The phenology of many species has been shown to shift under climate change. However, because species respond at different rates, ecological communities may be disrupted leading to species extinctions and loss of ecosystem services. Hence, there is a need to monitor and understand phenological change.
Population data, gathered by standardised monitoring schemes, can be used to this end. However, such schemes require significant organisation and financial resources. Distribution data (georeferenced biological records with dates) are easier and cheaper to collect and may be an unexploited resource for phenology analyses. This would allow analysis of more taxa from more regions of the world. However, distribution data are potentially biased due to the unstandardised behaviour of biological recorders.
Here, the ability of distribution data record dates to accurately predict phenology is investigated by using the British butterfly fauna as a model system. We used the total number of distribution records per unit time across Great Britain as a proxy for butterfly abundance. Phenology metrics of mean flight date and flight period length were then calculated from the resulting abundance–time relationships for each year in a 15-year time series. These estimates were validated against those generated from a standardised-effort population monitoring scheme.
We analysed 1 078 328 records from 30 British butterflies and found that distribution data accurately predicted the mean flight date for 22 of the 30 species tested. Flight period length was only predicted accurately for seven of 30 species.
We found a nonlinear but consistent positive relationship between the accuracy of mean flight date estimates and sample size (number of records) at both inter- and intraspecific scales. Our results suggest that a threshold sample size of c. 6500 distribution records (430 per year) is a pragmatic compromise between accuracy and recording effort, leading to little loss of accuracy in phenology predictions (an average decrease in accuracy of 2·9 days was observed).
The results suggest that distribution data are a potentially useful resource for phenology research. This may allow practitioners to monitor particular regions and previously unstudied species relatively cheaply using existing mapping schemes.
During recent decades, it has become clear that climate change is having a significant effect on the phenology of many species (Parmesan 2006; Hill, Griffiths & Thomas 2011). These changes occur in the direction predicted under global warming scenarios (Parmesan 2006) and are likely to disrupt existing ecological communities as individual species respond at different rates (Walther et al. 2002; Root et al. 2003; Thackeray et al. 2010). Ultimately, this may lead to widespread extirpation and extinction (Thomas et al. 2004; Tylianakis et al. 2008; Willis et al. 2008). Consequently, predicting and monitoring the effects of climate change on phenology are a key issue for 21st century biologists (Visser 2008; Miller-Rushing et al. 2010).
A major challenge concerning biodiversity monitoring schemes, many of which are designed to collect phenology data, is the considerable effort required on the part of professionals and volunteers to achieve adequate levels of temporal and spatial coverage that will allow large scale or long-term trends to be revealed (Thomas 2005; Fox et al. 2006). This problem is particularly true of invertebrates, which are often neglected by conservation biologists and funding bodies (Clark & May 2002; Leather 2009), yet is also present in a range of other taxa and geographical regions. Additionally, it is not expected that different species will adjust their phenology in the same direction or at the same rate (Visser & Both 2005; Doi, Gordo & Katano 2008), and so the narrow focus of existing schemes may be underestimating the consequences of phenological change.
Furthermore, the consequences of shifting phenologies need to be understood in the context of concurrent change in other species and environmental variables (Visser & Both 2005). This point may be particularly salient given that different trophic levels and interaction partners are known to respond to climate change at different rates (Van Nouhuys & Lei 2004; Memmott et al. 2007; Both et al. 2009; Thackeray et al. 2010). These unequal phenological responses are likely to influence species demography and ecosystem processes in novel ways that may not be fully understood without data on multiple nodes within the ecological web. Consequently, we have an incomplete picture of the global phenological response to climate change.
To address these gaps in our knowledge, distribution data may prove useful. Distribution, or ‘atlas’, data are spatially and temporally explicit information on a species occurrence and are commonly used to create regional distribution atlases of specific taxa (Robertson, Cumming & Erasmus 2010). Distributions are mapped from the presence of at least one recorded occurrence within a specified grid cell. These records may be obtained through the extraction of museum specimen data (Funk & Richardson 2002), the use of historical records (Hassall et al. 2007), the submission of casual species observations or through nationally coordinated surveys (Harding & Sheail 1992; Fox et al. 2006). Crucially, such data are a record only of species’ presence and, thus, are different from more detailed presence–absence distribution data obtained from intensive standardised surveys.
Whilst they contain less information than detailed population monitoring data, distribution data are available for a greater range of taxa and geographical regions and often for longer time periods (Thomas 2005; Robertson, Cumming & Erasmus 2010; www.gbif.org). Additionally, it may be logistically easier to collect meaningful volumes of this data type than adequately standardised and replicated population estimates. As distribution records have dates attached, they can be analysed in a temporal context. In theory, one might interpret the number of distribution records available for a species throughout a time series in an analogous fashion to population abundance data. Both data types may produce an abundance–time distribution of a species within a year from which phenology metrics may be drawn.
There has been some interest in the potential of distribution data to reveal phenological patterns (Hassall et al. 2007; Carroll et al. 2009; Altermatt 2010; Poyry et al. 2011; Kauserud et al. 2012), but no rigorous test of its utility in such a role. Validation tests are crucial as distribution data are likely to be highly biased in space and time. In space, data may be influenced by recorder effort (Dennis, Sparks & Hardy 1999), the visual apparency of target species (Dennis et al. 2006) and the expected species richness of a site (Dennis & Thomas 2000).
There are also a number of biases specific to the application of distribution data to phenological research. These will not be apparent when using distribution data for its original purpose and are related to the behaviour of biological recorders. Within years, there is often a drive amongst recorders to collect the first record of a species (http://www.butterfly-conservation.org). Recorders may also lose interest in a particular species as a season progresses and have renewed interest in unusual late events. Interannually, extreme weather and an expectation of shifting phenology may change the activity of biological recorders. These effects may bias phenology estimates derived from distribution data.
The magnitude and influence of these biases are unknown due to the lack of data on actual recorder effort and the implementation of standardised collecting protocols. A key strength of distribution data, however, is that their collection is not hampered by adhering to rigorous controls, and so spatial and temporal coverage can be much greater than for standardised surveys. In the UK, for example, the standardised monitoring scheme for butterflies covers around 1000 active sites. Butterfly distribution data, on the other hand, covers 3834 unique 10 km UK grid squares. The greatest potential strength of distribution data for phenology research, however, comes from its taxonomic scope. Distribution data are available for a much wider range of taxa than the Lepidoptera, birds and bats, which comprise the major population monitoring schemes. Despite the expected shortcomings in the application of distribution data to temporal research, the question remains over whether any phenological signal is strong enough to penetrate potential biases and produce reliable estimates. The British butterfly fauna provides an ideal system within which to answer this question. The UK has a spatially and temporally extensive butterfly distribution data set generated by the Butterflies for the New Millennium (BNM) project (Fox et al. 2006). The aim of the BNM is to map the national distribution of species and, thus, to assess changes over time. The BNM was launched in 1995 and has run continuously with three major drives of record collection activity occurring during 1995–1999, 2000–2004 and 2005–2009. Over 7·5 million records have been collated. The scheme is operated through a network of volunteers and local co-ordinators who feed data to Butterfly conservation. There is also a detailed and pioneering transect monitoring programme for butterflies: the UK Butterfly Monitoring Scheme (UKBMS) (Fox et al. 2006; Brereton et al. 2011). The UKBMS is a standardised scheme which has been in operation in some form since 1976. Volunteers undertake weekly transect walks which generate abundance measures for species on over 1000 sites across the UK (Brereton et al. 2011). Both these data sets have also played a large role in investigating the influence of climate change on butterflies (Parmesan et al. 1999; Roy & Sparks 2000; Menendez et al. 2006; Pateman et al. 2012).
We assume that phenology estimates drawn from UKBMS data will give an accurate baseline against which BNM estimates may be compared. The UKBMS is designed to detect a range of population indices, including phenology metrics (Botham et al. 2008), and is temporally standardised. The UKBMS itself, however, is not infallible. Issues concerning the visual apparency of species may apply to both the UKBMS and the BNM data sets (Dennis et al. 2006). Indeed, the UKBMS fails to routinely produce population trends for a number of rare or visually unapparent species (Fox et al. 2006). A further limitation of the UKBMS is that it is spatially and temporally restricted. Although there are c. 1000 active sites, these may not capture all of the warmest microclimates across landscapes and, therefore, very early and late individuals may be missed. This is also exacerbated by the fact that transect monitoring only starts in April and runs until the end of September. Consequently, an increasing proportion of species flight periods may occur outside of the monitoring period as phenology shifts with the warming climate. These caveats must be kept in mind when commenting on the relative accuracy of BNM phenology estimates.
In this study, two standard butterfly phenology metrics, mean flight date and flight period length (Stefanescu, Penuelas & Filella 2003), are calculated on both data sets for 30 univoltine species over a 15-year time period. We then compare the ability of BNM distribution data to predict phenology estimates from the standardised UKBMS recording scheme. The influence of distribution record sample size on the relative accuracy of phenology estimates is also investigated, and the potential application of distribution records in phenology monitoring is discussed. Due to the biases that may be present in distribution data, it is expected that flight period length estimates will not be predicted well by the BNM. This metric is more likely to be sensitive to non-uniform recording effort throughout a year. Mean flight date estimates are hypothesised to be more robust to these biases, and so are expected to be predicted well by the BNM.
Materials and methods
Data Collection and Preparation
Butterflies for the New Millennium and UKBMS data were supplied by the Centre for Ecology and Hydrology and Butterfly Conservation for the years 1995–2009 and the 30 univoltine species given in the Supporting Information. Analyses were restricted to univoltine species due to the problems involved in calculating the phenology metrics for multivoltine species (Botham et al. 2008). All UKBMS records present in the BNM were removed. For both data sets, the number of days since April 1st was calculated for each record. The UKBMS only monitors butterfly populations between April 1st and September 30th each year (Fox et al. 2006), and so the BNM was also restricted to this time frame to ensure fair comparison. The analysed BNM data set consisted of 1 078 328 records (from 30 species over 15 years). Data were then aggregated to give an abundance (UKBMS) or record count (BNM) per day for every year, species and data set. The phenology metrics of mean flight date and flight period length were calculated for each species in each year. These correspond to the weighted mean:
and standard deviation:
respectively (Stefanescu, Penuelas & Filella 2003), where x is the number of days since April 1st and w is the total abundance per day recorded by the UKBMS or the number of BNM records. These metrics are used as the Gaussian phenology curve is specified by the mean and standard deviation. Both metrics are also commonly used in the study of phenology (Brakefield 1987; Roy & Sparks 2000; Stefanescu, Penuelas & Filella 2003).
We expect no systematic bias between the phenology estimates of the two data sets. Both schemes have comparable latitudinal distributions, and there does not appear to be a pattern in the degree of accuracy of estimates through time (see Supporting Information).
Testing BNM Predictions
Observed estimates (UKBMS) were compared with predicted estimates (BNM) of each phenology metric and every species using type II major axis regression. Major axis regression was used as there may be error in both the x and y variables (Legendre & Legendre 1998). BNM predictions were not considered to be significantly different from UKBMS estimates if (i) a significant (>0·05) positive correlation existed between the two, (ii) the 95% confidence intervals of the regression intercept encompassed zero and (iii) the 95% confidence intervals of the regression slope encompassed 1 (Mesple et al. 1996). Meeting these three criteria indicated that there was a good match between phenology estimates derived from the BNM and the UKBMS. Significance of the correlation coefficient was assessed using 9999 permutations. Regressions were performed using the lmodel2 package in R (Legendre 2008).
Interspecific Variation in the Mismatch Between BNM and UKBMS Phenology Estimates
The average absolute value of the difference between the UKBMS and BNM phenology estimates was calculated. This gave the average mismatch in days of mean flight date for each species. The 95% confidence intervals around these means were also calculated. Sample size, defined as the total number of BNM records for a given species summed over all years, was extracted. The average mismatch for each species was then regressed against their sample sizes, as were the 95% confidence intervals. The absolute value of the confidence interval was used to represent the potential error above or below the mean. Variables were log transformed to meet parametric assumptions. This analysis was not performed for the flight period length estimates due to the poor ability of the distribution data to predict this metric (see 'Data Collection and Preparation'). Model predictions were compared across species with varying total sample sizes to locate a threshold number of BNM records that (i) did not predict an average maximum mismatch of >5 days, (ii) was smaller than the total record numbers of at least 50% of the successfully predicted species (see ‘Testing BNM Predictions’) and (iii) was greater than the total record numbers of at least 50% of the unsuccessfully predicted species. An accuracy of at least 5 days was chosen as reported phenological shifts in many butterfly species over similar time periods tend to be larger than this, ensuring that distribution data could detect phenological changes if present (Botham et al. 2008). Additionally, the second two criteria ensure that the threshold is reasonable with respect to the total record numbers which produced successful or unsuccessful species in the initial analysis.
We tested for phylogenetic autocorrelation in model residuals using Moran I tests with Geary randomisations in the ade4 R package (Paradis 2006; Dray & Dufour 2007). We used 1000 phylogenetic trees as described in Oliver et al. (2012), using closely related conger species where molecular sequences were not available for four of the species.
Intraspecific Variation in Mismatch with Altered Sample Size
Species that were predicted successfully by the BNM and had a sample size greater than the threshold size determined in the previous section were subsampled to further investigate the influence of decreased sample size. Subsets of decreasing size were randomly extracted from the original BNM data for each species. This procedure gave 20 levels of subsampling, decreasing from 100% in 5% increments. This randomisation was repeated 100 times to obtain the average mismatch in mean flight date and associated 95% confidence intervals for each subsampling level for each species. These mismatches and absolute confidence intervals were then regressed against the actual subsample sizes in the same way as described in the previous section. All analyses took place in R (R Development Core Team 2011).
Testing BNM Predictions
Mean flight date
Twenty-six of the 30 species tested had a significant linear relationship between the UKBMS and BNM estimates (P < 0·05) for mean flight date. Of these, 22 had 95% confidence intervals of the intercept that included zero and of the slope that included one. Thus, the yearly predictions of mean flight date derived from BNM data are not significantly different from the UKBMS data set for the majority of the British univoltine species. Figure 1 displays scatterplots and associated regressions for Anthocharis cardamines (Linnaeus, 1758) and Pyronia tithonus (Linnaeus, 1767), two example species randomly chosen from those which did not differ from a 1 : 1 line. Regression details and plots for all species are given in the Supporting Information, and slope estimates are displayed in Fig. 2.
Flight period length
Seven of 30 species had a significant linear relationship between the UKBMS and BNM estimates (P < 0·05) for flight period length. All seven had 95% confidence intervals of the intercept that included zero and of the slope that included one. This indicates that the yearly predictions of flight period length from each data set are divergent for the majority of univoltine species. Regression details for all species are given in the Supporting Information. Figure 3 displays scatterplots and associated regressions for a successfully predicted and unsuccessfully predicted species: A. cardamines and P. tithonus, respectively. Plots for all species are presented in the Supporting Information.
Interspecific Variation in the Mismatch Between BNM and UKBMS Phenology Estimates
Mean flight date
A significant positive linear relationship was found between log average mismatch between BNM and UKBMS predictions and log species' sample size (d.f. = 28, R2 = 0·28, t = −3·27, P < 0·01, Fig. 4a). The average mismatch in mean flight date estimates decreases exponentially with increasing sample size. A similar relationship is also seen between the log 95% confidence intervals of the mean mismatch and log sample size, indicating that the error about mean flight date estimates also decreases with increasing sample size (d.f. = 28, R2 = 0·43, t = −4·59, P < 0·01, Fig. 4b). No significant phylogenetic autocorrelation in residuals was apparent. Combined, these results suggest a threshold sample size below which prediction accuracy rapidly deteriorates. Based on the criteria given above, a threshold of 6500 records over 15 years predicts a maximum mismatch (mean mismatch + 95% CI) of 5·03 days. This threshold is also smaller than the sample sizes of 15 of the 21 species predicted successfully by the BNM and is larger than six of the eight species not predicted successfully by the BNM. The predicted mean mismatch and 95% CI at this threshold are 3·65 ± 1·38 days.
Intraspecific Variation in Mismatch with Altered Sample Size
Mean Flight Date
By reducing records through subsampling, an increase in the average mismatch between BNM and UKBMS predictions was observed, alongside an increase in the 95% confidence intervals. This is consistent with the results from the previous section. For example, A. cardamines showed a significant negative relationship between log subsample size and log average mismatch (d.f. = 18, R2 = 0·64, t = −5·73, P < 0·01, Fig. 5a), as did P. tithonus (d.f. = 18, R2 = 0·87, t = −10·78, P < 0·01, Fig. 5b). The only exception was one species, Aphantopus hyperantus (Linnaeus, 1758), which showed no relationship. Details for all species are presented in the Supporting Information. Species also showed a negative relationship between the size of the 95% confidence intervals around the mismatch and subsample size – similar to the interspecific analysis. Pyronia tithonus illustrates the general pattern of the results (d.f. = 18, R2 = 0·82, t = −9·2, P < 0·01). Details for all species are presented in the Supporting Information.
As species were progressively subsampled, they showed little increase in mismatch between BNM and UKBMS predictions of mean flight date, whilst sample sizes were above the 6500 record threshold. Across the 15 species that had sufficient initial sample size for this analysis, the average increase in mismatch after subsampling to 6500 records was 2·01 ± 0·28 [SE] days (Table 1).
Table 1. Estimates for the mismatch in mean flight date between Butterflies for the New Millennium (BNM) and UK Butterfly Monitoring Scheme data sets using either the original BNM number of records or subsampled to 6500 records. For the subsampled data, estimates of mismatch and 95% confidence intervals in days are calculated from loglinear models of mismatch and 95% confidence interval vs. subsample size over 100 iterations
Average mismatch ± 95% CI
Per cent reduction in record number
Original record number
Threshold record number
2·26 ± 0·94
2·3 ± 0·99
2·52 ± 0·47
2·52 ± 0·57
4·31 ± 1·22
4·34 ± 1·3
1·65 ± 0·52
1·74 ± 0·63
3·4 ± 1·55
3·48 ± 1·57
2·45 ± 0·93
2·52 ± 0·96
2·13 ± 0·94
2·16 ± 0·96
2·94 ± 1·41
2·94 ± 1·43
3·59 ± 1·15
3·61 ± 1·19
2·45 ± 0·67
2·47 ± 0·77
2·25 ± 0·91
2·37 ± 0·96
1·24 ± 0·55
1·37 ± 0·61
6·26 ± 2·49
6·21 ± 2·47
0·55 ± 0·23
0·68 ± 0·29
1·21 ± 0·35
1·26 ± 0·47
In this study, distribution data were able to accurately predict the mean flight dates derived from a standardised-effort recording scheme for the majority of species (22 of 30). Less successful, however, was the prediction of flight period length. In only seven of 30 species did distribution data accurately predict flight period. This is in accordance with our hypotheses, as a standard deviation, flight period length is more likely to be influenced by non-uniform recording effort. Furthermore, this study has shown that there is a consistent relationship between the degree of mismatch which may be expected between distribution data and a standardised-effort recording scheme, vs. the number of distribution records being used. This relationship suggests a threshold sample size of c. 6500 records beyond which prediction accuracy deteriorates rapidly.
The ability of distribution data to match the UKBMS predictions of mean flight date for the majority of species tested is remarkable given the temporal biases that are expected to be present within the BNM data set. As discussed previously, these are largely related to the uneven distribution of records throughout time due to specific aspects of recorder behaviour. This study suggests that either these biases do not exist in a substantial form or that the phenology signal is strong enough to be seen through them. Regardless, the results highlight that distribution data are a potentially valuable resource for phenology research.
Of the eight species whose BNM mean flight date estimates do not match those generated from the UKBMS, six had a sample size below 6,500. At these low sample sizes, the potential biases in recorder behaviour may have become pronounced enough to overcome the strength of the phenology signal. The issues associated with rarity may, however, equally be influencing the UKBMS estimates. If this is the case, then the mismatches may be explained by the different magnitudes or directions in which the data types are influenced by rare species. For example, the UKBMS does not routinely generate population trends for Carterocephalus palaemon (Pallas, 1771) or Melitaea cinxia (Linnaeus, 1758). Both of these species are not predicted well by the BNM and have small sample sizes. Alternatively, the better geographical coverage and fewer constraints placed on recorders may mean that the BNM provides a better estimate of phenology for some species. Further testing of population monitoring methods would be able to differentiate between these two possibilities.
The BNM does not match the UKBMS estimates for Thymelicus lineola (intercept = −27·51, slope = 1·23) despite a relatively large sample size of 21 947 records. The reasons for the increased mismatch between the data sets for this species are unclear, and further work investigating patterns of recording and monitoring in relation to species traits may go some way to explaining the mismatch between the BNM and the UKBMS. However, generally speaking, the mean flight dates of widespread and relatively common species are predicted well by the BNM.
The second key finding of this study is that flight period length is not predicted accurately by the BNM. For the flight period length analyses, the majority of species (16 of 30) displayed a regression slope <1 (see Supporting Information). Whilst most of these slopes were not actually significant, this trend indicates that the BNM estimates of flight period length tended to be greater than those of the Butterfly Monitoring Scheme (BMS). This trend can be explained if recorders are oversensitive to a species outside of its peak abundance period. Individuals that are seen either early or late in the season could attract an inflated number of submitted records due to the novelty or unexpectedness of being sighted. The flight period length will then be overestimated. This offers only a tentative explanation for the inability of the BNM to predict flight period length, yet it is grounded in the temporal biases likely to be present within distribution data. In addition, the marginally broader latitudinal range of BNM records (Fig. S4) and/or a greater range of microclimates sampled by the BNM may lead to greater variance about the mean and result in greater estimates of flight period length.
Thirdly, our study finds a consistent pattern in the relationship between the average mismatch between distribution and transect data, and sample size. With decreasing sample size, there is a trend for the average mismatches and their associated 95% confidence intervals to increase at both inter- and intraspecific scales. The true applied consequences of this, however, may not be appreciated without reference to the magnitude of the observed change. There is very little change in the expected accuracy as sample size is decreased above the threshold of 6500 records. For example, at its original sample size of 132 647 records, A. cardamines has an average mismatch of 2·26 ± 0·94 days. This increases to a predicted 2·3 ± 0·99 days at a subsample size of 6500. This is a 95·1% reduction in sample size with an average increase in mismatch of 0·04 days and a potential maximum increase of 1·97 days. Similar patterns are seen for the other subsampled species (Table 1). This fact not only highlights the robust nature of the methodology, but also its potential use. An average of 430 records per year (the average number of records from 6500 records over 15 years) could be an achievable target for a range of widespread taxa and geographical regions. For example, data from the NBN Gateway suggest that around 60% of butterfly species, 20% of moths and 50% of dragonflies meet the threshold number of records (Fig. 6). These numbers are encouraging but emphasise the continued need for large scale citizen-science schemes, especially if we wish to understand the phenology of less well-studied taxa across the entire breadth of the ecological web.
Additionally, the possible range over which distribution data estimates may deviate from transect generated estimates for those species succeeding the 6500 threshold is on the scale of 7·29 (Satyrium w-album Knoch 1782) to 0·64 days (P. tithonus). This margin of error tends to be below recorded long-term (i.e. several decade) phenological changes for butterflies (Roy & Sparks 2000; Forister & Shapiro 2003; Stefanescu, Penuelas & Filella 2003), birds (Crick et al. 1997) and plants (Fitter & Fitter 2002). This suggests that distribution data could play a role in investigating long-term changes. From our analyses, we suggest that species with a total sample size >6500 records over 15 years (430 per year) should be appropriate for phenology analyses.
Whilst these results are encouraging for the use of distribution data in detecting and monitoring butterfly phenology, they should not be limited to this taxon. In this study, only a single assumption has been made regarding the life history of the organisms. This is that the phenology event in question (e.g. butterfly emergence periods) follows a Gaussian distribution. Such a pattern is common for a wide range of phenological phenomena, a few individuals are relatively early or late, whilst the majority fall towards the middle. This lack of complicating assumptions makes the application of distribution data to phenological research easily generalised. A caveat may be that the behaviour of butterfly recorders is fundamentally different from other biological recorders. We see no strong reasons, however, why this should be the case. We encourage the use of distribution data at smaller spatial scales as long as the threshold sample size is reached but caution that at larger scales, the spatial synchrony of phenological event under study should be considered. There is a high degree of butterfly phenological synchrony in the UK, which may have helped distribution data perform on this scale (Roy & Asher 2003). We urge future practitioners to consider these points carefully, and perhaps apply their own validation tests, as the 6500 record threshold, we suggest here may be unsuitable for some scenarios.
In summary, this study illustrates the utility of using distribution data to predict aspects of a species phenology. Crucially, mean dates can be predicted well, whilst the range of time over which species are apparent (flight period lengths in this butterfly example) are not. These mean flight date predictions are robust and require relatively small sample sizes to achieve adequate levels of accuracy. This, coupled with the ease with which these types of data may be collected, suggests that distribution data can make a valid contribution to the continued monitoring and study of phenology in a range of taxa and locations.
We thank the many thousands of volunteer recorders who have gathered UKBMS and BNM data over the 15-year period of this study and Graham French for supplying NBN data. We thank Albert Phillimore and anonymous reviewers for valuable comments on an earlier draft. TRB received financial support from the Grundy Educational Trust. The BNM and UKBMS projects are funded by Countryside Council for Wales, DEFRA, Environment and Heritage Service, Forestry Commission, Joint Nature Conservation Committee, Natural England, Northern Ireland Environment Agency, Redwing Trust, Scottish Executive Environment and Rural Affairs Department and Scottish Natural Heritage.