Applying the stress- gradient hypothesis to curb the spread of invasive bamboo

1. The stress- gradient hypothesis (SGH) provides a conceptual framework for ex plaining how environmental context determines the nature of biotic interactions. It may be also useful for predicting geographic variability in the effect of manage ment interventions on biological invasions. We aimed to test hypotheses consist ent with the SGH to explain context dependency in bamboo invasion of secondary forests in Japan, and establish a predictive understanding of forest management impacts on invasion. 2. We use a priori physiological knowledge of invasive giant bamboo, Phyllostachys bambusoides , to generate hypotheses consistent with the SGH. We modelled vari ation in giant bamboo occupancy within 810 secondary forest plots across the broad environmental gradients of Japan using a national vegetation database. 3. Consistent with the SGH, we find that the


| INTRODUC TI ON
World-wide, invasive species threaten the capacity of ecosystems to sustain biodiversity and deliver critical ecosystem services (Pyšek et al., 2012). To effectively prevent and control their negative impacts now and in the future, a predictive understanding of invasion success is required (González-Moreno et al., 2014). However, predicting invasion success is challenging due to its contingency on myriad interdependent abiotic and biotic drivers (Elliott-Graves, 2016;Pyšek et al., 2020). While abiotic conditions such as climate can explain the distribution of invasive species at macro-ecological scales (Ficetola et al., 2007), patterns of invasion often appear geographically idiosyncratic even within hospitable climatic ranges (Holmes et al., 2010). Such patterns are suggestive of interactive effects of broad-scale environmental conditions with local biotic interactions (González-Moreno et al., 2014). Both negative (e.g. competition) and positive (facilitation) biotic interactions are well-recognised in invasion ecology (Cavieres et al., 2005;Lucero et al., 2019), and the nature of these interactions have been shown to vary across broad extents (Early & Keith, 2019). However, the drivers of this variability, or 'context dependence', have yet to be characterised for a range of taxa and ecosystems (Chamberlain et al., 2014).
The stress-gradient hypothesis (SGH) provides a unifying conceptual framework that can help explain how environmental context determines whether native species inhibit or facilitate invasion (Bruno et al., 2003). The SGH predicts that the relative frequency or strength of facilitation and competition will vary inversely across abiotic stress-gradients (Bertness & Callaway, 1994;Callaway & Walker, 1997). Under increasingly stressful conditions, the presence, size and density of neighbouring 'nurse' species can become increasingly important in facilitating the persistence of a species in conditions that would otherwise not be tolerated, enabling the latter species to expand its realised niche (Bruno et al., 2003). For instance, cushion plants can facilitate recruitment of other herbaceous species in alpine environments by buffering temperature extremes (Cavieres et al., 2005). Conversely, under benign physical conditions, interspecific competition with neighbours may increase, such that otherwise facilitating 'nurse' species become competitors instead (Bertness & Callaway, 1994;Brooker et al., 2008;Callaway & Walker, 1997). While the SGH has successfully explained variability in biotic interactions in stressful settings, such as alpine, intertidal and early successional environments (He et al., 2013), there have been few applications, however, to applied questions relevant to ecosystem management, except for the deliberate use of nurse plants as restoration tools in harsh environments (Guignabert et al., 2020). However, management efforts that alter the abundance or distribution of nurse species could also have unintended consequences for invasive species, which the SGH could help to mitigate.
Here, we apply the SGH to predict plant invasion success in response to forest management across the broad environmental gradients of Japan. Specifically, we use the SGH and a priori physiological knowledge of invasive giant bamboo Phyllostachys bambusoides, to generate hypotheses about bamboo occupancy, and apply this to fine-scale forest survey data collected across the broad environmental gradients of Japan. Importantly, we show how the SGH could help elucidate where species interactions should be considered when formulating local forest management practices that involve changes to the abundance of species facilitating or competing with invaders.
Bamboo invasion is an ideal case study for testing the applicability of the SGH in an applied context. First, the distributional limits of Phyllostachys species in both their native and non-native ranges are strongly temperature-dependent (Song et al., 2017;Takano et al., 2017). Within these limits, its occupancy within secondary forests is highly variable (Isagi & Torii, 1997; examples in Appendix S1), which could reflect interactions between biotic and abiotic drivers. Second, physiological studies of multiple Phyllostachys species under controlled settings have demonstrated that photoinhibition (light-induced decline of photosynthesis) occurs under moderate light intensities at low temperatures, but can be ameliorated at warmer temperatures or under shade (Lou et al., 2018;Van Goethem et al., 2013. In other words, physiological understanding of Phyllostachys points to interactions among multiple abiotic drivers (He & Bertness, 2014;Maestre et al., 2009), suggesting observed fine-scale geographic variation in its invasion success might be consistent with the SGH. Specifically, we hypothesised that in cool and high-light environments, the shade cast by forest canopies could exert positive (facilitative) effects on bamboo invasion under stress (Hypothesis 1, Figure 1c). However, under benign conditions corresponding to warmer temperatures, we hypothesised that dense canopies might exert negative effects due to competition for space, nutrients or available light (Hypothesis 2, Figure 1d). We made the explanatory prediction (sensu Mouquet et al., 2015) that when under high light, effects of photoinhibition would manifest as a shift in the thermal niche of bamboo along a gradient of canopy cover, consistent with the SGH.

| Study system and taxa
Giant bamboo (Phyllostachys) species were introduced to Japan from China around three centuries ago for their culm woods and edible shoots (Isagi & Torii, 1997). Bamboo forests were widely planted across western and central Japan as part of production landscapes known as 'satoyama', mosaics of secondary broadleaved forests managed as coppices, residential areas, paddy fields and grasslands (Takeuchi, 2010). Rapid economic growth, the introduction of new fuels and large-scale urban development during the 1950s-1960s resulted in widespread satoyama abandonment. Without active management, bamboo plants extend their rhizomes laterally, driving bamboo forest expansion into adjacent secondary forests that were historically managed as coppices, but are now abandoned (Fukamachi et al., 2001). Carbohydrates from the clonal adult root system can sustain the growth of new bamboo shoots, that are initiated every summer and emerge above-ground the following spring, making them shade-tolerant and able to invade forests with dense canopies (Wang et al., 2016). Bamboo invasion can inhibit regeneration, decimate understorey plant communities and alter soil chemistry (Nakai & Kisanuki, 2006).

| Ecophysiology of Phyllostachys spp
Photosynthetic and growth rates of bamboo are strongly temperature-dependent. P. bambusoides can achieve 90% of its F I G U R E 1 Conceptual figure representing hypotheses consistent with the stress-gradient hypothesis. (a) The effect of canopy cover will change in direction across the stress-gradient, determined by light intensity and temperature (b). (c) Hypothesis 1: Under low stress (low light intensity + any temperature, or high light intensity + warm temperature), dense canopies may cause light to become a limiting resource, thus limiting bamboo establishment in secondary forests. Forest thinning should therefore increase the probability of invasion. (d) Hypothesis 2: Under stressful conditions (high light that can cause photoinhibition + low temperature), the shade cast by dense canopies may facilitate establishment by protecting bamboo from light stress. Under these conditions, forest thinning should reduce the likelihood of invasion

Factorial experiments have demonstrated strong photoinhibition of
Phyllostachys species under unshaded cold conditions (3-10℃), but weak photoinhibition in both unshaded warm (13-20℃) conditions and shaded cold conditions (Van Goethem et al., 2015). In contrast to warm conditions, thermal dissipation of excess excitation energy is not possible under cold conditions. Phyllostachys species photosynthesise year-round (Wang et al., 2016), with photosynthetic capacity increasing from a spring minimum, due to the combination of low air temperatures and increasing light irradiation, to an autumn maximum (Gratani et al. 2008;Van Goethem et al., 2013). These results suggest that the degree of shade cast by dense forests might have important facilitative effects at low temperatures, protecting Phyllostachys from photoinhibition. We therefore set out to test the hypothesis that the effect of canopy cover on bamboo invasion probability would depend on, that is, interact with, light intensity and temperature (Figure 1c,d).
In regions with warmer temperatures, where growth is less limited by photoinhibition, canopies could plausibly exert negative effects due to competition for space, nutrients or light.

| Vegetation data
We retrieved bamboo incidence data from the vegetation survey database of Japan, provided by the Biodiversity Center of Japan (http:// gis.biodic.go.jp/webgi s/files/). This dataset consists of vegetation and geospatial data for >55,000 plots surveyed between 2000 and 2015, compiled from a range of sources, comprising species presence/absences, in addition to the percentage cover and height of vegetation. For plots located in forest habitats, consultants recorded the percentage cover of canopy, and soil properties. We focussed on P. bambusoides, which was the most prevalent in secondary deciduous forests, and refer to it as 'bamboo' from hereon.

| Environmental data
We selected potential explanatory variables of bamboo occupancy based on existing empirical studies (Someya et al., 2010;Suzuki, 2015; Suzuki & Nakagoshi, 2008; Table S2). Climatic variables were available at a 1-km resolution (Japan Meteorological Agency, 2010), and included temperature, rainfall, solar radiation, sunshine duration and snow depth based on annual averages between 1971 and 2000. For temperature, we calculated average annual values of monthly minima, means and maxima, in addition to obtaining values for the coldest and warmest months. We calculated two distance metrics to characterise the likelihood of propagule dispersal by vegetative spread. Distance to forest edge was expected to drive down invasion probability, as bamboo invades by spreading from bamboo forest into adjacent habitats, resulting in bamboo forest expansion. Distance to road serves as a proxy of secondary forest abandonment, because secondary forests and bamboo forests are less likely to be managed if far from roads and relatively inaccessible (Kobayashi & Nakamura, 2018). Indeed, distance to road has been shown to be positively correlated with expansion rates (Suzuki, 2015;Suzuki & Nakagoshi, 2008).
Local topographical variables have been shown to affect bamboo spread, including slope and measures of aspect (Suzuki & Nakagoshi, 2008), and were calculated using a 30-m digital elevation model (Table S2). Invasion of secondary forest by bamboo is more likely to occur in abandoned areas where bamboo and secondary forest management have ceased (Suzuki, 2015). In the absence of fine-resolution forest management data and assuming that bamboo and secondary forest abandonment coincide with agricultural abandonment, we used the extent of abandoned agricultural land (paddy fields, farmland, pasture) available at a 5-km resolution (Osawa et al., 2013), following Suzuki (2015). It has been suggested that stand structure might be important in determining the likelihood of giant bamboo spread (Isagi & Torii, 1997), so measures of forest canopy cover and forest type (i.e. whether dominated by an evergreen or deciduous tree species) were extracted from the plot database.
We also considered survey year as a potential explanatory variable, as a correlate of time since widespread landscape abandonment.
Minimum residence time (Wilson et al., 2007) is not relevant in this case study because all bamboo plantations were deliberately established and commercially productive before abandonment such that their residence time was effectively uniform across the country. The more relevant time period, therefore, is year of abandonment, which we have represented with distance to road as a proxy.
We translated our hypotheses concerning regional and local variables, and their interactions, into statistical models (Spake, Bellamy, et al., 2019). Consistent with the SGH (Figure 1), we hypothesised a three-way interaction between solar radiation, temperature and canopy cover. Aspect and slope can moderate the amount of solar radiation and hence the microclimate of a locality (Ashcroft & Gollan, 2013). Suzuki and Nakagoshi (2008) found differences in the importance of slope in limiting bamboo expansion across two regions varying in precipitation and temperature variables (Suzuki & Nakagoshi, 2008). We therefore hypothesised interactions between these topographical and climatic variables.

| Vegetation plot stratification
We required a subset of plots, surveyed in secondary forests, which were plausibly at risk of invasion by bamboo over the sampling period. This was to ensure that our absence records used in our models were 'real' absences and not simply due to plots being outside of the dispersal range of the species. First, we selected plots with complete data available for climate variables, canopy dominant species and canopy cover (n = 16,556), and excluded plots from the southern subtropical islands due a low sample size of plots occurring at outlying temperature values, and differences in management in these regions (n = 16,156; Figure S3.2).
Second, we selected plots that were plausibly at risk. The susceptibility of an ecosystem to invasion depends on propagule pressure, abiotic conditions and biotic interactions (Catford et al., 2009 Figure S3.3). Plot elevation ranged between 0 and 2,735 m. We selected all plots below an elevation of 750 m, to identify plots near to residential areas, that is, at elevations below which bamboo forests were extensively planted; 99% of plots containing bamboo were below 750 m. This yielded 10,287 plots ( Figure S3.3).
The environmental characteristics of a site must be hospitable for invasion to occur (Catford et al., 2009), so we further selected plots with abiotic characteristics within the ranges that were hospi-  (Kattge et al., 2020). Only deciduous plots were selected to limit covariation between temperature and phenology type in Japan (Kira, 1991;Suzuki et al., 2015), which in turn could be confounded with an effect of canopy cover on invasibility (n = 5,906, Figure S2.4). Next we selected plots dominated by tree species within which bamboo was present in at least 1% of plots dominated by that species, in order to select forest types, with corresponding microclimates and management regimes, at risk. Finally, we stratified plots based on the range of abiotic variables observed within plots where bamboo was present. We did this by selecting plots that occurred within the range of plots that contained bamboo for mean annual temperature, rainfall and solar radiation ( Figure S3.6).
In total, this multi-step plot selection procedure yielded 810 plots dominated by deciduous broadleaved tree species, of which 67 contained bamboo, distributed across Japan. Hereon we refer to these as secondary forests sensu lato, due to the history of deciduous broadleaved forest exploitation in Japan (Totman, 1989).

| Statistical analysis
All analyses were carried out using R version 3.5.3 (R Core Team, 2019). Data exploration was carried out following the protocol described in Zuur et al. (2010;see Appendix S4). Covariation among survey year and temperature suggested that vegetation survey campaigns were conducted progressively across space in successive sampling years. To avoid collinearity, we excluded survey year from the global models (we retained temperature, reflecting our hypotheses, Figure 1). Snow depth was also excluded due to moderate collinearity with temperate and solar radiation. Moreover, the more relevant temporal variable for bamboo invasion is year of abandonment, which we have represented with distance to road as a proxy (which did not covary with temperature). Distance measures (to forest edge and nearest road), and agricultural abandonment required log transformation to reduce the effect of outliers.
To quantify how the probability of secondary forest invasion by bamboo varies with regional, landscape and local-level drivers (Table S2) and their interactions (Figure 1), we fitted generalised linear models against a binomial distribution with a clog-log link function to the bamboo occupancy (presence/absence) data, due to unbalanced presences and absences (Zuur et al., 2009). Our modelling approach had two stages: exploration and multi-model inference (detailed in Appendix S5).
2.6.1 | Stage 1: Exploring the shapes of predictor relationships While we had formulated a priori hypotheses about the effects of environmental drivers (Figure 1), we did not know the exact shape of relationships, that is, whether to include a quadratic term for temperature, or whether annual temperature means, minima or maxima would provide superior fit, necessitating model comparison .
We created global models that contained the additive main effects of slope, northness and eastness, temperature variables, distance to road, distance to forest edge, in addition to a three-way interaction between temperature, solar radiation and canopy cover, and an interaction between solar radiation and slope as hypothesised (see Appendix S6 for details of the global model). A log transformation of plot size was included as an offset to adjust for variation in plot size.
Global models differed in the substitutions of temperature variables, whether annual means minima or maxima were included. They were not included in the same model due to high collinearity (e.g. mean and maximum annual temperature had ρ > 0.8, Appendix S5).
We also ran models with and without an additional quadratic term for temperature. From these global models, we generated full, 'all subsets' sets of nested models, with the dredge function from package MuMIn (Bartoń, 2018; see Appendix S5), to be compared with Akaike's Information Criterion (AIC). All variables were centred and scaled to unit S.D. prior to model fitting (Grueber et al., 2011).
This initial model exploration demonstrated that the hypothesised interactions, and a main effect of a quadratic term for mean annual temperature, were supported by the data, that is, were included in models with ΔAIC < 2. Prior to multi-model inference, we therefore restricted the model set to models containing the hypothesised interactions, and both a linear and quadratic term for mean annual temperature. This was done because the sign and magnitude of regression coefficients can change considerably across models depending on the presence or absence of interactions and polynomial terms, precluding multi-model inference (Cade, 2015).
Natural averaging was done instead of zero averaging, where parameter estimates (and errors) of zero are substituted into models where the given parameter is absent, because our aim was not to rank the importance of predictor variables (Nakagawa & Freckleton, 2010), but to model biologically meaningful responses.
We graphed the relationships between covariates and predicted values of occupancy using the predict.glm() function, using the aver- Model validation indicated all assumptions were met (Appendix S7).
Generalised variance inflation factors were calculated to check for collinearity. Because predictors were centred and scaled, we interpreted the relative magnitudes of averaged regression coefficients as a measure of relative, standardised effect sizes (Schielzeth, 2010).

| Application to forest management in Japan
To identify where biotic interactions (i.e. shade from forest canopy) should be considered when implementing forest management, we mapped the spatial heterogeneity of the effect of thinning (i.e. reductions in canopy cover) on bamboo invasion across Japan. To do this, we first mapped the predicted occupancies of bamboo in space using 1 × 1 km pixel-level values of the climate variables that were important in predicting occupancy. Occupancies were predicted for 100% and 60% canopy cover, corresponding to a typical reduction in stand volume that is incurred during secondary forest thinning in Japan (Spake, Yanou, et al., 2019). Occupancies were only predicted within the sampled range of predictor variables, for pixels within

| RE SULTS
We found that both regional climate, landscape and local variables and their interactions explained variation the probability of bamboo occupancy in secondary deciduous forests across Japan (details of multi-model inference and parameter estimation in Appendix S6).
Models within the top model set (ΔAIC < 6) explained 21%-24% of variation in bamboo occupancy (Table S6.1). The averaged model predicted occupancy to decline weakly with rainfall and distance to road (Figure 2). Occupancy declined with hill slope, with a more strongly negative relationship observed in regions with higher solar radiation ( Figure 2). Occupancy was best explained by annual means of monthly average temperatures ('average temperature' from hereon), with a main effect of a quadratic term.
A three-way interaction between average temperature, solar radiation and canopy cover explained variation in bamboo occupancy The interactive effects of canopy cover with solar radiation and temperature translated to high spatial heterogeneity in the effect of canopy cover-and hence the effect of thinning-on bamboo occupancy across Japan (Figure 4a-c).

| D ISCUSS I ON
Our study of P. bambusoides invasion in Japan demonstrates that the stress-gradient hypothesis (SGH) provides a valuable framework for applied invasion ecology: It can help explain the context dependence of invasion across locations, and can be used to guide management. Implicit to the SGH is the integration of ecophysiology, species interactions and biogeography, which-although inherent to the niche concept (Helmuth et al., 2005)-are rarely integrated (Godsoe et al., 2017;Wiens, 2011). Here we have documented a thermal niche shift of bamboo occupancy under high solar radiation (Figure 3b): When under dense canopies, the optimal niche of bamboo shifts to several degrees lower. At low temperatures, bamboo cannot establish in secondary forests with sparse canopies where it is prone to light stress. At higher temperatures, where light stress is ameliorated, dense canopies can be limiting to its spread (consistent with hypotheses in Figure 1).
This apparent contradiction explains why bamboo invasion success is highly variable among deciduous secondary forests in Japan (Isagi & Torii, 1997). Our study suggests that the SGH holds much potential for applied fields of ecology that aim to inform conservation and management plans.
The SGH has been subject to debate since its conception, motivating numerous refinements since its conception (e.g. Maestre et al., 2009). Observed exceptions from the SGH have been attributed, among other factors, to multiple, uncorrelated stressgradients that can 'cancel out' SGH effects (Maestre et al., 2009).
Controlled, factorial experiments have therefore been widely advocated to identify actual exceptions to the SGH rather than those due to failures to meet its underlying assumptions (He & Bertness, 2014;Maestre et al., 2009). So while the SGH makes predictions about species' niches-and is therefore inherently a question of biogeographic scale (Wiens, 2011)-most studies have been conducted across limited extents (Ziffer-Berger et al., 2014), with communities compared across subjectively defined 'high' and 'low' stress conditions (He et al., 2013). Here we show that applying the SGH to field data spanning broad gradients helps explain context dependence in bamboo invasion, a ubiquitous phenomenon in invasion ecology (Catford et al., 2019;Pyšek et al., 2020). This yielded two main benefits. First, the SGH's focus on the niche encourages integration of F I G U R E 2 The probability of secondary forest invasion by bamboo (occupancy) consistently decreased with rainfall (a) and distance to road (b). Invasion declined with slope, this effect was stronger in regions with higher radiation (135 MJ/m 2 ) than lower radiation ( physiological knowledge into analyses, so that models are appropriately parameterised with necessary interactions. A priori physiological insight and hypotheses concerning the SGH led us to examine and reveal a three-way interaction among solar radiation, temperature and canopy cover. This allowed us to detect a thermal niche shift for bamboo under dense canopies (Figure 3). Such insights are difficult to obtain from typical correlative species distribution modelling, as these require little knowledge of the mechanistic links between an organism's physiology and their abiotic environment (Helmuth et al., 2005). Indeed, both physiological insight and biotic interactions are typically omitted from models that approximate species occupancy or abundance (Kearney & Porter, 2009;Staniczenko et al., 2017).
Our findings agree with experimental evidence and make sense mechanistically, with the degree of shade cast by dense forest canopies facilitating invasion at low temperatures, where thermal dissipation of excess excitation energy is not possible. Nevertheless, before implementing our findings into management, we suggest that manipulative mesocosm experiments in the field, designed to emulate this three-way interaction in forests, would be a useful intermediary to ground truth that shading cast by forest canopies exerts facilitative and competitive effects under different abiotic conditions, and to confirm whether forest trees are competing with the bamboo for resources unmeasured in this study. This study could be carried out on existing bamboo forests, as annual expansion rates can be reliably estimated by rapid, one-time measurements of bamboo culms spreading from source bamboo forests .
Second, the application of the SGH permits the characterisation of geographic variability in the magnitude, and even direction, of a widespread management intervention, forest thinning ( Figure 4).
While biotic interactions can challenge forecasts of community responses to climate change (Hillerislambers et al., 2013), these interactions also present management opportunities: the SGH can help characterise spatial variability in the effect of management.
Across Japan, forest thinning is widely used to enhance the biodiversity and recreational value of forests (Spake, Yanou, et al., 2019).
Incorporating the SGH into our analytical design revealed potentially adverse effects of forest thinning in cooler areas. Given that F I G U R E 4 Spatial variation in Phyllostachys bambusoides occupancy, predicted by climate variables solar radiation, mean annual temperature and rainfall, for secondary forests under (a) dense canopies (100% cover) and (b) more open canopies (60%). (c) Spatial variation in the effect of reducing canopy cover from 100% to 60% on predicted bamboo occupancy probabilities [the risk difference, that is, occupancy difference between (a) and (b)]. Positive values (red) correspond to increases in the probability of bamboo occupancy following a reduction in canopy cover, while negative values (blue) corresponded to a decrease in occupancy probability. Colour shading represents the strength of the effect. Inset are marginal changes in occupancy with canopy cover from two locations differing in solar radiation, rainfall, and temperature. Shown occupancy probabilities are averaged predictions from each model across the top model set (ΔAIC < 6), using values of regional climate variables and other covariates held at their median Canopy cover (%) Canopy cover (%) Probability of occupancy Probability of occupancy Risk difference giant bamboo is projected to expand its range northwards in Japan with rising temperatures (Takano et al., 2017), thinning might reduce spread at these northern limits.
We expected distance to road to relate positively with invasion probability, due to its representation of the likelihood that a bamboo and/or secondary forest has been abandoned (Kobayashi & Nakamura, 2018), and because distance to road has previously correlated positively with expansion rates in Japan (Isagi & Torii, 1997;Suzuki & Nakagoshi, 2008). On the contrary, we found a negative relationship with distance to road. This is likely due to differences in study design: Here we investigated bamboo occupancy in secondary forest plots, while previous studies have measured annual expansion rates of existing bamboo forests into neighbouring vegetation (Appendix S1). Distance to road likely correlates with the occupancy of source bamboo forests, because bamboo forests were typically planted near roads and residential areas (Fukamachi et al., 2001). We did not detect an effect of our measure of agricultural abandonment, possibly due to this variable's coarse resolution (5 km) and consequent poor representation of bamboo and secondary forest abandonment, which varies at a finer scale than 5 km.
We found that hill slope increasingly related negatively to forest invasion with increasing solar radiation ( Figure 2). Slope can modify the amount solar radiation received by a surface (Ashcroft & Gollan, 2013), but can also covary with soil types (Kamoshita, 1955). Suzuki and Nakagoshi (2008) compared drivers of bamboo forest expansion rates in two regions of Japan, and concluded that slope exerted a more strongly negative effect in the region with drier, less developed soils. We were unable to include soil variables in our analyses, as available soil type information was too inconsistent across plots to serve as a useful covariate. Soil type and moisture are likely to directly drive bamboo occupancy and abundance (Takano et al., 2017), and also has the potential to mediate the effects of temperature extremes on plant communities (Fridley et al., 2010). We recommend that future vegetation surveys in Japan adopt a systematic soil classification scheme, to permit the use of this covariate.
Tests of the SGH have been applied much less to diffuse, rather than species-specific interactions, despite their wider prevalence (Brooker et al., 2008), and rarely in forest ecosystems, which are not classically considered as extreme environments (He et al., 2013).
However, our findings are consistent with tests of the SGH for species-specific interactions conducted across smaller extents. For example, Cavieres et al. (2005) found that seedling survival and netphotosynthetic rates of invasive Taraxacum officinale, were higher when growing within native cushion plants, Azorella monantha, than in open areas in the high-Andes (Cavieres et al., 2005). This suggests that the SGH may be broadly applicable across ecosystems and communities, and in invasion ecology. Given that rising temperatures are expected to increase the spread of invasive species, and have stronger positive effects on the performance of invasive than native plant species , future applications of the SGH present a promising way forward for predicting future species invasions-and how to mitigate them.

AUTH O R S ' CO NTR I B UTI O N S
R.S. and M.S. designed the initial analysis and collated the data; R.S., M.S., J.A.C. and F.E. further developed the conceptual ideas; R.S.
analysed the data and wrote a first draft. All authors contributed substantially to revising the manuscript.