1 Introduction

Climate change and increased nitrogen deposition are two aspects of current global change, which are known to have consequences for the functioning of ecosystems and biodiversity (IPBES 2019). In the future, it is expected that drought events will become more frequent and severe (Spinoni et al. 2018), and it is still uncertain how plant communities will respond to this change and how possible drought effects may interact with nitrogen availability (e.g. Stampfli et al. 2018; Van Sundert et al. 2021; Yu et al. 2022).

Plant species may be adapted to severe drought events by having deep roots and/or xeromorphic (i.e., drought-resistant) leaves (Buckland et al. 1997). Xeromorphic characters of leaves are negatively correlated to specific leaf area (SLA) and positively correlated to leaf dry matter content (LDMC) (Jung et al. 2014; Wright et al. 2001). These leaf plant traits are readily available in plant trait data bases, and a number of studies have demonstrated that drought events may alter the distribution of plant traits (Jung et al. 2014; Vitra et al. 2019; Wellstein et al. 2017). On the other hand, in non-drought years increased nitrogen availability selects for fast growing competitive species with high resource acquisition rates and increased shoot/root ratios (Johnson et al. 2008). Thus, leaf traits that were positively selected during drought events are selected against in normal years with increasing amounts of nutrients, i.e. increasing nitrogen availability is positively correlated with SLA and negatively correlated with LDMC (Craine et al. 2001; Garnier et al. 2007). Such temporally varying community selection forces stimulate important plant ecological research questions. For example, do possible community selection forces during a drought event depend on the level of nutrients? And do the community resistance and resilience to a severe drought event deteriorate with increasing nitrogen availability (de Vries et al. 2012; Kiene et al. 2023; Ploughe et al. 2019; Yu et al. 2022)?

Both water and nitrogen availability are known to affect plant growth and important life history characteristics, such as the root-shoot ratio, and, consequently, are expected to be important factors in plant competitive interactions (e.g., Liang et al. 2022; Lütke Schwienhorst et al. 2022; Meyer-Grünefeldt et al. 2015; Ploughe et al. 2019; Stampfli et al. 2018). For example, the effects of drought and nitrogen availability on the competitive interactions between Lilium bulbiferum and Secale cereale were studied in a mesocosm experiment (Lütke Schwienhorst et al. 2022). In the absence of nitrogen addition, the drought effect on L. bulbiferum, measured by the percent of dead tissue, decreased with the abundance of S. cereale—suggesting a positive/facilitative effect of Secale cereale on L. bulbiferum under drought stress. However, when nitrogen was added, the outcome of interaction changed towards a negative/competitive interaction as the drought effect on L. bulbiferum increased with the abundance of S. cereale (Fig. 3, Lütke Schwienhorst et al. 2022). This shows that the effect of drought on species interactions may depend on the level of available nitrogen (Lütke Schwienhorst et al. 2022). In a study of a Leymus chinensis meadow ecosystem, the effect of nitrogen addition varied with drought severity (Shi et al. 2018). At the peak of the drought, assimilation, stomatal conductance and transpiration of L. chinensis were all reduced, both with and without nitrogen addition, but the reduction was smaller when nitrogen was added.

While it is important to study the detailed ecological mechanism underlying the demonstrated three-way interactions among drought, nitrogen and inter-specific competitive interactions, it is equally important to study the effect of drought and the mediating effect of nitrogen at the community level, such that more general ecological conclusions may be obtained and possibly linked to community selection patterns of plant traits and plant functional types during drought events (Buckland et al. 1997). For example, Stampfli et al. (2018) studied the effect of drought under rainout shelter along a gradient of land-use intensity (a proxy for nutrient availability) on grassland vegetation. They found a synergistic impact of drought and land-use intensification on the grassland vegetation composition and that the more resource-acquisitive grasses increased in abundance at the expense of the deeper-rooted forbs. Meng et al. (2021) found a reduction in aboveground biomass during drought from 31.7% in unfertilized plots to 47.6% in fertilized plots, and, in experimentally fertilized natural grasslands in the Great Plains. Bharath et al. (2020) found that perennial grasslands fertilization reduced drought resistance. Soil type, soil chemical properties, competition, plant age, plant life type, response in root-shoot biomass and grazing pressure are probably important factors for determining the combined effect of drought and nitrogen (Britton et al. 2003). Moreover, the stochasticity of extreme drought events generally augments the temporal variation of the environment, which may lead to higher diversity through the storage effect (Chesson et al. 2014).

There are no consistently used ecological definition of drought, but here we will use the terminology of Ploughe et al. (2019), who introduce the concept of an ecological drought where functional changes and plant mortality are observed. An ecological drought is more prolonged than a meteorological drought and may eventually lead to long-term relative community changes if the drought event persists.

In this study, the effect of a drought event during July and August 2018 was investigated in a replicated long-term field experiment that was set-up in 2001 to study the combined effect of nitrogen availability and the herbicide glyphosate in a grassland agro-ecosystem (Bruus Pedersen et al. 2004; Holst et al. 2008; Strandberg et al. 2012). The plant cover of all higher plants was measured in permanently placed subplots in 2018, 2019 and 2021. We assume that the effect of the drought on observed plant cover was limited at the 2018 summer census performed June 20th, which therefore was used as a measure of the status of the pre-drought plant community. The effect of the drought was clearly visible in late summer 2018, and the 2019 census was therefore used to measure the effect of the drought event on the plant community. Finally, the 2021 census was used to measure the possible recovery at the plant community level.

We used these three years of data collection to investigate the immediate effect of a drought event (comparing 2018 with 2019), the subsequent recovery (2021) and the potential mediating effect of nitrogen availability on the community composition and overall diversity. The community composition was investigated using model-based ordination of pin-point cover data (Damgaard et al. 2020), and from the resulting two-dimensional ordination plot the importance of plant traits on community composition was assessed. Furthermore, to address the multivariate nature of the data, where responses are likely correlated across species, we have built upon the model-based ordination approach and estimated the relationship between plant traits and the observed selection forces. This method allows us to estimate a two-dimensional trait selection response that is based on each species position in the two-dimensional ordination space.

2 Materials and methods

2.1 Experimental site

The experimental site was a former forest plant nursery on dry, nutrient-poor, sandy soil that had been laid fallow a couple of years prior to the start of the experiment (Bruus Pedersen et al. 2004; Strandberg et al. 2012). The site is quadrangular and surrounded by small parts of forest on two sides (south and west) and separated from the neighboring fields by 5-meter broad hedgerows on the other sides (Fig. S1). In 2001, the area was deep-plowed down to 60 cm to eliminate establishment from the soil seed bank and prepared for the experiment by harrowing and rolling. Thirty-one grassland species covering different life form strategies (CRS strategies sensu Grime 2001) were sown in the spring 2001. Each plot was 7 m × 7 m with a buffer zone of 1.5 m surrounding the plot. A buffer zone of 10 m separated the experiment from the surrounding vegetation. The buffer zones were also sown with the seed mixture. Every year in March, the woody species were cut to slow down the establishment of species such as Pinus spp. and Quercus spp.

In the summer of 2018, there was a severe drought event at the site where a large proportion of the plants wilted, which was followed by the death of aboveground part of most plants. The drought index for the site calculated by the Danish Meteorological Institute was 9.9 in July measured on a scale where maximum drought is 10 (Scharling and Vilic 2009).

2.2 Experimental manipulations

The experimental manipulations with glyphosate and nitrogen (0, 25 and 100 kg N/ha) were set up as a complete randomized block design with 10 replicates. As the purpose of the present analysis was to investigate the effect of drought and the possible mediating effect of nitrogen availability, the plots receiving glyphosate were excluded from this analysis. All plots received phosphorus (53 kg/ha), potassium (141 kg/ha), sulfur (50 kg/ha) and copper (0.7 kg/ha) every year. The nitrogen was added as N27 CAN, calcium ammonium nitrate (NH4NO3 + CaCO3). Since 2001, the manipulations have been performed once every year in the spring (mid-ultimo May).

2.3 Plant cover data

The most common way to quantify plant species abundance in grass and herb dominated communities is to measure plant cover, which is the relative area of the species when projected onto the soil surface. In the summer of 2018 (June), 2019 (July), and 2021 (August), plant cover data was sampled in 0.5 m × 0.5 m quadrates with the pin-point method (Damgaard and Irvine 2019; Levy and Madden 1933; Lindquist 1931; Pellissier et al. 2014) using a horizontal frame with 25 intersections, i.e. a 5 × 5 regular grid with 10 cm distance between all strings that made up the grid. At each intersection, a sharply pointed pin with a diameter of 0.5 mm was passed vertically through the vegetation and the cover of a species was measured by the proportion of the inserted pins that touched the species. Nomenclature follows Hansen (1991). There were two permanently marked randomly positioned pin-point frames (subplots) in each plot.

2.4 Plant traits

To further understand the ecological mechanisms underlying the observed changes in relative species abundance, we retrieved selected plant traits of the species from databases. The specific leaf area (SLA) and leaf dry matter content (LDMC) were retrieved from the LEDA trait database (Kleyer et al. 2008), which is a homogenous trait database established from plant measurements in North-West Europe, which is geographically and climatically close to Denmark. Furthermore, the Ellenberg indicator values (Ellenberg 1979) of nutrients (EN) and humidity (EF) and an indicator of stress tolerance (GS) as measured by Grime’s plant species CSR strategies (Grime 2001) were used as a proxy for plant phenology and life strategy features that may play an important role during drought events. Finally, it was recorded whether the species belonged to the grass family (Poaceae). The used species trait values are given in Table S1.

2.5 Model-based ordination

To visualize the effect of drought and nitrogen availability on the plant community composition, the observed pin-point cover data for all observed higher plant species were transformed into a two-dimensional representation by an ordination technique. Recently, several ordination methods have been developed, where the distribution of the sampled species abundance is taken into account. Collectively, these new techniques are referred to as model-based ordination, which offer the possibility of adjusting the sampling distribution to better account for the inherent mean-variance relationship in the species observations (Hui et al. 2015; Warton et al. 2015).

Here, we model the plant community composition using latent variables (Damgaard et al. 2020; Hui et al. 2015; Niku et al. 2017; Warton et al. 2015, 2016). More specifically, the objective is to reduce the pin-point cover data with k samples and n plant species to a \(k*p\) matrix of latent variables, where \(p<n\) is the number of latent variables and the dimension of the ordination plot, which in this study is set to two. The effects of the sampled year and nitrogen manipulations are included into the mean structure, which means that the latent variables model the residual correlation between species, i.e., the covariation that cannot be explained by sampling year and nitrogen.

Following Damgaard et al. (2020), the mean cover of species j in subplot i, denoted here as \({q}_{ij}\), is modelled as

$$logit\left({q}_{ij}\right)={\left({\varvec{\alpha }}_{fixed\left[i\right]}+{\varvec{\gamma }}_{random\left[i\right]}+{\varvec{z}}_{i}\right)}^{{\prime }}.{\varvec{\theta }}_{j}+{\beta }_{j}$$
(1)

where all parameters in bold denote vectors of dimension p. The vector \({\varvec{\alpha }}_{fixed\left[i\right]}\) denotes the fixed effects (nitrogen addition and year) at subplot i. The vector was calculated from five vectors of p dimensions with parameters using a design matrix (Table 1).

Table 1 Design matrix of the linear model
Full size table

The five vectors, \({\varvec{v}}_{1}\dots {\varvec{v}}_{5}\) (Table 1) model the effects of year and nitrogen level. The first vector, \({\varvec{v}}_{1}\), models the effect of the drought event (2019 cover data) on the community composition in control plots without nitrogen addition. The second vector models the recovery (2021 cover data) of the plant community after the drought effect in control plots without nitrogen addition. The third vector models the pre-drought (2018 cover data) effect of nitrogen addition on the community composition. The fourth vector models the mediating effect of nitrogen addition on the effect of the drought event (2019 cover data in plots with nitrogen addition). Finally, the fifth vector models the mediating effect of nitrogen addition on the recovery (2021 cover data in plots with nitrogen addition) after the drought event.

More specifically, a design matrix D of the fixed effect vector \({\varvec{\alpha }}_{fixed\left[i\right]}\) in subplot i and where \({n}_{i}\) is the annual application of nitrogen (0 to 100 [kg N/ha]) in subplot i, is constructed (Table 1), and the mean fixed effect at subplot i is then calculated as \({\varvec{\alpha }}_{fixed[i,j]}={{({\varvec{v}}_{1},{\varvec{v}}_{2},{\varvec{v}}_{3},{\varvec{v}}_{4},{\varvec{v}}_{5})}^{T}}_{j}.{\varvec{D}}_{\varvec{i}}\), where \({\varvec{v}}_{1}\dots {\varvec{v}}_{5}\) are five vectors, each of dimension two, which contain the fixed effect parameters that need to be estimated, \({\varvec{D}}_{\varvec{i}}\) is a row vector of dimension 5, and j is an index that takes the values 1 and 2.

The vector \({\varvec{\gamma }}_{random\left[i\right]}\) models random effects at the level of subplot i, i.e. the location of subplots within plots. The random effects are assumed to be drawn from a common multivariate normal distribution with a zero mean vector and an unstructured p*p covariance matrix, \({N}_{p}(0,\varvec{\Sigma }\)), where for p = 2, \(\varvec{\Sigma }=\left(\begin{array}{cc}{\sigma }_{1}^{2}& \rho { \sigma }_{1} {\sigma }_{2}\\ \rho { \sigma }_{1} {\sigma }_{2}& {\sigma }_{2}^{2}\end{array}\right)\). The standard deviations \({ \sigma }_{i}\) are assigned an improper uniform positive prior, while the correlation coefficient \(\rho\) is assigned a uniform prior between − 1 and 1.

The vector \({\varvec{z}}_{i}\) denotes a vector of \(p\) latent variables for each subplot i and, as is standard, is assumed to come from a standard normal distribution \({N}_{p}(0,\mathbf{I}\)), where the zero mean vector and identity covariance matrix are used to avoid location and scale invariance and ensure that the parameters in the model are identifiable (Hui et al. 2015; Niku et al. 2017).

The vector \({\varvec{\theta }}_{j}\) denotes a \(p\)-dimensional vector of coefficients or loading for species j (e.g. e.g. Hui et al. 2015) and is, similarly to the fixed effects above, assigned a weak prior \({N}_{p}(0,100 \mathbf{I}\)). Note that in order to ensure that the parameters are identifiable, we further apply a standard constraint of assuming that the upper triangular portion of the \(n*p\) matrix of species-specific coefficients is constrained to be zero, while the diagonal elements are constrained to be positive (see see Hui et al. 2015; Niku et al. 2017). Finally, the quantities \({\beta }_{j}\) denote a vector of constants chosen a priori, which are required to ensure that \(logit\left({q}_{ij}\right)\) are real numbers.

The underlying motivation for using model-based ordination instead of traditional ordination is to take the distributional properties of the species abundance sampling into account, such that e.g. the correct mean-variance relationship is used in the modelling (Warton and Hui 2017). In this study, we consider the multi-species pin-point cover data, and it has previously been demonstrated that a reparametrized Dirichlet-multinomial distribution is a suitable candidate distribution to model multi-species pin-point cover data (Damgaard 2015; 2019; Damgaard et al. 2017). The advantage of using the reparametrized Dirichlet-multinomial distribution is that the degree of spatial aggregation in plant communities is taken into account and explicitly modelled by a parameter \(\delta\) (Damgaard 2018). More specifically, the observed hierarchical multi-species pin-point cover data, Y, is modelled by a mean cover vector of the n-1 first species, \({q}_{j},\) and a parameter \(\delta\), which measures the degree of intra-specific spatial aggregation, by,

$$Y\sim Mn\left( {\sum\limits_{n} {y_{i} } ,\left( {p_{1} , \ldots ,p_{{n - 1}} ,1 - p_{1} - \ldots - p_{{n - 1}} } \right)} \right)$$
(2)
$$\Lambda \left( {p_{1} , \ldots ,p_{{n - 1}} } \right) \sim Dir\left( {\frac{{q_{1} - q_{1} \delta }}{\delta }, \ldots ,\frac{{q_{{n - 1}} - q_{{n - 1}} \delta }}{\delta },\frac{{1 - \delta }}{\delta } - \frac{{q_{1} - q_{1} \delta }}{\delta } - \ldots - \frac{{q_{{n - 1}} - q_{{n - 1}} \delta }}{\delta }} \right)$$

At the limit when \(\delta \to 0\), the reparametrized Dirichlet-multinomial distribution degenerates into the multinomial distribution. Here, the prior distributions of \({q}_{j}\) and \(\delta\) are assumed to be uniformly distributed between (0, 1) and (0.01, 0.95), respectively. For more details on the properties of the reparametrized Dirichlet-multinomial distribution, see Damgaard (2018).

2.6 Estimation and inferences

The proposed latent variable model was estimated using Bayesian Markov Chain Monte Carlo (MCMC) methods using the Metropolis-Hastings algorithm with normally distributed candidate distributions. Specifically, a MCMC chain with a burn-in period of 170,000 iterations followed by 30,000 additional iterations that were used for inferences.

Trace plots of the sampling chains of all parameters and latent variables and Geweke diagnostics plots were examined to assess their mixing properties and convergence of the MCMC chain. Additionally, the overall fitting properties of the model were checked by examining the regularity and shape of the marginal distribution of parameters as well as the distribution of the deviance \((=-2{log}L\left(Y\right|\theta \left)\right)\). The efficiency of the MCMC procedure was assessed by inspecting the evolution in the deviance.

We constructed an ordination plot with the posterior means of the five fixed effect vectors, where the starting points were chosen such that the temporal dynamics were more easily visible, e.g., the end point of the 2019 vectors was used as the starting point of the 2021 vectors, in order to visualize the degree of recovery. Furthermore, the posterior means of the species coefficients, \({\varvec{\theta }}_{j}+{\beta }_{j}\) for each species were plotted. Note that the mean cover of \({q}_{n}\) is not estimated in Eq. 2. The units in the ordination plots are relative to the residual latent variables of the samples, which are fixed to a unit standard variation for ease of interpretation (and parameter identifiability). However, to get a better view of both the fixed effect vectors and the position of the species in the same ordination plot, the length of the fixed effect vectors and the corresponding credibility regions were rescaled. The length of the first two vectors were multiplied by three. The length of the last three vectors that modelled the effect of nitrogen addition, which are scaled in the units of kg N/ha, were multiplied by 100, which corresponds to the largest addition level (100 kg N/ha).

The credibility regions of the five fixed effect vectors enable us to make statistical inferences on the effect of the different treatments on species composition. If the 95% credibility regions of treatment vectors did not overlap with (0, 0), it was concluded that the observed changes in the relative species abundance were significant.

All calculations were done using Mathematica (Wolfram 2016).

2.7 Hill–Shannon diversity

The effect of drought and the mediating effect of the nitrogen application on overall species diversity were measured using the Hill–Shannon diversity index. Estimates of species richness are strongly influenced by the presence of rare species (Haegeman et al. 2013; Roswell et al. 2021), and it is therefore preferable to use species diversity indices, such as the Shannon index, where species occurrence is weighted with its relative local abundance. Moreover, it has been recommended to use the Hill diversity transformation of diversity indices, since they are on the same scale as species richness (Hill 1973; Jost 2006; Roswell et al. 2021).

Here, the Hill–Shannon diversity index of each plot was calculated using the pin-point cover data as the Bayesian updated Hill–Shannon diversity index, where the mean species cover from all plots was used to fit the prior beta distributions of each species (Damgaard et al. 2022a). The Bayesian updated Hill–Shannon diversity indices were analyzed in a linear model with sampling year and the manipulated level of nitrogen as explaining factors. The residual variation in the linear model was assumed to be normally distributed, and this assumption was checked using residual plots.

2.8 Selection and plant traits

While the community selection forces during the drought event and the following phase of recovery operate on the species level, there may be some general patterns of the selection forces that are more transparent when considering the traits of the different species. To examine such patterns, standardized species traits values were regressed in a linear model onto the mean species position on the plane as estimated by \({\varvec{\theta }}_{j}\), and a two-dimensional gradient vector was calculated for each trait value. The dot products of these gradient vectors and the fixed effect vectors were then used to infer whether the direction of the estimated fixed effect vectors corresponded with the gradient vectors of the different traits (Fig. 1).

Fig. 1
figure 1

Conceptual figure of the use of the dot product to determine the selection forces on plant traits. A: trait gradient vector, where standardized species traits values were regressed in a linear model onto the mean species position on the plane as estimated by \({\varvec{\theta }}_{j}\). B: fixed effect vector that summarizes the community selection forces of the fixed factor. C: The orange vector is the projection of the red trait gradient vector onto the yellow fixed effect vector, and the dot product is the signed length of this yellow vector. The magnitude of the dot product measures the importance of the plant trait for community selection. For example, if a trait gradient vector and a fixed effect vector are parallel, then it was concluded that the trait may play an important functional role in the observed changes in relative species abundance. Conversely, if a trait gradient vector and a fixed effect vector are orthogonal, then it was concluded that the trait only played a minor role in the observed changes in relative species abundance

Full size image

3 Results

The MCMC iterations of the model-based ordination demonstrated a considerable covariation among the parameters, but overall, the fitting properties were judged to be acceptable (Fig. S2). The median degree of spatial aggregation of the plant species as modelled by the parameter \(\delta\) was 0.23, with a credibility interval between 0.20 and 0.26. The marginal posterior distribution of the fixed effect vectors, their confidence regions and the position of the species are shown in an ordination plot (Fig. 2).

Fig. 2
figure 2

Map of latent variables, posterior means of the five fixed effect vectors, where the starting points were chosen such that the temporal dynamics were more easily visible, e.g. the end point of the 2019 vectors was used as the starting point of the 2021 vectors. Full red arrow: change in species composition from 2018 to 2019 without nitrogen; full blue arrow: change in species composition from 2019 to 2021 without nitrogen; grey stippled arrow: the effect of nitrogen in 2018; red stippled arrow: the effect of nitrogen on the change in species composition from 2018 to 2019, blue stippled arrow: the effect of nitrogen on the change in species composition from 2019 to 2021. Note that the length of the fixed effect vectors and the corresponding credibility regions were rescaled. Mean species coefficients, \({\varvec{\theta }}_{j}+{\beta }_{j}\) for each species are shown; green numbers are grasses and brown numbers are herbs and other taxa. 1: Convolvulus arvensis, 2: Agrimonia eupatoria, 3: Solidago virgaurea, 4: Agrostis capillaris, 5: Hypochoeris radicata, 6: Elytrigia repens, 7: Achillea millefolium ssp. millefolium, 8: Linaria vulgaris, 9: Senecio jacobaea, 10: Festuca ovina, 11: Holcus lanatus, 12: Pilosella officinarum, 13: Leucanthemum vulgare, 14: Galium mollugo, 15: Lichen, 16: Euphorbia esula ssp. esula, 17: Verbascum nigrum, 18: Bryophyta, 19: Hypericum perforatum, 20: Tanacetum vulgare, 21: Rumex acetosella, 22: Festuca rubra, 23: Lepidium campestre, 24: Poa angustifolia, 25: Viola tricolor, 26: Urtica dioica, 27: Agrostis gigantea, 28: Viscaria vulgaris

Full size image

Judged by the 95% credibility regions, only the pre-drought selection due to variable nitrogen availability in the long-term experiment had a significant effect on the relative species abundances, i.e., the fixed effect vector was significantly different from 0 (Fig. 2). Neither the severe drought in 2018 nor the recovery until the summer of 2021 had a significant effect on the species community composition (Fig. 2). Furthermore, this conclusion of a non-significant effect of drought and the following recovery was not altered by the manipulated variation in nitrogen availability (Fig. 2).

The scale of the random effects, i.e., the effect of subplots within plots, is relatively high (Table 2), and high random effects will, anything else being equal, make it less likely to detect systematic treatment effects. Given that the scale of the random effects is relative to the unit standard deviation of the residual latent variables, we conclude that the random effects of plot tended to be somewhat larger than the variation among the residual latent variables (Table 2, σ1 is significantly larger than one). There was no significant correlation between the two dimensions of the random effects (Table 2).

Table 2 Marginal credible intervals for the posterior distribution of the parameters
Full size table

Selected standardized species trait values were regressed in a linear model onto the mean species position on the plane as estimated by \({\varvec{\theta }}_{j}\), and a two-dimensional gradient vector was calculated for each trait value. The dot products of these gradient vectors and the fixed effect vectors were then used to infer whether the direction of the estimated fixed effect vectors corresponded with the gradient vectors of the different traits (Fig. 1; Table 3). However, since the only fixed factor vector that was significantly different from 0 was \({ \varvec{v}}_{3}\) (pre-drought selection due to variable nitrogen availability), only these results are highlighted here. Three plant traits were shown to be almost equally important among the plant species that were selected for with increasing nitrogen availability (Table 3, indicated by the percentage of the total of the absolute values). Not surprisingly, the plant species with relatively high Ellenberg N and low Grime S values were positively selected with increasing nitrogen availability (Table 3, positive and negative sign of the dot products, respectively), whereas increasing nitrogen availability selected against grass species (Table 3, negative sign of the dot product).

Table 3 The dot products of plant trait gradient vectors and the fixed effect vectors (see Fig. 1)
Full size table

The diversity was estimated as Hill–Shannon diversity, and the mean diversity across all plots was 3.82 with a standard deviation of 1.26. The analysis of the overall diversity using a linear model showed the same pattern as the more detailed analysis of species composition. Pre-drought selection due to variable nitrogen availability had a significant effect on diversity, such that plots with added nitrogen had significantly lower Hill–Shannon diversity (Table 4, Nitrogen). Furthermore, the severe drought event and subsequent recovery did not influence the Hill–Shannon diversity (Table 4, Year), and there was no significant interaction between the drought event and nitrogen availability (Table 4, Year*Nitrogen).

Table 4 ANOVA table of the effect of year and nitrogen on the Bayesian updated Hill–Shannon diversity (R2 = 0.14)
Full size table

4 Discussion

Contrary to our prior expectation, we could not detect any significant effects of the severe drought in the summer of 2018 on the plant community composition or on overall diversity the following year and three years later. This result was surprising, since we visually observed significant wilting and death of aboveground part of most plants at the study area, and the drought index was 9.9 in July on a scale where maximum drought is 10 (Scharling and Vilic 2009). Using the terminology of Ploughe et al. (2019), the studied drought event in 2018 may thus be classified as an ecological drought, which is more prolonged than a meteorological drought, without reaching the onset of long-term relative community changes. Furthermore, we did not detect any significant mediating effect of nitrogen on the effects of drought.

The present results are somewhat similar to the results found by Yu et al. (2022), in a study of an arid grassland, who observed no interaction effect of drought and nitrogen on community composition. Likewise, Grime et al. (2008) reported a high degree of long-term resistance to drought in an infertile grassland.

A possible explanation for the high resistance of the plant community to the severe drought event could be that the studied grassland is well-established, and most species are perennial plants with roots that have been developed to withstand drought as the soil at the site is sandy and regularly dries out also under normal climate conditions. Furthermore, over the years significant intraspecific spatial aggregation has developed (\(\delta\) was estimated to 0.23), which, everything else being equal, will reduce the effect of interspecific competitive interaction in the absence of seedling recruitment (Stoll and Prati 2001), and this reduction will, again, lead to reduced competitive advantages of the more drought-tolerant plant species. The possible effect of intraspecific spatial aggregation in reducing the effect of interspecific competitive interaction may be investigated by analyzing the glyphosate treated plots, since the vegetation in these plots is less well-established and does not form a closed vegetation cover, but this is outside the scope of the present study.

An alternative explanation of the observed drought resistance may be the added micronutrients in the manipulated plots. Particularly potassium plays an essential role in plant physiological responses, e.g. enzyme activation, protein synthesis, photosynthesis, osmoregulation, stomatal movement, energy transfer, phloem transport, cation-anion balance and stress resistance (Marschner 2012), affecting everything from plant phenology, biomass and resistance towards abiotic stressors (Hansen et al. 2023; Wang et al. 2013). As such, it is conceivable that the added potassium in the field study adds a buffer to the detrimental effects of drought, which plants in more natural ecosystems would not benefit from.

It would have been beneficial to have access to more detailed information on the effect of the severe drought event on the physiology of the different plant species. This would have enabled a more detailed analysis of the effect of severe drought on the competitive and facilitative interactions between species pairs (Lütke Schwienhorst et al. 2022). Unfortunately, the fitting procedure using MCMC takes a very long time, otherwise it may have been relevant to increase the dimensions of the ordination to more than two dimensions, which, in principle, could increase the sensitivity of the analysis, be it, however, at the price of reduced ease of interpreting the result.

As expected, we found significant effects of pre-drought selection due to variable nitrogen availability on both plant community composition and overall diversity (Damgaard et al. 2022b). Increasing levels of nitrogen availability were associated with a decrease in Hill–Shannon diversity. The observed community compositional changes with increasing nitrogen availability could partly be explained by selection for plant species with relatively high Ellenberg N, low Grime S, and tall herbal species. The selection for relatively high Ellenberg N and low Grime S with increasing nitrogen availability was unsurprising and a natural outcome of how the indices are constructed. While grass species were more resilient to increasing nitrogen deposition than forb species in a study of acid grassland (Duprè et al. 2010), dominance by tall-herb communities have been documented in studies of plant community responses to high nutrient content in restoration of wetland (Moeslund et al. 2022). Furthermore, the dominant grasses at low and high nitrogen levels, respectively, are different. At low nitrogen, Festuca ovina dominates, whereas Elytrigia repens dominates at high nitrogen levels. This is in accordance with the Ellenberg N for the two species being 1 and 7, respectively. F. ovina has anatomical features characteristic of drought tolerance such as acicular leaves and dense root structure and the wild species have been shown to be moderately drought tolerant and good recovery following drought stress (Qiu et al. 2020). Traits that seem to be lost in cultivated F. ovina that compared to other cold-season grasses such as Festuca arundinacea, Festuca rubra and Lolium perenne had low forage production and recovery after drought (Taleb et al. 2023).

An important issue encountered by ecologists in species distribution modelling is to describe species responses to the environment and how these species responses are mediated by species traits. Specifically, this interaction effect between species traits and environmental variables on species distribution and how to include it in the multivariate models has been dubbed the fourth corner problem (Legendre et al. 1997; Niku et al. 2021). To address this issue, we have built upon the model-based ordination approach and incorporated the trait-environment relationship as the dot product of the gradient vectors and the fixed effect vectors. This method allows us to estimate a two-dimensional trait selection response that is based on each species position in the two-dimensional ordination space.