Where have all the mosquito nets gone? Spatial modelling reveals mosquito net distributions across Tanzania do not target optimal Anopheles mosquito habitats

The study area included all of Tanzania, including the large islands in the Indian Ocean (Fig. 1). Tanzania occupies 886,100 km² [21] and its population exceeded 45 million people by 2012 [22]. Tanzania remains one of the poorest countries in the world, with a per capita average yearly income of $630 US as of 2013 [22]. Of all vector-borne diseases in the country, malaria still causes the highest morbidity and mortality rates [23]. Over 85 % of cases are caused by Plasmodium falciparum, the most dangerous malaria parasite [24].

The country has a tropical climate with wet and dry seasons that vary due to its topography. The northern and eastern regions experience two wet seasons, with short rains from October to December and long rains from March to May. The western, central, and southern regions mainly have one wet season that lasts from October to April or May. About 80 % of the population lives in rural arid regions [25] at very low population densities, with the remaining population in high-density urban centers, such as Dar es Salaam, which approaches 3133 people/km² [26]. Mosquito net coverage following massive ITN rollouts is not related to wealth but prioritizes pregnant women and families with children under the age of five [10].

The main malaria vectors in Tanzania are Anopheles arabiensis, Anopheles funestus, and Anopheles gambiae s.s. [27]. Over 400 published literature sources were reviewed for georeferenced occurrences of Anopheles mosquitoes in Tanzania. An observation for Anopheles species was recorded if collection localities were specified for any month and year of collection between 1934 and 2014. These data were supplemented by unpublished data collections (personal comm. Dr. Maureen Coetzee) and online database IR Mapper [28] and the Malaria Atlas Project database [29]. Each coordinate was verified using Google Earth (Version 7.1.2.2019, Google Inc.).

When complete, the database comprised 203 unique georeferenced observations of An. gambiae s.l. (35.5 %), An. arabiensis (32.4 %), An. funestus s.l. (16.2 %), and An. gambiae s.s. (15.9 %) across all years. Identification of mosquitoes within the An. gambiae s.l. complex relied on polymerase chain reaction or cytogenetics. The database details species, month and year of collection, citation source, village of collection, and trapping method.

Maxent software [30], v.3.3.3k, was used to predict the relative habitat suitability of Anopheles mosquitoes across Tanzania. Maxent combines presence-only species occurrence records with environmental data to create a model that predicts areas of relative habitat suitability for a given species. Maxent uses a background dataset that consists of pseudoabsences (i.e. a random sample of non-occurrences) instead of true recorded absences. The output estimates relative habitat suitability and not direct occurrence probability since areas where the species was not recorded are considered less suitable relative to where the species was actually recorded. Maxent niche models with high training accuracy have been constructed for arthropod vectors, including mosquitoes [4, 31‐33], ticks [34‐36], tsetse flies [37‐39] and sandflies [40‐42]. Despite Maxent’s tendency to be more conservative than other machine-learning techniques [43], it is one of the most reliable species distribution modelling methods, even with small numbers of species observations [44, 45]. However, internal training and testing accuracies evaluate the model’s fit to existing observations and external tests against independent measurements or temporal environmental changes provide stronger tests of model performance [46].

Since large-scale mosquito net distributions commenced in 2004, the niche model building was narrowed to the Anopheles coordinates collected between 1999 and 2003. Many historical records for anopheline mosquitoes in the 1999–2003 period do not distinguish species within the An. gambiae s.l. complex (e.g. only three geographically-unique points were recorded as An. gambiae s.s. compared to 54 recorded as An. gambiae s.l.). All recorded Anopheles species were combined to build the Maxent model. The focus of this study includes any of these malaria vectors, despite niche differences among species, relative to mosquito net ownership. All eight species within the An. gambiae s.l. complex can transmit malaria parasites except Anopheles amharicus, which does not occur in Tanzania, and Anopheles quadriannulatus, which has not been recorded in the country since 1968 [47]. The absence of spatially detailed data sufficient for these models prevented species-by-species evaluation of habitat suitability for the time period relevant to national ITN distribution. In total, 56 occurrence records were thus used to build the Maxent model, substantially exceeding minimum requirements (>5 observations) to produce informative predictions [44] (Fig. 2). Of these records, one was An. arabiensis, seven were An. funestus, 46 were An. gambiae s.l. and two were An. gambiae s.s., though more than one species was recorded at each location.

Multiple models were constructed using elevation, human population (Fig. 3), land cover, and variations of temperature and NDVI (average, minimum, and maximum). For assessment of model accuracy, occurrence records were randomly partitioned into 75 % for model training and 25 % for model testing. Ten replicates were run for each model, using tenfold cross-validation, each with a randomized partitioning of training and testing data, a technique that can reliably test spatial model skill for disease vector distributions [46]. The habitat suitability raster maps were then averaged to determine the relative probability of suitability per grid cell.

The model’s accuracy was determined using a threshold-dependent binomial omission test and a threshold-independent receiver operating characteristic analysis [30]. For the binomial test of omission, a fixed threshold for habitat suitability values of 0.1 was used, which is a threshold commonly set in other studies [4, 43]. For the threshold-independent analysis, model skill was evaluated using area under the receiver operating characteristic (AUC) [4, 48, 49]. AUC values approaching 1 indicate perfect discrimination between suitable and unsuitable areas for the target species, and values of 0.5 indicate performance no better than random. Each variable’s unique and shared contribution to model accuracy was evaluated using jackknife procedures [30]. Each variable’s relative contribution to overall model predictions was then calculated and presented as a percentage.

The histories of many mosquito-borne diseases, including malaria, indicate that climate is rarely the principal determinant of mosquito distributions in tropical regions and that shorter-term consequences of human activities often exert greater effects [50, 51]. Relevant climatic factors include land surface temperature and precipitation [4, 43]. Models incorporating satellite-based land cover and/or topography as indicators of mosquito habitats can pinpoint local variations in malaria risk [52]. Human population density is a critical determinant of relative malaria risk because of density-dependent transmission of parasites among individuals in the presence of anthropophilic Anopheles vectors [43].

Measurements of environmental conditions were used at 500–1000 m resolution across Tanzania derived using the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor on the Terra satellite. All 8-day Land Surface Temperature composites (LST; MOD11A2, 1000 m resolution), 12 monthly Normalized Difference Vegetation Index composites (NDVI; MOD11A3, 1000 m), and land cover observations (MCD12Q1, 500 m) were acquired for 2001. NDVI measures greenness and photosynthetic activity, which detects vegetation differences due to irrigation (such as rice-growing areas). Digital elevation data at 90-m resolution were obtained from the Shuttle Radar Topography Mission (SRTM; [53]) and mosaicked for the terrestrial area of Tanzania. Finally, 2001 human population distribution data were collected from the OakRidges National Laboratory LandScan database at 30 arc second (~1 km) resolution [54] (Fig. 3).

The MODIS data were unprojected from Sinusoidal to geographic coordinate system (GCS WGS 1984 datum) using the MODIS Reprojection Tool (MRT, v.4.0 from US Geological Survey). Human population and elevation layers were transformed to the same geographic coordinate system using ArcGIS v.10.1 (ESRI 2012, Redlands, CA). Some 8-day LST and NDVI mosaics pixels were flagged as missing or of low quality, usually due to atmospheric haze or cloud cover. Data were imported into R v.3.0.2 [55] and the raster brick function (package raster) was used to extract mean, minimum and maximum temperature and NDVI, omitting missing and low-quality pixels (see Additional file 1 for more details about the R script used for the raster brick function). Raster bricks for temperature and NDVI were then exported to ArcGIS 10.1 and all environmental layers were clipped to a land mask for Tanzania and converted to ASCII for use in Maxent.

This study used 2011–2012 data conducted by the AIDS Indicator Survey (AIS) throughout Tanzania, with permission from the Demographic and Health Surveys (DHS) Programme [56], funded primarily by the US Agency for International Development. AIS fieldwork was conducted between December 2011 and May 2012, implemented by Tanzania’s National Bureau of Statistics. This survey provided data on number of ITNs owned per household, including long-lasting insecticide nets (LLINs). All DHS procedures and questionnaires were reviewed and approved by the ICF International Institutional Review Board. Further details are provided in each survey’s final report.

To protect respondent confidentiality, the georeferenced cluster points (i.e. sample points that represented 17–18 surrounding households) were each displaced by the DHS, with shifts in latitude and longitude under set parameters [57]. Urban clusters were displaced 0–2 km and rural clusters were displaced 0–5 km, with 1 % (or every 100th point) displaced from 0 to 10 km. Shifts in direction and distance were both random. In addition, displacement of the georeferenced clusters was restricted at the district (admin2) level for the 2011–2012 AIS survey (Fig. 1). During the randomized cluster displacement process, clusters could cross localized administrative boundaries (e.g. ward boundaries), but not a DHS region or district boundary.

For the 2011–2012 AIS survey for Tanzania, 10,496 households were selected for sampling, of which 10,040 were successfully sampled and included in the final survey results. The survey’s GPS dataset comprised of 583 georeferenced survey cluster points, where each cluster point represented 17–18 surrounding households. Ten clusters were removed due to missing data. Of the remaining 573 clusters, 440 were in rural locations and 133 were in urban locations. Using this data set, mosquito net ownership was measured in two ways: first, the average number of mosquito nets owned per household and, second, the proportion of houses with ≥1 mosquito net (Fig. 4). While both offer distinct perspectives on mosquito net ownership, the latter is particularly relevant to the stated national target of 90 % of households with at least one mosquito net [13]. For each layer, the resulting calculated value from the surrounding households was assigned to their corresponding cluster coordinate.

Mosquito net ownership patterns were expected to be spatially autocorrelated and be used to estimate mosquito net ownership rates throughout unsurveyed areas of Tanzania. To test this possibility, extrapolations were attempted using kriging and inverse distance weighting techniques in ArcGIS. Predicted mosquito net ownership from extrapolated surfaces were compared to observed mosquito net ownership rates using tenfold cross validation, but model fit was very poor. Mosquito net ownership shows practically no spatial autocorrelation (measured using Moran’s I on cluster points = 0.07, p < 0.05, where Moran’s I scales as the Pearson product-moment correlation between 1 and −1, with 0 indicating no autocorrelation). Further analyses were restricted to point-based measurements.

Buffer zones were created surrounding each survey cluster, with urban clusters having a circular buffer zone with a radius of 2 km and rural clusters having a circular buffer zone with a radius of 5 km. Since cluster points were randomly relocated but did not cross district boundaries, buffer zones that crossed into an adjacent district could be made more spatially precise by clipping along the district boundary. Buffer zones were consequently retained only within the district particular to each survey locality. This processing resulted in a total of 912 mosquito net polygons. The area of the buffer zone was assigned the value attributed to its cluster point for the average number of mosquito nets and the proportion of houses with ≥1 ITN. For overlapping buffer areas, the averaged value of the individual buffers was calculated (see Additional file 2 for the R script used to create this layer).

To relate mosquito habitat suitability to mosquito net ownership, the habitat suitability map created in Maxent and the mosquito net buffer layers created in ArcGIS 10.1 were imported into R statistical software (v.3.0.2, [55]). Since each buffer zone had either a 2- or 5-km radius, and had been clipped depending on proximity to district boundaries, R was used to calculate average habitat suitability for each unique cluster and surrounding buffer (see Additional file 3 for the R script used for this process). Ordinary least squares (OLS) regression models were then constructed to test whether Anopheles habitat suitability predicted either average numbers of mosquito nets owned per household or proportions of households with ≥1 mosquito net. For each regression analysis, Levene’s tests were used to test whether variances differed between household clusters that were grouped by deciles according to predicted suitability for anopheline mosquitoes. Residual spatial autocorrelation and the associated Moran’s I were calculated using R. In addition, possible relationships were tested between the two measures of mosquito net ownership as a function of mosquito habitat suitability by performing quantile regression analyses in the 0.1 (the 10 % of household clusters with the lowest predicted suitability for mosquitoes) and 0.9 (the 10 % of household clusters with the highest predicted suitability for mosquitoes) deciles. Quantile regression is a method for providing a more complete view of possible causal relationships in ecological processes aside from the mean of the response variable distribution [58]. Using quantile regression to analyse the highest and lowest 10 % of the data may reveal changes in ITN coverage within these households as mosquito habitat suitability increases that could be undetected by common regression models through the mean of the response variable. If ITN ownership is consistently high, regardless of an area’s predicted suitability for anopheline mosquitoes, no trend is expected (slope = 0) among quantiles. Less favourably, ITN ownership rates could increase toward areas with high suitability for these vectors, suggesting that priority for ITN distribution had been given to localities with the greatest malaria risk. Finally, ITN ownership might decline toward areas with high anopheline suitability.

Elevation, human population distribution and land cover provided stronger predictions of mosquito habitat suitability than any temperature- or NDVI-based measurements (Table 1). Model fits were comparable whether the latter measurements were omitted (AUC = 0.872) or included (AUC = 0.862–0.875). Elevation made the greatest contribution to the model, followed by human population distribution and land cover. Jackknife analysis attributed the maximum training gain to elevation (0.828), which indicates this variable contains the most information that is not present in the other variables (Fig. 5). Concentrated areas of high habitat suitability are found in low elevation areas, mainly along eastern coastlines and extending into Kilombero Valley and south toward Mozambique (Fig. 6). Areas surrounding Lake Victoria and north of Lake Malawi also contain extensive areas with high habitat suitability for anopheline mosquitoes. Dense shrub land and cropland/natural vegetation mosaic land covers were associated with higher mosquito habitat suitability. Internal accuracy of the model through the threshold-dependent binomial omission test was also high, with an omission error rate of 7.14 % with a suitability threshold of 0.1 with the omission error rate remaining below 20 % up to a set threshold of 0.25 (n = 56). Removing elevation from Maxent models degraded model fit substantially (AUC = 0.765), increasing proportional contributions of human population distribution and land cover (65 and 25.1 %, respectively), with minor increases in temperature and NDVI (7.1 and 2.8 %, respectively). The final model excluding temperature and NDVI, but maintaining a high AUC value of 0.872, was retained.

The relationship between the average number of mosquito nets owned per house and mosquito habitat suitability was not significant (p > 0.05, Fig. 7a). Average mosquito net ownership per house did not significantly differ between rural (2.485 mosquito nets) and urban (2.491 mosquito nets) clusters (t = 0.13, p = 0.8967). Variance did not significantly differ among deciles of habitat suitability (F = 0.7821, p = 0.6331; Fig. 7b). There is no relationship between mean numbers of ITNs owned per household and predicted habitat suitability among households in areas among highest or lowest mosquito habitat suitability (the top or bottom deciles, or 10 %, of clusters, respectively) (p > 0.05; Fig. 7a).

There is a slight tendency for the proportion of households with at least one mosquito net to decrease toward areas with higher mosquito habitat suitability (R² = 0.051, p < 1 × 10⁻⁶; Fig. 7c). Overall, 93.41 % of surveyed households had at least one mosquito net (93.51 % in rural households and 93.05 % in urban households). Variance in ITN ownership rates across surveyed households with at least one mosquito net did differ significantly between deciles of habitat suitability (F = 3.0037, p = 0.0015; Fig. 7d). While the 0.9 quantile of the proportion of households with at least one mosquito net showed no significant trend relative to mosquito habitat suitability, the 0.1 quantile of the proportion of households with at least one mosquito net significantly decreased as mosquito habitat suitability increased (β = −0.0753, t = −3.38, p < 0.001; Fig. 7c).