Adaptive landscape genetics and malaria across divergent island bird populations

Abstract Environmental conditions play a major role in shaping the spatial distributions of pathogens, which in turn can drive local adaptation and divergence in host genetic diversity. Haemosporidians, such as Plasmodium (malaria), are a strong selective force, impacting survival and fitness of hosts, with geographic distributions largely determined by habitat suitability for their insect vectors. Here, we have tested whether patterns of fine‐scale local adaptation to malaria are replicated across discrete, ecologically differing island populations of Berthelot's pipits Anthus berthelotii. We sequenced TLR4, an innate immunity gene that is potentially under positive selection in Berthelot's pipits, and two SNPs previously identified as being associated with malaria infection in a genome‐wide association study (GWAS) in Berthelot's pipits in the Canary Islands. We determined the environmental predictors of malaria infection, using these to estimate variation in malaria risk on Porto Santo, and found some congruence with previously identified environmental risk factors on Tenerife. We also found a negative association between malaria infection and a TLR4 variant in Tenerife. In contrast, one of the GWAS SNPs showed an association with malaria risk in Porto Santo, but in the opposite direction to that found in the Canary Islands GWAS. Together, these findings suggest that disease‐driven local adaptation may be an important factor in shaping variation among island populations.

Studying pathogen-mediated selection has largely involved a candidate gene approach, where variation at genes with known, or predicted, host immunity function is investigated in relation to infection (Bernatchez & Landry, 2003;Netea, Wijmenga, & O'Neill, 2012). Many studies have focused on the major histocompatibility complex (MHC), a gene family that plays a key role in pathogen recognition in the adaptive immune system. However, a greater proportion of phenotypic variance in malaria response has been attributed to non-MHC genes (Jepson et al., 1997). Within the innate immune system (a first line of defense against infection), Toll-like receptors (TLRs) are a family of pattern-recognition receptors which have been linked to malaria resistance (Ferwerda et al., 2007;Mockenhaupt et al., 2006), and show evidence of pathogen-mediated balancing selection (Ferrer-Admetlla et al., 2008;Fisher et al., 2011;Gavan, Oliver, Douglas, & Piertney, 2015). TLRs therefore represent important candidates for investigating the role of pathogens in maintaining host genetic variation.
An alternative to the candidate gene approach is the use of genome-wide association studies (GWAS). These enable detection of single nucleotide polymorphisms (SNPs) throughout the genome that show statistical associations with pathogen infection. In addition to identifying associations at known immune loci (Fellay et al., 2007;He et al., 2015;Wong et al., 2010), GWAS approaches may reveal novel candidate genes (Fu et al., 2012;Ravenhall et al., 2018;Thye et al., 2010) for further study of the evolutionary dynamics between host and pathogen.
Pathogen communities on each island are shaped by chance colonization and extinction events, which can result in distinct pathogen assemblages and selection pressures between islands (Fallon, Bermingham, & Ricklefs, 2005;Olsson-Pons, Clark, Ishtiaq, & Clegg, 2015;Wang et al., 2017). Limited gene flow in and out of islands also allows for stable communities of hosts and pathogens (Spurgin, Illera, Padilla, & Richardson, 2012), which may facilitate strong coevolutionary relationships.
Malaria infection shows high spatial variability in this species, both F I G U R E 1 Map of Berthelot's pipit populations. Berthelot's pipits are found across all islands within the Madeiran archipelago (top left panel), the Canary Islands (bottom left panel), and the Selvagens archipelago, situated between the Canary Islands and Madeira between and within islands, making it a highly suitable model for investigating the role of spatial scale in pathogen-mediated selection.
Characterization of malaria throughout Berthelot's pipit populations (Illera, Emerson, & Richardson, 2008;Spurgin et al., 2012) found the highest prevalence of Plasmodium and Leucocytozoon infection on Porto Santo, whereas no infection was detected elsewhere in the Madeiran archipelago. Prevalence of malaria on Tenerife is influenced by a combination of climatic and anthropogenic effects (González-Quevedo et al., 2014;Padilla et al., 2017), with malaria undetected at high altitude. Associations between the distribution of MHC variants and environmental predictors of malaria infection have been detected (González-Quevedo, Davies, Phillips, Spurgin, & Richardson, 2016). No evidence of Haemoproteus infection has been found in this species.
Here, we test for associations between fine-scale patterns of genetic variation and malaria in Berthelot's pipits across two divergent populations (Tenerife and Porto Santo) to investigate the spatial scale of local adaptation in the presence of gene flow within a population. This study also allows us to test the repeatability of patterns of association across populations. These two islands, situated on different archipelagos, show high genetic divergence at neutral loci, with limited to no gene flow between them (Armstrong et al., 2018;Spurgin et al., 2014). Despite a sharp decline in overall genetic diversity associated with the initial colonization of Madeira (Armstrong et al., 2018), higher levels of TLR4 allelic and amino acid richness exist in Madeira compared to the Canary Islands (González-Quevedo, Spurgin, Illera, & Richardson, 2015). Furthermore, evidence of positive selection at TLR4 in Berthelot's pipits suggests it may be an evolutionarily important locus (González-Quevedo et al., 2015). In this study, we (a) test for associations between Plasmodium infection status (the only Haemosporidian genus commonly detected; >1% prevalence) in this species (Spurgin et al., 2012), and variation at TLR4 and two SNPs previously identified in a GWAS of malaria infection across Berthelot's pipit populations (Armstrong et al., 2018); (b) determine the environmental predictors of malaria risk on Porto Santo; (c) compare genetic associations with malaria risk in Porto Santo and Tenerife, utilizing the above measures of malaria risk for Porto Santo, and those previously calculated for Tenerife (González-Quevedo et al., 2016.  (Illera et al., 2007;Spurgin et al., 2012 (Cramp, 1988).

| Parasite screening
We used a nested PCR approach that detects Plasmodium and Haemoproteus to characterize malaria infection status (Waldenström, Bensch, Hasselquist, & Östman, 2004), with multiple positive and negative controls included in each PCR plate. Samples that successfully amplified at least twice were classified as infected. The strain of Plasmodium was determined by Sanger sequencing; Haemoproteus was not detected. We focused on Plasmodium as the most widespread and abundant haemosporidian found in Berthelot's pipits (Illera et al., 2008;Spurgin et al., 2012). In addition, the vector species of Plasmodium (mosquitoes; Culicidae) and Leucocytozoon (blackflies; Simuliidae) have different ecological niches (Harrigan et al., 2014;Imura et al., 2012), thus combining the two may confound results. to genotype all additional samples from Tenerife and Porto Santo (n = 577) at each TLR4 SNP, except for TLR4_5 which was excluded due to a very low minor allele frequency of <0.05. Assay design and genotyping were performed by LGC Genomics, Hertfordshire.

| Genetic analysis
Genotypes AT, CT, and TT at the triallelic TLR4_4 SNP were coded as missing data (Tenerife n = 12; Porto Santo n = 1), to treat this SNP as biallelic. We used DnaSP v6 (Librado & Rozas, 2009) to phase the TLR4 SNPs into haplotypes. To aid phasing, we included TLR4 sequences from all Berthelot's pipit populations, and each phased TLR4 haplotype previously detected in Berthelot's pipits (González-Quevedo et al., 2015). Samples with <90% phasing certainty were excluded from models that included TLR4 haplotypes as predictors.
We translated the phased TLR4 sequences originating from Sanger sequencing into protein haplotypes. This gave us the amino acid residues at each of the codons containing a SNP, from which we were able to infer the amino acids at each SNP for samples that were genotyped with KASP™ genotyping.
We used PLINK 1.9 (Chang et al., 2015) to calculate linkage disequilibrium (LD) between each pair of SNPs, and test for deviations from Hardy-Weinberg equilibrium with Hardy-Weinberg exact tests. Where frequencies ≤0.05 were found for SNP minor alleles or TLR4 protein haplotypes, these variants were not included as predictors in genetic analyses. All GIS analyses were performed in QGIS v2.18 (QGIS Development Team, 2017). MINTEMP and PRECIPITATION were obtained from WorldClim global climate data v2 (Fick & Hijmans, 2017) at a resolution of 30 arc-seconds (approximately 1 km 2 ). ALTITUDE was obtained from the Shuttle Radar Topography Mission (SRTM) 3 Arc-Second Global elevation data (srtm.csi.cgiar.org) at a resolution of approximately 90 m 2 . SLOPE and ASPECT were calculated from SRTM data. VEGTYPE was characterized using the CORINE Land Cover inventory (CLC 2012 v.18.5.1; http://land.coper nicus. eu/pan-europ ean/corine-land-cover/ clc-2012). We combined the land cover classes into six categories: arable, urban-associated, forest, rock-associated, grass, and shrub ( Table 1). Values of MINTEMP, PRECIPITATION, ALTITUDE, SLOPE, and ASPECT were calculated by taking the average value of each variable within a 100 m buffer around each sample location. In the case of VEGTYPE, the sample was assigned the category with the largest area within the buffer.

| Porto Santo GIS analyses
ASPECT was classified as one of eight categories: N, NW, W, SW, S, SE, E, and NE.
We calculated DISTWATER with polygons drawn on Google Earth satellite imagery over water sources encountered during sample collection and obtained from the OpenStreetMap data-filtering tool Overpass Turbo (https ://overp ass-turbo.eu) using the query "natural = water." Water sources were 76-28,190 m 2 . The presence of livestock was dependent on visual encounters, as farming census data were not publicly available. The type of livestock was used to differentiate between factors related to livestock farming that might cause aggregations of birds (DISTFARM), and the potential effect of poultry as a reservoir of avian malaria (DISTPOUL). DIST_URB was the distance to the nearest urban-associated area, as classified by VEGTYPE.
DENSITY was calculated as follows. A 1 km 2 grid was overlaid on Porto Santo, and a 1 km radius buffer was drawn around the centroid

| Environmental predictors of malaria risk
We with the lowest AIC (and therefore the highest likelihood) to calculate weighted averages of parameter estimates and the relative importance of model predictors (Burnham & Anderson, 2002). We report AICc, a modification of AIC that is recommended for small sample sizes (Hurvich & Tsai, 1989).
We used variance inflation factors (VIFs) calculated using the R package car (Fox & Weisberg, 2011) to test for collinearity between environmental variables, using a threshold of >3 to indicate unacceptably high collinearity (Zuur, Ieno, & Elphick, 2010 (Table 3), we tested whether the inclusion of an interaction term improved the fit of a binomial GLM with the two variables as main effects and malaria infection as the response. The interaction DENSITY*DISTWATER gave the largest improvement in AICc (main effects only, AICc = 103.3; main effects and interaction, AICc = 90.8) and was therefore included in model selection (Table 3).

| Model selection and model averaging
Fitting all combinations of the six environmental variables and one interaction term (see above) as predictors of malaria infection using binomial GLMs, we performed model selection and model averaging following Grueber, Nakagawa, Laws, and Jamieson (2011) using the R package MuMIn (Bartoń, 2018) to obtain the best-supported models for explaining occurrence of malaria infection. Prior to analysis, we used the R package arm (Gelman & Su, 2018) to standardize the input variables to a mean of zero and a standard deviation of 0.5  (Gelman, 2008;Grueber et al., 2011). The model selection process calculated ΔAICc, the difference in AICc between each model and the "best" model (the model with the lowest AICc), and the Akaike weight, which quantifies the likelihood of each model having the best explanatory power within a set of models (Burnham & Anderson, 2002). Using the R package DescTools (Signorell, 2018), we calculated the McFadden-adjusted pseudo-R 2 (the likelihood of a logistic regression model relative to an intercept-only model, adjusted to account for the number of predictors in the model; McFadden, 1974).
A threshold of ΔAICc ≤ 7 is recommended to retain models that have sufficient support, without dismissing models which still provide some explanatory power (Burnham, Anderson, & Huyvaert, 2011). We applied model averaging over this set of models to calculate weighted averages of parameter estimates and the relative importance of each predictor (the sum of Akaike weights for models which include that predictor). We used the zero method of model averaging to avoid biasing results toward predictors with low explanatory power (Burnham & Anderson, 2002;Lukacs et al., 2007).

| Spatial autocorrelation
We tested for spatial autocorrelation in model residuals as this may lead to spurious associations between predictor and response variables (Dormann et al., 2007;Lennon, 2000). We created Moran's I correlograms at distance class intervals of 750 m and 1,000 m using the R package ncf (Bjornstad, 2018), with 1,000 permutations to test the significance of Moran's I at each interval. After correcting for multiple testing using the Holm correction (Holm, 1979;Legendre & Legendre, 2012), there was no evidence of spatial autocorrelation in the model residuals (all adjusted p values > .05). Correcting for spatial autocorrelation was therefore not required for the estimation of malaria risk in Porto Santo.

| Malaria risk scores
We used the predicted values of the best model identified by model selection as a malaria risk score between 0 and 1 for each sample location. This represented the probability of an individual at that location being infected with malaria, as a result of the environmental conditions. An earlier study determined that malaria infec-

| Genetic associations with malaria infection
Genetic variation was classified in three ways: (a) SNP genotype, encoded as 1 for heterozygotes and 0 or 2 for each of the homozygotes; (b) presence (1) or absence (0)  We tested the genetic variables as predictors of malaria risk in general linear models (LMs) for Tenerife (2011) and Porto Santo (2016). As malaria risk was derived from spatially varying environmental predictors, genetic variation alone was unlikely to account for all spatial autocorrelation in malaria risk. We therefore included distance-based Moran's eigenvector maps (dbMEMs) as spatial predictors of malaria risk. dbMEMs are used to identify gradients of spatial variation (spatial structure) in a response variable, across multiple potential scales from broad to fine, calculated by eigenvector decomposition of distance matrices based on the spatial coordinates of samples (Borcard & Legendre, 2002;Dray, Legendre, & Peres-Neto, 2006). Hence, the use of dbMEMs as predictors of malaria risk accounts for spatially autocorrelated variation in malaria prevalence that would otherwise be explained by environmental conditions, transmission dynamics, and/or unmeasured genetic gradients. We calculated dbMEMs for each island using the R packages adespatial (Dray et al., 2018) and vegan (Oksanen et al., 2018), retaining dbMEMs with positive eigenvalues, representing positively autocorrelated spatial variation. dbMEMs were ranked by descending R 2 values in single-predictor LMs of malaria risk and sequentially added into each model of the genetic associations with malaria risk outlined above, until additional dbMEMs no longer improved AICc. With each iteration, we checked whether spatial autocorrelation in model residuals had been controlled for, to find the minimum required number of dbMEMs. The inclusion of dbMEMs resulted in VIFs < 3, indicating that any collinearity between dbMEMs and genetic variants was acceptably low. We performed hierarchical partitioning using the "lmg" method in the R package relaimpo (Grömping, 2006) to calculate the proportion of variance in malaria risk explained by genetic variants.  (Ortego et al., 2007) was detected (Illera et al., 2008;Spurgin et al., 2012 Excluding sampling years where no juveniles were caught, the prevalence of malaria was significantly higher in adults than in juveniles, both in Porto Santo (test of equal proportions χ 2 = 25.6, p < 0.001) and in Tenerife (χ 2 = 3.1, p = 0.039). As juveniles were present in much lower numbers than adults (Table 4), we removed juveniles from further analysis. Final sample sizes are shown in Table 4. As amino acid substitutions could potentially alter TLR4 function (Schröder & Schumann, 2005), we classified TLR4 variation into four protein haplotypes (denoted with the prefix "TLR4_P"; Table 5 and Figure 2b). TLR4_P2 was translated from two haplotypes differing at the synonymous SNP TLR4_3. The TLR4_P3 and TLR4_P4 haplotypes were absent from Tenerife, and TLR4_P3 was at low frequency (<0.05) in Porto Santo.

| Sequencing
We tested for deviations from Hardy-Weinberg equilibrium at Tenerife R 2 = 0.62). We found moderate LD in Porto Santo between SNPs TLR4_2 and TLR4_3 (R 2 = 0.39) and between TLR4_2 and TLR4_4 (R 2 = 0.37). All other combinations of SNPs had low levels of LD (R 2 < 0.1).

| Malaria risk models
Model selection of the environmental predictors of malaria infection found 17 models with ΔAICc ≤ 7 relative to the "best" model, which contained VEGTYPE, ALTITUDE, DISTWATER, DENSITY, and DISTWATER*DENSITY (Table 6) (Table 7). We used the predicted values from the best model as our estimate for malaria risk for

| Genetic associations with malaria infection
As we found high levels of LD between TLR4 SNPs, we calculated VIFs for models of genetic associations with malaria infection and

| Genetic associations with malaria risk
We tested for associations between genetic variants and malaria risk on Porto Santo (2016) and Tenerife (2011). The results are summarized in Table 9. On Porto Santo, increasing numbers of T alleles at SNP 5239s1 (estimate = 0.69, SE = 0.27, p = 0.011), and A alleles at SNP TLR4_2 (estimate = 1.03, SE = 0.42, p = 0.016), were associated with increased malaria risk. However, the residuals of this model were highly spatially autocorrelated. To control for this, we included seven dbMEMs with high R 2 in single-predictor models of malaria risk (Figure 7), chosen from a set of dbMEMs which gave the lowest AICc in a multipredictor model of malaria risk. After controlling for autocorrelation, SNP 5239s1 was still associated with malaria  Figure 8) but TLR4_2 was not (p = 0.423). We did not find an association between TLR4_4 and malaria risk, either before or after controlling for autocorrelation.
Hierarchical partitioning of the above models (Table 10) found that SNP 5239s1 explained 5.2% of the variance in malaria risk before controlling for autocorrelation, and 3.3% of the variance after the addition of dbMEMs. Despite having a non-significant association with malaria risk in the model containing dbMEMs, TLR4_2 explained a greater proportion of the variance in malaria risk compared to 5239s1, both before (7.4%) and after (4.3%) controlling for autocorrelation.
Before taking autocorrelation into account, there were no significant associations between SNP heterozygosity and malaria risk on Porto Santo; however, after including seven dbMEMs to remove spatial autocorrelation in model residuals, heterozygosity at SNP 5239s1 was strongly associated with increased malaria risk (estimate = 0.74, SE = 0.25, p = 0.004). Heterozygosity at other SNPs was not associated with malaria risk.
Prior to controlling for autocorrelation, the presence of protein haplotype TLR4_P1 was associated with reduced malaria risk We did not find any significant associations between SNP genotype, SNP heterozygosity, or TLR4 protein haplotypes and malaria risk on Tenerife. We were unable to remove spatial autocorrelation in model residuals through the addition of dbMEMs as model predictors.

| D ISCUSS I ON
We used previously identified candidate SNPs linked to malaria infection across populations (from a GWAS analysis performed on RAD-seq SNPs; Armstrong et al., 2018) and TLR4 SNPs (González-Quevedo et al., 2015) to investigate the relationship between potentially adaptive genetic variation and avian malaria within two island populations of Berthelot's pipits. In addition to testing for associations with infection status, we calculated the malaria risk at each sampling location, predicted by modeling fine-scale environmental showed associations with malaria risk in Porto Santo, but not in Tenerife, where malaria risk was lower.

| Genetic associations with malaria
We have previously used RAD-seq SNPs to detect genetic variants that were associated with LK6 infection in Berthelot's pipits in the Canary Islands (Armstrong et al., 2018). The strongest association was found for SNP 5239s1, ca. 2,000 bp from interleukin-16, a proinflammatory cytokine that moderates the expression of other cytokines associated with malaria infection (Kern, Hemmer, Damme, Gruss, & Dietrich, 1989;Lyke et al., 2004;Mathy et al., 2000). In the present study, SNP 5239s1 was a predictor of malaria on Porto Santo, with the lowest infection and risk found in samples with the AA genotype. Remarkably, this was the opposite relationship to that found in the Canary Islands (Armstrong et al., 2018), where increased incidence of the T allele was associated with reduced infection. This may be indicative of pathogen-mediated balancing selection, which can arise from heterozygote advantage (Doherty & Zinkernagel, 1975), rare-allele advantage (Slade & McCallum, 1992;Takahata & Nei, 1990), and local adaptation to fluctuating pathogen selection pressures (Hill et al., 1991). When controlling for spatial autocorrelation, we found an association between SNP 5239s1 heterozygosity and malaria risk on Porto Santo, although contrary to the heterozygote advantage model, heterozygotes were associated with greater malaria risk than homozygotes (an effect which was largely driven by the decline in risk found with AA genotypes). Berthelot's pipit populations on the Madeiran and Canary Islands archipelagos have been isolated from each other for at least 8,500 years (Spurgin et al., 2014). Different populations may therefore be undergoing independent coevolutionary cycles with the same malaria strain, with alternative alleles conferring an advantage between divergent populations (Bonneaud, Pérez-Tris, Federici, Chastel, & Sorci, 2006). Alternatively, undetected genetic and phenotypic differences within the LK6 strain could potentially drive local adaptation between the archipelagos, with different alleles favored in different populations (Alcaide, Edwards, Negro, Serrano, & Tella, 2008;Loiseau et al., 2009). We used a single genetic marker, the mitochondrial cytochrome b locus, to classify the malaria strain. Several genes on the Plasmodium genome with relevance to infection success have shown greater genetic variation than at cytochrome b (Jarvi, Farias, & Atkinson, 2008;Lauron et al., 2014). It is possible that Berthelot's pipits on separate archipelagos could be adapting to different malaria strains within LK6, although this remains to be tested. We did not find evidence of associations between SNP 5239s1 and malaria infection or risk on Tenerife, despite this population being included in the previous GWAS (Armstrong et al., 2018). SNPs that are related to individual-level variation in parasite burden do not necessarily show the same associations at the landscape scale (Wenzel, Douglas, James, Redpath, & Piertney, 2016). It is possible that with the comparatively low malaria risk in Tenerife, gene flow is overriding landscape-scale associations between SNP 5239s1 and malaria risk (Forester, Jones, Joost, Landguth, & Lasky, 2016;Lenormand, 2002).
The previous GWAS result could have been driven by other populations such as Lanzarote and Fuerteventura, where malaria infection rates were higher (Illera et al., 2008;Spurgin et al., 2012).
Polymorphisms in immune genes can alter the effectiveness of their proteins for detecting and responding to pathogens (Lazarus et al., 2002;Sommer, 2005). The TLR4 SNPs sequenced here are situated within the ligand-binding region, which plays a key role in TLR pathogen recognition (Werling, Jann, Offord, Glass, & Coffey, 2009).
Evidence of positive selection in birds or mammals has been detected at each of the codons identified as polymorphic in Berthelot's pipits (Areal, Abrantes, & Esteves, 2011;Králová et al., 2018;Wlasiuk & Nachman, 2010), suggesting that these sites may be important for the evolution of pathogen recognition. On Tenerife, the presence of the TLR4 protein haplotype TLR4_P1 was associated with decreased malaria infection prevalence in 2011, but not across all sampling years. In earlier years, approximately half of the samples were collected from the high-altitude (>2,000 m above sea level) plateau of El Teide. Malaria has not been found in Berthelot's pipits in this location (González-Quevedo et al., 2014;Illera et al., 2008;Spurgin et al., 2012), although a survey of passerine communities on Tenerife found malaria at low frequency in high-altitude habitats . The relationship seen in 2011 between TLR4_P1 and infection may be masked in other sampling years by the increase in uninfected individuals from areas of low malaria abundance. We did not find a relationship between TLR4_P1 presence and malaria risk, potentially due to the explanatory power of the Tenerife malaria risk model (McFadden-adjusted pseudo-R 2 = 0.10). On Porto Santo, both TLR4_P1 and the SNP TLR4_2 were associated with malaria risk, although these relationships were no longer significant after including dbMEMs to remove autocorrelation. Both of these genetic variants showed significant associations with dbMEM1, which itself explained 22% of the variance in malaria risk, making it difficult to disentangle the real effects of these variants from any spurious associations arising from residual autocorrelation.

| Environmental predictors of malaria risk
We modeled the environmental predictors of malaria distributions in Porto Santo to understand fine-scale spatial differences in malaria TA B L E 9 Summary of linear models of the association between genetic variants and malaria risk in Berthelot's pipits on Porto Santo (PS) and Tenerife (TF) higher vector abundance and malaria are found in proximity to water (Ferraguti et al., 2018;Ganser et al., 2016;Illera et al., 2017).
In the present study, distance to urban areas was removed prior to model selection due to a positive collinearity with distance to water sources. Therefore, we cannot rule out the importance of additional sources of standing water that may be associated with urban environments. Other studies have found links between urbanization and increased malaria and/or vector abundance (Alemu, Tsegaye, Golassa, & Abebe, 2011;Li et al., 2014), although this appears to vary between vector species, with some favoring more natural habitats (Ferraguti et al., 2016). Pipit density was positively associated with malaria risk on Porto Santo, although the model-averaged parameter estimate was relatively small. There was, however, a strong negative interaction between distance to water and pipit density on this island, likely due to aggregations of mosquitoes and hosts around water sources, which may increase disease transmission rates (Begon et al., 2002;Greer, Briggs, & Collins, 2008;Le Menach, McKenzie, Flahault, & Smith, 2005;Raghwani et al., 2011).
Vegetation type was associated with malaria prevalence on Porto Santo. The highest abundance of malaria was found in arable and grassland habitats, with lower malaria in rock-associated habitats. However, this result should be interpreted with caution due to small sample sizes, as only six pipits were caught on rock-

ACK N OWLED G M ENTS
We would like to thank the many people who helped us in the field, CA, CG-Q, and MD performed laboratory work and analyzed se-