Challenges of DHS and MIS to capture the entire pattern of malaria parasite risk and intervention effects in countries with different ecological zones: the case of Cameroon

Cameroon is a central Africa country, bordered with Nigeria to the West, Chad to the North, Central African Republic to the East, Congo, Gabon and Equatorial Guinea to the South. The country is decentralized and organized around 10 regions, 58 divisions and 360 communal areas. English and French are the official languages. Yaoundé is the political capital and Douala is the economic town. The global surface of the country is 475,650 km², population is around 22 million inhabitants [14, 15] and index of human development is 0.512 in 2015 [16]. The percentage of the population living in urban areas is 49%. Children under 5 years old represent 17% of the population [3, 17]. Despite of the presence of natural resources as oil, gas, iron, gold, and favourable climatic situations for agriculture, the national income per inhabitant is still low (< 2000$ per year) with important disparity between urban and rural areas [18]. The country has different geographic and ecological zones, that generate six epidemiologic facets of malaria transmission [19‐21], corresponding to different ecological systems: the dry Sahelian in the Far North region and the Sudano-guinean in the North region where malaria transmission period is between 4–6 months; the highlands of Adamawa and West regions with length of malaria transmission between 7–12 months; the equatorial forests which includes Centre, East and part of South regions where the transmission is stable; and the Atlantic coastal covering the Littoral and a part of South and South-West regions where the malaria is perennial with seasonal variations. The malaria transmission in the North part of Cameroon is characterized by seasonal pattern linked to rainy season which cover the period from August to October. Like many Africa countries, Plasmodium falciparum is also the predominant species and responsible of more than 95% of confirmed infection cases in this study [22, 23].

Environmental and climate predictors were extracted from satellite sources (Table A.1 in Additional file 1). In particular, data were compiled on land surface temperature during the day and night (LSTD, LSTN), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), land cover surface and permanent water bodies obtained from the moderate resolution imaging spectroradiometer (MODIS) terra satellite. rainfall estimates (RFE) and altitude were retrieved from FEWS (or Famine Early Warning Systems Network) and SRTM (or Shuttle Radar Topographic Mission) web sites, respectively [25, 26]. Climatic proxies with weekly and bi-weekly temporal resolution were averages over the 1 year period prior to the survey.

Socio-economic data were included in both, the DHS and the MIS surveys. Two socio-economic proxies were used, education of women in reproductive age and household asset index. The education level was treated as a categorical variable with three levels (primary, secondary and university). The household asset index was included in the database and used in categorical form, grouped into quintiles corresponding to the poorest, poor, middle, rich and richest segments of the population. Rural and urban area information was available in the database for the observed survey locations and it was extracted from the GRUMP (or Global Rural and Urban Mapping Project) database at the locations of predictions [27].

To capture the effects of interventions at national level, output indicators were generated using data available in the DHS and MIS, according to the household survey indicators tool for malaria control developed by Roll Back Malaria and partners. In particular, the following coverage indicators of use and access to ITN interventions were created: (a) proportion of children under 5 years old who slept under an ITN the previous night; (b) proportion of households in the cluster with at least one ITN; (c) proportion of households in the cluster with at least one ITN for every two people; (d) proportion of population with access to an ITN within their household. Furthermore, a health system performance indicator was calculated to measure the proportion of children under 5 years old with fever in the last 2 weeks who seek treatment at hospital, tested and treated with recommended ACT [28].

Bayesian geostatistical binomial models fitted on cluster level malaria aggregated data were used to estimate parasitaemia risk at high spatial resolution based on climatic predictors (Model 1). Climatic variables were categorized in groups with cut-offs defined from quintiles and exploratory analysis. Geostatistical variable selection was carried out to identify the most important climatic and environmental predictors, including their best fitting functional form [29, 30]. For each predictor a categorical indicator was introduced with values 0, 1 and 2 corresponding to exclusion of the predictor from the model or inclusion in linear or categorical form, respectively. It was assumed that the indicator arose from a multinomial distribution with probabilities defining the variable-specific exclusion/inclusion probabilities (in linear/categorical forms) in the model (Additional file 2). A threshold of 50% was considered for the probability of inclusion (i.e. posterior inclusion probability) into the predictive geostatistical model. In the final model, the effect of a predictor was considered to be statistically important if the 95% Bayesian credible interval (BCI) of the coefficient did not include the one on the odds ratio scale. Validation of Model 1 was performed to assess the model’s predictive performance. In particular, the sample was divided into a training set which included 80% of the data which was used for model fit and a test set consisting of the remaining data. Model validation compared the mean error between the observed parasitaemia at the locations of the test set with the model-based predicted risk. The model predictive performance was also evaluated by calculating the proportion of test locations correctly predicted within the 95% of BCI. Bayesian kriging was applied using Model 1 to predict the parasitaemia risk over a gridded surface of 117,192 cells and obtain pixel-level risk estimates at 2 × 2 km² resolution [31, 32].

Geostatistical variable selection was also applied to select the most important coverage indicators of malaria interventions. A Bayesian geostatistical Bernoulli model was fitted on the parasitaemia status of each child to estimate the effect of selected malaria interventions (Model 2) after adjusting for potential confounding effects of the climatic factors used in Model 1 and of the socioeconomic factors. The same methodology was employed separately on the DHS and the MIS data. Model fit and prediction were conducted in R [33] and OpenBUGS version 3.2.3 (Imperial College and Medical Research Council, London, UK) [34, 35]. Convergence of parameters was assessed by the Geweke statistic and by visually inspecting the traceplots [36]. Computations were performed in the parallel scientific computing (sciCORE) platform of Basel University. Different maps were produced by ESRI’s ArcGIS version 10.2.1 for Desktop (http://​www.​esri.​com/​).

The observed parasitaemia risk in children under 5 years old was 30% at national level, 37% in the rural and 20% in the urban areas. In urbanized cities, such as Yaoundé and Douala the malaria parasite risk was among the lowest in the country, i.e. 12 and 13%, respectively. Five percent of the population had access to an ITN within their household and 21% of children slept under an ITN during the night preceding the survey. The percentage of children under 5 years old with fever in the last 2 weeks that treated with ACT was 6%. The proportion of children from the poorest and poor quintiles was 63% and the proportion of mothers with at least primary education was 80% (Table 1).

The geostatistical variable selection identified NDVI and altitude (in categorical form) as the most important predictors of parasitaemia risk, using the cluster level model (Model 1). The proportion of the population with access to an ITN in their household and the proportion of children under 5 years old with fever in the last 2 weeks treatment-seeking at hospital, tested and treated with ACT, were selected from Model 2 (Table 2).

Posterior estimates of the model’s parameters are shown in Table 3. The climatic, cluster level model confirmed known relations between malaria parasite risk and climatic predictors, i.e. a positive association with NDVI and a negative relation with altitude. A malaria parasite risk map was generated using the climatic predictors identified from the DHS data (Fig. 2).

The individual level model (Model 2 in Table 3) shows that children in urban areas or those living in households with higher socioeconomic level were less affected by malaria. Children to mothers with high educational level or aged below 12 months had low malaria parasite risk. The proportion of population with access to an ITN in their household was able to capture a statistically important effect on parasitaemia risk.

The national observed malaria parasite risk was 33% with substantial disparities between rural (43%) and urban (19%) areas. The North-West region and the towns of Yaoundé and Douala had registered low prevalence of 10, 6 and 16%, respectively. According to the survey data, 9% of the population had access to an ITN within their household and 15% of households possessed one ITN per two persons. The percentage of children under 5 years old with fever in the last 2 weeks that received ACT was 12%. The percentage of children from the poorest and poor quintiles was 68% and the proportion of mothers with at least primary education was 76% (Table 1).

The geostatistical variable selection applied to the cluster level model (Model 1) had identified NDVI, the categorical forms of EVI and of distance to water, the presence of forest and altitude as the most important predictors of parasitaemia risk. The individual level model estimated high posterior inclusion probabilities for the following ITN coverage proxies, i.e. the proportion of population with access to an ITN in their household, the proportion of children who slept under an ITN in the previous night and the proportion of households with one ITN per two persons. However, the paired correlations between the above ITN indicators were ranging from 0.6 to 0.8; therefore, the indicator included in the final model (Model 2) was the last ITN coverage measure which had the highest inclusion probability (Table 2).

Parameter estimates for Model 1 and 2 are shown in Table 4. The cluster level predictive model indicated that malaria parasite risk was positively related with NDVI, EVI and presence of forest, and it was negatively associated with altitude. A malaria parasite risk map was drawn with climatic predictors selected by the MIS (Fig. 3).

The individual level model (Model 2 in Table 4) showed that children in rural areas as well as those living in households with lower socioeconomic status are more vulnerable to parasitaemia risk. Children aged below 12 months have low risk. The educational level of mother was not statistically associated with malaria parasite risk. ITN coverage was statistically important and had a negative effect on malaria parasite risk.

The proportion of test locations falling into the BCIs of the predictive posterior distributions with probability coverage varying from 50 to 95% was comparable for both surveys (Model 1), but the accuracy of estimates was higher for the DHS data as it is shown by the smallest BCI width (Fig. 4) which indicates the difference between the upper and the lower values of the interval. Validation of the cluster level Model 1 on the DHS and MIS data showed a mean absolute error of 0.050 and 0.038, respectively.