Summary 1. How the environment influences the transmission and maintenance of disease in a population of hosts is a key aspect of disease ecology. The role that environmental factors play in host-pathogen systems has been well studied at large scales, i.e. differences in pathogen pressures among separate populations of hosts, or across land masses. However, despite considerable understanding of how environmental conditions vary at fine spatial scales, the effect of these parameters on host-pathogen dynamics at such scales has been largely overlooked.
2. Here we used a combination of molecular screening and GIS-based analysis to investigate how environmental factors determine the distribution of malaria across the landscape in a population of Berthelot’s pipit (Anthus berthelotii) in Tenerife.
3. Anthropogenic factors were better predictors of malaria distribution than natural factors in this population, with proximity to water reservoirs and poultry farms being the key predictors of infection. Contrary to predictions, climatic differences across the landscape did not appear to influence the distribution of malaria. This may reflect the scale at which the study was performed.
4. These results suggest that levels of malaria infection in this endemic species are artificially elevated by the impact of humans.
5. Studies such as the one described here improve our understanding of how environmental factors, and their heterogeneity, affect the distribution of pathogens within wild populations. The results demonstrate the importance of measuring fine scale variation, and not just regional effects, in order to understand how environmental variation can influence wildlife diseases. Such understanding is important for predicting the future spread and impact of disease and may help inform disease management programmes, as well as the conservation of specific host species.
Key words: Avian malaria, Berthelot’s pipit, environmental predictors, GLM, model selection
Introduction Understanding how ecological variables influence the prevalence and transmission of disease in a population is a key issue in ecology (Hudson 2002). This is especially important at a time when climate and anthropogenic habitat changes are dramatically affecting the distribution of pathogens and their hosts (Harvell et al. 2002; Lafferty & Gerber 2002; Garamszegi 2011), and the number of emerging infectious diseases of wildlife is increasing (Daszak 2000). Establishing how biotic and abiotic factors influence the distribution of disease is important if we are to predict the spatial patterns of future disease threats and effectively manage their impact on biodiversity (Murray et al. 2011). Some recent studies have assessed how environmental variables affect pathogen distribution across landscapes in wild populations (e.g. Kleindorfer & Dudaniec 2006; Wood et al. 2007; Murray et al. 2011) and identified relationships have been used to predict where pathogens are likely to survive and/or disease outbreaks to arise (Ron 2005; Lerdthusnee et al. 2008; Murray et al. 2011). However, such studies have normally been done at coarse scales (but see Eisen & Wright 2001 and Wood et al. 2007 for exceptions), thus the effects of fine scale environment variation on the distribution of pathogens has been largely overlooked. Studies at finer spatial scales are now required to provide insight into why pathogens typically have patchy distributions within a landscape (Eisen & Wright 2001; Salje et al. 2012).
Haemosporidian parasites, such as species of the genera Plasmodium, Heamoproteous and Leucocytozoon (hereafter termed malaria for simplicity), are intracellular protozoan blood parasites transmitted by blood sucking invertebrates that occur in every continent apart from Antarctica (Valkiunas 2005). In humans, malaria is a major public health problem with more than 200 million cases and more than one million deaths each year (Chuang & Richie 2012). Malaria also infects other vertebrates including reptiles, turtles, birds and other mammals, reducing their survival and fitness (Martinsen, Perkins & Schall 2008).
The vector-transmitted nature of malaria and the need for specific conditions to complete parasite development means that transmission, and therefore patterns of infection, are highly dependent on environmental factors (Guthmann et al. 2002). A number of key abiotic variables can affect the distribution of malaria (Lapointe, Atkinson & Samuel 2012). Temperature has a considerable effect (Xiao et al. 2010; Sehgal et al. 2011) because sub-optimal temperatures are associated with slower parasite development (Valkiunas 2005; LaPointe, Goff & Atkinson 2010) and reductions in vector density (Gilioli & Mariani 2011). Elevation influences malaria through its negative correlation with temperature (Balls et al. 2004; LaPointe, Goff & Atkinson 2010; Lapointe, Atkinson & Samuel 2012). Water availability is important because of the role it plays in vector larval development (Gilioli & Mariani 2011) and rainfall has been shown to be correlated with malaria prevalence (Briet, Vounatsou & Amerasinghe 2008; Galardo et al. 2009; Zhou et al. 2010), as has distance to bodies of water independent of rainfall (Wood et al. 2007; Haque et al. 2009). In areas where precipitation is low, topography may also play an important role in water accumulation and therefore vector abundance (Balls et al. 2004; Clennon et al. 2010). Biotic variables, including host life-history traits and demographic factors, may also affect the distribution and transmission of malaria (Ortego & Cordero 2010; Gilioli & Mariani 2011). Malaria prevalence has been shown to be positively associated with host density (Ortego & Cordero 2010), but negatively correlated with non-host species density because of the dilution effect these have on the transmission of malaria (Mutero et al. 2004; Nah, Kim & Lee 2010). Additionally, the presence of other diseases could make hosts more susceptible to malaria by modifying behavior or by affecting host immune response (Van Riper et al. 1986; Jarvi et al. 2008).
The factors described above all relate to ‘natural’ environmental variables, but anthropogenic activities such as animal husbandry and urbanisation can also affect the prevalence of malaria (Patz et al. 2000). One such factor that has been little explored is the influence of livestock on the transmission of malaria. If livestock serve as a reservoir of malaria, their presence might increase prevalence within taxonomically similar local fauna - though this will depend on the host-specificity of the malaria strains involved (Yotoko & Elisei 2006; Beadell et al. 2009). On the other hand, livestock could dilute malaria transmission by providing non-host blood meals to vectors (Nah, Kim & Lee 2010). At another level, livestock farms could increase prevalence if they alter the local environment, for example by providing water reservoirs suitable for vector development or by facilitating transmission through aggregations of wild animals attracted to the farms. Urbanisation may also have an impact on malaria prevalence, contingent on the degree of urban adaptation of host species (Bradley & Altizer 2007), and on the extent to which such areas provide suitable habitats for vector development (Guthmann et al. 2002; Omumbo et al. 2005; Antonio-Nkondjio et al. 2011).
Assessing and comparing the roles of the environmental variables potentially influencing malaria prevalence is challenging, and their importance may vary between different vector species, strains of malaria, and hosts species. Avian malaria occurs in most bird species (Valkiunas 2005) and can have a heterogeneous spatial distribution even within a single host population (Eisen & Wright 2001; Wood et al. 2007; Lachish et al. 2011) Infection. infection can have negative effects on individual fitness (Marzal et al. 2005; Marzal et al. 2008; Knowles, Palinauskas & Sheldon 2010), and is thought to have contributed to the decline and extinction of several bird species (Van Riper et al. 1986). As such, it represents an important disease threat to wild bird populations. Finally, because of their relatively high density, distribution throughout the landscape (often occurring in both pristine and anthropogenic habitats) and the ease with which they can be sampled, avian species provide excellent models in which to investigate the causes and consequences of malaria transmission in natural systems (Fallon et al. 2004; Ricklefs, Fallon & Bermingham 2004). .
Berthelot’s pipit (Anthus berthelotii) is a passerine endemic to the Macaronesian archipelagos. The large and widespread population on Tenerife exists across a range of habitats and ecological gradients, but is isolated from all other conspecific populations (Illera, Emerson & Richardson 2007). Importantly it is host to significant levels of a restricted number of avian malaria strains (Illera, Emerson & Richardson 2008; Spurgin et al. 2012) -thus making analyses tractable. This population, therefore, provides an excellent study system with which to assess how ecological and landscape factors influence malaria prevalence in a wild population.
The aim of the present study was to use a combination of molecular disease screening of individuals and fine scale GIS analysis of the habitats in which they were sampled to: (i) determine how the environment modulates the prevalence of malaria, and (ii) assess the relative importance of anthropogenic, natural, biotic and abiotic environmental factors, on the presence of malaria in pipits across the landscape.
STUDY SPECIES AND SAMPLING
Berthelot’s pipit (Anthus berthelotii) is a sedentary passerine that lives on all the islands in the Atlantic archipelagos of the Canary Islands, Selvagens and Madeira (Cramp & Perrins 1977). On Tenerife (Fig. 1) the pipit is numerous and widespread, inhabiting open, semi-arid habitats from sea level up to mountainous habitats at elevations of 3700 m. To obtain a representative sample of the pipit across its entire range and all environmental gradients on Tenerife, a 1 Km2 grid was laid over a map of the island obtained from Google Earth in ArcGIS version 10 (Esri 2011, Redlands, CA, www.esri.com). Each accessible Km2 that contained habitat suitable for pipits was visited and whether or not pipits were present was recorded; where present, an attempt was made to catch at least one pipit per Km2using clap nets baited with Tenebrio molitor larvae. Each captured bird was ringed and morphological measurements (wing length, tarsus length, bill length, bill height, bill width, head length and mass) taken. Blood samples were taken from all birds by brachial venipuncture and stored in 100% ethanol in screw cap micro-centrifuge tubes at room temperature.
Genomic DNA was extracted from blood using a salt extraction method (Richardson et al. 2001). The sex of each bird was determined by polymerase chain reaction (PCR) as described in Griffiths et al. (1998). Only DNA samples that successfully amplified the sex specific markers were used in the malaria screening. To screen for avian malaria infection a nested PCR was used following (Waldenstrom et al. 2004). An initial PCR amplified a 580 bp fragment of the cytochrome b gene of the Haemosporidian genome, then, nested PCRs were performed to amplify Haemoproteus, Plasmodium or Leucocytozoon following the methods described in Illera, Emerson & Richardson (2008). All samples were screened at least twice (3 times in the case of any ambiguous results) and only samples that amplified twice and were verified as malaria through sequencing were taken to be infections. PCR products were sequenced using BigDye terminator reaction kit (Perkin Elmer Inc. Waltham, MA) and products were run on an automated sequencer (ABI PRISM 3700, Applied Biosystems, Carlsbad, USA). Sequences were visually checked using FinchTV (http://www.geospiza.com/finchtv/) and aligned using BIOEDIT version 7.0.9 (Hall 1999) to sequences from the National centre for Biotechnology Information (NCBI) GenBank database and the MalAvi database for avian malaria (Bensch, Hellgren & Perez-Tris 2009).
Environmental variables were selected and assigned to one of four different categories as follows: (1) Natural abiotic: minimum temperature of the coldest month (MINTEMP), precipitation (PRECIP), aspect (ASPECT), slope (SLOPE) and altitude (ALT), (2) Natural biotic; vegetation type (VEGTYPE) and pipit density (DENSITY), (3) Anthropogenic abiotic; distance to water reservoirs (DISTWATER), distance to urban site (DISTCONST) and distance to livestock farms (DISTFARM), and (4) Anthropogenic biotic; distance to poultry farms (DISTPOUL). DISTFARM was calculated as an alternative predictor to DISTPOUL to account for effects derived from farm characteristics, and not the animals bred in them per se, while DISTPOUL was included to investigate the effects that birds bred in these farms might have on the transmission of malaria.
All environmental variable calculations and resampling were carried out in ArcGIS version 10 and R (R Development Core Team 2011). The environmental variables, MINTEMP, ALT, SLOPE, ASPECT, PRECIP, VEGTYPE and DENSITY, were calculated within 50 m, 100 m and 200 m radii buffers around each point where a bird had been caught in order to carry out a sensitivity analysis into the scale-dependence of their effects (see below). In each case an area-weighted mean was calculated within each buffer. Climatic variables (MINTEMP and PRECIP) were obtained from the WorldClim database (Hijmans et al. 2005) at a resolution of 30 arc seconds (1 Km). Topographic variables (ALT, SLOPE, ASPECT) at a resolution of 90 m were calculated from digital elevation models obtained from the Shuttle Radar Topography Mission Digital Elevation Database version 4.1 (Consortium for Spatial Information, www.cgiar-csi.org). Vegetation data were obtained from GRAFCAN (Cartográfica de Canarias S.A., www.grafcan.com (Del-Arco et al. 2006) and were used to calculate proportional areas of each of five categories of VEGTYPE (forest, grass, shrub, rock associated and urban associated vegetation) within each buffer and each bird was assigned the majority vegetation type within the surrounding buffer.
Distance variables (DISTWATER, DISTCONST, DISTFARM, and DISTPOUL) were calculated by overlaying the layer for pipit location points over polygon layers for: artificial water reservoirs, urban areas; and the position, species and census of livestock farms from the government of Tenerife (Plan de ordenamiento territorial, Cabildo de Tenerife, http://www.tenerife.es/planes/). For each variable the ‘proximity’ tool of the analysis extension of ArcGIS 10 was used to calculate the distance to the nearest relevant feature for the variable concerned.
An index of pipit density (DENSITY) was calculated as the number of pipits per square Kilometre, based on our geo-referenced records of pipit presence, using the ‘density’ tool of the spatial analyst extension in ArcGIS 10, with a neighbourhood size of 2500 m radius around the centre of each square Kilometre sampling cell. This index was used to reflect the size of the subpopulation of pipits found in the same area as the sampled pipit and thus to provide a measure of the local conspecific host population available for malaria infection.
The relative importance of natural abiotic, natural biotic, anthropogenic abiotic, and anthropogenic biotic predictors in influencing prevalence of malaria was assessed using both non-spatial binomial generalised linear models (GLM), and spatial autologistic models (Augustin, Mugglestone & Buckland 1996). We implemented model selection approaches (Burnham & Anderson 2001) to compare the relative fit of competing models, or sets of models, using Akaike’s information criterion (AIC) as the measure of model fit. We performed three sets of modelling procedures, nested within each of our two modelling methods (non-spatial binomial GLMs and spatial autologistic models), one for each buffer radius (50m, 100m, and 200m), hence performing a sensitivity analysis of potential scale-dependent effects of buffer radius on our results. For each of our three model sets, distance based environmental variables (DISTWATER, DISTCONST, DISTFARM, and DISTPOUL) remained invariant. The results obtained at these three sampling scales were very similar (supplementary table S1), therefore we chose to report only the results using the 100-m radius buffer, since this best approximates the territory size of Berthelot’s pipit (Juan Carlos Illera Pers Comm.).
In each of our three model sets, nested within our two modelling methods, the same series of modelling steps were repeated. First we compared AICs for single-predictor models, where each predictor is tested separately. Prior to running multi-predictor models, co-linearity between each pair of predictor variables was evaluated using pairwise bivariate correlations in PASW Statistics version 18 (SPSS Inc. 2009, Chicago, IL, www.spss.com). When a pair of variables had a correlation coefficient > 0.7, the variable with the highest single-predictor AIC (lowest fit) was dropped from our set of predictors. We then ran all possible combinations of 9 predictors resulting in 511 models in total and recorded the AIC, AIC (the difference between the best model’s AIC and that of the model in question) and Akaike weights (a measure of the relative explanatory value of the model, compared to all possible ones). We considered models with AIC ≤ 2 as having sufficient support (Burnham & Anderson 2004). All possible model subsets that included biotic, abiotic, anthropogenic and natural predictors were additionally assessed to determine which of the four environmental categories had most influence on malaria infection and in order to identify the best-fit model within each subset.
For our binomial generalised linear models (GLM), we checked for spatial autocorrelation in the model residuals by calculating Moran’s I coefficients at 1000 m distance classes and generating a correlogram using the package ncf in R (Bjornstad 2012). Autologistic regression modelling (Augustin, Mugglestone & Buckland 1996) was implemented to account for the observed spatial autocorrelation by including an autocovariate to assume spatial autocorrelation up to a maximum of 1000 m. This autocovariate was calculated following Dormann et al. (2007) using the R package spdep (Bivand 2012). Residual spatial autocorrelation was found to be absent from these autologistic models. The R package fmsb (Nakazawa 2012) was used to calculate the Nagelkerke R2, a logistic analog of R2 in ordinary regression.
In total 388 Berthelot’s pipits were sampled between January and April 2011. Malaria was detected in 156 out of 388 individuals (40.2%). Of these 156 individuals, 14 (9%) were infected with Leucocytozoon, while Plasmodium was detectedin 148 (95%), with six birds (3.8%) infected with both genera. Three strains of Plasmodium were detected; LK6 and LK5 -first described in the Lesser kestrel (Falco naumanii) (Ortego et al. 2007) - were detected in 139 and seven individuals, respectively, while KYS9 - first isolated from Culex pipiens mosquitoes (Inci et al. 2012)- was found in two individuals. Two strains of Leucocytozoon were detected; REB11 in 12 individuals and ANBE1 (Spurgin et al. 2012) in two. To avoid confounding the analyses by including different genera/strains of protozoa, only birds infected with the most commonly detected strain, Plasmodium LK6, which accounted for 139 out of 156 (89.1%) of all infections, were included as infected in the analyses.
MODELS AND SPATIAL ANALYSES
ALT was highly correlated with PRECIP (Spearman’s rho = 0.838, p < 0.001), and with MINTEMP (Spearman’s rho = -0.869, p < 0.001). PRECIP was also correlated with MINTEMP (Spearman’s rho = -0.923, p < 0.001). Since these three predictors fall into the same natural abiotic category, PRECIP and ALT were removed on the basis that MINTEMP had the lowest AIC of the three among single-predictor models (Table 1). Single-predictor GLMs further showed that DISTWATER followed by MINTEMP and DISTPOUL best predicted malarial infection in pipits. Autologistic models resulted in a better fit, but the relative importance of predictors remained the same (Table 1). In general, there was a relatively low amount of variation explained by each predictor as revealed by the Nagelkerke R2 values (0.085 - 0.139). In the following sections only the results from the spatial autologistic models are presented.
In the multiple-predictor spatial models, the best fit model contained DISTWATER and DISTPOUL, both being negatively correlated with the presence of malaria (Table 2). In 12 other models, all of which contained DISTWATER and DISTPOUL, AIC was less than or equal to 2. Of the other predictors VEGTYPE was present in five, while six other predictors were each present in three or less of the best fit models. Out of 511 possible models, 263 models had a probability greater than 0.95 of including the best model, indicating that the 95% candidate set was too broad to be informative. Therefore we investigated the relative importance of each predictor individually, by calculating the summed Akaike weight of all possible models where the predictor is present. Models containing DISTWATER had a summed Akaike weight of 0.86; DISTPOUL had a summed Akaike weight of 0.81; and MINTEMP had a summed Akaike weight of 0.44. The remaining five predictors had summed Akaike weights lower than 0.41 (Table 3).
The results of all possible autologistic models representing different categories of environmental variables (biotic, abiotic, anthropogenic and natural) are summarised in table 4. The best fit model (lowest AIC) was the anthropogenic model containing DISTWATER and DISTPOUL which explained 16% of variation in malarial infection, followed by the abiotic model that included MINTEMP and DISTWATER and the natural model with only MINTEMP. The predictor set with the least fit was the biotic model including DISTPOUL and VEGTYPE.
Discussion By measuring variables at a fine landscape scale in an avian malaria-host system, we found evidence that specific environmental factors influenced the distribution of malaria in a wild population. While the climatic variables we analysed were not keypredictors of the prevalence of Plasmodium LK6 in pipits in Tenerife, anthropogenic factors, such as distance to artificial bodies of water and distance to poultry farms were.
Abiotic natural factors have been shown to play a major role in the prevalence and transmission of vector-borne diseases (Sleeman et al. 2009; Linthicum et al. 2010), including malaria (Van Riper et al. 1986; LaPointe, Goff & Atkinson 2010; Xiao et al. 2010; Sehgal et al. 2011). However, contrary to predictions based on previous studies, and earlier work on pipits indicating a very low prevalence of malaria at high altitudes (> 1600 m, Spurgin et al. 2012), temperature and altitude were not the best predictors of malaria in the present study. MINTEMP was not present in the best fit multi-predictor model and the combined Akaike weight of models containing MINTEMP (0.44) was low compared to the values obtained for DISTWATER (0.86) and DISTPOUL (0.81). Why this is the case we are not sure. It is possible that minimum temperatures are not sustained long enough for vectors and/or parasites to be affected. Another possible explanation is that malaria transmission occurs only during the warm summer months, so minimum temperature (which occurs in the winter months) wouldn’t have an effect on the overall prevalence of the disease. Other abiotic natural variables that have been identified as predictors of malaria are the topographic variables of aspect and slope, which affect the presence and persistence of the wet habitats required for vector productivity (Balls et al. 2004; Githeko et al. 2006; Cohen et al. 2008; Atieli et al. 2011). However, in the present study we found no indication that either slope or aspect was important for malaria prevalence. This could be due to the volcanic nature of soils in Tenerife (Fernandez-Caldas, Tejedor-Salguero & Quantin 1982), which are highly permeable and unlikely to hold water long enough to allow for larvae development.
Malaria prevalence has often been shown to be closely correlated with precipitation levels (Galardo et al. 2009; Zhou et al. 2010; Bomblies 2012), but this is not the case in our study. In Tenerife rainfall is scarce and the land is steep and porous, consequently water does not naturally remain in pools long enough to provide habitats for vector reproduction. There are, however, many artificial water pools and small canals that have been created for agricultural and other purposes. Thatdistance to the nearest reservoir is a predictor of malaria infection while rainfall is not, suggests that these reservoirs provide suitable habitats for vector larvae development thus facilitating malaria transmission. Previous studies have shown artificial water reservoirs, irrigation canals, and dams to be important for the production of malaria vectors (Fillinger et al. 2004) and that water bodies are closely associated with malaria (Wood et al. 2007; Zhou et al. 2010; Lachish et al. 2011).
Various biotic factors may also influence the prevalence of malaria. Vegetation type did not have a effect on malaria infection in our dataset. Berthelot’s pipits inhabit open areas and are not found in closed canopy forests, such as the laurisilva present in the wetter northern parts of the island. Hence, the low apparent importance of habitat may be a result of focusing on a specific host species, rather than a pattern general to avian malaria. Nevertheless, it is also possible that the vectors of malaria in Tenerife are not so much constrained by the vegetation cover and might be equally abundant across areas.
Host density is also expected to affect malaria prevalence because it modifies vector-host contact rates. However, while some studies support this prediction (Ortego & Cordero 2010), others, including the present study, fail to find a correlation (Bonneaud et al. 2009). It may be that our estimate of pipit density, based on the presence/absence of pipits in a square Kilometre grid is not sufficiently accurate. This indirect measure was designed to reflect density at the appropriate scale, however it is possible that pipit density is more highly localised than thought. Furthermore, Although adult pipits tend to hold the same breeding territories from year to year (Illera & Diaz 2008), we do not know whether the spatial structure of pipit density is constant year-round. Patterns of movement and juvenile dispersal, which are not taken into consideration by the present study, could also have important implications for the transmission of infectious diseases such as malaria, (Altizer, Oberhauser & Brower 2000; Jones et al. 2011). Direct measures of host density may provide a better estimate of the effects that pipit density has on malaria prevalence; however, such measures would be extremely difficult and time consuming to calculate, requiring counts of abundance within each km2 across the year.
The local density of all vertebrate hosts, not just the focal host species, could have an effect on malaria prevalence. Within host communities, some species act as key hosts harbouring parasitic fauna, thus altering prevalence in other host species (Hellgren et al. 2011). The malaria strain detected in the pipits, Plasmodium spp. LK6, has also been reported in blackbirds and canaries (Phillips 2009), species that co-occur with pipits in many areas of Tenerife. Unfortunately, we have no data on densities of these two species in order to investigate whether they have an effect on pipit malaria prevalence. A comprehensive study of the prevalence of avian malaria in the bird community on Tenerife would be needed to understand the role other bird species might play in the prevalence of malaria.
Human activities have been shown to affect vector-borne diseases (Patz et al. 2000; Friggens & Beier 2010; Gottdenker et al. 2011), including malaria (Serandour et al. 2007). Factors such as deforestation, animal husbandry, construction and artificial water management modify the ecological balance within which vectors and their parasites develop and transmit disease (Patz et al. 2000). Livestock farms have been shown to influence the transmission of infectious diseases by facilitating atypical aggregations of wild birds including infected individuals both of the focal, and other species (Carrete et al. 2009). Conversely, such farms have also been shown to dilute the effect of malaria transmission by reducing biting rates on susceptible hosts (Mutero et al. 2004; Liu et al. 2011). In the present study, distance from the nearest poultry farm at which a pipit was caught had a significant negative effect on the probability of malaria infection. This suggests either that; i) poultry farms provide suitable habitats for vectors, ii) aggregations of wild birds that facilitate transmission occur as a result of these farms, or iii) poultry are themselves reservoirs of malaria. That there was no effect of distance to nearest non-poultry livestock farm, suggests the third explanation is most plausible, as all livestock farms should equally support vectors and encourage bird aggregations. The specific LK6 lineage has not been reported in poultry (though screening for such lineages in poultry has rarely been undertaken), but various other Plasmodium lineages have been reported in both poultry and wild birds (Bensch, Hellgren & Perez-Tris 2009), thus providing evidence that poultry can be a reservoir of avian malaria. Direct screening of poultry farms on Tenerife would be needed to confirm the association.
Other anthropogenic activities, such as urbanisation, can be important predictors of vector-borne diseases (Bradley & Altizer 2007) including malaria (Guthmann et al. 2002). For example, mosquito species can quickly adapt to urban environments (Antonio-Nkondjio et al. 2011; Kamdem et al. 2012) and urbanisation has been shown to increase human malaria prevalence (Alemu et al. 2011). In our analyses, distance to the nearest constructed site had a negative effect on malaria prevalence in the individual predictor non-spatial model, however, the effect disappeared after accounting for spatial autocorrelation. It may well be that the broad classification of ‘urbanisation’ used in this study lacks the resolution to identify the particular characteristics of urbanisation that influence malaria prevalence. Given that urban expansion is happening around the world, further work to understand its impact on wildlife disease prevalence is warranted.
While our models have identified key environmental variables associated with malaria infection, considerable variation (83%) remains unexplained. This confirms the view that wildlife diseases, such as malaria, are complex and that many different factors, including ones not closely linked to environmental gradients, can influence their spatial distribution within a host population. (Hawley & Altizer 2011). It is especially important to note that host related factors, such as individual immunity (and immune variation in the population), host movement patterns and stochastic processes that might influence the epidemiology of malaria were not accounted for in this study. Furthermore, as the specific vectors that transmit avian malaria in Tenerife have not yet been described (but see Bensch, Hellgren & Perez-Tris 2009), we were unable to incorporate an understanding of the ecology of these vectors into our analysis. Finally, although DISTWATER did predict the prevalence of malaria, the resolution of the GIS layer used in our study was not able to capture the presence of very small water bodies that might be equally important in malaria transmission. Consequently we may have underestimated the explanatory power of this predictor.
Assessing the role of the environment in the transmission of pathogens in the wild is crucial to our understanding of disease dynamics and of the causes and consequences of host-pathogen coevolution. The evidence from our study supports previous work which suggests that the prevalence of malaria can vary over small spatial scales (Lachish et al. 2011). Contrary to other studies which found that climatic variables were strongly associated with malaria prevalence (Balls et al. 2004; Briet, Vounatsou & Amerasinghe 2008; Garamszegi 2011), we found that anthropogenic environmental variables, namely proximity to artificial water reservoirs and poultry farms, were the most important predictors of malaria in pipits across Tenerife. This may, at least in part, reflect the scale at which the study was performed – when measured across greater scales the influence of locally important predictors of disease may be swamped by regional differences (Balls et al. 2004; Briet, Vounatsou & Amerasinghe 2008; Garamszegi 2011). This study demonstrates the importance of measuring local fine scale variation, and not just regional effects, in order to understand how environmental variation can influence wildlife diseases.