This study uses data from the UK Breeding Bird Survey (BBS). The BBS uses a formal sampling framework where 1-km squares are randomly selected according to a stratified sampling design, and each square is surveyed annually. Fieldwork involves two early morning visits to each square where birds together with large and easily detectable mammal species are counted along two 1-km transects. This includes the three most widespread and abundant deer in England, Reeves’ muntjac Muntiacus reevesi Ogilby 1839, roe deer Capreolus capreolus L. and fallow deer Dama dama L. Red deer Cervus elaphus L., sika deer Cervus nippon Temminck 1838 and Chinese water deer Hydropotes inermis Swinhoe 1870 also occur in England but are relatively localised (Dolman et al. 2010) and poorly monitored by the BBS (Wright et al. 2009). Whilst observer counts will underestimate the true number of deer and birds present, there is good evidence for birds (Buckland, Goudie & Borchers 2000) and deer (reviewed in Morellet et al. 2007) that such standardised monitoring produces measures of abundance that correlate well with other methods of intensive density estimates. In this study, we use BBS data from England for the period 1995–2006, during which all three most widespread deer species increased significantly in abundance (Wright et al. 2009).
We consider the impact of the three most widespread and abundant species of deer on 15 bird species in lowland England, exclusively passerines that are associated with woodland habitat (Table S1, Supporting Information). Eleven of these species (dunnock Prunella modularis L., nightingale, song thrush Turdus philomelos L., willow warbler, willow tit, marsh tit, bullfinch, blackcap Sylvia atricapilla L., chiffchaff Phylloscopus collybita L., and blackbird Turdus merula L.) depend on dense understorey for nesting and/or foraging. Four of the species (blue tit Cyanistes caeruleus, nuthatch Sitta europaea, robin Erithacus rubecula and chaffinch Fringilla coelebs) do not depend so strongly, if at all, on the field or shrub layers but are included as control species here. The 11 understorey species were chosen on the basis that they are widely believed to be dependent on dense understorey habitats, are potentially vulnerable to habitat modification by deer browsing (Fuller 2001) and had been recorded on sufficient BBS sites for change in relative abundance to be monitored (Baillie et al. 2009).
Deer have relatively large home ranges, though this varies between species, so the surveyed sites may not be at the best scale at which to measure the local abundance of these species. In addition, because deer are not very detectable in woodland, the observed rates of change of deer are likely to have many errors associated with them. For these reasons, we used generalised additive models (GAMs) with logarithmic link and quasi-Poisson error structure to remove some of this variability by producing a smoothed relative abundance surface for each deer species across England (after Wood 2006). GAMs have the advantage over GLMs for this purpose in that they incorporate a nonparametric smooth, enabling us to explore spatial patterns of relative abundance which are not constrained to have a specific form. The amount of smoothing was optimised automatically within the model-fitting procedure using generalised cross validation (GCV) scores. A gamma value of 1·4 was included in each model to reduce the possibility of overfitting (Wood 2006). We fitted a model to the data for each year separately, in which easting and northing were expressed as a two-dimensional smooth. Ten landcover variables (Centre for Ecology and Hydrology, CEH aggregate classes; Fuller et al. 2002) were also included as predictor variables. We conducted the analyses in R (R development core team, 2009) and used the ‘mgcv’ package (Wood 2006) to fit the GAMs.
The residuals and Moran’s I values provided no evidence that spatial autocorrelation was important in these models. Examination of the residuals and calculation of the percentage of deviance explained by each model suggested that model fits were good. Additional model evaluation was carried out for each deer species and year by removing 10% of site counts, using the remaining 90% to predict counts for the removed 10%, and repeating for each of 10 nonoverlapping samples of 10%. Correlations between observed and predicted counts were examined to assess predictive ability of the models. Whilst the correlation coefficients for these species were not particularly high, they were considered satisfactory considering the small 1-km scale of the predictions and that we were predicting counts rather than presence/absence (Table S2a, Supporting Information). In addition, these values compare favourably to a comparison of observed values produced using the whole data set (Table S2b, Supporting Information). Comparing the observed values in three pairs of years (1995/1996, 1999/2000 and 2004/2005) reveals a high degree of inherent variability in observed counts (Table S2c, Supporting Information). This allowed us to determine the proportion of the natural variability which the model validation described, which was high (Table S2c, Supporting Information). The fitted models were used to produce year-specific modelled deer abundances across England for each deer species for each surveyed BBS site. These predicted deer abundances were used as predictors in the following analyses.
Changes in the abundance of birds and deer may be driven by different, but concurrent processes, and for this reason, it is easy to misinterpret correlations at a large spatial scale. To examine whether there is a relationship between deer and bird populations, we used an analytical approach developed by Freeman & Newson (2008), which increases the efficiency of data use and the precision of parameter estimates over previous methods for examining the importance of ecological interactions in monitored populations. Freeman & Newson (2008) modelled a change in the abundance of one species as a response variable, in relation to the abundance of another species at the site level. In our case, a change in abundance of a focal bird species in relation to the abundance of roe deer, fallow deer and Reeves’ muntjac. Several biotic and abiotic covariates were also included to control for environmental variability and to provide greater context to our question regarding the influence of deer abundance upon change in bird numbers. To increase the focus of the study to sites where deer may reasonably have an impact on bird populations associated with dense understorey habitats, we restrict the analyses to 1936 woodland and farmland BBS squares in lowland England. This was done by (i) excluding 1-km squares classified as marginal upland or true upland by the CEH Land Classification System (Bunce et al. 1996) and (ii) using CEH Landcover Map 2000 data (Fuller et al. 2002) to exclude any squares with 25% or more urban/industrial cover (landcover type 21), i.e. major urban areas. Considering the mobility and landscape use by deer (Dolman et al. 2010), the impact of deer on bird population change may vary depending on the extent of woodland and farmland in the wider area around a 1-km BBS square. For each BBS square, we calculated the percentage coniferous woodland, broadleaved woodland and farmland (landcover types 14–23) in the wider area within and around 1-km squares of interest (1-km square of interest and neighbouring squares, i.e. nine 1-km squares). The interactions between these three habitat types and each deer species were included as covariates in the following models.
To determine the impact of a single deer species on a bird species, it was necessary to control for the impact that other deer may have on the bird species of interest, by including all three abundant deer species in the model. The impact of deer would be mainly through an indirect effect on the understorey vegetation, which may not have an impact on the bird species immediately. We therefore considered three sets of models that considered either no lag, a 1-year or a 2-year lag for the effect of deer. Because there is a loss of 1 year of data with each year lag, further year lags were not considered. If we were to consider all combinations of lagged and not lagged for each deer species, we would have to fit 33 = 27 models for each bird species, which is clearly impractical and would considerably increase the problem of multiple testing and interpretation. For this reason, we fitted the three sets of models, with the same number of years used in each model and Akaike Information Criterion (AIC) used to identify the best fitting model for subsequent analysis. Model selection was then carried out on the following variables using only the best fitting time-lag.
Weather may influence bird abundance independently of deer, so we included changes in temperature (mean daily minimum temperature) and rainfall (days with rainfall ≥1 mm) during the preceding breeding season (using species-specific breeding seasons, see Table S1, Supporting Information). For resident species, weather variables from the preceding winter (December–February) were also included as covariates in the model. We used spatial monthly weather data provided at a 5-km resolution by the Meteorological Office through the UK Climate Impact Programme matched to the appropriate 1-km square (UKCIP, http://www.ukcip.org.uk). Multiple deer species were incorporated into the Freeman & Newson’s (2008) model as follows: Suppose that Ni,t is the count of the focal bird species, Pi,t the modelled count of the deer species at site i in year t, Rt is the instantaneous rate of change of the focal bird species population during the period t − 1 to t in the absence of any covariate effect and α is the effect on that rate of change of the deer species. This allows us to separate out the contribution that individual deer species make to change in the bird population of interest α, over and above other potential unidentified drivers of bird population change Rt that are not accounted for by deer and other covariates included in the model. We assume that the observed counts Ni,t have a Poisson distribution and account for large-scale annual changes and the local effect of the deer species. The deer–bird relationship adopted here models the rate of change in the focal bird species abundance as a linear function of the deer abundance, with α as the slope and year-specific intercepts, Rt. Parameter α therefore measures the response in rate of change of the focal bird species per unit of deer abundance; if it is negative, it indicates that the instantaneous rate of change of the focal bird species declines with deer abundance and vice versa. The general model where the effect of deer abundance is not lagged is shown below (a model with 1- or 2-year time-lags is produced by modifying the t subscripts to t − 1 or t − 2 subscripts for the deer terms). This gives us the following model. For illustration, we present the interactions between each habitat variable and each deer variable as a single term in equation 1, rather than specifying each of the 9 habitat*deer interaction terms.
- (eqn 1)
Models were implemented by fitting generalised linear models using the genmod procedure in sas (SAS Institute 2001) with Poisson errors. Nonsignificant weather and habitat*deer interaction terms were removed by stepwise backwards deletion. We accounted for variation between sites by adopting fixed site effects as well as year effects, and treated consecutive observations at a site as independent Poisson-distributed observations. Examination of the residuals suggested that serial correlations were minimal, and model fits were good.
To help interpret the importance of any significant negative relationships identified here, we estimated the impact of deer on national populations of these species by considering the deer coefficients in relation to the change in modelled deer abundance on BBS squares. This was done by rearranging equation 1 to estimate the annual rate of change in the focal bird species at each site as a result of changes in deer abundance as shown in equation 2. Specifically, this equation estimates the average rate of change for one site i in the presence of one deer species from year t to year T.
- (eqn 2)
Excluding BBS squares where deer were never recorded for each of the remaining sites, the modelled deer abundance was inserted into the equation along with the model coefficients describing the effect of deer abundance α, and the Rt’s, describing the effect on annual growth rate in the absence of deer. The equation could then be used to estimate the effect of each deer species P on the population growth rate of each focal bird species at each site, by setting α = 0 to estimate the growth rate with no deer present. An average effect of each deer species on each bird species was then calculated across sites.
We also considered whether any broad pattern emerged across the full suite of deer coefficients by calculating a weighted mean of deer coefficients across bird species for understorey and nonunderstorey specialists separately. This provided a way of evaluating the relative impact of the different deer species.