tools can be fine-tuned over time to enhance their performance [
16].
Discussion
Current anti-malarial prophylaxis and treatment, and vector control using insecticides are susceptible to the emergence of resistant malarial parasites and vectors. Hence, there is a pressing need for other interventions incorporated into the programme that can delay the onset of resistance. There is also a need for new drugs and insecticides and a malaria vaccine, coupled with improved monitoring and surveillance [
28]. Mapping areas where people are at an elevated risk of infection and
P. falciparum parasitaemia is important for the design and implementation of district-based malaria control interventions.
Here an integrated approach for spatial risk profiling of
P. falciparum parasitaemia was used, building on previous research pertaining to the mapping and prediction of helminth infections in the Man region, western Côte d'Ivoire [
29]. Reasons why this approach is termed 'integrated' are as follows. First, a diversity of data (demographic, environmental and socioeconomic) was obtained from different sources, including cross-sectional questionnaire and epidemiological surveys and remote sensing. Second, data covered different spatial scales. For example, RFE, LST and NDVI data were collected by remote sensing at a large spatial scale. At a small spatial scale, data on proximity to standing water (e.g., swamps and irrigated agricultural fields) were obtained from questionnaires addressed to school directors and from digitized maps. Third, the data were collated, stored and managed using a GIS. Finally, Bayesian geostatistical models were employed to produce smoothed risk maps of
P. falciparum parasitaemia, and to compare model outcomes assuming either stationary or non-stationary dependence. Age, socioeconomic status, sleeping under a bed net, bed net coverage and different environmental factors - both small-scale (e.g., close proximity to standing water) and large-scale (e.g., LST, NDVI and RFE) - were significant risk factors for
P. falciparum parasitaemia. Interestingly, after introducing spatial correlation into the regression analyses, age, bed net coverage and - depending on the type of the model - mean RFE over the malaria transmission season, and distance to rivers appeared to be significant risk factors for
P. falciparum parasitaemia. Appraisal of model performance revealed no difference when comparing stationary with non-stationary models. However, the non-stationary model with ecological subregions showed that the geographical variability is different between subregions.
Two shortcomings of the present study should be noted. First, school-aged children are usually not the most severely affected group with malaria in highly endemic areas. Since the western part of Côte d'Ivoire is holoendemic for malaria [
15,
17,
18,
21], it is likely that school-aged children have acquired some kind of immunity to malarial parasites [
30,
31]. However, parasitaemia levels in school-aged children might be higher than in younger children. Underlying reasons are that school-aged children in high endemicity areas are mainly asymptomatic carriers, they might be more exposed to mosquito bites due to their behaviour, they are less likely to be treated because of a lower incidence of clinical malaria, and hence they might harbour considerably more parasites than preschool-aged children. Second, due to the possibility of sequestration mechanisms of infected erythrocytes from peripheral blood, as well as partially acquired immunity, microscopic examination of only a single finger prick blood sample might have underestimated the true prevalence of infection, and
P. falciparum parasitaemia might have been slightly different [
32‐
34].
Notwithstanding these shortcomings, several risk factors were found to be associated with
P. falciparum parasitaemia, including demographic factors (e.g., age), socioeconomic factors, personal preventive measures (e.g., sleeping under a bed net and bed net coverage) and a host of environmental factors. As expected, children who reported sleeping under a bed net were less likely to have a high malaria parasitaemia as were children from schools with a bed net coverage >25%. A study from rural Tanzania revealed that people from poorer households were less likely to access preventive measures [
35]. A similar result has been reported for the population under study here [
19]. Based on these observations and the common belief that the poorest population segments would share the highest burden of malaria, the current results surprisingly point in the opposite direction: schoolchildren from better-off households were more likely to have a higher parasitaemia than their poorer peers. This result is in accordance with previous work focusing on spatial risk profiles of
P. falciparum prevalence in the same group of children [
15] and consequently warrants further investigation.
For the current mapping of
P. falciparum parasitaemia, a similar geostatistical approach was used as before when modeling
P. falciparum prevalence data [
15] and common helminth infections [
22,
23,
36]. Importantly, the statistical significance of several covariates changed once spatial correlation had been taken into account. For example, children's socioeconomic status, sleeping under a bed net and several environmental factors - most notably LST, NDVI, close proximity to standing water and presence of pasture - were not significant anymore in the spatial models. This issue might be explained because omission of spatial correlation, when analysing spatially-explicit data, overestimates the significance of the regression coefficients [
13]. In contrast to previous spatial analyses of
P. falciparum prevalence data, it was found that environmental factors such as rainfall during the main malaria transmission season and distance to the nearest permanent river were significant predictors for
P. falciparum parasitaemia. These environmental covariates are related to the presence and abundance of malaria vectors, including
Anopheles gambiae and
Anopheles funestus, which are the key vector species as found in previous work in the nearby forest and wet Savannah zones of Côte d'Ivoire [
37,
38] and the medium-sized town of Man located in the centre of the current study area [
20]. As shown in a study from Burkina Faso these vectors breed in small pools (
An. gambiae) and larger semi-permanent water bodies (
An. funestus) [
39]. In previous research pertaining to
P. falciparum prevalence data, most of the environmental factors included had a large spatial scale and none of the environmental covariates was found significant [
15]. Hence, it was concluded that environmental data at a small spatial scale are necessary for more precise spatial risk profiling at the district level where decisions are usually made for the control of malaria and other infectious diseases. Indeed, including information obtained from interviewing the directors of schools about the proximity of residential houses to standing water revealed a number of significant environmental covariates in the non-spatial analyses, although there was a lack of statistical significance in the spatial models. At a more local or regional scale, only distance to rivers, which was used as a proxy for standing water, was significant in one of the spatially-explicit models. Further ground-based investigations are required, since only data derived from questionnaires and digitized maps were used rather than ecological surveys to explore small-scale environmental features. It will also be interesting to determine the use of topography-derived wetness indices, which have been linked to household malaria risk at small spatial scale in two communities in the Kenyan highlands [
40]. Perhaps somewhat surprising at first, the present spatial analyses showed that RFE during the main malaria transmission season, which is rather a broad scale indicator, indicated the spatial heterogeneity of parasitaemia in the study area. This observation might be explained by the distinct climatic conditions, i.e., higher precipitation in the mountainous northern part of the study area.
Comparing the performance of different models did not reveal any significant difference in the predictive ability between stationary and non-stationary models, and hence the predicted parasitaemia risk maps were similar. Interestingly though, the non-stationary model with ecological subregions predicted a slightly larger area with high parasitaemia in the north-eastern part of the study area. The corresponding standard deviations of the map showed that uncertainty was particularly high in this subregion. A likely explanation of this observation is that there were fewer sampled locations in that specific subregion (Figure
1). However, uncertainty in the north-eastern part of the study area was also elevated (though to a lesser extent) when employing a stationary and a non-stationary model with fixed subregions. Of note, the spatial parameters in the non-stationary model with ecologic subregions revealed that geographic variability differed between subregions. Consequently, this would rule in favour of using non-stationary models for predicting
P. falciparum parasitaemia. Previous spatial analyses of
P. falciparum prevalence in the same area revealed that non-stationary models performed somewhat better than stationary models [
15].
An important aspect of the current study is that the statistical model approach influences not only the spatial parameter estimates, including the prediction maps and standard deviations of the prediction, but also the significance of malaria risk indicators. Depending on the statistical model chosen, i.e., stationary or non-stationary, the significance of several environmental factors changed. For example, in the stationary and the non-stationary models with ecological subregions, mean RFE during the main malaria transmission season was significantly explaining the geographical heterogeneity, whereas in the non-stationary model with fixed subregions, this covariate was not significant. Instead, distance to rivers appeared as a significant covariate in the non-stationary model with fixed subregions. Such differing results have also been reported by others when comparing stationary and non-stationary models for the risk of malaria across Mali [
9]. The covariate mean RFE during the main malaria transmission season had the lower BCIs near 1 in both stationary and non-stationary models with ecological subregions, and the increase in odds due to increased rainfall was only 0.28 and 0.24, respectively. In contrast, the non-stationary model with fixed subregions seems particularly promising, as the upper BCI for the covariate distance to rivers was not close to 1 and the parasitaemia risk decreased by over a third with increasing distance from rivers.
Employing a spatially-explicit risk profiling approach, demographic, environmental and socioeconomic risk factors were identified that govern the geographic distribution of
P. falciparum parasitaemia in a high endemicity area at the district level. This information can be utilized for designing and implementing malaria control interventions. In particular, at the time of the study in 2001/2002, virtually no malaria control interventions were carried out in the region of Man. The very low frequency of schoolchildren reported sleeping under a bed net (< 10%) documents this issue [
19]. Although bed nets were available for purchase from local dispensaries and the district hospital in the town of Man, the price was perceived as too high. It is speculated that the malaria situation in this region has not improved, partially explained by an armed conflict starting in September 2002 that also hit the region of Man and resulted in a collapse of the health care delivery systems [
41,
42]. Available information supports this claim; coverage of bed nets (ITNs) was reported below 5% in Côte d'Ivoire at a national scale [
43] and in the Man region in particular [
44]. The results further suggest that health-seeking regarding prevention and treatment of malaria at dispensaries was weak, as no statistical significance was found with regards to distance to a health post.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
GR contributed to the conception and design, participated in the data collection, carried out the spatial analyses and interpretation of the data and drafted the manuscript. KDS was involved in the data collection, quality control, data analyses and drafting of the manuscript. PV contributed to the analysis of the data and drafting of the manuscript. BHS was involved in the interpretation of the data and critical revision of the manuscript. AY was involved in the acquisition of data. MT contributed to the conception and design. JU contributed to the conception and design, interpretation of the data and drafting of the manuscript. EKN was involved in the conception and design as well as the critical revision of the manuscript. All authors read and approved the initial submission and the revised version of the manuscript.