Site Selection and Population Description
Malaria outcome, socioeconomic, and demographic data were obtained from a cross-sectional study of 5,602 participants across 24 communities in four regions of Madagascar (Figure 1) (34, 36). We conducted larval vector surveys and subsequent analyses in these four study regions. The Rice et al., 2020 study included data from a second sub-region of the east coast (their NE region) but malaria prevalence estimates and study design differed and thus this region is omitted here. Each region will be referred to by their geographical location throughout the paper, these are listed here along with their corresponding administrative districts: southeast, SE = Vatovavy Fitovinany: Mananjary district; southwest, SW = Atsimo Andrefana: Toliara II district; west coast, WC = Atsimo Andrefana: Morombe district; and high plateau, HP = Amoron’i Mania: Ambositra, Ambatofinandrahana, and Fandriana districts.
Sampling regions were selected to represent the different typologies of malaria transmission patterns in Madagascar, ranging from consistent, endemic transmission in the east, to episodic, epidemic transmission in the central highlands (22, 29). These zones correspond to Madagascar’s distinct ecological regions, with differing precipitation and vegetation patterns and include humid forest, grasslands, dry spiny forest-thicket, and others (38). The landscapes in these regions are a mosaic of original, natural vegetation and human modified areas where secondary forest, cultivated land, and pasture dominate. Ecotype descriptions were adapted from Goodman et al. 2018 and Moat & Smith 2007 (27, 37). The SE region falls under the ecotype humid forest, characterized by degraded humid forest, wooded grassland-brushland mosaic, lowland moist evergreen forest, secondary forest, and secondary grasslands and experiences year-long, endemic malaria transmission. The SW region falls under the ecotype dry spiny forest and is characterized by dry spiny forest-thicket, degraded dry spiny forest-thicket, mangroves, secondary forest, and secondary grasslands and experiences endemic, but seasonal malaria transmission. The WC region falls under the ecotype dry deciduous forest and is characterized as dry forest, dry spiny forest-thicket, mangroves, secondary forest, and secondary grasslands and experiences endemic, but seasonal malaria transmission. Finally, the HP region ecotype is a wooded grassland-bushland mosaic, characterized by plateau grassland-bushland mosaic, secondary forest, and secondary grasslands and experiences seasonal, episodic malaria transmission. The SW and WC regions denoted in this study are both part of the Atsimo Andrefana region of Madagascar, but these regions are nearly 200km from each other and are characterized by distinct climatic and environmental profiles, thus providing us motivation to consider them as separate.
In this study, communities are defined as clusters of households (the typical rural settlements in these areas, tanàna or tanàna kely) and the adjacent peri-domicile areas. Sites are defined as communities and the surrounding mosaic of cultivated land, uncultivated human-modified land, and undisturbed land. In all communities, sampling was performed by randomly selecting ~50 households with two selection criteria: 1) a child five years of age or younger; and 2) a woman of reproductive age (15-49 years) (See (36) for more details on enrollment). Six communities were sampled in each region, three more proximate to an urban area (defined as <20km) and three more distant (defined as >20km) from an urban area. 33-53 households (168-304 individuals) were sampled per site and 242-309 households (1,461-1,665 individuals) were sampled per region, though only complete cases were used for the analysis.
Malaria outcome and associated socioeconomic and demographic variables
Health and survey data were collected at the individual and household level (36). In total, 10 of these variables were used (see Supplementary Table 1). One key variable included, hamlet ownership, is defined as a household owning a separate home away from the community proper that serves as a base for easier access to agricultural settings. Age of individuals included in the analysis was divided into four groups: under 5 years (reference category), 5-14 years, 15-64 years, and 65+ years. This breakdown was motivated by the literature (38-40) that often show distinct differences in malaria prevalence across these age groups. Many data sources specifically sample children under 5, because they often bear the highest malaria mortality (41). Therefore, we use this age group as reference to aid in comparison of estimates here with other studies. Rapid diagnostic tests (RDTs) were used to diagnose malaria (SD Bioline Malaria Ag P.f/Pan RDT). False negativity and the contribution of infections by other species such as Plasmodium vivax are known limitations of RDT based surveys (42), however, we justify using RDT positivity as a reliable proxy for malaria infection as previous studies using these RDTs in Madagascar found high agreement when comparing between RDT results and molecular detection via PCR (over 87% sensitivity and specificity) (42-44). Regarding Plasmodium species, over 96% of malaria cases in Madagascar are due to P. falciparum (41, 45) and in previous molecular confirmation of RDT positive cases from the east coast of Madagascar, over 98% were P. falciparum infections (43). A total of 776 individuals were positive by RDT for malaria, varying from 6-381 individuals per region.
Ecological variables and Anopheles larval sampling
To complement the data described above, ecological surveys were conducted in the 24 study communities to determine ecological risk factors for malaria transmission between the months of May 2017 – August 2017. Larval surveys were conducted during this period to sample vectors during, or as close to peak transmission season as possible, in all zones. Seasonal transmission dynamics exist in these regions, but we attempt to circumvent challenges presented by them through overlapping sampling times. The overlap between larval sampling and RDT testing in the SE and SW differed by up to five and up to three months, respectively. The peak transmission period in these regions has been found to be from January to May or June (20), so all larvae were collected during the peak transmission season for each region. Thus, we would not expect the mismatched sampling times to bias our results (For further seasonal analysis, see ESM2 Table 2)..
All potential and positive larval habitats were mapped with a Garmin Oregon 550t prior to larval collection. All habitats within a 25m radius of households or the community perimeter were geocoded, and mosquito larvae were sampled from each. We also conducted transect mapping to identify larval habitats and species composition of the local ecotype of each research site (all mosquito sampling methods described in detail in ESM2).
All larvae were sorted by genus and instar prior to identification by morphological examination. Anopheles gambiae is a complex of species represented by Anopheles gambiae sensu stricto, Anopheles arabiensis, and Anopheles merus in Madagascar. These three species are indistinguishable by morphology, and necessitate molecular identification (46). As such, larval specimens identified by morphology will be collectively referred to as in Anopheles gambiae complex. All third/fourth instar larvae were identified morphologically to the lowest possible taxonomic level using the taxonomic key presented in Grjebine (1966) at Institut Pasteur Madagascar (47). Some Anopheles larvae were unidentifiable to species because they were damaged in the sampling process which removed features necessary to distinguish species.
In Madagascar, aquatic forms of agriculture are used for farming predominantly rice, but this method is also used for growing other types of plants, and some spaces remain empty while still pooling water. As such, we define aquatic agriculture here as a combination of all forms of agriculture reliant on a flooded field. Aquatic agriculture cover was determined through ground truthing and remote sensing. Distance from the household waypoint to the edge of the nearest aquatic agriculture was calculated. For each household, we calculated the percentage area that was aquatic agriculture within a 1km buffer. Finally, we enumerated all possible mosquito breeding habitats within a 25m buffer around all households. All spatial variables were calculated using ArcMap and all satellite imagery was purchased from the AIRBUS Spot 6 and 7 Satellites at 1.5m spatial resolution (48).
Statistical Analyses
To assess ecological correlates of the odds of mosquito larvae presence in sampled habitats, we modelled five outcomes using simple logistic regression: (A) Presence/absence of Anopheles larvae, (B) Presence/absence of Anopheles gambiae complex larvae, (C) Presence/absence of Anopheles mascarensis larvae, (D) Presence/absence of Anopheles coustani larvae, and (E) Presence/absence of Anopheles squamosus larvae. For model (A), data were restricted to habitats containing any genus of mosquito larvae. For models (B)-(E), habitats included in the model were only those with Anopheles mosquito larvae present, which was a subset of all habitats. Predictors included in model (A) were the Normalized Difference Vegetation Index (NDVI, proxy of green vegetation cover calculated via remote sensing of satellite imagery; value calculations apply to the month in which larval sampling took place) at each habitat (49), habitat type (aquatic agriculture, ponds, containers), percent of forest cover in a 1km radius around the centroid of each site (10), soil moisture of each site (50), temperature of each site (mean monthly, (51)), and percent of aquatic agriculture in a 1km radius around the centroid of each site (52) (ESM 1 lists variables/sources/tests for correlation of variables). Variables included in models (B)-(E) were identical to model (A), with indicator variables for the presence of each Anopheles species. We analyzed these data using logistic mixed-effects models with random effects at the site level.
To assess the ecological, socioeconomic, and demographic correlates for the odds of malaria infection across the four study regions, we employed four multilevel logistic regression models, one for each region, with random effects for households nested within communities. Our outcome of interest was malaria infection status, as diagnosed by RDT. We employed separate models for each region due to the high variation in malaria prevalence (34)and distinct ecology of each region (27).
For the SE and WC regions we began with the same full model, and for SW and HP we began with slightly modified models. The difference in models was due to household hamlet ownership in SW and HP having none or limited heterogeneity. We excluded incomplete cases from analyses (n=214). To test for bias induced by complete case analysis, we assessed correlations between remaining individuals and the outcome of interest. The correlation estimates from these data were within the confidence bounds of the correlation estimates from the full dataset, and thus we conclude no significant change in estimates after exclusion of individuals with missing data. All statistical analyses were conducted in R v3.5.1 with the lmerTest package (53, 54).