Implications of farmland expansion for species abundance , richness and mean body mass in African raptor communities 2 4

Globally, conversion of natural habitats to farmland poses the greatest extinction risk to birds, its 20 consequences being especially pervasive in the case of large predators and scavengers, whose declines may trigger extensive cascading effects. Human population growth in sub-Saharan Africa 22 is expected to drive a vast expansion in agricultural land by 2050, largely at the expense of pastoral land and savanna. In East Africa, the greatest expanse of suitable land yet to be converted to 24 agriculture lies mainly in South Sudan, DRC and Tanzania. To gauge the effects of land conversion on raptor populations in this region we used road survey data from neighbouring Uganda, from 26 which we determined linear encounter rates (birds seen 100 km-1; n = 33 species), and species richness (53 species). Encounter rates were much lower in pastoral land than in protected savanna 28 (median difference: -41%; 23 species), and lower still in agricultural land (-90%; 24 species). These disparities were influenced by diet and body mass. For large eagles and vultures, encounter 30 rates in agricultural land were 97% lower than in protected savanna (median of 12 species), whereas for smaller raptors they were 30% lower (12 species). Large, apex consumers were thus 32 more vulnerable to farmland expansion, and this was reflected in the mean body mass of species encountered in savanna (1,740 g), pastoral (995 g) and agricultural land (856 g). Body mass 34 differences remained significant when vultures were excluded. Since threat status is linked to body mass, encounter rates for globally threatened and near-threatened species likewise showed a more 36 pronounced deficit in farmland than those of least concern. Accordingly, pastoral and agricultural transects were less species-rich (10.6 and 6.7 raptor species 100 km-1, respectively) than savanna 38 transects (13.2 species). Our findings suggest that the projected expansion of agricultural land in sub-Saharan Africa is likely to reduce raptor populations in pastoral land and savanna by c. 50% 40 and 90%, respectively. We propose that conservation efforts focus on identifying the causes of raptor population deficits in farmland, and on safeguarding tracts of unprotected, intact savanna, 42 together with existing protected areas. 44


Introduction
Land use conversion is considered to be the single biggest driver of biodiversity loss in the tropics 50 (Foley et al. 2005, Jung et al. 2017. In particular, the expansion of cropped and pastoral land within natural ecosystems is the most important form of land conversion, by area (Lambin & 52 Meyfroidt 2011). Farming is more damaging to wild nature than any other sector of human activity (Balmford et al. 2012) and poses the greatest extinction risk to birds, especially in developing 54 countries (Green et al. 2005). In much of sub-Saharan Africa the expansion of agricultural habitats, particularly cultivated land, has occurred mainly at the expense of natural grassland, savanna and 56 forests (Brink & Eva 2009), with profound effects on their ecological assemblages (Newbold et al. 2017). Similarly, the replacement of wild herbivore communities with domestic livestock has had 58 substantial impacts on a range of ecosystem processes, contributing towards increased woody cover and a rise in herbivore methane emissions (Hempson et al. 2017). 60 While land use change has impacted severely on the extent, continuity and quality of terrestrial habitats, the loss of predators, scavengers and other apex consumers may have an equally 62 pervasive influence on the natural world, due to the extensive cascading effects that follow their disappearance (Estes et al. 2011, Dirzo et al. 2014. In Africa, these effects include the potential 64 loss of ecosystem services provided by vultures and other avian scavengers, which are likely to inhibit disease transmission, through the rapid disposal of carcasses (Ogada et al. 2012). The loss 66 of this service in India has been described in a well-documented trophic cascade, wherein the collapse of vulture populations was followed by a substantial rise in the feral dog population, 68 which in turn contributed to a $34 billion increase in healthcare costs associated with rabies treatment in humans (Sudarshan et al. 2007, Markandya et al. 2008. 70 For many African raptors the impacts of farmland conversion have been intensified through a range of anthropogenic effects, which include incidental and deliberate poisoning, linked mainly to 72 the illegal killing of livestock predators and elephants (Otieno et al. 2010, Virani et al. 2011, Ogada 2014, Ogada et al. 2015, 2016, Monadjem et al. 2018. In West and Central Africa, large raptors are 74 also killed for bushmeat (Buij et al. 2016), while trade in raptor body parts for traditional medicines is widespread, occurring in at least 19 African countries (McKean et al. 2013, Williams et 76 al. 2014. Human disturbance can also adversely affect both tree-and cliff-nesting species (Borello & Borello 2002;Monadjem & Garcelon 2005;Bamford et al. 2009), while energy infrastructure 78 poses a significant, growing threat to larger species, through collisions and electrocution (Jenkins et al. 2010, Rushworth & Krüger 2014, Kibuule & Pomeroy 2015. 80 The impacts of these pressures have attracted considerable attention, reflecting their scale, the graphic evidence they generate and their recent dramatic rise, particularly in the case of vulture 82 poisoning (Ogada et al. 2016). In contrast, the effects of land use change on African raptor populations are more diffuse, and perhaps more difficult to quantify. Much of the transition to 84 agriculture coincided with the colonial period, and hence pre-dates the standardised collection and analysis of biological survey data. Furthermore, for many observers the extent to which land use 86 has changed may be obscured by shifting baseline syndrome, each generation viewing the conditions they encounter as the new norm, and focusing only on the extent to which these have 88 changed over their own lifetime (Papworth et al. 2009).
While the effects of land use change on biodiversity may be difficult to quantify, its scale, and that 90 of human population growth, are comparatively well documented. Between 1960 and 2016 the human population of sub-Saharan Africa increased by 0.8 billion (Canning et al. 2015, World Bank 92 2017a. During part of that period  the area of agricultural land in sub-Saharan Africa increased by 57%, mainly at the expense of natural vegetation, which contracted by 21%, 94 with a loss of almost 5 million ha of forest and non-forest natural vegetation per annum (Brink & Eva 2009). The human population is projected to increase by a further 1.8 billion during 2016-96 2060(Canning et al. 2015, generating an unprecedented surge in the demand for food. While the FAO has estimated that some 80% of this demand may be addressed through higher yields and 98 increased cropping intensity (Bruinsma 2009), the shortfall will have to be met through farmland expansion. In sub-Saharan Africa the expected increase in arable land alone has been estimated at 100 64 million ha by 2050 (Bruinsma 2009).
Despite the geographic scale of land use conversion in Africa there have been few long-term 102 studies quantifying its impacts on bird communities. Notable exceptions are the raptor road surveys conducted in West Africa (Thiollay 2006a,b,c) and Northern Botswana (Herremans &  104 Herremans-Tonnoeyr 2000, Garbett et al. 2018), which reported substantial declines, both within protected areas (PAs) and farmland, and across a range of feeding guilds. Not surprisingly, raptor 106 encounter rates in both regions were higher in PAs than in surrounding farmland, particularly for eagle and vulture species. These and other effects have been examined in West Africa by Buij et al. 108 (2013), who concluded that while some Palearctic raptors may benefit from cropland expansion, the majority of Afrotropical and insectivorous Palearctic raptors are likely to decline in the face of 110 further agricultural intensification. Declines are likely to be particularly severe among larger raptor species, reflecting the pattern of extinction risk evident among avian scavengers and 112 mammalian predators; larger species being disproportionately threatened and among the first to disappear (Fritz et al. 2009, Di Marco et al. 2014, Dirzo et al. 2014, Ripple et al. 2014 Şekercioğlu 2016).
In East Africa, the most extensive areas of land suitable for agricultural conversion, by virtue of 116 being non-forested, unprotected and supporting a low human population density, lie in South Sudan, the Democratic Republic of the Congo (DRC) and Tanzania (Lambin & Meyfroidt 2011). As a 118 step towards evaluating the likely impacts of farmland conversion on raptors, we assessed the abundance, species richness and mean body mass of raptors in relation to land use in neighbouring 120 Uganda. In common with most African countries, Uganda has undergone significant changes in land use over recent decades. 2013), we expected the majority of raptor species surveyed to be more abundant in protected savanna than in either farmland type. Second, we expected species richness (the number of species 132 detected over a given distance) to be higher in protected savanna than in pastoral or agricultural land. Third, we predicted that disparities in encounter rates in relation to land use would be more 134 pronounced in the case of large, resident raptors than for smaller, migratory species.

Data collection 138
We recorded the number of individuals of each diurnal raptor species seen while driving a series of transects along roads and tracks in Uganda, during January (86% of surveys), February (10%) and 140 March (4%), 2008-2015 (Tables A1, A2). Owl species were likely to be substantially underrecorded, and hence were excluded from the survey. Forty transects, of 9-122 km in length 142 (recorded by odometer), were surveyed at a mean of 33 km hr -1 on public roads, and 25 km hr -1 in National Parks. Most transects were surveyed once per annum over the eight-year period, and the 144 total distance surveyed was 11,188 km. 950-1100, 1150-1400 m), rainfall band (800-950, 1000-1150, 1200-1400 mm) and tree cover (categorical). We specified transect length (log transformed) as an offset, and used a log-link 190 function. Most transects were surveyed annually, yielding 226 transect-surveys in which the factors listed above were all recorded. To control for the effects of pseudo-replication, we entered 192 'transect identity' and 'year' as random terms.
Count data for scarce species typically follow a Poisson distribution, but one in which the amount 194 of variation per sampling unit (e.g. per transect-survey) may be higher than expected, or overdispersed (Linden and Mantyniemi 2011), in which case a negative binomial model may give an 196 improved fit. We therefore fitted models with both a Poisson and a negative binomial distribution, calculating the variance for the latter either as φµ ('NB1') or as μ(1+μ/k) ('NB2') (Linden and198 Mantyniemi 2011, Brooks et al. 2017). For each of these three models we ran a zero-inflated and a non-zero-inflated version, yielding six model types ( function to derive the number of encounters predicted for each transect-survey. We then calculated the predicted encounter rate for each transect-survey (from the length of the transect), 204 and the mean encounter rate predicted for that species within each land use type.
Differences in encounter rates for a given species in each land use type could reflect variation in 206 both its detectability and its abundance. We therefore compared detection patterns (the proportion of detections made in each distance band) of a given species in different land use types, 208 e.g. contrasting the pattern of detections made in protected savanna with that in pastoral land. We applied Kruskal-Wallis tests to identify, then exclude, species whose detection patterns in one land 210 use type differed significantly from that in another. Where fewer than 20 detections had been made in the two land use types being considered, we pooled observations for the relevant genus, 212 and excluded the species in question if the detection patterns shown by members of its genus differed significantly between the two land use types being considered (Kruskal-Wallis test).

214
Species retained for pair-wise land use comparisons are identified in Table A1.
The difference between the mean encounter rates predicted for a species in two land use types was 216 expressed as a proportion of its encounter rate in protected savanna, in pairwise comparisons between savanna and either pastoral or agricultural land. Similarly, the difference in encounter 218 rates between pastoral and agricultural land was expressed as a proportion of the rates recorded in the former. We used linear mixed-effects models to test whether these differences varied in 220 relation to diet, mass, migratory status or threat status.  (2013), and global threat status (threatened/near-threatened vs least concern) was obtained from BirdLife International (2018). 226 For each pair of land use types we entered the proportional difference in the species' encounter rate as the dependent variable. Since sample sizes were small, we limited each model to one fixed 228 factor (median body mass, diet, migratory status or threat status). Because some genera (e.g. Gyps) were represented by multiple species, we included 'Genus' as a random effect (Table A3: Model 2).

230
We selected the model with the lowest AIC score for each pairwise comparison. Probability estimates for the effects of each explanatory variable were calculated using the Kenward-Roger 232 approximation (Halekoh & Højsgaard 2014).

Body mass 234
To further compare body mass differences in relation to land use, we calculated the total mass of all individuals seen on each survey of a given transect (using mass values given in Table A1), and 236 divided this by the number of individuals, to give the mean mass of individuals seen per transectsurvey. We assigned these values to 250 g intervals, to examine their frequency distribution in 238 relation to land use. We then used a linear mixed-effects model to generate predicted values, specifying a natural log transformation of the mean mass as the dependent variable, and the 240 following fixed factors: transect length (km, log transformed), the presence/absence of 'outside' observers (binary), land use category, tree cover, mean altitude band (m) and mean annual rainfall 242 band (mm) (categorical) ( (±SE) predicted mass value for each land use type.

Species richness 248
To investigate the relationship between species richness and land use we plotted the cumulative number of species encountered during successive transect-surveys within each land use type, 250 against the cumulative distance travelled. We excluded transect-surveys with missing data for tree cover or other factors. This approach shows the pattern of change in the number of 'new' species 252 encountered over the (cumulative) distance surveyed, which totalled 5,031 km (protected savanna), 2,315 km (pastoral land) and 2,635 km (agricultural land). Since the pattern observed 254 may have been influenced by factors other than land use, and involved the repeated sampling of transects, we further examined the relationship between species richness and land use using a 256 linear mixed-effects model. In the model, we entered the number of species encountered on each transect-survey as the dependent variable, and the following variables as fixed effects: transect 258 length (km, log transformed), the presence/absence of 'outside' observers (binary), land use, altitude, mean annual rainfall and tree cover (categorical) ( Table A3: Model 4). Since we expected 260 the number of species encountered to vary both in relation to land use and transect length, we included an interaction between these two variables, which improved the model fit (ΔAIC > 2).

262
'Transect identity' and 'year' were entered as random terms, to control for repeated sampling of the same transects and years. All (53) raptor species seen were included in this analysis (Table  264 A4). To calculate the number of species predicted by the model we specified constant (modal) values for: 'outside' observers (present), altitude band (950-1100 m), rainfall band (1000-1150 266 mm) and tree cover ('light').

Encounter rates in relation to land use 270
Over the eight survey years, 6,708 individuals of 53 raptor species were detected. Thirty-three species were seen in sufficient numbers to enable us to model encounter rates in relation to land 272 use and other factors (Table A4). Of 23 species whose detection patterns (in relation to distance from the transect) were comparable in protected savanna and pastoral land (Table A1), 15 were 274 less abundant in the latter. The median difference in the rate at which they were encountered was -41% (quartiles +40% to -80%; n = 23 species; Wilcoxon matched-pairs test: P = 0.088). A much 276 greater disparity was evident between protected savanna and agricultural land; 19 out of 24 species were less abundant in the latter, and the median difference in their encounter rates was -278 90% (quartiles -31% to -100%; n = 24; Wilcoxon matched-pairs test: P < 0.003). Encounter rate differences between pastoral and agricultural land were also significant; 10 out of 14 species were 280 less abundant in the latter, with a median difference of -52% (quartiles -2% to -83%; n = 14; Wilcoxon matched-pairs test: P = 0.025) (Fig. 1). 282 When the same comparisons were made with unmodelled data, median differences in encounter rates were broadly similar to those obtained from modelled data: a median of -48% between 284 protected savanna and pastoral land, (quartiles +16% to -74%; n = 23 species; Wilcoxon matchedpairs test: P = 0.041); -88% between protected savanna and agricultural land (quartiles -45% to -286 98%; n = 24; Wilcoxon matched-pairs test: P < 0.001); and -41% between pastoral and agricultural land (quartiles -12% to -70%; n = 14; Wilcoxon matched-pairs test: P = 0.013) (Fig. 1). 288 In separate models, encounter rate differences for the same species in savanna and pastoral land were correlated with body mass and threat status. Heavier species and those of conservation 290 concern showed a greater drop in abundance on pastoral land than lighter species and those of least concern. The first of these models (incorporating body mass) provided the better fit (Table 1, 292 Fig. 2). Similarly, encounter rate differences in savanna and agricultural land were significantly correlated with diet, body mass and threat status, with the former model (incorporating diet) 294 providing the best fit (Table 1, Fig. 2) . Species specialising in predating small mammals or reptiles were significantly more abundant in agricultural land than in savanna, compared with generalist 296 species (Table 1). Body mass, diet, migratory-and threat status had no significant influence on encounter rate disparities between pastoral and agricultural land. 298

Body mass in relation to land use
The mean body mass of all raptor individuals encountered during transect-surveys varied more 300 widely in protected savanna than in pastoral or agricultural land, the two farmland types supporting a much more homogenous raptor community, with regards to size (Fig. 3). Predicted 302 average body mass values for birds seen from transects through pastoral land (mean = 995±25.6 g(SE); n = 47 transect-surveys) and agricultural land (mean = 856±18.1 g; n = 42) were 43% and 304 51% lower than those seen in protected savanna (mean = 1,740±63.0 g; n = 131). Since vultures are heavier than most other raptors, and were more abundant in protected savanna, we re-306 examined the relationship after excluding vulture species from the model. The pattern observed was broadly similar, however, body mass averaging 933 g (±27.6 g; n = 47 transect-surveys) in 308 pastoral, 824 g (±16.6 g; n = 42) in agricultural land and 1,332 g (±37.0 g (SE); n = 130) in protected savanna. Results from unmodelled data were similar with regards to body mass 310 variation in relation to land use (Table A5).
Disparities in encounter rates between protected savanna and both pastoral and agricultural land 312 were thus linked to body mass. To test this further, we examined encounter rate differences for small species (1 kg) and large species (1 kg) within the three pairwise land use comparisons. 314 Small species were more abundant (median difference: +44%; n = 10) and large species significantly less abundant in pastoral land than in protected savanna (median: -76%; n = 13; 316 Kolmogorov-Smirnov test: D = 0.615; P < 0.03). Both size classes were less abundant in agricultural land than in protected savanna, but to differing degrees. The median disparity for small species (-318 30%; n = 12) was less pronounced than that for large species (-97%; n =12; Kolmogorov-Smirnov test: D = 0.667; P < 0.01). Body mass effects on encounter rate differences in pastoral and 320 agricultural land were not significant (small species: -63%; large species: -40%).
In a linear mixed effects model restricted to large (1 kg) species, the disparity between encounter 322 rates in protected savanna and pastoral land increased significantly in relation to body mass (disparity = -0.547*log(mass) + 3.712; n = 13 species; P<0.02). This indicates that the disparity 324 widened by a further 18 percentage points for each 1 kg increase in body mass.

Threat status 326
Nine of the species examined were of global conservation concern, being listed as Critically Endangered (four species), Endangered (two), Vulnerable (one) or near-threatened (two species) 328 (BirdLife International 2018). Species of global conservation concern were heavier on average (4,075 g; n = 9) than those of least concern (976 g; n = 21; Kolmogorov-Smirnov test: D = 0.794; P < 330 0.001). Since heavier species were significantly less abundant in farmland than in protected savanna, similar disparities were evident with respect to threat status. Species of conservation 332 concern showed a significantly greater drop in encounter rates between protected savanna and pastoral land (median difference: -87%; n = 5), than those of least concern (median difference: -334 5%; n = 18 species; Kolmogorov-Smirnov test: D = 0.689; P = 0.049). Similarly, encounter rate differences between protected savanna and agricultural land were much greater for species of 336 conservation concern (median difference: -100%; n = 8), than for those of least concern (median difference: -42%; n = 16 species; Kolmogorov-Smirnov test: D = 0.750; P = 0.005). Only one species 338 of conservation concern was likely to have benefitted from farmland conversion; Hooded Vultures were recorded 58% more frequently in agricultural land than in pastoral land. Encounter rates for 340 this species in pastoral and agricultural land were both higher than in protected savanna, however pairwise comparisons were confounded by differences in the species' detectability in the latter. 342

Species richness
We recorded 48, 42 and 31 diurnal raptor species in protected savanna, pastoral and agricultural 344 land, respectively, partly reflecting differences in the cumulative distances surveyed in these land use types (Fig. 4A). Over the first 100 km surveyed, the number of species encountered had already 346 begun to diverge, being 13.6 species (by interpolation) in savanna, 9.4 in pastoral land and 6.1 in agricultural land. By 2,000 km, disparities in species numbers were proportionally less 348 pronounced: 45 and 38 species in savanna and pastoral land, 27 species in agricultural land. In protected savanna, species number levelled off after a cumulative survey distance of c. 3,500 km, 350 but showed no indication of doing so within the (shorter) distances surveyed in pastoral and agricultural land (Fig. 4A). 352 A linear mixed-effects model was used to control for the effects of survey-and habitat variables, and for the repeated sampling of transects (Table A3: Model 4). This confirmed that the number of 354 species encountered on each transect varied in relation to land use and length, yielding predicted totals of 13.2 species in protected savanna, 10.6 in pastoral land and 6.7 in agricultural land, on 356 transects of 100 km (Fig. 4B).

Discussion
We show that raptor encounter rates were 41% lower in pastoral land and 90% lower in 360 agricultural land than in protected savanna. In addition, encounter rates in agricultural land were 52% lower than in pastoral land, despite the latter being already depleted, mainly through the loss 362 of large, scavenging species. This disparity is of particular relevance, since the 64 million ha expansion in agricultural land required to meet growing food demands by 2050 (Bruinsma 2009) 364 is likely to be achieved mainly through the conversion of land already supporting pastoralism to some degree (Lambin & Meyfroidt 2011). Our findings suggest that such areas are likely to 366 experience a median decline in raptor abundance of the order of 50% if converted to agriculture. In areas still largely comprising intact savanna, raptor abundance is likely to decline by a median of c. 368 90%, or higher in the case of large eagles and vultures.
Similar abundance patterns have been observed elsewhere in Africa. In West and southern Africa 370 the relationship between raptor abundance and land use is influenced both by body size and migratory status; large, resident species are more sensitive to land use change than small 372 Afrotropical or Palearctic migrants (Herremans & Herremans-Tonnoeyr 2000, Thiollay 2006c, Anadón et al. 2010, Buij et al. 2013). It has been suggested that non-breeding migrants are better 374 able to tolerate the disturbance associated with farming activities than resident species, which tend to remain on their territories year-round, and hence avoid areas subject to disturbance when 376 they are breeding. Furthermore, larger species are more likely to suffer from hunting pressure, through direct persecution (for bushmeat) and through the loss of their prey base (Thiollay 2006c) 378 or of large trees in which to nest. Since large species tend to require larger territories, they are also less likely to persist in small fragments of suitable habitat. Our results were broadly consistent 380 with these findings, in showing a link between abundance disparities, diet and body mass.
A more direct analysis of the effects of land conversion has been made in the Serengeti ecosystem, 382 Tanzania, where Sinclair et al. (2002) compared bird species abundances in protected savanna with those in adjacent areas converted from savanna to cultivated land in the 1950s. Some 50 384 years later, insectivores and granivores/frugivores were 77% and 60% less common in the farmland plot than in the adjacent protected savanna. Furthermore, while the study recorded 104 386 individuals of 15 raptor species in protected savanna, only four individuals of three raptor species were recorded in the neighbouring farmland. 388 These deficits are broadly similar to those reported here, and consistent with findings reported by Child et al. (2009) in South Africa, in suggesting that African raptor species are particularly 390 sensitive to farmland conversion. Child et al. (2009) showed that among nine functional bird groups examined, scavengers and raptors most often suffered a decrease in richness within 392 agriculturally dominated landscapes.
An underlying assumption of the current study is that raptor populations within Uganda's four 394 savanna national parks represent a baseline from which farmland communities have departed. Survey transects within these PAs overlapped extensively with those in farmland, in terms of 396 altitude, and received a similar level of rainfall to those in pastoral land (1,019 vs 950 mm). Within agricultural areas, however, mean annual rainfall was slightly higher (1,178 mm), and the land 398 perhaps more likely to have once supported a mosaic of savanna and forest (Langdale-Brown et al. 1964). In addition, the public roads surveyed lay mainly in the southern half of the country (Fig.  400 1A), and hence might not have accurately reflected raptor abundances further north, where there are larger, continuous expanses of pastoral land. 402 Disparities between encounter rates in savanna and farmland could have been magnified by the greater disturbance effects associated with public roads in farmland areas. Species deterred by 404 traffic disturbance, housing and the higher human population densities associated with public roads may have been more abundant at greater distances from these roads. That is, our approach 406 may have under-estimated species abundances in farmland, where road-related disturbance levels are likely to have been higher than in protected savanna, where traffic volumes and human 408 numbers are low.
Our paired land use comparisons were restricted to species whose detectability did not differ 410 significantly between the two land use types under comparison, and were made using values predicted from GLMMs, which controlled for the effects of potentially confounding variables, and 412 for repeated sampling of transects. For heavier species and those of conservation concern, encounter-rate disparities between protected savanna and each farmland type were significantly 414 more pronounced than among lighter species and those of least concern. Furthermore, surveys within protected savanna yielded more raptor species (over a given distance; Fig. 4), showing a 416 wider variation in body mass (Fig. 3). The much greater uniformity in body mass evident in pastoral and agricultural land only partly reflected the near-absence of vulture species from these 418 landscapes.

Conservation management
While African farming systems typically involve simpler, non-mechanised methods and fewer 422 chemical treatments than in Europe and North America, their impacts on bird species adapted to savanna or wooded habitats can be profound (Sinclair et al. 2002, Child et al. 2009 2013, Renwick et al. 2014). Uganda is unusual within Africa, in that much of its land has already been converted to crop production, the impacts of which have recently become the focus of agri-426 environmental research (Hulme et al. 2013, Renwick et al. 2014. As in western countries, mitigation efforts are likely to follow either of two contrasting approaches: land sharing, in which 428 low-yield, 'wildlife-friendly' farming is promoted, at the expense of semi-natural land; and land sparing, in which farmers strive for higher yields, while leaving aside larger fragments of semi-430 natural land (Hulme et al. 2013). Theoretically, land sparing should ensure that more of the original savanna is retained in perpetuity, affording a refuge for species poorly adapted to 432 synanthropic conditions, and a benchmark against which to gauge the effects of human interventions elsewhere (Sinclair et al. 2002). 434 In this study, raptor abundance and species richness in agricultural land were such that typical land sharing measures are unlikely to prove effective in retaining or re-establishing viable 436 populations, except in the case of synanthropic species. Thus, on land deemed suitable for farmland conversion, including large tracts of South Sudan, DRC, Tanzania and Mozambique 438 (Lambin & Meyfroidt 2011), conservation efforts should focus instead on identifying and safeguarding the largest remaining expanses of unprotected, relatively intact savanna. Here, and 440 within existing protected areas, efforts should focus on retaining intact raptor communities. In Uganda, such efforts would include the following. First, exclude or minimise anthropogenic 442 disturbance of protected areas (e.g. pollution from an ongoing oil exploration programme in and around Murchison Falls NP). Second, allow pastoral areas bordering PAs to revert to savanna, 444 particularly where they might form a bridge or corridor between PAs supporting globally threatened resident species. These might take the form of community-run conservancies or private 446 game reserves, which have proved successful in boosting game populations elsewhere in East Africa and in southern Africa. Third, factors contributing to the observed disparities in raptor 448 abundance among the three land use types examined here should be identified and addressed. Together, such initiatives could help to counteract the biological impoverishment associated with 450 farmland expansion, and ensure the survival of intact raptor communities. 452

Acknowledgements 454
We thank Uganda Wildlife Authority for granting us permission to make road counts in protected 456 areas, and for providing their most experienced rangers to assist with the survey. Several volunteers from NatureUganda acted as recorders. Our thanks go to Taban Bruhan, Judith 458 Mirembe, Roger Skeen, Lydia Tushabe and Lilian Twanza. Roger Skeen also kindly provided count data for Kidepo Valley National Park. Murn Campbell and Darcy Ogada made valuable comments 460 on an early version of the manuscript. We also thank Greame Cumming, Johannes Kamp and two anonymous reviewers for their very helpful comments. We are especially grateful to Will Cresswell 462 for providing guidance on modelling in R. The cost of fieldwork was generously covered by The Peregrine Fund (USA) and the Royal Society for the Protection of Birds (UK).     Table 1 620 Variables associated with differences in raptor encounter rates within pairs of land use types. A negative effect, e.g. with respect to body mass, for 'Savanna vs Pastoral', indicates that the mean body mass of species detected on 622 pastoral transects was lower than on protected savanna transects. Since sample sizes were low, potential explanatory variables (body mass, migratory status, threat status and diet) were examined in separate models.

624
Parameters from models showing statistically significant effects are shown below, the best fitting model being that with the lowest AIC value (ΔAIC = 0.00

652
indicates that correspondingly fewer individuals were seen in the latter. Globally threatened or nearthreatened species (black symbols) tended to be heavier and showed a greater drop in abundance than 654 most species of least concern (grey symbols). Diamond symbols indicate scavenging vultures; circles indicate other raptor species    5 Species whose encounter rates were included in a pairwise comparison between Savanna and Pastoral land, i.e. their detection patterns,

718
with respect to distance from the transect, did not differ significantly between these two land use types.    The difference between the mean fitted encounter rate for a species in two land use types, expressed as a proportion of its encounter rate in protected savanna (in pairwise comparisons between savanna and either pastoral or agricultural land) or pastoral land (in pairwise comparisons between pastoral and agricultural land) 738 3 A natural log transformation of the mean body mass of individuals of all species encountered on a given transect survey 4 Six transect-surveys on which no raptors were seen have been excluded 740 26 Table A4 Unmodelled and modelled encounter rates (birds seen 100 km -1 ) for 53 and 33 raptor species, respectively. Modelled encounter rates are the rates predicted from GLMMs, after controlling for the effects of variation in transect length, the presence of outside observers, land use, altitude, rainfall and tree cover, and the repeated sampling of the same transects and years. Differences in the sample sizes from which modelled and unmodelled estimates were drawn reflect missing values for some of the variables used in the models. Encounter rates were modelled only for species with at least 10 sightings  Table A5 A. The mean body mass of all individuals of (53) raptor species encountered during transect-surveys through protected savanna, pastoral and agricultural land, calculated from unmodelled data. The mean body mass was calculated for each transect-survey, from the number of individuals seen per species, and the species' median mass (Table A1; from del Hoyo et al. 2017). Mean body mass was much higher in savanna than in pastoral or agricultural land, and these differences remained when vulture species were excluded.

B.
Pairwise comparisons of encounter rates for large (>1 kg) vs small species, in relation to land use, using unmodelled data. Each comparison shows the median difference in encounter rates between protected savanna and either pastoral or agricultural land. A negative value indicates that the encounter rate in savanna was higher. Large species showed a significantly greater drop in abundance than smaller species A.

All species
Excluding vultures