Background
Malaria remains a leading cause of morbidity and mortality in tropical and subtropical regions of the world. There are an estimated three billion people at risk of this disease and more than half a billion episodes of clinical
Plasmodium falciparum occur each year, killing over one million individuals annually [
1,
2]. It has been estimated that the global burden of malaria exceeds 40 million disability-adjusted life years (DALYs) [
2,
3] and the disease drains the social and economic development of affected regions [
4,
5]. High-risk groups are children under the age of five years and pregnant women, with sub-Saharan Africa particularly affected. Indeed, this part of the world accounts for a striking 90% of the global burden of malaria [
6,
7].
Predicting the abundance and spread of malaria in endemic settings, in order to develop locally-adopted malaria control strategies to lower the burden of the disease is a pressing public health issue. Recently, an audacious goal has been announced in Seattle, USA during a meeting led by the Bill and Melinda Gates Foundation, namely to eradicate malaria [
8]. This goal – so the claim – has become a realistic hope thanks to new scientific advances, including the development of novel antimalarial drugs, vaccines and integrated control efforts through insecticide-treated nets (ITNs), prophylactic treatment and indoor residual spraying (IRS), in the face of a growing political will and financial support for malaria control initiatives [
8]. A deeper understanding of the spatial distribution of malaria is pivotal so that appropriate local elimination efforts can be designed and rigorous monitoring implemented.
Advances made with geographical information system (GIS), remote sensing and geostatistical modelling to predict the spatial and temporal distribution of malaria and
Anopheles vectors have opened new avenues in this field of research. In particular, modelling disease and disease-related data within a Bayesian framework allows fitting of complex models in quite a flexible way. Additionally, Bayesian approaches provide computational advantages over traditional frequentist approaches via implementation of Markov chain Monte Carlo (MCMC) simulation [
9‐
11]. Recent studies made use of the advantages offered by Bayesian methods for spatially-explicit modelling of malaria [
12‐
18].
In Côte d'Ivoire, malaria is one of the primary public health concerns. This is illustrated by a study carried out in the savannah zone that documented malaria being responsible for at least 60% of the consultations in hospitals and 46% in paediatric clinics [
19]. In 2005, Côte d'Ivoire ranked at position 13 among countries with the highest rates of under-five mortality and estimates at the time suggested that only 4% of children under five years of age slept under an ITN [
20]. In the present study, small-scale patterns and spatial risk factors of the prevalence of
P. falciparum among schoolchildren in a rural part of western Côte d'Ivoire were explored, using Bayesian geostatistical models.
Discussion
The purpose of this study was to assess risk factors and small-scale spatial patterns of P. falciparum infection prevalence among schoolchildren in a highly endemic area of rural western Côte d'Ivoire. The following covariates were significantly associated with infection: age, socioeconomic status, sleeping under a bed net, distance to health care facilities and a number of environmental factors. However, after accounting for spatial correlation, only age remained a significant risk factor for P. falciparum prevalence, whereas NDVI showed only 'borderline' significance. The predictive ability of the spatial models was examined using a training sample of 78% of the schools, with the non-stationary model performing better than the stationary one.
There are a number of shortcomings worth discussing. First, only a single finger prick blood sample was collected from each child for microscopic examination. Hence it is conceivable that some infections, particularly those with a low parasitaemia, were missed [
28,
29]. Second, it should be noted that school-aged children in highly malaria-endemic areas are not at highest risk of disease-associated morbidity and mortality. The prevalence in children below the age of five years might have been even higher than the observed
P. falciparum prevalence of 64.9% among six to 16-year-old children. Third, the parasitological survey was carried out over a period of several months due to the large number of schoolchildren subjected to interviews and finger prick blood sampling, which might have introduced a bias in the observed prevalence from one school to another due to seasonality. Fourth, in the absence of high-resolution data to compute distances to small standing water bodies that might serve as
Anopheles breeding sites, information from digitized maps was used to obtain the distance to rivers as an indication for the distance to breeding sites. The most likely vector in this area is
Anopheles gambiae and, to some extent,
Anopheles funestus. The former vector species breeds in transient, sunlit and generally small pools, whereas the latter has been associated with larger, semipermanent bodies of water containing aquatic vegetation and algae [
30].
The analysis presented here showed that schoolchildren from wealthier households were more likely to be infected with
P. falciparum compared to schoolchildren from the poorest households. This result is surprising given that the common expectation would be that the poorest of the poor are at highest risk of malaria [
31]. Several studies have shown that the burden of malaria is elevated among the poorest population segments, probably because they are at a higher exposure to malaria vectors and have fewer means for personal protective measures. For example, a study carried out in a rural community in Cameroon found a significant relationship between malaria and low protective housing conditions, such as living in wooden plank houses [
32]. Surprisingly, no significant association between the risk of a
P. falciparum infection and housing conditions was evident in the present study. It is conjectured that issues related to exposure were associated to socioeconomic status, which calls for further investigation. Previous research conducted in rural Tanzania, for example, found that lack of access to health care and preventive measures, including ITNs, was associated with people's socioeconomic status [
31]. Interestingly, the current study confirms that children from poorer households were less likely to sleep under a bed net. Furthermore, children who reported sleeping under a bed net were at a decreased risk of having a
P. falciparum infection. Additionally, it was found that the risk of a
P. falciparum infection was associated with distances to health care facilities. Nevertheless, after taking into account spatial correlation, the covariates socioeconomic status, distance to the nearest health care facility and sleeping under a bed net showed no significant association anymore, and hence other factor must explain the observed spatial heterogeneity of
P. falciparum.
Several environmental factors, namely NDVI, RFE and distance to rivers, were significantly associated with a
P. falciparum infection in the bivariate non-spatial models. These findings are in accordance with previous studies that showed significant associations between malaria and NDVI, rainfall and distance to rivers at a broader spatial scale [
33‐
35]. It is conceivable that these environmental factors are related to the presence and abundance of malaria vectors, which is governed by suitable breeding and resting sites of
Anopheles. An interesting observation in the present study was that children from schools that were located in close proximity to rivers (<500 m) were at a lower risk of a
P. falciparum infection compared to more distant schools (between 500 m and 1000 m). Children from schools with distances <500 m were significantly more often reporting to sleep under a bed net, suggesting that the former observation might be partly confounded by a higher level of bed net coverage and usage due to nuisance from mosquitoes near rivers. Children enrolled in schools located at distances >1000 m of rivers were less likely to be infected with
P. falciparum, which might be related to the flight range of mosquitoes, which is, on average, below 1 km [
36]. Interestingly, none of the environmental covariates showed a statistical significant association to
P. falciparum prevalence after accounting for spatial correlation. Hence, the current results demonstrate the importance of accounting for spatial correlation when analysing malaria prevalence data at small spatial scales as reported here. Indeed, omission of spatial correlation would have underestimated the standard errors of the covariate coefficients [
37]. Furthermore, in contrast to previous work focussing on helminth infections in the same study area [
11,
21,
24,
38], no risk map and corresponding uncertainty map have been presented, since none of the environmental factors investigated was significant in the spatially-explicit model. The results therefore suggest that at small spatial scales, individual-level factors (e.g. age) determine the spatial distribution of the
P. falciparum infections rather than coarser environmental factors. These observations suggest that environmental factors are particularly salient for malaria prediction at larger spatial scales.
In geostatistical modelling, the standard assumption is that there is a stationary spatial dependence in the data, which implies that the spatial correlation is a function of the distance between points and independent of the location. Bayesian non-stationary geostatistical models were employed before for the prediction of helminth infections in the same study area [
24,
38]. Gosoniu and colleagues were the first to use Bayesian non-stationary geostatistical models for malaria risk, in their recent research on Mali [
16] and West Africa [
39]. The authors' underlying assumption was that local characteristics related to human behaviour and environment, including vector ecology, influenced spatial correlation differently at different locations over large areas, i.e. an entire country. The results presented here suggest that the use of non-stationary models may also be required at a smaller spatial scale (i.e. at the district level), since the non-stationary model performed better than the one assuming stationarity. The current work on
P. falciparum can be integrated with our previous work on helminth infections for mapping
P. falciparum-helminth co-infections using multinomial regression models for the simultaneous targeting of malaria and helminthic diseases [
11]. School-aged children are at the highest risk of such co-infections and data suggest that co-infections with
P. falciparum and hookworm have an additive impact on anaemia, implying that those high-risk groups would greatly benefit from integrated malaria and helminth control [
40].
Authors' contributions
KDS contributed to the conception and design of the study, collected the data, was responsible for quality control issues of malaria slide reading, assisted with the analysis of the data and editing of the manuscript, GR contributed to the conception and design of the study, collected the data, analysed and interpreted the data and drafted and edited the manuscript, AY was involved in the collection of the data and supervision of the field work, PV contributed to the analysis of the data and editing of the manuscript, MT contributed to the conception and design of the study, EKN and JU oversaw all aspects of the study, including conception, design, execution of the field work, interpretation of the data and editing of the manuscript. All authors read and approved the final version of the manuscript.