Statistical analysis
Predictor variables were tested for collinearity. Distance to coast, which was highly correlated to elevation (r = 0.75), was excluded from further analyses (Appendix Table A1).
The native species richness and alien species richness per plot at grain sizes of 9 m2 and of 1 m2 were response variables in statistical models. We included ecologically relevant interaction terms in our analyses. Underlying geology can influence soil moisture (Huang et al., 2016, Kopec, 1995), and the effect of elevation on species (Gerdol et al., 2017). Furthermore, we expected that northness and elevation interact, since north-facing slopes receive more sunlight in the southern hemisphere (Saremi et al., 2014), which might particularly be beneficial to high elevation sites which are generally cooler. Hillshade could affect soil moisture, with wetter soils occurring in areas receiving less solar radiation (Najafifar et al., 2019). In addition, elevation also affects the biotic interaction effect of A. selago on co-existing species, as the cushion plant facilitates plants at high elevations but competes with them at low elevations (le Roux and McGeoch, 2010). Therefore, the interaction terms geology*TWI, geology*elevation, northness*elevation, hillshade*TWI, and presence/absence of A. selago *elevation were included.
We also investigated the factors correlated with the differencein species richness between the large (9 m2) and small (1 m2) spatial grains (henceforth ‘Δ9-1’) to assess what may be shaping differences in patterns and drivers of richness at different grain sizes. To calculate the difference in species richness between grain sizes, the species richness of the nested central 1 m2 subplot was subtracted from that of the larger 9 m2 plot. We then tested which of the above predictor variables and the interaction terms were related to Δ9-1. By identifying the drivers of the difference in richness between large and small grain, we could determine whether local turnover occurs and what factors could be contributing to turnover.
We tested for spatial autocorrelation in the response variables using Moran’s I . Moran’s I values were small, but significant, suggesting some spatial autocorrelation (Table A2). Therefore, to account for spatial autocorrelation, we opted for simultaneous auto-regressive models (SAR) (Kissling and Carl, 2008), particularly the SARerr model . We tested the effects of the explanatory variables on species richness at each grain size, and on Δ9-1, using SAR models. These statistical analyses were run assuming a Poisson distribution. Best subset models with the best set of predictors were created using the dredge function from the MuMIn package, with model selection based on the lowest Akaike information criteria (AIC) values (Barton and Barton, 2015).
Alien species, especially where they are abundant and/or have high cover, may mask the effects of environmental drivers on native species richness. This prompted us to (a) initially repeat the above analyses only using a subset of the plots where alien species were absent. However, on Marion Island, alien species tend to be more prevalent in coastal areas with biotic inputs, than in inland areas or areas far from any bird or seal colonies (Haussmann et al., 2013, Figure 1). Most coastal plots had some alien species present (Figure 1); therefore, this analysis led to us excluding many of the coastal plots with the highest native species richness. We thus present these results in the supplementary materials but do not further discuss them. Instead, we repeated these analyses using a subset of the data where (b) alien cover was less than 10% in the plots, assuming that such low alien cover is unlikely to significantly affect patterns of richness of native plants.
For alien species, analyses were conducted on the full dataset, and also for a subset of plots where (c) at least one alien species was present because we assumed that some alien species may not have reached niche equilibrium with the environment (i.e., range occupancy of all suitable sites) on Marion Island (le Roux et al., 2013b) and that if there is at least one alien species in a plot, it is likely that more alien species may have had time to colonise the plot.
All statistical analyses were run in R, version 4.2.0, using functions from the spdep (Bivand et al., 2006), car (Fox et al., 2012), mass (Ripley et al., 2013), nlme (Pinheiro et al., 2017), ape (Paradis and Schliep, 2019), MuMIn (Barton and Barton, 2015) and ggplot2 (Wickham et al., 2021) packages.