The selection of DVS has been previously described 24 . Briefly, a series of authoritative review articles were consulted 25 26 27 28 29 , and those Anopheles vectors identified as main, dominant or principal were assembled to give an inclusive list of 52 species globally. This list was further refined in consultation with a Technical Advisory Group (TAG) of vector experts to give a final list of 41 species and species complexes worldwide, nine of which are found in the Americas (Table 1) 24 .

The Malaria Atlas Project (MAP) parasite rate data library was searched first 30 , augmented with a systematic search, conducted in January 2007, for published, peer-reviewed literature detailing primary anopheline vector occurrence data. Online scientific bibliographic databases (PubMed 31 and Web of Science 32 ) were searched for post 31 December 1984 articles using "Anopheles" as the keyword search term. In addition, relevant electronic archives were also searched, including AnoBase 33 , the Walter Reed Biosystematics Unit (WRBU) Mosquito Catalog 34 , the Disease Vectors Database 35 , archives of MalariaWorld 36 and Malaria in the News 37 , and a review of selected bibliographies 26 .

The resulting citation library was then reviewed and refined, retaining all references that met the following criteria for inclusion: (i) the reported study was undertaken after December 1984 (longitudinal surveys that began prior to but continued past this date were included); (ii) the surveys provided location information to a precision of administrative unit level one or higher; (iii) the surveys reported primary data; (iv) the surveys provided species-level information at the studied location; and (v) the surveys reported the presence of at least one DVS. Electronic mail alerts were then set up on various malaria information bulletins (Malaria in the News 37 , Malaria World 36 and the Malaria Bulletin 38 ) and by using "Anopheles" as a key word to generate article alerts via PubMed 31 . Content notifications for specific high yielding journals including Malaria Journal and Parasite and Vectors (via BioMed Central 39 ), Vector-Borne and Zoonotic Diseases 40 and the Indian Journal of Medical Research 41 were also set up. Results from these searches were included until 31 October 2009.

Globally, the literature search resulted in 3857 publications or reports containing potential data to be reviewed. Of these publications, 2276 fulfilled the inclusion criteria, providing data for 147 countries worldwide. A total of 366 sources detailed surveys conducted across 25 countries in the Americas that are further summarized in the results. These data were then augmented with species records from the WRBU MosquitoMap 42 .

Expert opinion (EO) maps were digitised from exhaustive searches of published distribution maps (Table 1). These were then refined by the TAG of Anopheles experts (see acknowledgements) during a meeting held in Oxford (23-25 September 2009). A compressed file containing these EO distributions in ArcGIS format is provided as Additional file 1: Expert opinion distribution maps for the nine DVS of the Americas.

The data were extracted into Microsoft Excel datasheets using a protocol provided elsewhere 24 . In brief, from each source article, the first author, date of publication, publication type (published article, thesis, unpublished report, etc.) were recorded. Vector- and survey-specific data abstracted included: survey location(s) including country- and administration-level information where given; whether the survey was conducted in an urban or rural location (as defined in the data source); the presence of forest or rice growing areas within the survey area (as defined in the data source); month, year and duration of the survey including information relating to multiple or longitudinal surveys where reported; species data, including sibling information where given; all vector sampling techniques employed in the survey (e.g. human landing indoors and outdoors, animal bait, resting indoors and outdoors); all identification methods (e.g. morphology, polymerase chain reaction (PCR) methods, cross-mating); and any control programs or methods currently in place in the survey location(s) (e.g. insecticide treated bednets (ITNs), insecticide residual spraying (IRS) or larvicide applications). No assumptions were made in the data abstraction, with all reported data accurately reflecting the level of detail given in the data source. For example, where a methodology described a series of monthly surveys conducted over a year, but the results were amalgamated and presented for the whole survey period, these data were abstracted reflecting the whole survey period. Monthly data were only recorded where survey results specifically reported the presence of the vector species in question during that month's survey.

The data then underwent a three-level checking procedure, with the first level check conducted by a different abstractor to ensure an independent assessment of the assembled data. All aspects of the data were reviewed to ensure the information had been correctly collated and the sites geo-referenced accurately. Detailed notes were included to describe how each line of data had been checked and where changes had been made. The second-level check (by MES) incorporated suggested changes with an emphasis on geo-referencing and the examining of evidence given in support of the choice of coordinates.

Once these first- and second-level checks were completed, the Excel sheets were migrated into a web-based PostgreSQL database 43 with a custom-designed Python interface (based on Django 44 ). The Python interface is important because it allows automated production of the occurrence data maps for review, and communication between the database and the PyMC statistical language ultimately used to produce the predicted ranges.

Third-level checks were implemented using the database to identify: (i) inconsistent or non-standard spellings within fields that would affect query summaries; (ii) blanks in mandatory data fields; (iii) any occurrence data that fell in the sea or other major water body (those falling in water bodies less than 1 × 1 km from the land were automatically adjusted, any points falling a distance greater than 1 × 1 km from the land were manually checked and corrected); (iv) any suspect geo-locations that fell outside the allocated country boundary; (v) all multi-point locations within a point or wide-area area type (e.g. where a survey listed three geo-referenced villages, but the data were reported across the whole survey area) in order to identify the point most central to the entire survey area.

Preliminary maps were produced, displaying the EO ranges and the occurrence data. Any points that fell outside the EO outline were checked and accepted as valid, or rejected, based on information provided by the TAG. The final occurrence data and EO maps were collated to show the extent of the geographical distribution of each species and to allow gaps in the data retrieved to be identified 24 . These are shown as thumbnails in the predictive maps in Additional file 2: Predictive species distribution maps for the nine DVS of the Americas.

A wide range of approaches have been developed for empirical modelling of species distributions given data on point observations of occurrence. In this study, the Boosted Regression Tree (BRT) method 45 46 was chosen to generate predictive maps of Anopheles species distribution. This selection was based on a number of factors: first, in a review of 16 species modelling methods, BRT was one of the top performing methods evaluated using the Area Under the receiver operating characteristic Curve (AUC) and correlation statistics 47 ; second, the method is flexible in being able to accommodate different types of predictor variables (e.g. continuous or categorical data); third, it is easy to understand and uses reliable, well documented and freely available R code; and fourth, the resulting maps are simple to interpret and include a ranked list of environmental predictors.

A full description of the BRT method is given in Elith et al. 45 . Briefly, BRTs combine regression or decision trees and "boosting". Regression trees use binary recursive partitioning to iteratively split the data into partitions. Put simply, the model uses the data (in this case presence and pseudo-absence (see below) of a mosquito species) and, in a series of steps, identifies the threshold of each input variable that results in either the presence or the absence of the species. It allows the input of continuous and categorical variables and different scales of measurement amongst the predicting variables.

Boosting is a machine-learning algorithm that increases a model's accuracy 45 46 48 . This is applied to the regression trees to help improve their predictive performance. Boosting is based on the principle that in machine learning, a set of "weak" rules can create a single "strong" or accurate rule, and it is easier to create many of these weak rules than it is to find a single accurate rule. For example, it is easy to generalise and state that, as a rule of thumb, a certain anopheline species is never found above a certain altitude, however, accurately specifying the exceptions to this rule across the species' range is considerably more difficult. In the case of BRT, a set of regression trees, each describing the data to a greater or lesser extent, are combined, or boosted, to create a single model with a higher predictive performance than each of its composing parts.

The boosting process within the BRT attempts to minimise a loss function, in this case deviance (two times the log of the ratio of the likelihood of the data, given the fitted model, and the likelihood of the data, given a perfectly fitting model) that indicates the poor performance of a weak model. Starting with the regression tree within the set that explains the most deviance, additional trees are added that best fit the residuals, or unexplained deviance, of the first tree. The residuals of this new two-tree model are recalculated and the next tree is added to fit these updated residuals, and so on. This stage-wise method could theoretically continue until the model is completely over-fitted, and predictive performance lost. To prevent such over-fitting, shrinkage (reducing the contribution of each tree) is applied to each new tree as it is added which involves the optimisation of the number of trees within the final model, the learning rate of the boosting (which is inversely related to the number of trees in the model), and the final tree complexity. This optimisation is achieved using cross-validation, where the model is tested using a sub-sample of withheld data, before the final model is run using all the data, applying the predetermined optimal settings.

Deviance (see above), Correlation, Discrimination (AUC) and Kappa (κ) summary statistics were calculated to accompany each map combining classic accuracy metrics (e.g. Kappa) and those more specific to, and calculated by, the BRT method and recommended for use in Elith et al. 45 (e.g. Deviance). These statistics are used here as a guide to the predictive performance of the maps, but not to compare one map to another, as they were each generated using different datasets.

During cross-validation, BRT repeatedly partitions the data into training and testing subsets and performs cross-validation on each of the different partitions (or folds). It then calculates the mean and standard error for each statistic. The deviance metric is the prime measure of accuracy for the BRT method, and indicates the proportion of deviance still unexplained by the model. The lower the deviance value (minimum value = 0), the better the model is at predicting the hold-out sets. Correlation simply correlates the predicted probability of presence against the occurrence data, giving a value between -1 and 1. The area under the curve (AUC) of the receiver-operator characteristic is also presented as an accuracy metric. The AUC provides values ranging between 0 and 1, where 1 is a perfect prediction, and is an index of the area under the curve of a plot of sensitivity (the proportion of the testing set that is correctly classified as a presence) against specificity (the proportion that is correctly classified as an absence) as the classification threshold is increased 49 . Kappa is an index of the proportion of agreement of predicted versus observed positive and negative samples, calculated from an error matrix that cross references the number of observed and the number of predicted pixels categorised as presence or absence 49 50 51 . It can provide values ranging from -1 to 1, where values less than zero indicate that the model is worse than random, zero shows that the model is no better than random, and greater than zero indicates that the model performs better than random.

Each of the environmental or climatic data grids described below have undergone a number of processing steps prior to being used in the BRT. Each grid was post-processed to ensure that the pattern of land and sea pixels corresponded exactly across all grids, in order to provide a suite of spatially identical grids. This required trimming or gap-filling via nearest-neighbour interpolation to reconcile minor misalignments in coastline definitions or, for example, small data gaps due to cloud error.

Covariates were chosen based on factors known to influence anopheline ecology and therefore include elevation 52 , climatology surfaces interpolated from networks of meteorological stations 53 , and remotely sensed data from Earth observation satellites in their raw form and categorised into global land cover maps 54 55 . Where remotely sensed imagery was available as multi-temporal data, temporal Fourier analysis (TFA) was used to ordinate the data by decomposing the temporal signal into an additive series of harmonics of different seasonal frequencies 56 . The TFA algorithm 56 generated seven products for each temporal variable: the overall mean, maximum and minimum of the data cycles; the amplitude (maximum variation of the cycle around the mean) and the phase (the timing of the cycle) of the annual and bi-annual cycles. The grids are all described in detail below.

The Shuttle Radar Topography Mission (SRTM), consisting of a synthetic aperture radar on board the Space Shuttle Endeavour, flew an 11 day mission to create a near-global high spatial resolution digital elevation model (DEM) 52 57 . The DEM grids used here are processed by applying land/sea masks and the data are projected into the MODIS Land (MODLAND) tile system. The MODerate Resolution Imaging Spectroradiometer (MODIS) is a NASA satellite sensor which provides 1 × 1 km spatial resolution satellite imagery globally 58 .

The Worldclim database consists of a freely available set of global climate data at a 1 × 1 km spatial resolution compiled using weather data collected from world-wide weather stations 53 . The data spans from 1950-2000 and describes monthly precipitation and mean, minimum and maximum temperatures. From these data, interpolated climate surfaces have been produced using ANUSPLIN-SPLINA software 59 .

The Advanced Very High Resolution Radiometer (AVHRR) 8 × 8 km products are available over a 20-year time series, and a limited series of 1 × 1 km resolution data are only available for April to December 1992; January to September 1993; February to December 1995 and January to April 1996. Both data series were downloaded (Goddard Space Flight Center's Distributed Active Archive Center on the Global Land Biosphere Data and Information Web Site 54 ) and processed for use here 60 .

The AVHRR grids used include the normalized difference vegetation index (NDVI), land surface temperature (LST) and middle infrared radiation (MIR). The NDVI numerically indicates the level of green, photosynthesizing, and therefore active, vegetation derived from the spectral reflectance of AVHRR channels 1 and 2 (visible red and near infrared wavelength, respectively) 61 62 . The LST index uses thermal infrared radiometry to measure land temperature, corrected for atmospheric influences, such as water vapour, aerosols, carbon dioxide or ozone 61 63 . Finally, the MIR data is applied to discriminating land cover. It is able to penetrate more fully than shorter wavelengths through aerosol particles, including atmospheric water, and it is considered better able to distinguish between vegetation, soil, rock and water 61 64 .

The Globcover project 55 provides satellite-derived land cover maps from the MERIS spectrometer on board ENVISAT. The data produced are at a 300 × 300 m resolution and the satellite imagery goes through a number of pre-processing and classification steps prior to map production, which include cloud screening and shadow detection, water reclassification and atmosphere (including aerosol) correction. The final map products include global land cover mosaics for the period from December 2004 to June 2006, providing 22 land cover classes; and regional mosaics, which detail up to 51 land cover classes. Due to the limited areas covered by the regional mosaics, the global mosaics are used here. They include land-cover classes particularly relevant to mosquito habitats, for example, post-flooding or irrigated croplands, rain-fed croplands, urban areas and numerous categories for forests, including those regularly flooded with fresh or saline/brackish water. The 22 categories were available to the BRT individually and also grouped into three land cover types: flooded areas, forested areas and dry areas. Finally, to produce grids of an equivalent resolution to the other environmental and climatic variables applied in the vector mapping (5 × 5 km), the 300 m mosaics were re-sampled using a majority filter where the most common class in each pixel subset was identified and used to redefine the new, larger pixel.

The covariate sources detailed above provide a range of spatial, high resolution environmental or climatic open source data freely available for species mapping. All grids were available to the BRT mapping process allowing the model to identify which variable, or suite of variables most accurately described the species distribution.

Numerous model iterations were run with varying combinations of buffer size, number of pseudo-absences and source data (hybrid data: occurrence data plus pseudo-presences randomly assigned from within the EO species range; EO data: pseudo-presences randomly assigned from within the EO range; and occurrence data only). As each of these categories required the use of different data inputs to the BRT, statistical comparison using the evaluation metrics was not possible. Therefore the "optimal" settings chosen are subjective and based on visual examination and comparison of the various maps guided by, but not relying on, the evaluation statistics.

The lack of true absences in the occurrence data is a common problem encountered with species mapping and can be addressed by the production of pseudo-absences. These can be created in a number of ways 65 including creating background samples within the region of study 47 66 to represent the set of conditions available to the species, though not necessarily areas where the species is genuinely absent. Alternatively, and to minimise the risk of including unidentified presence data, pseudo-absences can be assigned within a buffer zone surrounding the region of study 49 67 . This second method is particularly useful where the extent of the species range to be mapped is known or can be reasonably estimated, for example with expert opinion maps. The buffer in such cases represents an area that is not geographically implausible for the species to reach, but maybe outside the limit of the environment that the species can tolerate. The creation of pseudo-absences in the buffer zone was used here.

The size of the buffer zone can have implications for the accuracy and viability of the final mapping outputs 67 . A buffer that is too large, allowing pseudo-absences to be generated far from the presence data, may result in a model that is defining regional or coarse geographical differences rather than the fine-scale variables that influence the presence or absence of a species. Conversely, a buffer that is too small, which relies on pseudo-absences generated too close to the presence data, risks incorporating unknown areas of presence, severely limiting the ability of the model to distinguish presence or absence. Both scenarios result in poor model performance and variable selection, and hence poor predictive ability 67 . To identify the optimal buffer size, maps were run for all species with buffer sizes of 100 km, 500 km, 1000 km and 1500 km, using 1000 pseudo-absences (see below). These maps were then visually and statistically compared.

The number of pseudo-absences generated in the mapping process can also influence the final mapping product. As a rule of thumb, a proportionally greater number of pseudo-absences to presence data should be used to overcome the possibility that within the randomly selected absences, some may fall on true, yet currently unidentified, presence points 49 . To establish the optimal number of pseudo-absences to use in the mapping, maps for all nine DVS were created and compared using 1:1, 2:1, 5:1 and 10:1 ratios of pseudo-absences to presence data, and additionally using a constant number of 500 and 1000 pseudo-absences for all species. These maps were then visually and statistically compared.

The predictive value of the expert opinion maps were tested using two methods: (i) for each species, 500 pseudo-presences, randomly distributed within the expert opinion species range boundary, were created and run through the BRT mapping alongside 500 pseudo-absences from within the buffer area. Using these pseudo-data, maps were created to identify whether the boundaries given in the expert opinion maps would be recreated by the BRT, such as, for example, the dry area indicated in the north-east of Brazil, identified as too dry for the forest/riverine species An. darlingi to survive. (ii) For each species, maps were created using the real occurrence data and, firstly, 500 pseudo-absences randomly selected from the buffer zone and of the same weighting as the true data, and, secondly 500 pseudo-presence points randomly assigned within the expert opinion range but with a weighting half that of the true data plus 500 pseudo-absences created as before. Maps created using both methods were again compared.

All mapping trails were conducted using environmental/climatic variable grids of 5 × 5 km resolution, allowing the production of high quality maps within feasible periods of time.

The need to understand the behaviour and life history characteristics or "bionomics" of the DVS is clear. The impact of interventions, such as ITNs, and control, such as IRS, is closely related to the behavioural characteristics of the local DVS. In addition, such information can help rank the relative importance of a species in an area with regards to transmission and be of major importance to the parameterisation of mathematical models. No recent comprehensive summary of these information exist.

Bionomics information was summarised for each of the nine American DVS (See Additional file 3: Bionomics protocol, for full methodology). A sub-library for each species was created by searching the citations listed within the MAP library by species name. Where a library contained 30 or more citations, it was filtered using the following terms: "behaviour", "behavior", "larva", "biting", "resting" and "habitat", in order to refine the bibliography to include only studies detailing behavioural/bionomic information (Table 2).

Each of the articles listed in each sub-library was read, and all relevant behavioural information extracted and summarised (Additional file 3: Bionomics protocol). The TAG were consulted and also provided a summary of their knowledge of the bionomics of each DVS which was combined with the relevant literature summary to detail general larval site characteristics (Table 3), known larval habitat types (Tables 4-5), and adult biting and resting behaviour (Table 6).

The bionomics review does not include any detailed information relating to insecticide resistance. This was purposely omitted as this highly dynamic and important aspect of the vectors is being addressed in detail by other groups including the Innovative Vector Control Consortium (IVCC) 68 and groups at the Liverpool School of Tropical Medicine.

Data were collected in 25 countries in the Americas (Table 7) from which 5408/5942 occurrence points and 2889/3257 sites have been geo-referenced (summary numbers generated in April 2010). "Occurrence" includes all temporal data, where given, and "sites" refers to unique locations. For example, if a study reports the presence of a species every month for a year, this gives 12 incidences of occurrence, but only one geo-referenced site. Data were categorised into the area types that define the size of the area sampled 30 . There were 1509 point locations (≤10 km²), 88 wide areas (10-25 km²), 42 small polygons (25-100 km²) and 369 large polygons (>100 km²) (Table 7) in the Americas.

Adult collections were reported from 1203 locations, larval collections from 633 and a combination of larval and adult collections from 377 sites. A total of 595 sites reported the presence of more than one DVS, with a maximum of five co-existing species found at San Pedro de Uraba, Colombia. The most popular sampling method of all the reported studies was outdoor landing catches using human bait (MBO). This method was used in 110 studies and at 411 sites. Across the Americas, human landing (indoors and outdoors) sampling techniques were used in 39.3% of studies and at 24.8% of sites compared to the 2.7% of studies and only 1.9% of sites that used animal bait collections. Similar numbers of studies searched for resting mosquitoes inside and outside houses (5.1% resting indoors and 5.2% resting outdoors). However, a greater number of sites were searched for outdoor resting mosquitoes in these studies (16.4% compared to 5.1% resting indoors), although this is a consequence of two large studies that searched for An. quadrimaculatus in the USA: Reinert et al. 69 , 235 sites searched; Seawright et al. 70 , 148 sites searched. Light trapping or carbon dioxide bait was only used at 4.4% of sites. Morphological identification methods were used at 1102 sites, with PCR techniques applied to identify samples from 178 sites. Longitudinal data were recorded where given, and within these data, 139 studies reported DVS occurrence sampled for over a year, relating to 431 sites. A single sampling period was reported from 85 studies, relating to 1235 point-location sites, whilst 139 studies reported multiple (two or more) sampling periods from 975 point-locations.

Anopheles darlingi was reported from the greatest number of geo-positioned point locations (488) and had the greatest number of recorded temporal occurrences (859) (Table 8). In comparison, An. freeborni was only reported from 31 site point locations, with only 35 overall occurrence records. The greatest number (377) of point-location sites was recorded from the United States. Guyana had the lowest reported data overall, with only a single point-location reported at the gold-mining town of Mahdia in the Amazon interior, where both An. aquasalis and An. darlingi were recorded along with a number of other secondary vector species.

All results for each mapping trial are given in Additional file 4: Summary tables showing evaluation statistics for all mapping trials and final BRT environmental and climatic variable selections for the final, optimal predictive maps. Optimal mapping categories were evaluated visually and using the deviance and AUC statistics, with the caveat that these could only be used as a guide rather than a true indication of predictive performance, as different data sets were used for each map produced.

The optimal buffer size within which pseudo-absences were randomly selected was determined to be 1000 km, with five out of nine species maps judged to perform better at this buffer size than those produced using the alternative sizes. A ratio of 5:1 pseudo-absences to presence data was applied, taking into account 250 (500 half weighted) pseudo-presences. This level of pseudo-absence input provided enough data to counteract any potential interference that may have been caused by the random inclusion of unidentified presences from within the absence data.

The EO mapping test indicated that where random pseudo-presences are created within the EO range, and no real occurrence data were included, the model would predict a high probability of presence within the whole expert opinion range and calculate a high deviance value for all species, indicating overall, a poor predictive performance. Where the hybrid method was used that incorporated both real occurrence data plus 500 half-weighted pseudo-presence points randomly assigned within the EO range, the mapping performance was greatly improved. Maps created using only the real presence data, produced a low deviance value, but visually, predictive performance was judged to be poor, possibly due to a paucity of data for some species. As with all species mapping, occurrence data will be biased to areas that are accessible for sampling, or as succinctly stated by Coetzee 71 , reflect the distribution of entomologists rather than mosquito species. As such, all occurrence-based mapping and mapping evaluations must be treated with some level of caution, especially where the data is sparse. It was thus considered that the hybrid maps performed better overall and are presented here.

The BRT species maps for all nine American species are given in Additional file 2: Predictive species distribution maps for the nine DVS of the Americas, and are summarised in Table 9. Spatial constraints prevent all the species being discussed in full here, however Anopheles darlingi, considered one of the most important malaria vectors in the Neotropical Region 23 , is reviewed below.

Anopheles darlingi is a lowland, riverine, forest dwelling species (see below and Table 4), unable to survive in dry climates, including for example, north-eastern Brazil (Rio Grande de Norte, Paraiba etc.) as is approximated in the EO map (Figure 1, inset). Figure 1 shows the predicted species distribution using hybrid data (318 occurrence data plus 500 pseudo-presences weighted at half that of the occurrence data and randomly selected from within the EO range) and a 5:1 ratio of pseudo-absence to presence data, taking into account the pseudo-presence points. Also indicated are the environmental variables identified by the BRT as being the most influential amongst those applied to the model in describing the species distribution.

Maximum precipitation is the variable identified as having the highest relative influence (36.31%) on the presence of An. darlingi compared to 9.04% for the second-rated variable (maximum LST) (Figure 1; Table 9). Precipitation is selected twice more within the top five (third: mean; and fifth: phase of the bi-annual cycle) and LST was selected once more (fourth: phase of the bi-annual cycle). Aside from precipitation and LST, the only other two variables selected within the top ten list were elevation (DEM) at seventh place with a relative influence of only 3.61% and MIR (mean), listed tenth and with a relative influence of 3.04%. Land use appears to have little effect, with Globcover channel 160 (closed to open (>15%) broadleaved forest regularly flooded (semi-permanently or temporarily), fresh or brackish water) only listed as the thirteenth variable with a relative influence of 1.87%.

The variables selected by the BRT correspond well to the bionomics of Anopheles darlingi. Larvae of this species are most often found in slow flowing rivers, associated with floating debris 72 73 74 . Rain will impact on river current and height, potentially flushing away the larvae and the river debris they inhabit or causing them to become stranded once the rain lessens or stops and the river recedes 73 75 . Land Surface Temperature does not have such a clear cut impact on An. darlingi and could be influencing its distribution in a number of ways. For example, it may be indicating the presence of hotter, dryer environments away from the forests, such as those that limit the distribution of An. darlingi in the north east of its range. Alternatively, LST may be identifying the higher temperatures associated with its lowland range, although there is recent evidence that this species is able to survive at higher altitudes than previously suspected (see below).

This study has focused on describing the distribution and bionomics of the main malaria vectors in the Americas. However, there are numerous secondary or local vectors that may also play an important yet often forgotten role in malaria transmission. In the Americas, secondary vectors include the various species of the subgenus Kerteszia (e.g. An. cruzii, An. bellator and An. neivai), whose larvae characteristically inhabit water contained in bromeliads, as well as An. (Anopheles) vestitipennis, An. (Ano.) neomaculipalpus, An. (Nyssorhynchus) braziliensis, An. (Nys.) triannulatus, An. (Nys.) strodei, An. (Ano.) intermedius and members of the An. (Nys.) oswaldoi complex. Their "secondary" vector status is a factor of location, distribution, vectorial capacity and, occasionally, history (i.e. a species once considered primary, but now relegated to a secondary role through changes in the local environment). Circumstantial evidence, or the identification of Plasmodium circumsporozoite proteins in females, can lead to a species being incriminated as a vector, where in reality it may have little to no impact 23 . However, some "secondary" species are proven and potent vectors within their local range.

Both An. cruzii and An. bellator are identified by White 29 as main malaria vectors, yet their larval habitats (bromeliads) restrict their role in malaria transmission to areas where such plants are abundant, for example in the rainforests of Brazil where these two species are considered to be primary local vectors. Deforestation may have reduced the availability of habitats for these species but they ought not to be overlooked as their range extends along the eastern coast of South America, from Guyana to the southernmost tip of Brazil. Moreover, there are reports of these species being found in bromeliads and artificial containers in urban and peri-urban sites 175 176 with An. cruzii specifically described as an aggressive biter throughout the day and night 177 . Anopheles neivai also makes use of bromeliads during its larval stage and is an important vector of human malaria in the Pacific coastal areas of Colombia 178 179 . It has been responsible for outbreaks of malaria in the Venezuelan Andes at altitudes above 1000 m 180 181 .

Anopheles vestitipennis is considered to be a secondary vector species of major importance within its range. In Belize, it has been described as a primary vector, ousting An. albimanus which plays a more secondary role there 182 183 184 185 . Achee et al. 183 found it to be positive for both P. falciparum and P. vivax and Loyola et al. 186 found it to be the most abundant species (>80% of those collected on human bait) in their study area within the Lacandon rainforest of Chiapas, Mexico, and the only species in the area to be positive for the P. vivax antigen. They also demonstrated that it readily bites humans both indoors and outdoors. Anopheles vestitipennis is also found in the Caribbean islands where it may have been involved in malaria outbreaks in Cuba 187 and Haiti (Rubio-Palis, unpub. obs.).

Malaria parasites have also been detected in several other species. De Oliveira Ferreira et al. 188 found natural infections of Plasmodium in An. triannulatus, An. braziliensis, An. strodei and An. oswaldoi in Rondonia, Brazil. Anopheles triannulatus and An. strodei were only infected with P. vivax, however the very small sample sizes (the largest being five specimens of An. triannulatus), prevents the drawing of any conclusions regarding any refractory characteristics.

Anopheles oswaldoi has been confirmed as a malaria vector in Brazil 119 188 189 190 , and it is the principal vector in the State of Acre in the Brazilian Amazon where it has been found in large densities and with relatively high sporozoite rates for P. falciparum (3.41%), P. vivax-210 (2.26%), P. vivax-247 (1.22%) and P. malariae (0.42%) 189 . Anopheles oswaldoi has also been incriminated as a vector of P. vivax in Colombia 191 , Peru 192 and Venezuela 193 .

The role of An. neomaculipalpus in malaria transmission is not clear but it has recently been found to be positive for P. vivax in Venezuela 194 195 and positive for P. falciparum in Colombia 195 196 and it is considered to be a highly anthropophilic species ( 195 ). Moreno et al. 195 suggested that this species may play a role as a secondary vector of "frontier malaria" in areas of forest subject to recent human activity that can increase vector diversity with the creation of new larval sites. De Oliviera Ferreira et al. 188 reiterated this hypothesis and suggested that the presence of infection amongst these secondary vectors may be linked to the extensive environmental changes that may reduce the populations of sylvatic animals in an area whilst simultaneously increasing the number of humans, which will inevitably cause a higher level of vector-human contact.

The predictive maps presented here are unlikely to be perfect representations of the full distributions of the DVS. The model has been applied to species presence data from one of the largest, most comprehensive, contemporary databases of DVS occurrence available, yet the limitations of these opportunistic data for species mapping are evident. Biases in collection location, limited data for some species and a lack of consistency in sampling methodology all contribute to modelling and output uncertainty. Despite these caveats, the maps represent the first attempt to model DVS distributions in the Americas using extensive occurrence data, contemporary EO range maps and a consistent methodology across all species. Previous maps have either focussed on a single species or subgroup 19 , single countries 20 21 , been based solely on EO 22 23 , or, specimens collected over a large period of time 18 . The maps are best considered a starting point in a continuing process of describing and understanding DVS distributions, and not as an end product. Therefore, and in accordance with the open access principles of the MAP, all the data compiled here will be made freely available on the MAP website 197 and efforts made to improve the maps as more data become available.

The bionomics review highlights DVS behaviour and life-history characteristics that are relevant for mosquito control, but also clearly indicates the marked behavioural plasticity of each species. The influence of human behaviour such as insecticide use, environmental disturbance to a greater or lesser extent, or host activities in the evening and night also drive local variation in species bionomics. Moreover, concerns regarding species identity also add to the uncertainty in categorising species behaviour and thus local, expert knowledge must be consulted when interpreting or acting on the data summarised here.

This is the first in a series of three publications describing the distribution and relevant bionomics of the global DVS of P. falciparum. The remaining two publications will detail the DVS of Africa, Europe and the Middle-East (Sinka et al: The dominant Anopheles vectors of human malaria in Africa, Europe and the Middle East: occurrence data, distribution maps and bionomic précis, unpublished), and the DVS of the Asian Pacific region (Sinka et al: The dominant Anopheles vectors of human malaria in the Asia Pacific region: occurrence data, distribution maps and bionomic précis, unpublished). Together, these three publications are intended to provide a baseline set of data and maps and summarise the current knowledge of the bionomics of the 41 species (DVS) identified as the primary vectors of P. falciparum and P. vivax malaria.