Three MIS studies have taken place in Madagascar since 2011, providing a valuable source of information about the status of malaria from a standardised data collection protocol applied across a nationally representative set of locations. A broad range of indicator metrics are included in the MIS reports. This present study looks at the parasitology data in particular to assess how the prevalence of malaria infection in children 6 to 59 months old has changed in recent years. Given the limitations of the routine health data reporting chain discussed previously [
4], this study provides a complementary perspective into the recent status of malaria endemicity in Madagascar, identifying important increasing trends since 2011 despite a relatively small fall in prevalence between 2013 and 2016.
Comparison with reported MIS results
The results presented in MIS reports are summaries of the raw survey data, with individuals weighted in such a way as to ensure appropriate national representation in the overall summary statistics. However, the results are not adjusted for the temporal lag in survey dates across years (Additional file
1: Figure S1). By 2016, the survey time window was pushed to 2 months later than in 2011, corresponding to a delay away from the peak transmission period in most regions (Additional file
3: Figure S3 and Additional file
4: Figure S4). The model presented here specifically accounts for this temporal lag, allowing for meaningful comparisons between years. Seasonal trends in transmission [
4] mean that a delayed survey period will underestimate endemicity relative to the original survey time period. This is reflected in the raw MIS results, which suggest a 6.2% national
PfPR
6–59mo in 2011 and 7% in 2016, a 13% increase over that period. In contrast, the spatio-temporal model presented here identifies a 127% increase, resulting in an estimated national mean 9.3% annual
PfPR
6–59mo in 2016.
The MIS specifically excludes three highland districts considered to be malaria-free (the cities of Antananarivo-Renivohitra, Antsirabe I and Fianarantsoa I) as well as communes above 1500 m in altitude. Despite not being represented in the mapping input dataset, model predictions are derived for these areas, which then inform the ecozone- and national-level summaries. Without prior exclusion of these zones, the model-based approach risked over-inflating the estimates of endemicity. Environmental covariates, however, appear to have discerned the unsuitability of the highland urban habitat, and the malaria-free districts are predicted to have <0.5% infection prevalence in all 3 years. The overall ecozone-level mean annual PfPR6–59mo for the central highlands was 0.6% in 2011, 2.0% in 2013 and 1.6% in 2016, which were of the same order of magnitude as those in the MIS summaries of 0.8%, 1.1% and 0.9%, respectively. The excluded MIS sampling zones do not, therefore, seem to have impacted the spatio-temporal model predictions.
Comparison with previous prevalence maps
While geostatistical methods have been previously applied to infection prevalence datasets to predict spatially continuous prevalence maps of Madagascar, this has been within the context of continental or global-level mapping [
37,
39]. The country-specific modelling approach applied here allows more freedom to the environmental covariates to adapt to Madagascar’s specific ecological context, with the model selecting those covariates most pertinent to the island’s environmental diversity. The continent-level spatio-temporal prediction cube developed by Bhatt et al. ([
39], reproduced for Madagascar [
4]), indicates a much coarser granularity than the present predictions, with little variation between years. Nevertheless, the continental maps do allow malaria endemicity in Madagascar to be viewed in its broader context, showing the relatively low prevalence of infection in Madagascar relative to many countries in sub-Saharan Africa (incidence rate ranked 13th lowest out of 43 countries in 2015 [
39]). Prevalence mapping analyses are also valuable in evaluating trends in malaria prevalence over time.
Comparison with health metrics information system data
In this study, we map the parasite reservoir, as detectable by microscopy. In parallel, clinical case numbers are collated by the routine health metrics information system [
4,
40,
41]. While the two metrics report different characteristics of malaria, spatio-temporal trends from both are similar, with a comparable geographic distribution of the burden of disease and an important increase in burden particularly along the west coast. A major increase in clinical cases was reported between 2014 and 2015, which subsequently reduced in 2016 (the third MIS year) following a mass distribution of bed nets treated with insecticide at the end of 2015 [
41]. The longitudinal nature of the data from the health metrics information system allows extreme events to be identified, such as epidemics, which may drastically affect the annual case totals, as in 2015, but which may not be distinguishable as exceptional, or even captured by cross-sectional surveys that assess prevalence at isolated time points.
A 2017 World Health Organization (WHO) report estimates that only around 31% of all clinical cases are reported through the surveillance system to the central level in Madagascar. This is likely primarily driven by the population’s low rates of seeking treatment. Of mothers seeking treatment in 2016 for their febrile children aged 6 to 59 months, 35.8% did so at public health facilities and 46.2% did so at any source (including public health facilities) [
11]. Such low capture of the overall case burden, therefore, throws into question the system’s capacity to adequately quantify changes in the malaria burden over time.
The two indicators, therefore, have their strengths and limitations, but together corroborate general trends of an increased burden between 2011 and 2016, punctuated with reductions, such as those identified between the 2013 and 2016 MIS surveys, or between 2015 and 2016 by the routinely reported case data [
40,
41]. This temporal heterogeneity may be partly attributable to reductions in NMCP activities caused by Global Fund disbursement delays in 2014 [
40]. In addition, recent evidence of the variable quality and durability of the insecticide-treated bed-net brands distributed across the country means that over their 3-year lifespans (mass distribution campaigns in Madagascar are triennial), protection from nets will be inconsistent [
42]. MIS campaigns in Madagascar have all been timed to take place in the transmission season following mass bed-net distribution campaigns, which may, therefore, capture snapshots of prevalence at its lowest.
Limitations to the approach
In this study, we have used data on the prevalence of infection to assess the current spatio-temporal trends of malaria infection in Madagascar. This metric is independent of clinical symptoms, and instead quantifies the extent of the parasite reservoir across the population. The value of such a metric lies in its simplicity and standardised collection methods, and, in the context of the MIS, repeated national representation. A recent review by Cohen and colleagues [
43], however, has argued for a multi-component mapping process that will adequately identify the underlying drivers of transmission, to enable NMCP to target control measures optimally. Elimination requires both the reduction of the parasite reservoir and the prevention of transmission. Prevalence data alone cannot fully characterise the local epidemiology without a parallel understanding of a site’s historical context, the significance of imported cases, the impact of recent control efforts, entomological and host behavioural/genetic factors, and so on [
43]. A practical interpretation of the prevalence map, therefore, requires insight into the underlying factors driving the observed infection rates. The map suite presented here is one component in understanding malaria in Madagascar, but ought to be interpreted in association with a broader set of malariometric data when used to determine control intervention policy. Malaria transmission is also highly dynamic both spatially and temporally, meaning that predicted maps based on single time-point snapshots of prevalence may simplify the true underlying situation. Maps such as those presented here provide insight into general trends, and the validation statistics presented in Additional file
2: Figure S2, Additional file
3: Figure S3 and Additional file
4: Figure S4 indicate the level of variability that might be expected around these predictions.
Malaria prevalence is strongly influenced by intervention coverage levels, including rates of insecticide-treated net (ITN) ownership [
39] and treatment seeking [
13,
14]. Including these covariates in the modelling framework could help inform the model about the patterns of endemicity but this was not considered feasible in this present analysis. The coverage of indoor residual spraying and ITN use rates, for instance, have complex non-linear relationships with malaria prevalence. For example, indoor residual spraying in Madagascar is carefully targeted to the highest (as an emergency response to reduce mortality during outbreaks) and lowest (prevention of reintroduction and subsequent autochthonous transmission) endemicity districts only [
8]. ITN coverage is strongly skewed to areas where malaria is endemic. The highland areas into which malaria is mainly imported are not covered by routine ITN distribution. The limited temporal window considered in this present study does not allow for the protective effect of high ITN coverage to be learnt by the model, and instead the coarse learnt association is that ITN coverage increases as prevalence increases. Treatment seeking was excluded for reasons of sample sizes. MIS data on treatment seeking from individual cluster locations are limited to mothers with infants who suffered from fever in the 2 weeks preceding the MIS interviews. Sample sizes at the cluster level are, therefore, very small, producing spurious results when analysed at the high resolution of the present analysis. Despite these barriers to including intervention covariates in the model, the suite of environmental and socio-demographic variables that were used allowed robust predictions of malaria prevalence, so this was not considered a major limitation to the mapping model presented here.
Madagascar is noted for its mosaic landscape of ecological habitats, with land cover varying across short distances [
44]. The critical importance of the environmental covariates in the modelling process is evident from the methods described here, with a strong predictive role associated with vegetation cover that explains differences in malaria prevalence between different areas (Table
2). The covariates associated with each MIS site describe the local conditions associated with the observed malaria prevalence at the time and location of sampling. However, MIS datasets are geopositioned with a deliberate degree of spatial uncertainty (displacement) to promote anonymity of up to 2 km in urban areas, 5 km in rural areas and 10 km for 1% of rural points [
45]. This spatial displacement, therefore, introduces uncertainty into the associations between reported
PfPR and their attributed covariate values, which could impact the model’s predictions. For this study, we assumed that while the island is ecologically heterogeneous, the impact of this spatial uncertainty will be acceptably low (with ecological conditions similar for most points even at a distance of 5 km or 10 km), and similar enough to allow a prevalence signal to be identified. The model validation statistics corroborate this assumption.
A further limitation of the MIS datasets that informed the current mapping analysis stems from their sampling design and sample sizes. The broad range of indicators included in the MIS activities present conflicting demands on sample sizes, which are further constrained by financial and logistical considerations. Sample sizes cannot, therefore, be optimised for all indicators, but instead are focussed on a limited number of these. A recent retrospective model-based analysis of the 2011 and 2013 Madagascar MIS datasets estimated that sample sizes were under-powered by 17% and 36%, respectively, to reach effective sizes for infection prevalence rates [
46]. This was based on rapid diagnostic test results and not microscopy (as considered in this present study), meaning that it may be a slight overestimate. The approach followed here, namely to consider samples from the three MIS datasets within a common Bayesian hierarchical modelling framework and to draw on a broad range of associated covariate surfaces, provides a solution to sample size limitations.
Furthermore, at very low transmission levels, such as in the central highlands and parts of the central fringe regions of Madagascar, microscopy-based cross-sectional surveys of parasite prevalence risk are underpowered to detect rare and low parasitaemia infections adequately [
47]. In these areas, higher sensitivity diagnostics (such as nucleic-acid amplification-based approaches) and alternative indicators (such as serological markers) are required to monitor malaria epidemiological trends more effectively [
47‐
50]. Particular interest is in the use of serological panels targeting both short- and long-lasting antigenic responses [
51]. When applied to appropriate sentinel populations [
52,
53], these tools can be powerful probes of changing transmission intensity or reintroduction events in the pre-elimination setting where the majority of infections at the time of survey may be microscopically sub-patent [
54].