Population responses of farmland bird species to agri-environment schemes and land management options in Northeastern Scotland

1. The decline of farmland birds across Europe is a well- documented case of biodiversity loss, and despite land stewardship supported by funding from agri-environment schemes (AES), the negative trends have not yet been reversed. 2. To investigate the contribution of AES towards farmland bird conservation, we compared abundance of five farmland bird species across 13 years and 53 farms (158 farm years = AES, 72 farm years = non AES) in Northeastern Scotland (UK), a region with relatively mixed farmland. 3. Between 2003 and 2015, on both AES and control farms, skylark ( Alauda arvensis ) showed a nonsignificant decline, and tree sparrow ( Passer montanus ) and yellowhammer ( Emberiza citrinella ) nonsignificant increases, whereas reed bunting ( Emberiza schoeniclus ) and linnet ( Carduelis cannabina ) populations remained relatively stable. 4. We did not detect a significant association between AES and avian abundance or population trends for any of these species, but there were positive associations with some AES management options. 5. Possible explanations for the lack of a significant AES- bird abundance association include poor uptake of the best AES

Comprehensive AES audits depend on assessing scheme performance across the full range of landscape complexity, but most AES studies have been in places where farming systems are relatively intensive (e.g., Bright et al., 2015;Merckx et al., 2010, but see Concepción, Díaz, & Baquero, 2008;Santana et al., 2017 for studies along a land-use intensity gradient). For example, autumn-sown wheat and oilseed rape, two intensively managed crops, cover 44% and 16% of arable land in England, compared with just 10% and 7% in Northeastern Scotland, where less intensively managed spring-sown barley is the dominant crop (55% of arable land, cf. 7% in England) and most lowland farms are mixed arable-livestock enterprises (FERA, 2018, Scottish Government, 2010. Although some studies have considered regional variation in the efficacy of English AES in high-intensity farmland (e.g., Davey et al., 2010), a key remaining question is how AES implementation affects biodiversity in heterogeneous, low-intensity farmland.
Here, we examined the association between long-term AES implementation and abundance and population trends of five farmland bird species in Northeastern Scotland, an area dominated by relatively low-intensity mixed farming. Specifically, we hypothesised that: (a) AES has a positive effect on species abundance, (b) the effect becomes stronger over time under AES treatment, (c) AES farms have more positive population trends, and (d) AES management options that match species' ecological requirements have a stronger positive effect on abundance. We present a novel application of a multimembership model, which allows us to estimate the effects of different management options, whilst also accounting for differences in the area they cover and the different combinations in which they are implemented. Finally, we also conducted a post hoc power analysis to determine the power of our survey design to detect different effect sizes of AES on population trends in low-intensity farmland.

| Study sites
Our study sites were 53 lowland farms in Northeastern Scotland ( Figure 1). There were two farm treatments-AES (mean farm area = 120 ha ± SD 56) and control (conventional farming, mean farm area = 103 ha ± SD 58). For details on site selection see Perkins, Maggs, Watson, & Wilson, 2011. Land-use types were similar between farms, with 95% of study sites supporting mixed farming practices (Supporting Information, Table S1). AES farms participated in the national Rural Stewardship Scheme (RSS, 2001, Scottish Government, 2006, its successor Rural Priorities (RP, 2009, SRDP, 2014, and Farmland Bird Lifeline (FBL), a local intervention scheme targeting corn buntings. AES management varied between farms, with participants selecting several management options from 33 available in RSS, 49 in RP and 7 in FBL (Supporting Information   Table S2). Some farms switched treatment between years due to entering or leaving AES agreements which ran for 5 years in the national schemes and were renewed on an annual basis in FBL (see Supporting Information Table S1 for yearly sample sizes). The average treatment duration was 5 years ± SD 4 for AES and 4 years ± 3 for control farms. Twenty-one farms remained under AES management for the entire study duration (13 years).

| Data collection
Farms were visited two or three times (70% of data points based on three visits) between May and mid-August of 2003, 2004, 2006, 2008, 2009, and 2015, as part of a corn bunting (Emberiza calandra) monitoring project during which other farmland birds were also counted (Perkins, Maggs, Stephan, Corrigan, & Wilson, 2016;Perkins et al., 2011). Survey routes scaled positively with farm size and passed within 250 m of all points on the farm, largely following field boundaries. Since the surveys were designed for corn buntings which nest from May to August and favour open landscapes, some late visits (mid-July to August) and survey routes might not have been optimal for all study species, e.g., detection rates for skylarks are low once they stop singing in late summer, and tree sparrows frequently occupy farm woodlands which survey routes often avoided. Nevertheless, the survey design, combined with highly experienced fieldworkers, was considered sufficient to detect a high proportion of birds of each study species. During each visit, bird location and activity were recorded on a 1:10,000 map. The maps were superimposed onto each other to determine the number of breeding birds from clusters of registrations.
We chose recording units based on species behaviour in a manner that optimises detectability. For yellowhammer, reed bunting and skylark, we recorded territorial males, based on singing birds. For tree sparrow and linnet, two species which sing less frequently and are semicolonial, singing birds underestimate abundance, so we recorded numbers of apparently breeding pairs (birds displaying breeding behaviour in suitable nesting habitat).

| Data analysis
All statistical analyses were conducted in r v. 3.4.2 (R Core Team, 2017). Overall bird abundance, trends, and AES effects were estimated using generalized linear mixed models (GLMMs), assuming a Poisson distribution (due to the count nature of the data) and using Bayesian inference via the package MCMCglmm (Hadfield, 2010).
GLMMs have been widely used in ecology, particularly because they can efficiently model multiple predictors whilst also accounting for variance introduced by the structure of the data and the experimental design (e.g., records from the same farm and/or year might be more correlated than those among farms/years, Bolker et al., 2009;Harrison et al., 2018). Models were run for 600,000 iterations, with a burn in period of 20,000 iterations, a thinning factor of 10, and parameter expanded priors which improve model convergence.
Convergence was assessed using visual inspection of the autocorrelation in posterior distributions, and the effective sample size was in excess of 1,000 for focal parameters.

| Overall AES associations with species abundance and number of years in AES
To test overall associations between AES and species abundance, we modelled farm treatment type (AES/control), whilst accounting for the confounding effects of farm area (log transformed and mean centred), location (latitude and longitude mean centred), and survey effort (visits: 2/3). We included latitude and longitude as fixed effects to account for a potential latitudinal land-use intensity gradient in Scotland (decreasing towards the north) and a longitudinal climate gradient going from continental to coastal climate. Farm identity and observation year were treated as random effects to account for some of the spatial and temporal nonindependence of data points. It is common for studies assessing temporal trends to include year as a fixed effect but not as a categorical random effect. However, where there are multiple observations for each year, observations made in a single year may be subject to the same year-specific effects arising from drivers not included in the model (e.g., weather). Including year as a random effect takes this pseudoreplication into account, whereas a model that does not include year as a random effect will tend to underestimate SEs (thereby inflating type I errors) of coefficients estimated for predictor variables that change over time.
We worked with the full models, and did not execute step-wise term deletion and model comparison which have been criticised for increasing the probability of type I errors (Mundry & Nunn, 2009;Whittingham, Stephens, Bradbury, & Freckleton, 2006). Reported b values are the slope estimate (effect size), and p values refer to MCMC p values. A predictor was considered "significant" when the 95% credible intervals (CI) for the corresponding model parameter did not overlap zero.
We examined whether AES effects become more pronounced as treatment duration increases by including "AES years" as a continuous fixed effect to the models testing for an association between AES treatment and bird abundance. "AES years" was set to zero for farms that never entered AES management, i.e., control farms, and equalled one or above for AES farms, depending on the number of years each AES farm was in a scheme.

| Population trends on AES and control farms
We modelled avian population trends during the study period (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) by adding year (mean centred) as a continuous fixed effect to the models outlined above (one model per species). To determine if population trends differed between AES and control farms, we ran five additional models which also included the year by treatment interaction term as a fixed effect.

| Sensitivity of species abundance to specific AES management options
To examine the associations between specific AES management options (Supporting Information Table S2) and overall bird abundance, we conducted an analysis using a multimembership modelling framework through the package MCMCglmm (Hadfield, 2010). This analytical approach is particularly appropriate for data with complex structure, where records belong to more than one level of classification (e.g., bird abundance on farms where multiple AES management options were implemented simultaneously). Our model structure included farm area (log transformed and mean centred), location (latitude and longitude mean centred), and survey effort

| Spatial autocorrelation analysis
To test for spatial autocorrelation in species' abundance, we ran spa-  Table S9 and Figure S1).

| Offset analysis
Conservation biologists often fit area as a linear offset to convert counts into densities. If survey effort scales perfectly with area, and abundance increases linearly with survey area (i.e., a slope of 1), estimating density this way requires one fewer degree of freedom. However, survey design, habitat configuration and/or species' ecology may cause the observed slope to depart from 1, and in such cases treating area as an offset will impact on the coefficients of additional predictors included in the model (Helzer & Jelinski, 1999). This is especially likely in farmland, where crop fields and grass are separated by narrow strips of semi-natural habitat, meaning that variation in farm size will affect habitat availability for crop-nesting and boundary-nesting species differently.
To examine how including area as an offset would impact on our inferences, we repeated the analyses using area as an offset rather than as an estimated coefficient. Slope estimates of log farm area as a fixed effect term in MCMC models showed high variability between species and all except reed bunting were well below 1 (0.48-0.72, Supporting Information  Information Tables S5 and S10). Although the two modelling approaches gave similar results, inclusion of an offset when the relationship with area departs from 1 has potential to generate incorrect inferences for other model terms. Therefore, we recommend that the effect of area should be estimated instead of fixed.

| Overall AES associations with species abundance and number of years in AES
We found no significant association between AES and the abundance of five bird species, and model predictions for bird abundance were similar on AES and control farms (Figure 2a, Supporting Information Table S4). The coefficients for the effect of the AES treatment relative to control were small, ranging from 0.19 to −0.06 on the log scale or a population size difference of +20% for tree sparrow and −6% for skylark ( Figure 2b). There was no significant effect of AES treatment duration on species abundance (Supporting Information Table S7), and model coefficients for the effect of years in AES were very small (Figure 2c).

| Population trends on AES and control farms
Between 2003 and 2015, the five bird species we monitored did not experience net directional changes. We observed directional albeit nonsignificant trends for three species-skylark declined in abundance, whereas tree sparrow and yellowhammer abundance increased ( Figure 3, Supporting Information Tables S5 and S6). Overall, we did not detect a significant difference between linear population trends on AES and control farms, although for some species (e.g., skylark between 2004 and 2008), trends on AES and control farms diverged (Supporting Information Table S6).

| Sensitivity of species abundance to specific AES management options
We found few positive (and no negative) associations between bird abundance and particular AES management options (Figure 4).

| Power analysis
A post hoc power analysis revealed that our survey design (53 farms surveyed in 6 years [2003, 2004, 2006, 2008, 2009 and 2015], with 62%-79% of farms in AES in any 1 year) had sufficient power (i.e., >80% probability of detecting a significant effect) to detect strong AES associations should they have been present (e.g., AES having 25% higher overall bird abundance than the control treatment) for two of five study species (skylark, yellowhammer, for which more records were available, since they were more common, Supporting Information Table S11). Distinguishing subtler AES effects for all study species (e.g., AES having only 10% higher bird abundance than the control treatment), however, poses a challenge, as detecting such small effects sizes and effects for rarer species would require monitoring of a substantially larger sample size of farms with equal representation of AES and control treatments.

| D ISCUSS I ON
We found no association between AES participation and bird abun- of landscape complexity on species' responses to AES is driven by metapopulation dynamics (Durell & Clarke, 2004), and habitat and resource availability on farms and the surrounding area (Batáry et al., 2011;Concepción et al., 2008). Implementing AES in high-intensity farmland creates a high ecological contrast, thus increasing their additionality, and resulting in an easier to detect relative effect Josefsson, Berg, Hiron, Pärt, & Eggers, 2017;Scheper et al., 2013). When implemented in lowintensity farming landscapes that tend to be more complex, AES might lack additionality and their effects might be harder to detect (Concepción et al., 2008). Our power analysis confirmed that the monitoring in this study was adequate to detect relatively large effects of AES (e.g., 25% higher overall abundance on AES than control farms), but underpowered for detecting weaker effects, and effects for patchily distributed species such as reed bunting (Supporting Information Table S11). We recommend that the design of future studies monitoring AES efficacy is informed by power analysis, such that they have power to detect weak effects, in particular when assessing AES policies in low-intensity farmland.
Second, overspill from AES farms into the wider landscape can boost bird abundance on control farms, making potential AES effects hard to detect in a comparative study .
Many farmland birds are vagrant during the winter and move in response to feeding opportunities, so it is possible that some individuals breeding on control farms use resources on AES farms. Previous winter monitoring of our study sites supports this, with flock sizes using AES unharvested crops often far exceeding local breeding numbers (Perkins, Maggs, & Wilson, 2008), whilst colour-ringing revealed corn bunting movements up to 15-20 km between wintering and breeding sites (RSPB unpublished data). Another Scottish study showed that tree sparrows and yellowhammers ranged over several kilometres during winter (Calladine, Robertson, & Wernham, 2006). We detected very little spatial autocorrelation of abundance among our study sites, suggesting few summer movements between the sampled farms, but the limited number of very close farms constrained our ability to accurately measure spatial autocorrelation over short distances. Complete survey coverage over larger contiguous survey areas (e.g., 100 km 2 ) could enable stronger testing of spill-over effects, especially if AES farms were entirely surrounded by controls, but this would require full control over which farms participated in AES. shorter monitoring durations (e.g., Perkins et al., 2011, 7 years andBright et al., 2015, 5 years) have documented AES effects on similar farmland bird species, suggesting that our monitoring duration of 13 years is sufficient time to allow any potential effects of AES management to manifest themselves. We did not have information about farm management and bird abundance prior to 2003, so our study had weak "before-after" design. However, post hoc analysis of the 12 farms that switched from control to AES revealed no significant differences in bird abundance before/after AES had been implemented (Supporting Information Table S12).
Crucially, AES offer farmers a wide range of management options to choose from, and deployment varied substantially between our study sites. We found that bird abundance varied significantly with the area deployed for particular AES options, suggesting that schemes were more effective when option deployment was wellmatched to species' ecological requirements ( Figure 4). For example, reed bunting abundance was enhanced by the creation of water margins which provide tall dense vegetation next to watercourses, the species' preferred nesting habitat (Brickle & Peach, 2004 (Perkins et al., 2011). However, farmers tend to choose AES options that maximise financial income whilst minimising disruption to their current management (Burton & Schwarz, 2013;. In most AES, farmers are paid according to the quantity rather than quality of land dedicated to certain stewardship measures, so they have little incentive for ensuring optimal option deployment (Canton, De Cara, & Jayet, 2009;Quillérou & Fraser, 2010). An alternative approach that has been successful in Germany is result-oriented schemes, where farmers are paid for management only after certain conservation targets are met (Matzdorf & Lorenz, 2010). Adopting a similar practice in future UK schemes may encourage better implementation of AES options and increase their effectiveness.
In summary, we did not detect an overall association between AES and species abundance or population trends. The five farmland bird species we studied in Northeastern Scotland did not experience net directional change. Four species showed signs of stable or increasing populations, with just skylark exhibiting a nonsignificant decline. The regional population trends we documented for linnet, skylark, and yellowhammer (no net change, i.e., stable populations) contrast with UK-wide declining trends from 1995 to 2016 (Harris et al., 2017), and might reflect greater resource provision for granivorous passerines in more diverse and less intensive mixed farming systems. We suggest that lack of AES associations was due to low additionality of schemes relative to conventional farmland habitats during the breeding season.

| Management recommendations
There were specific management options such as species-rich grasslands, water margins, and wetland creation that appeared to enhance bird abundance. These options all provide similar habitat, i.e., relatively undisturbed herbaceous or grassland vegetation that is not managed for agricultural production. Undisturbed vegetation can provide safe nest sites for ground-nesting birds like yellowhammer and reed bunting, and also insect-rich foraging areas during the breeding season. For farms lacking such habitat, AES are currently the main incentive for farmers to create them. Similarly, AES unharvested crops provide bespoke seed food resources for farmland birds during winter, and although we failed to detect associations with breeding abundance, heavy usage during winter (Perkins et al., 2008) supports our recommendation for farmers to routinely select this management option. To improve scheme effectiveness, we recommend better targeting and management of AES options at the farm and field-scale to match the ecological requirements of target species, and to use AES to fill "resource gaps" where specific habitats are lacking. Finally, we recommend that future AES studies of bird population change carefully consider landscape context and likely effect sizes, and use power analyses to design monitoring schemes that will provide adequate tests of whether these expensive policies are having their intended effect.

ACK N OWLED G EM ENTS
The surveys were undertaken by RSPB Scotland, with funding support from Scottish Natural Heritage. We thank fieldworkers Amanda Biggins, Chris Bingham, and Bertrand Couillens, and farmers and landowners for their participation. We thank Isla Myers-Smith, Jeremy Wilson, Will Peach, and two anonymous reviewers for providing very helpful comments on the manuscript.