The current headline indicator for ecosystem health and sustainability incorporates a geometric mean of relative abundances of breeding birds. Recently, a family of diversity measures (λ-measures) has been proposed as a novel instrument to separate diversity trends in dominant and rare species. This makes them an ecologically informative complement to current composite diversity indices.
Using both a geometric mean and the set of λ-measures, we study habitat-specific temporal trends in the diversity of British breeding birds. The analysis employs abundance estimates corrected for variation in detectability between individuals from different species to reduce bias. Applying generalized additive models, we predict long-term trends. We locate significant changes in these diversity trends.
While the geometric mean reveals overall diversity trends by habitat type, supplementing these by the λ-measures provides a more nuanced picture of trends: a positive trend in the geometric mean may hide predominantly declining trends among the rarer species, which is then revealed by trends in the λ-measures.
Synthesis and Applications. Bird populations are seen as useful indicators of the health of wildlife and the countryside because they occupy a range of habitats, they tend to be towards the top of the food chain, and data is provided by long-term surveys. Hence, many countries apply wild bird indicators (WBIs), quantifying trends in biodiversity, to monitor environmental health. The UK's WBI, for example, has become one of the government's headline indicators of sustainable development. Understanding the population changes underlying the estimated trends is indispensable if we are to allocate limited resources more effectively. Employing a novel set of measures alongside the traditional geometric mean index, we analyse diversity trends among British breeding birds. It reveals that species that are scarce, but not yet in the focus of conservation action, may be the ‘losers’ in biodiversity action plans. This suggests that additional resources should be devoted to species showing long-term decline before they reach the low population levels that currently trigger large-scale species-specific rescue projects.
We are faced with an unprecedented decline in biodiversity at a time when the pressure on the Earth's ecosystems is growing (Butchart et al. 2010). International responses to the Convention on Biological Diversity (CBD) call for large-scale assessment of biodiversity trends (de Heer, Kapos & ten Brink 2005; Pereira et al. 2010). To monitor progress towards such goals, data should be representative on at least a national scale and should be amenable to analyses that minimize potential bias, extract long-term trends, and determine whether the rate of biodiversity loss has been reduced (Magurran et al. 2010).
In the UK, biodiversity is recognized as a resource and an integral part of a ‘healthy and resilient natural environment’ and should impact decision-making (PSA 2007). Government actions include setting regulations and establishing grant schemes such as the Environmental Stewardship (Garrod et al. 2012) and the English Woodland Grant Scheme (Urquhart, Courtney & Slee 2010). In the UK (as in several other countries), an indicator of trends in wild bird populations [the UK Wild Bird Indicator (WBI), DEFRA 2011] serves as a measure of progress towards the political targets. Common breeding birds are seen as good indicators of ecosystem health (Gregory et al. 2003, 2005; Gregory & van Strien 2010): they occur across different habitat types and extensive monitoring programmes exist in many countries.
Adopting a nationwide randomized stratified sampling design, the UK Breeding Bird Survey (BBS; Risely et al. 2011) provides high quality data for statistical analysis that is tailored towards the biodiversity targets. Summary statistics such as the WBI are based on a geometric mean of relative abundances of species, that is, they summarize within-species trends in abundance relative to each species’ abundance in a baseline year (Gregory et al. 2005; DEFRA 2011). The geometric mean also reflects the evenness component of species diversity (Buckland et al. 2011b). It has advantages over more traditional measures of species diversity (Buckland et al. 2011b) and is therefore preferred as a headline index (Buckland et al. 2005; Lamb et al. 2009; van Strien, Soldaat & Gregory 2012). However, as for any summary, it can conceal finer details in diversity trends.
Studeny et al. (2011) proposed a family of measures (λ-measures) for quantifying biodiversity trends, more precisely changes in evenness. A free parameter λ controls the sensitivity of the measures in this family to either rare or more common species. This allows us to tailor diversity assessment with respect to different degrees of rarity within the chosen group of species and provides a tool for analysing diversity trends in more detail, which offers additional ecological information relevant for ecosystem management.
As pressures on species are often habitat specific, the ability to quantify trends in diversity at the habitat level is essential for effective conservation management (Newson et al. 2009; Renwick et al. 2012). Farmland birds, for instance, experienced a marked decline in the UK in the 1980s (Siriwardena et al. 1998). However, an overall decline does not necessarily affect all species equally. Recent studies of woodland birds, for example, have shown that specialist and generalist birds respond differently to change (Vickery et al. 2004; Davey et al. 2012), while bird species associated with human habitats such as Passer domesticus house sparrow, Sturnus vulgaris common starling and Apus apus common swift are the focus of growing concern (Baillie et al. 2010). Conservation managers therefore need to pinpoint trends for both rare and common species, and to place these in the context of the overall trends seen across the different habitat groups.
In this study, we evaluate large-scale trends in biodiversity across five major habitat types (farmland, grassland, urban species, wetland and woodland) of UK breeding birds. The new λ-measures in conjunction with a geometric mean index allow us to separate trends in less common species from those in abundant species. This provides a showcase for the application of these methods as an informative supplement to current headline indicators. We obtain a more robust and informative assessment of temporal trends in diversity than is possible by analysing species-specific population trends alone. As international biodiversity schemes are to be coordinated and extended over the coming years according to the CBD's action plan, this study provides relevant information at an early stage.
Materials and methods
We analyse data from the British BBS. Organized by the British Trust for Ornithology (BTO) and carried out by volunteer observers, it has been conducted yearly since 1994 (Greenwood et al. 1995; Risely et al. 2011). Our data comprise years 1994–2008, except for 2001 when access to many survey sites was restricted due to an outbreak of foot and mouth disease.
The BBS follows a stratified random design where sampling units are 1 km2. These are allocated randomly within strata based on regions corresponding closely to UK counties. In each region, the sampling rate depends on the number of available volunteers. Observers visit their survey square twice a year, once in April or early May, and once in late May or June. In these analyses we focus on data from the first visit to minimize the possibility of including juvenile birds, except for late breeding birds, such as summer migrants (see Supporting Information). In their allocated square, the volunteers walk two parallel transect lines of 1 km each while assigning each record to one of four categories (0–25 m from the line, 25–100 m, >100 m, and flying over). In accordance with Newson et al. (2008), we consider data from the first two categories only.
Using habitat information recorded along with species counts, a classification method based on Jacobs’ preference index (Jacobs 1974; Newson et al. 2008) assigns each species to one of six main habitat groups (coastal, farmland, grassland, urban, wetland, woodland), according to its primary habitat use. Here, we exclude coastal species as well as all nocturnal species as they are not adequately surveyed by BBS methods.
Analyses of long-term national biodiversity trends should be based on a list of species that is comprehensive and representative, although it must in practice be restricted to those species for which there are adequate data. We began by considering the entire suite of bird species recorded through this scheme. However, the geometric mean cannot be calculated if a species’ index of abundance equals zero, and its precision is often diminished by rarely recorded species. This constraint applies to all studies that use the geometric mean to assess diversity (including those adopting the Living Planet Index – Loh et al. 2005; Buckland et al. 2005, 2011b). Sufficient data from the combined set of survey sites is necessary to estimate UK abundance reliably, but sample size for individual sites may be small or even zero. Setting the threshold as low as possible, we use data for species that were sufficiently widespread (>10 sites overall) and abundant (>15 records on average per year) to allow estimation of a year effect in detection probabilities. In addition, two wetland species, Recurvirostra avosetta pied avocet and Limosa limosa black-tailed godwit, are excluded despite being classified as sufficiently common, because no individual was recorded at any site for several years.
Occurrence by stratum of some grassland and wetland species correlates negatively with sampling effort. Both habitat groups also comprise waders for which flocks might be recorded during the breeding season. Following the recommendations of Field & Gregory (1999), who found that BBS population trends of waders were influenced by large counts of nonbreeding individuals at a small number of sites, we identified outliers (flocks of birds, birds outside their usual breeding range). Again aiming to retain as much data as possible, we excluded only nine extreme records (see Table S2, Supporting information). In addition, one survey square that covers the Abbotsbury Swannery was omitted as not representative for Cygnus olor mute swan records.
Apart from these outliers, a further 10 rare species were excluded from the grassland and wetland communities. Initial inclusion of these species led to asymmetric and wide confidence intervals for some diversity estimates and reduced their information content substantially. We felt that those few species unduly compromised precision, potentially caused by a mis-fit between the survey design and their heterogeneous distribution. Their exclusion resulted in smaller and more stable variances, while the overall trends were largely unaffected. These removals left 98 species across all habitats (see Table S1, Supporting Information, where excluded species are also listed).
We seek to identify long-term trends in biodiversity and to determine years in which the rate of change in trend changes. This is done with a headline indicator in mind and achieved by establishing composite summary statistics within each habitat group and analysing their trend curves, rather than following individual species’ trends. Different composite summaries, namely a geometric mean of abundances and a recently proposed family of evenness measures, are evaluated. The combined use of these diversity measures and their interpretation provide more nuanced insights into diversity trends.
We follow a model-based approach where, for each species, we (i) estimate detection probabilities including year effects to account for potential trends and (ii) estimate total UK abundance for each species. (iii) Long-term (smooth) trends in species density are derived by fitting a generalized additive model to the point abundance estimates. Within each habitat group, (iv) diversity summary statistics are evaluated based on both the point estimates and the smoothed trends. Finally, confidence intervals for these trends and the location of potential changes in trends are determined simultaneously by (v) a nonparametric bootstrap of survey sites, respecting the stratified design. The following provides details on the different parts of the analysis.
The existing UK WBI is based on a geometric mean of relative abundances. Given a list of S species, we calculate this geometric mean from the estimated abundances for each species i in each year j as
The geometric mean meets many of the requirements of a headline indicator (Buckland et al. 2005; Loh et al. 2005) and has been adopted by policy makers for these reasons. It summarizes species-specific trends in abundance as well as evenness (Buckland et al. 2011b) as it gives equal weight to all species, in the sense that rarer species contribute as much to the sum as more common ones. Because of this, it is sensitive to fluctuations in abundance of rare species that can increase its variance.
Traditional diversity measures are often based on species proportions Nij/Nj, where Nj = ΣiNij, and focus on dominance, such as Shannon's and Simpson's indices (Magurran 2004; Maurer & McGill 2011). These classical indices reflect largely the same properties, that is, a combination of species richness and evenness (Hill 1973). They remain unchanged if all species decline at the same rate, which is one of their disadvantages compared with the geometric mean (Buckland et al. 2011b).
A family of measures was proposed by Studeny et al. (2011) for quantifying evenness and is here referred to as the ‘family of λ-measures’. It corresponds to a family of goodness-of-fit statistics (Read & Cressie 1988) and is related to a family of inequality measures used in economics (Cowell 1980), where in both cases deviance from a uniform distribution (of abundances) is quantified. It is given by
While the geometric mean combines trends in abundance and evenness, the λ-measures summarize evenness only. The free parameter λ allows us to weight the biodiversity measure towards either rare (λ < 0) or dominant (λ > 0) species, while still providing a summary over all species. As λ goes from large negative to large positive, the weight given to rare species decreases, while that given to dominant species increases. We consider values between −1 and 2, which include the following special cases: for λ = 0 and λ = 1, linear transformations of Shannon's index and the log-version of Simpson's index can be derived, respectively. Like the geometric mean index, the λ-measures for negative λ necessitate the exclusion of very rare species: they cannot be computed if an annual abundance estimate for any species is zero (as this would result in a division by zero in the expression in brackets). Because low values of the measures in this family correspond to high evenness, we consider their reciprocals for easier comparison with other measures.
Correcting for differences in detectability
Measures based on species proportions, such as the λ-measures, are biased when they are calculated from count data assuming equal detection probabilities independent of species identity (Yoccoz, Nichols & Boulinier 2001; Buckland et al. 2011a). By contrast, the geometric mean of counts is unchanged by species’ differences in detectability of individuals; it is only affected if there is a trend in detectability with time (in which case, it is more severely affected than are measures based on species proportions). Hence, where possible, detectability should be taken into account explicitly when estimating diversity, independent of the measure of choice. Here, the survey design allows us to estimate detection probabilities and consequently estimate abundance correcting for detectability. To reduce potential bias, we evaluate the diversity measures based on these abundance estimates. The BBS is conducted using line transect sampling (Buckland et al. 2001). We model the fall-off in detectability with distance from the line by fitting a half-normal model to the binomial count data (numbers of birds counted within 25 m of the line and between 25 and 100 m of the line). To allow for trends in detection probabilities, year is incorporated as either a continuous covariate or as a factor in the scale parameter (Marques & Buckland 2003). For each species, we fit a detection function that is assumed independent of year, together with one in which year is a continuous covariate. For those species recorded at more than 10 sites in every year, we also fit a model with year as a factor. Finally, the model with the smallest AIC is selected. For each species, this provides the estimated probability of detection of an individual of species i in year j. The following total UK abundance (point) estimates take the original survey stratification into account:
where for year j, nijr is the number of records of species i in region r, a =0·4 km2 is the survey area covered per plot (two strips of length 1 km and width 200 m), mjr is the number of plots visited in region r, and Ar is the size (i.e. total number of available squares) of region r.
We anticipate some bias in these abundance estimates. For some species, the assumption of perfect detection of birds on the line is possibly violated. Further, the ideal line would be placed independently of habitat features. The actual line is often an approximation, avoiding arable crops or bypassing obstructions. However, biodiversity measures derived from these abundance estimates should be appreciably less biased than ones calculated from uncorrected counts (Buckland et al. 2011a).
Estimating long-term trends
The point estimates of abundance, driven by short-term fluctuations, typically show variation over time. To establish underlying long-term trends, we smooth the yearly fluctuations applying a scatterplot smoother (Hastie & Tibshirani 1990). Following Fewster et al. (2000) we use generalized additive models (GAMs) to do so. A GAM is fitted to mean counts for each species, with year as the independent variable. The mean is calculated as an average of the stratum means, weighted by the size of the region. Including an offset term for detectability, the density estimates from this model are
where a =0·4 km2 is the survey area covered per plot and f(•) is a smooth function. Finally, the predicted abundances are . As with generalized linear models, GAMs involve choosing an error distribution. Here, the obvious choice (for a mean of counts) would be a Gamma distribution, but it cannot be applied for years where a species was not recorded. Tweedie distributions can model the mean-variance-relationship of over- or underdispersed data flexibly through a parameter p (Jørgensen 1997), and include the Gamma distribution as a special case (P =2). By choosing a Tweedie error distribution with P =1·9, we are close to a Gamma distribution, but can include years with missing observations. We fit GAMs with a Tweedie error distribution using the mgcv library (Wood 2006, 2008), implemented in the statistical software r (R core development team 2011), and thin-plate spline regression with degrees of freedom selected by in-built cross-validation, but with an upper limit of three.
For a randomized survey design, precision of biodiversity measures can be quantified using nonparametric bootstrapping, with sites as the resampling unit (Fewster et al. 2000). Respecting the stratified design, we sample visited sites within each region with replacement to get the original number of sampled squares in each region. By recalculating population estimates and re-evaluating the diversity measures for each of 999 resamples, 95% confidence limits for the annual diversity index are derived by the percentile method (Buckland 1984).
Precision of the geometric mean can be low (Buckland et al. 2011b). By definition, the index is unity with zero variance in the baseline year. The further we move from the baseline, the wider the confidence intervals become as autocorrelation decreases, rendering diversity estimates less informative over time. Precision for a subsequent year is driven by the variance in the baseline year (here 1994, the first year of the BBS survey) as well as in the subsequent year. Low effort in the early years of the survey increases uncertainty for the entire time series. This effect could be avoided by changing the baseline to a later year; however, this would follow trends by looking backwards and could be deemed counter-intuitive. Using the first year as a baseline can also be problematic because of first-time observer effects, as in the North American BBS (Kendall, Peterjohn & Sauer 1996). Studying potential observer effects for the British BBS, Eglington et al. (2010) found no significant effects on estimated abundance trends. Supporting this, we found no consistent pattern of lower detectability for new observers.
Changes in temporal trends
The rate of change in diversity is reflected by the slope of the estimated trends and it is of interest to locate trend changes. Changes can be for the better (a decrease in biodiversity that slows or reverses or an increase that accelerates) or worse (a decrease in biodiversity that accelerates or an increase that slows or reverses) and are captured by the second derivative of the trend curve (second derivative >0 in the first case and <0 in the second).
Fewster et al. (2000) and Buckland et al. (2005) successfully used numerical evaluation of the second derivative in combination with bootstrap confidence intervals (see above) to identify years as significant change points in trends. We apply this method here for the habitat-specific diversity trends. The resulting confidence intervals are independent of choice of baseline year for the geometric mean and do not increase in time (Buckland et al. 2011b). In addition, this method does not require a long time series. This makes it a valuable instrument to link changes in diversity to effects of conservation actions.
We summarize habitat-specific trends in diversity comparing the results from the different measures, and demonstrate how they can complement one another and hence increase the value of single (‘headline’) indices. We relate and contrast this to individual population trend summaries.
Summary of individual population trends
Figure 1 summarizes changes in individual species’ population sizes by comparing the estimated abundance index in 2008 to that in the first year. This summary can help interpret the differences in trends shown by different composite indices. For example, for farmland and woodland species, we see that most common species are now doing better than in 1994. For species near human habitation, and grassland and wetland species results are less clear, but in all cases a substantial part of the community has declined. These plots also show that the grassland and wetland communities comprise mostly rare species, while the most abundant species can be found in the woodland community and near human habitation. However, these summaries do not follow continuous trends in time nor identify years in which trends change. To achieve this, a composite index, such as the geometric mean, is used. In combination with the λ-measures, we are able to differentiate between rare and common species while retaining a summary of yearly trends across species.
Trends vary between habitat groups as does precision of the diversity estimates (Figs 2 and 3). Rare species increase uncertainty. Hence, indices that are sensitive towards rare species, such as the geometric mean index and λ-measures for negative parameter values, show lower precision and generally have less power to detect trend changes.
Neither the geometric mean nor the λ-measures for λ = 0 and λ = 1 indicate an increase or a decline in trend. However, the index for λ = −1, which focuses on the less abundant species, is declining with a significantly lower value at the end of the time series. Second derivatives and their confidence intervals do not show a significant change in trend. Hence there is an indication of a continuing decline in evenness. With no significant decline in the geometric mean, these results suggest that scarce species tend to be declining, while dominant species are faring better. This is in accordance with the summary of individual species population trends in Fig. 1; the most common species show an increase in population size between 1994 and 2008, while more than half of the rare species have declined, two to substantial degrees (<50% of the population size in 1994).
Grassland and wetland birds
Neither the geometric mean index nor the λ-measures give evidence for a trend in diversity for grassland species. Potentially due to low sampling effort at the beginning of the survey, precision for the geometric mean is low and trend estimates hence of little use. A negative change in trend appears in 1999. The λ-measures draw a similar picture. For positive λ, the second derivative indicates a negative change around 1998. For the more common species, further positive changes are picked up – the first in 2002 reversing the previous negative direction, and a second in 2005 (λ = 2), indicating an improvement in the rate of change of diversity. As for grassland species, estimated precision for all diversity indices is low for wetland birds and neither the geometric mean nor the λ-measures reveal a significant trend or significant changes in trend.
Species near human habitation
All indices show roughly the same pattern, but to different extents. According to the geometric mean, diversity increased significantly between 1994 and the early 2000s (CIs above one). The second derivative suggests a change for the worse in 2001/02. The λ-measures confirm this, but show a stronger upwards trend from the beginning of the survey. This suggests that this trend is primarily driven by trends in evenness rather than in abundance. As for the geometric mean, the second derivative indicates that this increase slowed down for common species (λ > 0) between 2000 and 2003 with little change thereafter. For the common species, diversity in 2008 (the last year considered here) is well above the value for 1994. For negative parameter values for the λ-measures and hence less common species, the initial upward trend is less pronounced, followed by the same negative change around 2002, which leaves diversity just below the level of 1994 in 2008. This provides further clarification of why the geometric mean, which gives equal weight to rare and common species, shows less of an upward trend in the first half of the period than do the λ-measures weighted towards the most common species. Unlike the individual population trend summary in Fig. 1, the composite indices show that across all species there has been an upward trend up to 2002, however to different degrees.
The geometric mean registers an increase in diversity, with the value in 2008 being about 25% [14–40%] higher than at the beginning of the survey period. No significant changes in this trend are evident. The trend in the λ-measures is less positive. Only if weighted towards the most common species, does an upwards trend appear towards the end of the time series (CIs being just above the level of 1994). Positive changes are identified for the more dominant species (positive parameter values) in 2002 and for the following 4 years for λ = 2. While there is no significant change for negative λ, the trend curve suggests a decline in diversity for the less common species.
Differences between the various indices are most noticeable for the woodland community. The strong upward trend in the geometric mean reflects the fact that most species in this community have increased in abundance compared to 1994 (see also Fig. 1). The trends in λ-measures, accounting for evenness in species proportions, are quite different: evenness declines when the index concentrates on rare species (λ = −1) and increases only slightly if common species receive more emphasis (λ ≥ 0). (However, there is an indication of a positive change in trend in 2002). Comparing the different summary statistics, we conclude that, while abundance has increased, proportions for common species have changed at similar rates, leading to little change in evenness unless we weight in favour of rare species.
National and international biodiversity management relies on methods to determine large-scale trends and to assess whether the rate of loss of biodiversity is successfully slowed down, halted or even reversed (Walpole et al. 2009).
Here, we present a comprehensive study of habitat-specific diversity trends in British breeding birds that are suitable for monitoring schemes worldwide. As indicator species of ecosystem and environmental health, birds are currently contributing to national headline indices to evaluate progress with regards to international biodiversity targets. A single headline index is bound to concentrate on selective aspects of the biodiversity concept. Importantly, the choice of index can influence whether and even what kind of trend (positive or negative) is identified. In addition, a positive change might not be representative of all species. We complement the current headline index, a geometric mean of relative abundances, by a new set of composite λ-measures. Their ability to separate effects for rare and common species reveals why different indices pick up contrasting trends. In particular, they can focus explicitly on trends for rare species as long as the species are still common enough to be included in an analysis. This latter group appears to be the ‘losers’ in terms of diversity trends. Positive trends in our results are mostly associated with the more abundant species. The UK WBI, confirmed by the geometric mean here, suggests that the negative trends could have stabilized for farmland birds and even have reversed for woodland species. However, the λ-measures reveal continuing negative trends for the less common species in both groups. This is confirmed by single species trends in abundance (see Figs S1 and S2, Supporting Information). This indicates a weakness in monitoring programmes: general surveys cover the most abundant species well and conservation programmes set in place single species surveys for the rare, endangered species (not included here). Those species that are neither abundant nor (yet) endangered (e.g. Poecile montanus willow tit, Phylloscopus sibilatrix wood warbler) are not monitored well by any scheme. This is worrying and emphasizes the necessity for a more nuanced assessment and care when monitoring schemes are developed.
However, less common species pose a problem from a statistical point of view. Their low numbers result in less data and greater short-term fluctuations. Uncertainty in the estimated detection probabilities for such species is high. As composite indices are summary statistics across a set of species, rarer species introduce uncertainty to the diversity estimates. While traditional indices concentrate on dominant species and hence show higher precision, confidence intervals for the geometric mean and the λ-measures for λ ≤ 0 tend to be wider. Hence, there is a trade-off between the inclusion of as many species as possible and precision of the diversity estimates. Similarly, analyses of subgroups are affected if they comprise mainly rare species or species with a very heterogeneous spatial distribution, such as the grassland community. Results could potentially be improved by conducting targeted surveys for some of the rare species and combining them with the general results from surveys such as the BBS. This is already done for some species, especially raptors and very rare species. However, such surveys have to be designed carefully for the results to be included in a statistical analysis. Particularly rare species might not be homogeneously distributed. Hence data collected in pristine locations to assure sampling success will not be representative on a larger scale. Summaries of individual species’ population trends can help classify species as rare or common and help target conservation actions or monitoring efforts. This can improve trend analyses based on headline composite indices, as shown here. Only the latter enable us to continuously monitor trends across species, and in particular, to identify changes in these trends with time.
The modelling approach used here could be extended to identify ‘hotspots’ of biodiversity change, both spatially and temporally. For example, a GAM could smooth across both space and time and potentially identify areas where the amount of spatial turnover (β-diversity) is changing through time. Habitat and climate information could also be incorporated into the GAMs to identify how these covariates affect spatio-temporal trends in biodiversity. For instance, using a related approach, Davey et al. (2012) found increasing homogenization in breeding birds linked to the warming climate in Britain.
As monitoring schemes similar to the UK BBS are developed to assess biodiversity trends internationally, the methods and results presented here will become useful in a wider context. Different composite measures should be interpreted together to gain more information. The λ-measures offer the possibility to investigate evenness in abundance weighted by rarity. This can inform management decisions and survey planning, in particular for less common species. Furthermore, our analysis shows that the less common but not yet endangered species should not be neglected in conservation plans. This was demonstrated here for breeding birds but is likely to hold equally for other taxa.
We are grateful to the BBS volunteers. The BBS is a partnership between the BTO, Joint Nature Conservation Committee (on behalf of Countryside Council of Wales, Natural England, Council for Nature Conservation and Countryside and Scottish Natural Heritage) and Royal Society for Protection of Birds. We thank Chris Elphick, Jon Bart and an anonymous referee for their helpful comments and suggestions that led to a much improved manuscript. A.C.St. thanks the University of St Andrews for funding. A.E.M. acknowledges the ERC (project BioTIME 250189) for funding and JBI financial support from Research Councils UK. P.J.H. thanks Scotland's ClimateXChange (www.climatexchange.org.uk) for funding.