Explaining variation in adult Anopheles indoor resting abundance: the relative effects of larval habitat proximity and insecticide-treated bed net use

The Asembo region of Rarieda District (part of Siaya County) in western Kenya (Fig. 1) is a rural community of about 60,000 people covering an area of about 200 km². Most of the residents are subsistence farmers, with the landscape largely dominated by small-scale agriculture. Small plots of land generally surround family-based groups of houses, which are further arranged into villages. While the houses are highly dispersed within villages, the boundaries between the 79 villages in Asembo are often discernable only by residents [27]. This is apparent in Fig. 1, which shows the roughly 10,500 households georeferenced within Asembo as of 2009 [28, 29]. Asembo sits in the lowlands along the shores of Lake Victoria, with elevations ranging from 1100 to 1400 m above sea level and low topographic relief. Rainfall is seasonally bimodal but local convective events may occur year round.

Malaria transmission occurs throughout the year, but seasonal peaks of transmission occur in May–July and October–November, tending to follow rainfall patterns with a short time lag. The predominant species of malaria is Plasmodium falciparum. Parasitaemia rates in children under 5 were around 40% in 2009 [30], although entomological inoculation rates have been estimated at less than 15 infectious bites per person per year since 2003 (MN Bayoh, unpublished data). Three primary malaria vector species are found in the region: Anopheles funestus, Anopheles gambiae s.s. and Anopheles arabiensis [31‐34]. Larval Anopheles habitats are numerous and widespread in Asembo, but they are heterogeneously distributed as patches of varying size and may surround a house from multiple directions [35, 36]. Household-level ownership of ITNs and LLINs has been relatively high in Asembo (i.e. over 80%) since it was the site of a randomized, controlled trial of ITNs from 1997 to 1999, followed by periodic distribution of nets through national programmes [27, 33].

Houses were sampled for indoor-resting adult Anopheles between 16 May and 24 June 2011 using the pyrethrum spray catch method (PSC) [19, 37]. This period was chosen to coincide with seasonal peaks of Anopheles populations, which occur after the March–May rainy season [31, 32, 38]. Weather data for the period of this study were downloaded from the National Climatic Data Center’s Global Summary of Day (GSoD) database, using records of temperature and precipitation from a weather station at the Kisumu Airport (about 40 km east of Asembo). For missing data (10 out of 211 days for precipitation, 2 out of 211 days for temperature), the inverse distance weighted mean of surrounding GSoD weather stations (within 250 km) was used.

All Anopheles collected during PSC sampling were morphologically identified according to Gillies and Coetzee [39]. Anopheles gambiae s.s. and An. arabiensis were identified by PCR [40]. The heads and thoraces of all female Anopheles were tested for P. falciparum circumsporozoite proteins by ELISA [41] using the P. falciparum sporozoite ELISA reagent kit (MRA-890, MR4, ATCC, Manassas, VA, USA). The blood meal hosts of all fed and half-gravid female Anopheles were identified by direct sequencing of the vertebrate mitochondrial cytochrome B gene [42].

Houses for PSC mosquito sampling were selected from within an 8 by 8 km study area (Fig. 1) to avoid edge effects when assessing the effect of larval habitat proximity (described below) on the number of adult Anopheles collected. Houses were selected for sampling using two-stage cluster sampling. First, the 8 by 8 km area was divided into 1 km² quadrats. Forty of these quadrats were randomly selected, and one house in each quadrat was randomly selected as the starting point for cluster sampling in that quadrat (i.e. 40 clusters of houses were sampled). Cluster sampling order was also randomly determined. On a given day of mosquito sampling, two teams of three field assistants each started at 0700 h by locating the selected house for their respective quadrat that day. After sampling at the selected house, the field team proceeded to sample at successive nearest-neighbour houses until 1000 h. In practice, the teams were able to sample from ten to twenty houses per cluster.

House-level variables that potentially influenced the number of Anopheles in the sampled houses were assessed at the time of mosquito sampling through a visual inspection of the house and via a standardized survey. The first of these was LLIN use. Ownership of LLINs was assessed through visual inspection, while use was self-reported during the survey. Specific numbers of LLIN users were counted in the surveys, and houses were later categorized into three groups of LLIN use: (a) houses where everyone who slept in the house the previous night used an LLIN; (b) houses where some residents had slept under an LLIN the previous night while other residents had not; (c) houses where no one had slept under an LLIN the previous night. Five additional house-level variables that potentially influenced the number of Anopheles in houses were included in the analyses based on findings from previous studies: presence of cattle near the house the previous night [16]; whether the inhabitants had cooked in the house the previous night [43, 44]; different wall types [15, 45, 46]; different roof types [15, 46]; and the number of people sleeping in the house the previous night [17, 45]. The status of the eaves (the gap between the top of the wall and the over-hanging roof) was also recorded as either open or closed [47, 48], but this variable was not included in any analyses, because fewer than 5% of sampled houses had closed eaves.

To characterize the relative juxtaposition of larval habitats to the houses sampled for adult mosquitoes, a predictive model of larval habitat locations was built as described in detail elsewhere [36], and model outputs (Fig. 2) were used to calculate several habitat proximity indices. First, the locations of larval Anopheles habitats were recorded in exhaustive ground surveys of thirty-one 500 by 500 m quadrats from 17 May to 4 July 2011. The surveyed quadrats were randomly selected, after spatial stratification, from a 10 by 10 km area within Asembo. Larval Anopheles habitats were defined as any standing body of water falling under the following habitat types: drainage channel, burrow pit, rain pool, runoff, cluster of hoof prints, stream bed pool, pond/reservoir, wet meadow, well and tire track [49]. The presence or absence of any Anopheles larvae was noted for each habitat, but all habitats were included in the final model regardless of whether Anopheles were present on the day of the ground survey. For a subset of habitats (up to 25 habitats per quadrat), Anopheles larvae and pupae were collected to confirm the presence of malaria vector species. All visible Anopheles larvae, up to a maximum of 20 per habitat, were collected using a 300 ml dipper or plastic pipette as appropriate according to the size of the habitat. The specimens were transported to the laboratory for species identification. Larvae were raised to fourth-stage instars for identification, while pupae were allowed to eclose as adults before identification. All identifications were done according to Gillies and Coetzee [39].

Using a topographic wetness index (TWI), land use/land cover (LULC), soil type, and distance to stream, the random forest statistical method [50] was applied to produce a landscape model of the probability of larval habitat presence at a 20 m resolution across the 10 by 10 km area [36]. A digital elevation model (DEM) of the study site was used to derive the TWI data. The 20 m resolution DEM was created using local universal kriging to interpolate 11,130 GPS elevation records previously taken within Asembo [28, 29]. The TWI data were then calculated using the ArcGIS extension TauDEM 5.0 (Tarboton, Utah State University). A satellite image from the IKONOS-2 sensor taken on 19 June 2007 was used to create the LULC classification at a 4 m resolution. Unsupervised classification was done using the K-means method [51] in ENVI 4.8 (Exelis Visual Information Solutions, Boulder, CO, USA). Classes were combined into a binary data layer of agricultural or non-agricultural land use. Soil data were taken from the 1:1,000,000 exploratory soil map of Kenya, compiled by the Kenya Soil Survey in 1980 [52]. All streams in Asembo were mapped using GPS units, and the Euclidean distance in metres to the nearest stream was calculated for each 20 by 20 m pixel of the 10 by 10 km area. The performance of the final model using these four variables was assessed by comparing the model output to holdout ground survey data, and calculating the area under the curve (AUC), sensitivity, specificity, percent correctly classified (PCC), and kappa.

This landscape model was used as a proxy for actual larval habitats locations in two ways (Fig. 2). First, the value at each 20 by 20 m pixel of the 10 by 10 km area was set equal to the probability of habitat presence as predicted by the model, and could take the value of any real number between 0 and 1. Second, the probabilities predicted by the model were converted to a binary value of either present (when P ≥ 0.020) or absent (when P < 0.020). Standard methods for converting modelled probabilities to a binary variable include using a threshold probability value (i.e. the value at which a probability equals presence rather than absence) that maximizes the agreement between observed locations (i.e. data collected for testing the model predictions) and locations predicted by the model, rather than simply using a fixed value (e.g. P = 0.5) [53]. In a previous study, a threshold value of P = 0.020 for the probability surface described above maximized this agreement [36].

Using these two outputs from the predictive model (probability of a habitat and presence/absence of a habitat) and the recorded geolocations of houses sampled for adult mosquitoes, the proximity of the houses to larval Anopheles habitats was quantified by calculating several proximity indices (Table 1). The first of these was the distance from the house to the nearest 20 by 20 m pixel predicted to have a habitat present in the binary output (i.e. distance to nearest habitat). Next, the number of pixels predicted to have a habitat (according to the binary output) within 50, 100, 300, 500 and 1000 m of the house was counted (i.e. number of habitats within n metres). Using the probability output, the mean value of all 20 by 20 m pixels within 50, 100, 300, 500 and 1000 m of the house was calculated (i.e. mean probability of a habitat within n metres). Both the minimum and maximum values for any one pixel within 50, 100, 300, 500 and 1000 m of the house were also extracted. Due to the high number of pixels with P = 0.000 across this landscape, all houses had at least one pixel with P = 0.000 within 300 m. Therefore, the minimum probability of a habitat within 300, 500, and 1000 m of a house was not used in further analysis. The range of distances over which to calculate the proximity indices was determined by assuming the average dispersal distance of Anopheles mosquitoes to be a few hundred metres while recognizing that it likely varies according to landscape patterns and human population density [1].

Regression with generalized linear models was used to quantify the association of the explanatory variables with the number of adult mosquitoes in houses, with separate analyses for each sex of An. arabiensis, An. gambiae s.s. and An. funestus. The explanatory variables were LLIN use, presence of cattle near the house the previous night, whether the inhabitants had cooked in the house the previous night, wall type, roof type, the number of people sleeping in the house the previous night, and a larval habitat index. As it was not clear a priori which of the larval habitat proximity indices described above would be associated with variation in adult Anopheles abundances, the larval habitat proximity index to use in each of the multivariate analyses was first determined by comparing univariate models with each of the habitat proximity indices to each other using Akaike information criteria (AIC). In the multivariate analyses, the single habitat proximity index from the univariate model with the lowest AIC for each respective sex and species was used. The presence of residual spatial correlation in the data was examined using the empirical variogram [54], on the Pearson’s residuals from a standard multivariate Poisson regression. In the final multivariate analyses, geostatistical Poisson regression with log-link functions was used to model the mosquito counts [55]. More specifically, the mean number of mosquitoes, \(\mu (x)\), at house location x was modelled aswhere: \(d(x)\) is a vector of explanatory variables; \(S(x)\) is a random effect that accounts for the spatial correlation between houses induced by unmeasured factors affecting mosquito abundance; and \(Z(x)\) is an unstructured random effect that accounts for extra-Poisson variation within houses. By including \(S(x)\) in the model, we explicitly do not assume independence among observations. In this case, \(S(x)\) was modelled as a Gaussian process with an isotropic Matern covariance function with variance parameter, σ ² , scale parameter, \(\varphi\) (the distance beyond which the spatial correlation is below 0.05), and shape parameter, \(\kappa\). Parameter estimation was done using Monte Carlo maximum likelihood, implemented in the PrevMap package [56] in the R software environment [57]. All maps presented were created using QGIS 2.14 [58]. All other figures, and all analyses, were implemented in R 3.3.2 [57].