Assessing the effects of malaria interventions on the geographical distribution of parasitaemia risk in Burkina Faso

Malaria is holoendemic in Burkina Faso with most transmission occurring during or shortly after the rainy season between July to December. Ninety-nine percent of infection is attributed to Plasmodium falciparum. The overall prevalence of infection in children aged 6–59 months is estimated at 66 % [Burkina Faso Demographic and Health Survey-Multiple Indicator Cluster (BFDHS-MICS 2010)] [1]. The under-five severe malaria attributable death has been dropped from 8.1 % in 2000 to 3.3 % in 2010 [2]. The government has made tremendous efforts to achieve the objectives of the 2006–2010 National Malaria Strategic Plan and implemented special programmes, such as home-based malaria management in 2008, universal coverage of insecticide-treated bed nets (ITN) in 2010, intermittent preventive therapy (IPT) for high-risk groups in 2005, piloting of the indoor residual spray (IRS) in certain health districts since 2010, larval control and sanitation programmes, introduction of effective tools for malaria control mainly the rapid diagnostic test (RDT) at all health facilities in 2010 and actual availability of artemisinin-based combination therapy (ACT) in health facilities in 2007 [2].

Burkina Faso carried out the first malaria indicator survey (MIS) in 2010, a nationally representative household survey in the country compiling malaria-related indicators. MIS surveys generate a number of indicators of malaria intervention coverage that can be used to assess progress towards the goals of the global malaria action plan (GMAP) [3]. These indicators measure ownership, use and access of ITNs, implementation of IRS, access to ACT and to intermittent preventive treatment for pregnant women (IPTp). MIS data have been used to assess effects of interventions [4‐10]. Some studies reported protective effects [4, 7] for specific interventions and others did not find any effect [9, 10]. In the analysis of the Senegal MIS data of 2008, using geostatistical variable selection, it appeared that among the various indicators of ITN ownership, only few were able to estimate a protective effect of ITN intervention on parasitaemia risk [4].

Intervention effects are likely to vary in space because there is often geographical variation in the intervention coverage and their effects are related to malaria endemicity [11‐13]. Recently, the effects of vector-control interventions on changes of malaria parasite risk were estimated at different spatial resolutions in six sub-Saharan African countries using Bayesian geostatistical models with spatially varying covariates. Results suggested that some interventions may not show any effect when looked upon at national level while they can have a protective effect at sub-national level [7].

Samadoulougou et al. [6] used the Burkina Faso MIS 2010 data to estimate the spatial distribution of malaria risk among children under five of age in Burkina. The authors included only ITN use as an intervention-related predictor and estimated an overall effect at country level, which was not statistically significant.

In this study, MIS data were analysed using Bayesian geostatistical models to assess the effects of different malaria interventions at national as well as sub-national level (50 health districts) in the country. Bayesian variable selection within geostatistical models allowed to screen different coverage measures for each intervention and spatially structured regression coefficients measured the effects of interventions at district level. Predictive maps of the disease burden adjusted for climatic effects were also produced.

A total of 5741 children aged between 6 and 59 months from 574 clusters were tested for parasitaemia using blood smear test results. The overall observed malaria parasite prevalence was 65 %. Thirty-four (5.92 %) clusters had no data thus reducing the actual number of clusters to 540 (Fig. 1).

The ratio of ITN to children under 5 years old ranges from 0.33 in district of Orodara to 1.52 in Nanoro district with 11 out of the 50 districts in the country having a ratio above 1. 80 % of the 50 health districts have IRS coverage less than 1 %. The maximum IRS coverage of 50 % was found in the districts of Diebougou and Toma. The percentage of households falling in the lowest social-economic quintile ranges from 1.6 % (in Tenkodogo) to 63.33 % (in Gorom-gorom). Table 2 gives a summary of the raw coverage measures per health district. The highest proportion of fever cases receiving ACT is 22 % and it is observed in the districts of Tenkodogo in the East central and Koudougou in the Central part of the country. Figure 4 (right) depict the ITNpU5, ITNsU5 and ACT coverage at health district level, respectively.

The results of variable selection are presented in Table 3. In particular, Bayesian geostatistical variable selection applied in the climatic model 1 indicated that the most important factors related to parasitaemia risk are LSTN (Night Land Surface Temperature) in linear form and the proximity to rice cultivation (categorical) with 74.10 and 74.50 % posterior inclusion probabilities, respectively. The place of residence (rural/urban) stood out as one of most important predictors with a posterior inclusion probability of 87.20 %. Geostatistical variable selection of the ITN coverage measures (model 2) showed that the proportion of children that slept under a net (ITNsU5) had the highest probability (equal to 32.02 %) to be included in the model among the ones that were assessed. The inclusion probability of less than 50 % indicates that ITNsU5 is less likely to have an important effect at national scale. However, ITNsU5 was included in model 3 as a spatially varying covariate to assess important effects at sub-national levels and identify health districts that the ITN interventions are associated with the parasitaemia risk.

The predictive performance of the model (model2) is shown in Fig. 2. Eighty-two (82 %) of test locations were falling into all credible intervals with probability areas greater than 50 %.

Parameter estimates of the three models are given in Table 4. In particular the climatic model 1 showed a negative correlation of the parasitaemia risk with LSTN and distance to rice fields. Moreover, living in rural areas increases the odds of being infected by about 3.91 times (95 % BCI: 2.88, 5.22). The minimum distance at which the spatial correlation is less than 5 % is equal to 10.61 km (95 % BCI: 2.39–20.82). Model 2 assesses the effects of malaria interventions on parasite risk after adjusting for climatic and socio-economic confounders. Results show that none of the health intervention measures is an important predictor of parasitaemia risk at national level. However, the model indicates that the odds of malaria infection decreases with better socioeconomic conditions reaching a 68 % reduction within the least poor group, OR = 0.32 (95 % BCI: 0.24–0.43). An increasing gradient of malaria risk was also observed with age with the oldest age group (4–5 years old) having odds of 2.08 times higher than infants. Furthermore, the model reveals a decreasing trend of parasitaemia odds with increasing mother’s education level although this decrease is not statistically important. The additional SES and malaria interventions related predictors in model 2 were able to explain some of the spatial correlation in the model. This is reflected in the estimate of the range parameter, which reduced to 3 km compared to 10.6 km in model 1.

Figure 3a–c depict the predicted parasitaemia risk maps (median (a), 2.5th (b) and 97.5th (c) percentiles of the posterior predictive distribution estimated from model 2) in children less than 5 years of age at 1 km² spatial resolution in Burkina Faso. Estimates show that malaria parasitaemia risk ranges from 36 to 71 % across the country while the median predicted prevalence is 59 %. The Southwest, Comoe, Cascade, East, Central-west, Boucle du Mouhoun and the Sahel regions bear the highest prevalence. The central, the North, the east central and regions appear to be the less burdened regions. The total number of infected children under 5 years old in the country was estimated to be 1097,296. Table 5 presents the population-adjusted and estimated number of infected under five children under 5 years of age per province and region.

Spatially structured coefficients of intervention coverage measures obtained by model 3 allowed estimation of the effect of ITN, IRS and ACT at health district level. Figure 4 presents the different coverage (right hand side maps) and intervention effects (left hand side maps) for each health district estimated from model 3 with the spatially varying covariates. ACT coverage appears to be an important health system component associated with decreased malaria parasitaemia risk in a number of the health districts. However the strongest effects are observed in the districts of Ouagadougou, Koudougou, Kaya, Zorgho, Koupela and Tenkodogo. ITN usage shows a protective effect in only four health districts namely Ouagadougou, Ziniare, Ouahigouya and Sebba. IRS which is in a pilot phase in the district of Diebougou (South-west) did not show any effect.

This is the first study in Burkina Faso to estimate the spatial effects of malaria interventions on the geographical distribution of parasite risk. Different ITN coverage measures related to bed net use and ownership were calculated and the effects of ITN, IRS and case management coverage on the spatial pattern of malaria risk were studied after taking into account disease variation due to climatic and socio-economic factors. Georeferenced estimates of the parasite risk were obtained as well as the number of infected children. This study analysed the Burkina Faso MIS survey data of 2010 and employed Bayesian geostatistical models with spatially varying covariates using geostatistical variable selection to identify the most relevant bed net coverage indicators. The geostatistical variable approach identified a list of the most important climatic and environmental predictors. Weiss et al. [20] have proposed a list potential predictors that could be used to improve malaria modelling.

The most important ITN coverage measure was the proportion of children under 5 years old sleeping under bed net, with however a low posterior probability to be included in the model equal to 32.02 % followed by far by the number of ITN per under five (2.46 %) and the number of ITN per household member (0.20 %). This finding support previous results showing that contingent upon the setting and the prevailing conditions, ITN ownership and ITN use may show different ability in capturing ITN intervention effects on parasitaemia [4].

Overall, the effect of interventions in Burkina Faso was not significantly associated with change in malaria parasitaemia risk at country level. These results are in line with the findings of Giardina et al. [7] in a multisite study. The lack of statistically important ITN intervention effects may be explained by the fact that at country-level, a sizeable percentage (37.97 %) of children under the age of 5 years still do not sleep under ITN. However when analysed at sub-national level (health district level) some interventions show protective effects in certain districts. For instance ITN use, which appeared to have not an important effect at national level proved to be an effective intervention in four health districts namely Ouagadougou, Ziniare in the central region of the country, Ouahigouya in the North and Seeba in the Sahel region. Among factors that might affect ITN effectiveness feature the inadequate coverage and the uneven distribution of ITN and the usage among households and health districts [21].

Indeed, malaria is holoendemic in Burkina Faso and the transmission occurs throughout the year. In such settings, even significant reductions in the total exposure would not necessarily warrant substantial reductions in parasitaemia [22‐25]. Furthermore, there is still low ITN utilization compliance (62.03 %) among children under the age of 5 years. This study found a protective effect of ITN usage in specific districts. This result supports findings reporting that the ITN intervention is expected to be more effective in low transmission settings with the highest ITN usage [26]. Districts where ITN use was found to be protective are located in low to mild transmission settings where coverage of ITN use ranges from 42 to 90 % [27].

ACT showed a protective effect in 19 health districts. However it is worth noting that except Gorom-gorom in the Sahel region, Solenzo in the North West (Boucle du Mouhoun), secteur 15 and 22 in the West (Hauts-bassins) and Diebougou in the Southwest, the remaining districts where ACT showed a protective effect are located in the central region. A close inspection of the coverage levels shows that ACT tends to be effective in districts where a minimum level of coverage is achieved (above 5 %). This finding supports the hypothesis that the effectiveness of a given intervention is related to both its coverage as well as the transmission levels. Indeed, the ACT effect is presumably related to the very low coverage [28]. High levels of transmission are also believed to limit the effect of ACT. Findings from a study conducted in Tanzania suggested that the percentage reductions in prevalence of infection and incidence of clinical episodes achieved by ACT were much higher in areas with low initial transmission [28]. ITN interventions, however, aim at reducing the malaria transmission intensity by reducing the chances that an individual will be bitten by an infective Anopheles mosquito [29]. However low compliance in ITN use may seriously reduce the potential impact of ITNs [27, 30]. Therefore, conjugated efforts to increase both the ACT coverage (in order to reduce the prevalence) and ITN use (to further reduce the transmission) are required to warrant a synergetic effect towards a better and effective control of the disease [31]. Findings from a continental study that used data from 32 malaria-endemic African countries showed that ITN intervention was the most important and effective malaria intervention accounting for an estimated 68 % decline in malaria parasite rate in 2015 [32]. The geostatistical model was able to identify districts with important protective ITN effects, although at country level the effect was not important.

IRS was not associated with malaria parasitaemia risk most probably owing to an extremely low percentage of houses sprayed within the last 12 months (0.92 %) prior to the survey.

Malaria is known to be a climate-driven disease and among the most important climatic factors features temperature. The model-based parasitaemia risk map depicts a strong spatial with lower parasitaemia risk estimated in the cities (urban settings) relative to rural settings. The results show a negative association between increased night temperature and malaria transmission. Laboratory experiments observed the shortest Anopheles gambiae s.s larval survival (<7 days) at 10–12 and 38–40 °C and the highest larval mortality occurring between 30 and 32 °C, with death (rather than adult emergence) representing over 70 % of the terminal events in mosquitoes originally from Lagos (Nigeria) [33]. In Burkina Faso, the monthly mean temperature in the hottest and driest period (March–May) is constantly well above 31 °C. Land surface night temperature, therefore appears to be an important predictor of malaria transmission. Furthermore, the behavioural high temperature avoidance experiment showed that An. gambiae, the most efficient malaria vector species in Burkina Faso, was more sensitive to increased temperatures than its sibling species, Anopheles arabiensis [33]. In nature, this probably results in short distance flights to seek cooler spots, typically the shaded resting sites under vegetation outdoors or cool dark comers indoors. The highly endophilic nature of An. gambiae protects the mosquito from the highly variable and more extreme external climate. This may explain the negative association between increased LSTN and the transmission because during the hot night spells local populations rest outdoor thus reducing knowingly or unknowingly the contact human-vector. The authors also found that female temperature avoidance was most pronounced in hungry females (which avoid temperatures above 25 °C), less strong in blood-feds (above 30 °C) and least strong in newly emerged females (above 32 °C). High night temperatures were also found to affect An. gambiae (one of the most predominant and effective malaria vector in Burkina Faso) behaviour and vectorial capacity [34, 35]. A significant negative association between temperature and malaria infection was also found in a previous study in Burkina Faso [6]; However the authors did not consider day and night temperatures separately and the climatic data considered in this study do not span the study period (April 2010–January 2011). The map of nighttime land surface temperatures (LSTN) is also provided (see Additional file 2).

The present study estimated a negative association between malaria parasitaemia and proximity with rice growing areas. The rice growing areas used in this study were extracted from the National land use database with cartographic scale coverage of 1/200,000 which features only large and economically relatively important rice growing areas. Furthermore, as exposed to an increased risk of malaria infection, surrounding populations receive relatively high attention from the local government including regular sensitization campaigns (Information Education and Communication). Consequently, as an income generating activity, the local population is relatively well off. This makes it easier to access health care and other protective measures. This effect is known in Burkina Faso as the “paddy paradox” defined as the occurrence of large populations of vectors but low amounts of malaria transmission where irrigated rice is grown.

The negative association has been reported in other studies in Burkina Faso, Ghana, Gambia and Tanzania [36‐39] The potential explanation might be that the irrigated rice fields are preferred habitats for An. arabiensis, which has a lesser vectorial capacity than other species [36].

Furthermore it is hypothesized that the “paddy paradox” is due to young pre-gravid mosquitoes dispersing more widely than gravid ones, not necessarily to low survival in the mosquito [37]. The map of the distance between the clusters and the nearest rice-paddy field in kilometre is also provided (see Additional file 3).

The predicted spatial distribution of malaria parasitaemia risk ranges from 36 to 71 % across the country while the median predicted prevalence is 59 %. The predicted parasitaemia map shows the higher risks in the Southwest, South-Central and the Eastern region of the country. The above mentioned regions coincide with the regions of country bearing the highest vegetation density and receiving the highest annual rainfall relative to the northern part which receives less rainfall and is more “desertic”. The predicted parasitaemia risk map shows similarities as well as discrepancies with the previous mapping efforts. Compared to the P. falciparum endemicity map of the Malaria Atlas Project (MAP), common patterns were found in the northern and northeastern parts of the country, which appeared to be less burdened [40]. Discrepancies were identified regarding the highest burden areas which MAP places in the northwestern part of the country while our map estimates in southwestern Burkina Faso. Samadoulougou et al. [6] indicated that the northern and southwestern regions have the highest and lowest malaria risk respectively. The above discrepancies may be explained by the different climatic/environmental and other covariates used in the predictive models.

This study estimated higher number of infected children in the cities despite the relatively low prevalence observed in the urban settings. This finding is consistent with the results from previous study that used the BFDHS-MICS 2010 data [6]. Differences were observed between raw and population-adjusted parasitaemia risk estimates which is explained by the low prevalence observed in densely populated areas. For example the province of Kadiogo, one of the smallest provinces with the highest population density and the lowest population-adjusted raw parasitaemia risk (34 %), shows an even lower parasitaemia risk adjustment for the population (4.04 %). Similar results were found using Senegal MIS 2008 data [4].

An increasing risk gradient with age was found. Infants had the lowest risk while older children had the highest risk. An association was observed between socio-economic status and malaria risk, with children within the least poor quintile being substantially at reduced risk. Similar results were observed in a previous study in Burkina Faso and in other malaria endemic areas [4, 6].

This study provides estimates of the effects of malaria interventions at country as well as at local scale. The estimated risk and intervention effect maps are valuable tools for identifying high-risk areas and areas with less effective interventions in order to improve malaria control in Burkina Faso. These outputs can serve as benchmarks to evaluate the effectiveness of future control interventions and progress of the efforts towards disease control.