Methods
Field site
The experiment was carried out at Whim bog in the Scottish Borders (3\({}^{\circ}\) 16’ W, 55\({}^{\circ}\) 46’ N). The site is at a transition between lowland raised bog and blanket bog, on 3-6m of deep peat. Mean temperatures of the air and soil (at 10-cm depth) were 8.6 \({}^{\circ}\)C and 7.7 \({}^{\circ}\)C respectively. The annual rainfall was 1092 mm (734-1462 mm range). On average, the water table was 10 cm below the peat surface, so relatively wet for most of the year. The peat was very acidic, with pH 3.4 (3.27-3.91 in water). The vegetation was mainly classified as a Calluna vulgaris - Eriophorum vaginatum blanket mire community \citep[M19 in the UK National Vegetation Classification,][]{Rodwell1998}, though M15, M17, M18 and M19 classes were all present. Replicate plots were highly variable and dominated by unmanaged Calluna vulgaris of variable age and stature occurring as mosaics containing Calluna vulgaris and Sphagnum capillifolium (Ehrh.) Hedw. hummocks and hollows containing S. fallax (H. Klinggr.) H. Klinggr. and S. papillosum Lindb. Other common species included Erica tetralix L. and the mosses Hypnum jutlandicum Holmen & E. Warncke and Pleurozium schreberi (Brid.) Mitt.
Experimental Treatments
Nitrogen was applied to the site using two different treatment systems, for dry deposition of NH\({}_{3}\) gas, and wet deposition of NH\({}_{4}^{+}\) and NO\({}_{3}^{-}\) in solution. Treatments commenced in June 2002 and continued all year round, except when temperatures were near freezing.
NH\({}_{3}\) deposition was manipulated using a free-air release system \citep{Leith2004}. NH\({}_{3}\) was supplied from a cylinder of pure liquid NH\({}_{3}\), diluted with ambient air and released from a perforated 10-m long pipe, 1 m off the ground. NH\({}_{3}\) was released only when the wind direction was in the south-west, between 180 and 215\({}^{\circ}\), temperatures exceeded freezing and wind speed exceeded 2.5 m s\({}^{-1}\). This produced a sector downwind wherein NH\({}_{3}\) decreased with distance from the fumigation source. NH\({}_{3}\) concentrations were measured 0.1 m above the vegetation using Adapted Low-cost Passive High Absorption (ALPHA) samplers \citep{Tang2001} at 8, 12, 16, 20, 24, 32, 48 and 60 m from the source along the transect. A detailed profile was measured to capture the concentration gradients both vertically and horizontally \citep{Leith2004}. NH\({}_{3}\) deposition was calculated from the concentration measurements, using the method of \cite{Cape2008}. The deposition at the permanent quadrat locations was interpolated using ordinary kriging, assuming the deposition velocity was spatially homogeneous.
Wet deposition of NH\({}_{4}^{+}\) and NO\({}_{3}^{-}\) was experimentally increased in a number of replicated plots in a randomised block design, using a water sprayer system \citep{Sheppard2004}. Using rainwater collected on a 178 m\({}^{2}\) pitched surface and stored in a 1.25 m\({}^{3}\) reservoir, concentrated solutions of either NH\({}_{4}\)Cl or NaNO\({}_{3}\) were diluted and transferred to each plot via lengths of 16-mm pipe. Each pipe terminated in a central sprayer with a 360\({}^{\circ}\) spinning disc that distributed the solution uniformly over the 12.8 m\({}^{2}\) plot. The volume of solution applied to each plot was monitored using a water meter on each supply line. Three treatment levels were applied, aiming to provide total nitrogen deposition rates of 1.6, 3.2 and 6.4 g N m\({}^{-2}\) y\({}^{-1}\), in addition to a control treatment which only received ambient nitrogen deposition (0.8 g N m\({}^{-2}\) y\({}^{-1}\)). The three treatment levels were achieved by applying either NH\({}_{4}\)Cl or NaNO\({}_{3}\) solution at concentrations of 0.57, 1.71 or 4.0 mmol dm\({}^{-3}\). In addition, phosphorous and potassium (PK) were added in the form of (K\({}_{2}\)HPO\({}_{4}\)) to the lowest and highest treatment levels, in the ratio 1:14 P:N, following the P:N ratio of amino acids. NH\({}_{4}^{+}\) and NO\({}_{3}^{-}\) treatments increased precipitation amounts by ca. 10%. Control plots receive the additional rainwater without any additional nitrogen. There were four blocks, with one treatment level in each, to give a total of 44 plots. The sprayer system was automatically triggered every 15 minutes, so long as there was sufficient rainwater in the collection tank, air temperature was above 0 \({}^{\circ}\)C and wind speed was above 5 m s\({}^{-1}\). This produced a realistic pattern of high frequency, extensive nitrogen deposition, with ca. 120 applications y\({}^{-1}\).
Soil water samples were extracted from dipwells in all plots approximately monthly. Concentrations of all detectable ions in the soil water were measured by ion chromatography following filtration. The detection limits were 0.014 and 0.062 mg l\({}^{-1}\) for NH\({}_{4}^{+}\)-N and NO\({}_{3}^{-}\)-N respectively.
Vegetation survey
Vegetation species composition was surveyed in all plots over the course of the experiment, usually every two years.
In each experimental plot, three permanent quadrats (40 \(\times\) 40 cm) were established before the start of treatment application in 2002. These were sub-divided into 16 sub-quadrats (10 \(\times\) 10 cm). At each survey, the percent cover of all species was recorded at the sub-quadrat level. For each species, the 16 sub-quadrat values were averaged to give a mean cover for each quadrat.
Statistical analysis
Univariate analyses
For the six most common species, the change in cover in each quadrat was analysed using a linear mixed-effects model \citep{Pinheiro2006}. We fitted fixed-effect terms for NH\({}_{3}\)-N deposition rate, \(F_{\mathrm{NH_{3}}}\), NH\({}_{4}^{+}\)-N deposition rate, \(F_{\mathrm{NH_{4}}}\), NO\({}_{3}^{-}\)-N deposition rate, \(F_{\mathrm{NO_{3}}}\), PK, and interactions between PK and \(F_{\mathrm{NH_{4}}}\) and PK and \(F_{\mathrm{NO_{3}}}\). Random-effect terms with a design matrix \(\mathrm{Z_{ijk}}\) were included to account for the repeated measures on each quadrat location, i, nested within each plot, j within each experimental block, k.
The analysis included the interaction effects of time, specifying four two-way interactions between time and NH\({}_{3}\), NH\({}_{4}^{+}\), NO\({}_{3}^{-}\) and PK, and two three-way interactions, between time, NH\({}_{4}^{+}\), and PK, and between time, NO\({}_{3}^{-}\) and PK.
Multivariate analyses
With ecological data, multivariate methods are commonly used to find those factors that best explain the differences in species composition between samples. Partial Least-Squares regression (PLS) is a multivariate technique, closely related to principal components analysis (PCA) and multiple linear regression. Multiple linear regression finds the maximum correlation between a multivariate matrix \(\mathbf{X}\) of independent variables and a response variable \(y\); PCA aims to capture the maximum variance in \(\mathbf{X}\) in a small number of new orthogonal components, calculated by re-weighting each of the original axes (species cover). PLS attempts to do both of these, by maximizing the covariance between \(\mathbf{X}\) and a multivariate response matrix \(\mathbf{Y}\). In this context, the response matrix comprises the plant species cover data, and the independent variables are the soil water chemistry data describing the changes to the physicochemical environment that accompany the nitrogen treatments. As in PCA, the method works by calculating loadings to re-weight the original axes to maintain orthogonal scores. In PLS, the components are obtained iteratively, starting with the single value decomposition of the crossproduct matrix \(\mathbf{S}=\mathbf{X}^{\mathsf{T}}\mathbf{Y}\), thereby including information on variation in both \(\mathbf{X}\) and \(\mathbf{Y}\), and on the correlation between them. Mathematical details are given by \citet{Wold1984} and \citet{Mevik2007}, with applications to ecological data discussed by \citet{Braak1993}.
Principal response curves (PRC) are a variant of redundancy analysis (itself a variant of PCA) which focus on the differences between the species compositions of the treatments and that of the control at the corresponding time. In this approach, \citet{VandenBrink1999} model the cover of each species as a sum of three terms: its mean cover in the control, a time-specific treatment effect, and an error. They denote \(T_{dts}\) to be the treatment effect of treatment \(d\) at time \(t\) for species \(s\). A first estimate of the treatment effect is the mean difference between the cover in treatment \(d\) and that in the control at each year. The response pattern of interest for each species \(s\) thus consists of the set of the treatment effects \(T_{dts}\). By further modelling the response pattern \(T_{dts}\) for each species as a multiple \(b_{s}\) of one basic response pattern \(c_{dt}\), i.e., \(T_{dts}=b_{s}c_{dt}\), the statistical model for the cover data becomes:
\begin{equation}
y_{tds}=y_{0ts}+b_{s}c_{dt}+\epsilon_{tds}\par
\\
\end{equation}
where \(y_{tds}\) is the cover of species \(s\) in treatment \(d\) at time \(t\), \(y_{0ts}\) is the mean cover of species \(s\) in time \(t\) in the control (\(d\) = 0), and \(\epsilon_{tds}\) is an error term with mean zero and variance \(\sigma_{k}^{2}\). Note that \(c_{0t}\) = 0 for every \(t\), because by definition \(T_{0ts}\) = 0 for every \(t\) and \(s\). The least-squares estimates of the coefficients \(b_{s}\) and \(c_{dt}\) can be found by a partial PCA model \citep[see Appendix 1 in][]{VandenBrink1999}.
PRC thus provides two sets of coefficients, which can be interpreted graphically. The first set consists of the treatment-time coefficients \(c_{dt}\) estimated for each combination of the treatment levels and time-point. \(c_{dt}\) represents the effect size of treatment \(d\) at time \(t\) relative to the control. \(c_{dt}\) values are depicted in the principal response curves, a line-plot of \(c_{dt}\) against time grouped by treatment. The second set of coefficients are the loadings for the species, \(b_{s}\). They represent the resemblance of species \(s\) to the overall response pattern specified by the principal response curves (i.e., the set of \(c_{dt}\) values). For the PRC analysis, quadrats in the NH\({}_{3}\) treatment were binned in groups matching the nitrogen deposition levels used in the NH\({}_{4}^{+}\) and NO\({}_{3}^{-}\) treatment, with an additional group where deposition was \(>\) 7 g N m\({}^{-2}\) y\({}^{-1}\).
In order to assess any effects of nitrogen deposition on community classification through time, plots were classified according to the the National Vegetation Classification (NVC) \citep{Rodwell1998} using Hill’s “G” metric \citep{Hill1989}. The best fitting communities were then tabulated to assess any changes over time and with nitrogen treatment. \citet{Hill1989} provides a rough guide to the goodness-of-fit of communities to the NVC classes, ranging from “Very good” (80-100) to “Very poor” (0-49).