Michelle Sims, Duke University Marine Laboratory, Nicholas School of Environmental and Earth Sciences, 135 Duke Marine Laboratory Road, Beaufort, NC 28516, USA (e-mail email@example.com).
1In recent years concerns have been raised regarding the status of the common guillemot Uria aalge in the UK. Numbers have declined in several regions, highlighting the need for continued monitoring of this internationally important population. However, the extent to which the current monitoring scheme is capable of detecting declines and options for improving efficiency has received little attention.
2We investigated the power of different monitoring design options for detecting long-term trends in abundance at a colony of guillemots. The ability to detect trends in abundance was reduced by the large temporal and spatial variability in colony attendance. Taking a linear mixed model approach, we obtained details on the sources and sizes of the variance components using count data collected from monitoring plots on the Isle of May, Scotland, and assessed how best to allocate sampling effort in the light of the count variability structure.
3Our results indicated that trend detection will be improved by counting birds in more plots rather than by increasing the number of counts at existing plots.
4The revisit pattern of counts at the monitoring plots during the seasonal counting period had little effect on trend detection power. However, given the practical issues associated with counting guillemots, alternative revisit patterns to the current approach are preferred.
5For a fixed number of visits per plot, power is strongly influenced by the choice of revisit design if the day-to-day variation in colony attendance is increased.
6Synthesis and applications. Aspects of the UK seabird monitoring scheme can be improved. Changes to the allocation of sampling effort and the plot-revisit pattern will improve both the statistical power to detect long-term trends and the efficiency of conducting the survey. We stress the importance of considering the structure and magnitude of the count variation in a power analysis because judicious design decisions depend on the relative magnitude of these variance components.
Monitoring changes in animal or plant populations, whole habitats, or ecosystems is integral for advising on the most appropriate management for their conservation. There may also be a legal obligation to monitor protected populations and habitats (Maxwell & Jennings 2005), for example the European Union (EU) Birds Directive (79/409/EEC, European Comission, 1979) requires all EU member states to protect certain wild bird species and their habitats. This includes establishing special protection areas and monitoring changes in the populations of qualifying species within them, to provide early warnings of unacceptable change.
The UK holds internationally important seabird populations, comprising 25 species of which 24 warrant protection under the EU Birds Directive as either an Annex I listed species or a regularly occurring migratory species. The Birds Directive requires that member states ensure habitat that is crucial to the survival and reproduction of such species is protected from deterioration. A key tool for delivering this requirement has been a detailed and comprehensive monitoring scheme that forms part of the UK's Seabird Monitoring Programme (SMP; http://www.jncc.gov.uk/seabirds, accessed May 2005). Since 1986, the SMP has collected annual data on breeding numbers and performance of seabirds at a representative sample of colonies across the UK.
The common guillemot Uria aalge Pontoppidan 1763 is the most abundant seabird breeding in the UK, with an estimated 1 million breeding pairs (Harris & Wanless 2004). The species is one of the most widely monitored by the SMP, with regular monitoring (annual and tri-annual counts) at c. 25 colonies. For the past two decades numbers of common guillemots at most colonies in the UK have been increasing and breeding success has been high and stable (Mavor et al. 2005). However, since 2000 declines in numbers and poor breeding success, in particular the breeding failure in 2004 at some of the UK's largest common guillemot colonies, have heightened concern over their wellbeing, as well as for other seabird species showing similar trends that appear to be a result of a changing marine environment (http://www.defra.gov.uk/environment/water/marine/uk/stateofsea/, accessed May 2005). It is therefore crucial that the monitoring scheme detects these effects on seabird numbers, yet there has been no evaluation of the current sampling method to judge how effective it is, or to advise how to refine the design so that trend detection improves.
Common guillemots are colonial, cliff-nesting seabirds, and monitoring of population numbers is carried out by counting birds at colonies during the breeding season. Developing a monitoring scheme for detecting long-term trends in abundance at a colony is difficult, as counting common guillemots is problematic. First, they build no nest and sometimes form highly dense aggregations, on average about 20 pairs m−2 (Harris & Birkhead 1985), thus inducing the potential for large counting errors. Secondly, although an egg or chick is rarely left unattended there is considerable temporal and spatial variation in colony attendance by the mates of the attending adults and additional non-breeding adults and immatures (Harris, Wanless & Rothery 1986). However, since it is nearly impossible to distinguish breeding birds from non-breeders, all birds attending the colony are counted. Consequently, counts can be extremely variable and this variability may cloud the detection of any trend in abundance that is present. As a result, monitoring annual abundance using the mean of repeated daily counts at sample plots is preferred to using a single annual count of the whole colony (Walsh et al. 1995).
Different designs for monitoring common guillemots can be compared by evaluating their statistical power for detecting trends. Power is sensitive to count variation but will improve as more precise information is obtained. Therefore, understanding how this variation affects our ability to detect trends plays a key role in the development of efficient monitoring designs. Count variation can be decomposed into numerous components each attributable to a different source or factor. The effect of each variance component on trend detection depends on the sampling design (Urquhart, Paulsen & Larsen 1998; Urquhart & Kincaid 1999; Larsen et al. 2001), and through careful design choices the effects of the most influential components can be minimized and hence power increased.
A good sampling design for monitoring trends in abundance of common guillemots must balance the statistical aspects of having a high power to detect trends and the practical constraints associated with counting. Much attention has been given to the development of a sampling design that fulfils both requirements (Harris, Wanless & Rothery 1983). The focus has mainly been on how to minimize the influence of day-to-day and within-day fluctuations in colony attendance on trend detection. However, there has been little consideration of the year-to-year fluctuations in attendance that is evident at monitored colonies (Rothery, Wanless & Harris 1988) when choosing sampling designs, and the relative influence of these sources of variation on trend detection.
We used data collected from a monitoring scheme of common guillemots on the Isle of May (56°11′ N, 2°33′ W) in south-east Scotland to estimate the size of the variance components. The Isle of May is a ‘key site’ seabird colony in the SMP because of the large numbers of several internationally important species breeding on the island, and long-term studies of changes in numbers, breeding success and survival rates have been conducted (Rindorf, Wanless & Harris 2000; Frederiksen et al. 2004a). The long time series of monitoring data provides detailed information on the structure of the variance components and allows the re-evaluation of the current allocation of sampling effort.
In this study we explored various monitoring approaches for detecting long-term colony trends in abundance of common guillemots on the Isle of May. The current sampling method for population monitoring in the SMP involves repeated counts of individuals in study plots randomly positioned within a colony. Details of the allocation of sampling effort, such as how many plots to survey and the timing and frequency of counts at a plot within a year, are described in Walsh et al. (1995) and are applied to many colonies around the UK. We conducted a power analysis to compare the performance of different temporal sampling designs and provided design recommendations that could be applied to monitoring at other colonies in order to improve trend detection in the future.
The population of common guillemots on the Isle of May has increased from 11 250 breeding pairs in 1981 to 20 185 breeding pairs in 2002 (Harris et al. 2000; Scottish Natural Heritage data, Edinburgh, Scotland). However, population growth was not constant over this period, with the population declining during the middle and late 1980s and subsequently increasing.
The portion of the data presented here came from annual monitoring counts made at 11 randomly positioned study plots delimited on large photographs on the Isle of May from 1981 to 1994 (Harris, Wanless & Rothery 1983; Charras & Parkinson 2003). Birds were counted from where the reference photographs had been taken between 09:00 and 12:00 h Greenwich Mean Time (GMT) on 9 or 10 days spread across the first three weeks of June (the chick-rearing period at this colony). This sampling window covered the times of day and breeding season when numbers vary least (Lloyd 1972; Birkhead 1978; Harris, Wanless & Rothery 1983). Counts later in the breeding season were not used because the population size changes dramatically as a result of the males taking their chicks to sea and fewer non-breeders visiting the colony (Harris, Wanless & Rothery 1986). Plots were always counted in the same order so that the timing of counts each day was fairly constant.
sampling design options
We varied two aspects of the current monitoring scheme, the sampling effort (i.e. number of plots, either 10 or 15, and number of repeat counts at each plot, either 1, 5, 10 or 15) and the temporal pattern of revisiting the plots (i.e. all plots counted on the same day vs. different days) and investigated their effects on the power to detect a trend. The current advice by the SMP is to select randomly as many plots as can be counted during the available time each day, although a minimum of five is required, and all plots are counted on 5–10 days each year. The number of plots and number of counts per plot were chosen to represent the current approach and realistic alternatives that could be applied on the Isle of May so that all plots could be counted on the same day and all counts could be made within the 3-week counting period in June each year.
We also explored different revisit patterns to plots during the seasonal counting period. The SMP recommends that all plots should be counted on the same day. However, the onset of bad weather during the course of the plot round may result in a count being aborted and sometimes data being discarded. So we compared the SMP approach, i.e. a single group of 15 plots all counted on the same day (Table 1, design 1), with an alternative multiple group design with 15 plots being divided into three groups of five plots (Table 1, design 2). Group designs such as these assume that only one group is counted per day, there are an equal number of plots per group and groups are counted for the same number of days each year. Compared with a single group design, having multiple groups involves counting fewer plots per day but there are more days counted in total. Decisions on which group design to consider for the Isle of May monitoring programme were restricted by the number of available count days. We made several assumptions when carrying out the power analysis: weather did not restrict the days that could be counted and counts on successive days were independent. These assumptions meant that there were 21 possible count days. To illustrate the power of a multiple group design, we used the example in Table 1 (design 2) because all plot counts could be made in the time available.
Table 1. Two examples of group revisit designs. In both designs, sampling effort consists of five counts made at 15 plots. Size refers to the number of plots in each group. Count-days refers to the number of days when a group is counted
Design 1: single group
Total = 5
Design 2: multiple groups
Total = 15
An alternative revisit pattern was to count plots on randomly selected days during the counting period. We compared the power of the single group design with this random-day revisit design for all levels of sampling effort. We assumed in the following results and discussion that the current monitoring design on the Isle of May had similar power to a single group design consisting of 10 counts made at 10 plots.
sources of count variation
Several sources of spatial and temporal variability in counts were considered.
Plot-to-plot variation ()
This is the variation in the number of birds among plots. This may reflect differences in availability of suitable nesting sites.
Year-to-year variation common to all plots ()
This can also be called synchronous variation (Larsen et al. 2001). It is the yearly variation in the number of birds that affect all plots in an identical way. For example, reduced food availability during the breeding season may result in a consistent drop in counts at all plots.
Year-to-year variation specific to each plot ()
This can also be called interaction variation (Urquhart, Paulsen & Larsen 1998). It is the yearly variation in the number of birds caused by localized effects influencing plots individually. For example, stormy weather can increase the chance of eggs or chicks being swept off cliff ledges. However, differences in the geographical locations mean that some plots are more exposed to these adverse weather conditions than others.
Day-to-day variation within years ()
This is the daily variation in the number of birds during the first 3 weeks of June This is often a result of day-to-day fluctuations in weather conditions encountered by all plots that affects the attendance of off-duty birds, non-breeders and immatures.
Variation in trend between plots ()
This is the amount by which the long-term trend at each plot differs from the overall colony trend. Localized predation may give rise to such differences between plots. Plots that are subjected to much predation might show a long-term decline in numbers while well-protected plots show an increase.
Residual variation ()
This comprises other sources of variation, such as observation error and variability in colony attendance during the daily time window for counting.
The precision by which a colony trend is estimated is affected by count variability. However, the size of each variance component reveals little about its effect on the variance of a trend; a component that is the largest in magnitude can have the smallest effect on precision. Rather, the influence of each component on trend detection depends on the sampling design and we can reduce the effect of a variance component on the variance of a colony trend by modifying the allocation of the sampling effort.
The variance of an estimated colony trend β from a linear regression analysis of the counts can be expressed in terms of the variance components (for examples of different sampling designs see Urquhart, Paulsen & Larsen 1998; Larsen et al. 2001). We have extended the formula to include the variation in trend between plots, the day-to-day variation within years and allow for different group revisit designs. Thus, suppose we have a group revisit design comprising annual counts at np plots for ny years. The plots are split into ng equal-sized groups and nc counts are made at each plot per year. We have confirmed numerically that the variance of the estimated colony trend for the group revisit design is:
Var(β) = Z1+Z2( eqn 1)
The term Z1 is the direct extension of Larsen et al. (2001) to our situation. The numerator of Z1 is constructed as the sum of variances of random terms contributing to comparisons between years and within plots, with each variance component divided by the number of independent error terms averaged to form the annual mean count. The denominator of Z1 comes from the usual formula for estimating the slope of a linear regression using annual means from ny successive years. The additional term, Z2, allows for variation in the average trend in the np observed plots about the mean trend across the colony. We can reduce the variance of β by changing the sampling effort, such as how many plots (np) to survey along with the number of counts per plot (nc), in the monitoring design. For example, if the variation in trend between plots (Z2) in equation 1 is large relative to Z1 in equation 1, the best design choice is to increase the number of plots. However, if the trend in each plot is the same () and the synchronous year-to-year variation () is large, then any design choice will have minimal effect on reducing the variance of β. The only option would then be to increase the duration of the study. As the same plots are counted each year, the influence of any additive plot-to-plot variation on the average yearly counts is consistent over time. Therefore, regardless of size, makes no contribution to the estimate of Var(β).
Besides modifying the sampling effort, other aspects of a monitoring design can improve the precision by which a trend is estimated and thus increase the probability of detecting a trend. Table 2 gives examples of the additional measures recommended by the SMP to reduce the size and influence of different variance components on the trend variance. These include imposing constraints on the weather conditions suitable for counting and restrictions on the timing of counts.
Table 2. Advice provided by the SMP to minimize the size and influence of the sources of variation on the trend variance
Avoid counting during strong winds/heavy rain or fog
Avoid counting during strong winds/heavy rain or fog
Day-to-day within years
Restrict seasonal window for counting to first 3 weeks of June
Restrict daily window for counting between 08·00 and 16·00 BST
Size of plots restricted to 100–300 birds and all plots are counted from land (rather than from a boat) to maximize counting accuracy
estimating the variance components
We fitted a linear mixed model to the Isle of May monitoring plot data to estimate the variance components. Let yi,j,k represent the logarithm of a count in plot i on day k in year j. Counts were log-transformed to produce a constant variance at all values of y. The linear mixed model fitted to the counts is:
( eqn 2)
where µ is the average plot count at the start of the monitoring scheme, β is the annual growth rate of the colony, xj is a continuous fixed effect for year, the residual variation is the random error term and the other five components of variation are random effect terms. Thus, αi, θi, γj, δj,k and κj,k represent the effects of plot, plot-trend, year (synchronous), year (interaction) and day-within-year, respectively.
Annual counts from 1981 to 1994 on the 11 randomly positioned plots were used to estimate the sizes of the different variance components (Fig. 1). The linear mixed model was fitted to these data using restricted maximum likelihood (REML; Patterson & Thompson 1971) in Genstat 7·1. Forward and backward stepwise regressions were performed using likelihood ratio tests to examine the importance of each variance component in explaining variation in the monitoring plot counts. All six variance components explained a significant proportion of the variability and were retained in the model.
Random effects for year (synchronous), year (interaction) and day-within-year were assumed to be independent and drawn from a normal distribution with zero mean and a constant variance. The plot and plot-trend effects may be correlated within each plot and we modelled the dependence by assuming the random effects are sampled from a bivariate normal distribution with a zero mean for each effect and a covariance matrix specifying the amount of correlation between the parameters. Hence, we assumed that:
estimating statistical power
We took a Monte Carlo simulation approach to calculating the power of different designs of monitoring schemes, as there is no simple formula such as in equation 1 for calculating the variance of trend estimates for a random revisit design. The Monte Carlo approach involved simulating a large number of data sets each representing monitoring plot counts from the Isle of May, and power was estimated as the proportion of data sets in which a significant trend was detected. Factors such as the sampling design, count variability, survey length and the size of the colony trend (β) influence what data are generated and we set these factors to study the power of detecting a known colony trend in a variety of situations.
To demonstrate the effect of the different monitoring designs on the power to detect a colony trend, we compared their power to detect an annual decline in abundance of 1·5% (equivalent to a 20% decline in 15 years), which is a conservative rate of change given that the actual measurements of plots on the Isle of May vary between −2% and +3% per plot per annum (see below). One thousand data sets were simulated from equation 2 using µ= 200, the recommended plot size (Harris, Wanless & Rothery 1983), and the variance components estimated from the Isle of May data. As β estimates the growth rate of the colony when counts are log-transformed, we set β=−0·015. The choice of model fitted to the simulated data sets to test for the presence of the colony trend depends on the revisit pattern of the plot counts during the sampling period. The model in equation 2 was used for the group revisit designs, while the same model was fitted for the random-day count designs but with the day-within-year term dropped. Models were again fitted using REML in Genstat 7·1. Because we were interested in detecting declines in the colony, we tested the departure of β from zero using a one-sided t-test with α= 0·1 and the degrees of freedom were estimated using Satterthwaite's (1946) approximation (see Appendix S1 in the supplementary material).
A power study using simulation techniques is subject to additional variation at the data-generation level. As a finite number of data sets is used to compute power, estimates depend on the random variation that is generated for those particular sets of data and a different power estimate might be obtained if we repeat the simulation process. Increasing the number of simulations will reduce the effect of this variation on estimates of power. In addition, minimizing the differences in the generated random variation among data sets of different sampling designs should reduce the influence of the simulation variability on any comparisons of the different sampling designs. At each simulation we therefore generated a master data set under a monitoring design of 15 counts to 15 plots over 25 years and, as all sampling designs under consideration were nested within this master design, data for a particular design were lifted from this array. For example, consider data for two identical sampling designs except that one was monitored for 11 years and the other 13. Both designs would have the same data for the first 11 years and the additional 2 years of data for the longer surveyed design are taken from the twelfth and thirteenth years of data in the master data set.
The estimated variance components from the monitoring plot data were high relative to the residual variance (Table 3). The plot variance was substantially larger than the other components, reflecting the marked difference in the numbers of birds between the plots at the start of the monitoring scheme (c. 100–400 birds). The direction of the trend varied among plots (Fig. 1) and estimates of the annual rate of change ranged from c.−2% to +3%. The remaining components were similar in magnitude across the different plots.
Table 3. Sources and estimated sizes of variation in the monitoring plot counts of common guillemots on the Isle of May
Source of variation
The power to detect trends for the three types of revisit designs with five counts made at 15 plots are compared in Fig. 2. The random-day revisit design and multiple group design had similar power, while the least powerful was the single group design, i.e. the one currently recommended by the SMP. However, the differences in power between the three designs were small.
The sensitivity of power to the number of plots and the number of counts per plot using a single group revisit design and a random-day revisit design are illustrated in Fig. 3. Clearly increasing sampling effort or survey length increased power but the relative improvement depended on the design choice. Higher powers to detect trends could be achieved by adding plots rather than increasing the number of counts per year. The effect of additional counts to a fixed number of plots on power was negligible and this insensitivity was most noticeable with the random-day revisit design. For example, the current monitoring design on the Isle of May would take 14 years to detect a 1·5% annual decline with 90% power, approximately 11 years with 80% power. This would drop by 2 years if the number of plots was increased to 15, less than 1 year with additional counts.
The trade-off between increasing the number of plots or the number of counts per plot is exemplified in Fig. 4, where the number of years taken to detect a 1·5% annual decline with 80% power is computed for a wide variety of sampling design options. For example, if we consider a multiple group design (Fig. 4a) with three groups each containing six plots, it would take 11 years to detect a trend whether there are seven or two counts made at each plot. However, a reduction in the number of plots per group from six to three would delay the detection of a trend by 4 years irrespective of the number of counts per plot. Note there was a limit beyond which any increases in either the number of plots per group or the number of counts per plot did not reduce the number of years taken to detect the trend (10 years).
Figure 5a shows how the day-to-day variation affects the power to detect trends. Power declined when the day-to-day variation increased in magnitude relative to the other components. The improvement in power using a random-day revisit design compared with the single group revisit design was more noticeable when the day-to-day variation was large. For example, if the day-to-day variation on the Isle of May increased by a factor of 10 in magnitude it would take more than 17 years to achieve 90% power compared with 15 years had the random-day revisit design been used. When the number of counts to each plot was reduced to a single visit every year (Fig. 5b) the difference in power between the two revisit designs was considerable.
We examined different sampling design options for monitoring common guillemots by evaluating the power to detect trends in abundance. Using a linear mixed model framework we estimated the sizes of the different sources of variation in counts for the current monitoring scheme on the Isle of May. Using information on the relative magnitude of the temporal, spatial and residual components of variation we compared the benefits of various monitoring designs. Design choices should be based on both the statistical power to detect trends and the practical issues associated with counting. The current monitoring scheme on the Isle of May proved an adequate design for detecting long-term trends. However, it is clear that some useful improvements could be made without a substantial increase in observer effort. The potential changes are described below.
If the costs of making an additional count at an existing plot are similar to counting at a new plot, then it is clearly better to put more effort into counting new plots at the expense of recounting existing ones. We therefore recommend that sampling effort should be directed at counting more plots rather than increasing the number of counts per plot. It is imperative that the same plots are then revisited each year, as prescribed by the existing scheme, as this eliminates the plot-to-plot variation from the estimates of trend precision. There are other yearly patterns of plot visits that could have been considered for detecting trends in multiple plot designs whereby plots are counted periodically (Urquhart, Paulsen & Larsen 1998; Urquhart & Kincaid 1999). The gains in power depend in part on the aims of the monitoring scheme, for example whether interest is in the total number of seabirds in the colony, or temporal colony trends or both. As we are using the monitoring scheme to detect colony trends alone, the sampling design in which the same plots are visited every year is superior to any other yearly plot visit pattern.
We assumed in our power calculations that counts on successive days were independent. However, counts on successive days are unlikely to vary independently (Harris, Wanless & Rothery 1983), in which case the standard error of the annual mean count based on counts on successive days could be underestimated. The current recommendation by the SMP to relieve this problem is to space out counts throughout the 3-week counting period. This is achievable with the results obtained from this study, which recommend having only a few counts per plot.
plot revisit design
There is little to be gained in terms of statistical power for changing the existing approach of a single group design to either a multiple group design or a random-revisit design. Practical issues should therefore govern what design to use. For example, a common problem that arises under the current approach is that bad weather may physically prevent a complete count of all plots in a day. Therefore, the multiple group design is an attractive alternative to the single group design because fewer plots need to be counted in one day. Although the random-day revisit design has the greatest power compared with the group revisit designs, it may be difficult to apply in practice because counts at the randomly selected dates may be prevented as a result of bad weather.
timing of counts
The existing approach of counting in the first 3 weeks of June, i.e. during the chick-rearing period before fledging starts, should continue. The recommendations given above on the sampling effort and revisit design are based on the variance structure of the monitoring plot counts on the Isle of May, in which the seasonal time window for counting is restricted to a period when attendance is relatively stable. This restriction reduces the day-to-day variation and consequently improves trend detection. The benefits of such a decision are clear when we look at the effect of increasing the day-to-day variation on the Isle of May on the power to detect a trend. This might occur had the seasonal window for counting been extended back into May or to the end of June and into July, periods associated with larger fluctuations in daily colony attendance (Harris, Wanless & Rothery 1983). Design decisions would change such that alternative revisit designs to the single group design become more appealing, having higher power for trend detection, and there should be more counts per plot to reduce the effects of the large day-to-day variance on the precision of a trend estimate. This further illustrates how sampling design choices depend on the relative size of the variance components. It should be noted that one of the commonest effects of climate change on populations is a shift in laying date (Crick et al. 1997). In the case of the common guillemot the trend is for breeding to get later (Frederiksen et al. 2004b). Clearly any changes in phenology at monitored colonies needs to be taken into account and the timing of the monitoring counts adjusted accordingly.
In our power calculations we assumed that information on changes in the numbers of guillemots at the randomly positioned plots are representative of the population being studied. While the randomly positioned plots will reflect changes in the originally defined population, if the population starts colonizing a new area of cliff or abandons previously occupied areas, the initial set of plots is unlikely to remain representative of the population. Harris, Wanless & Rothery (1983) have addressed this issue and suggest re-evaluating a monitoring scheme every 4–5 years.
This study highlights the importance of considering the sources of count variation when designing monitoring schemes for common guillemots. The results show that changes in the sampling design can minimize the influence of certain components of variation on the precision of a trend estimate and the power analyses provide a good evaluation of sampling design options for the Isle of May. Clearly design choices must consider the practical issues associated with counting common guillemots as well as optimize the power to detect trends. The findings from this study are likely to be relevant to monitoring schemes at other colonies. However, it should be recognized that the sizes of variance components may differ among colonies because of differences in weather conditions, food availability and disturbance, for example. Such differences will affect decisions on how best to optimize the allocation of sampling effort. While it may be impractical to carry out a detailed examination of the sources of count variation at every colony, it is clearly prudent (at the least) to carry out a basic assessment of whether the sources of count variation differ substantially from those reported here before making any changes to existing monitoring protocols.
The work was supported by funds from the Scottish Executive Environment and Rural Affairs Department. We thank Scottish Natural Heritage for permission to carry out the work on the Isle of May that underpins the analyses in this paper.