Assembly rules operate only in equilibrium communities: Is it true?

Spatial community assembly

We first examine spatial heterogeneity, that is, the mean point‐to‐point variation in species composition within the community. Prediction from theory here is complex (Gitay & Wilson 1995). Early in succession, heterogeneity might be high due to chance colonization. This effect would decrease through time, but heterogeneity might rise eventually due to sorting of the species into communities (Childress et al. 1998). If there is a positive‐feedback switch, in which the species in each patch modify their environment to their own advantage, giving alternative stable states, the later increase in heterogeneity in species composition and micro‐environment would be further reinforced (Wilson & Agnew 1992; Mason et al. 2007). Chaos theory predicts a similar result, suggesting that initial heterogeneity will sometimes be magnified with time (Stone & Ezrati 1996). Thus, theory indicates higher heterogeneity just after severe disturbance, and then again in undisturbed communities.

An additional aspect of heterogeneity is the spatial trend, that is, spatial autocorrelation. At the community level this is the ability to predict the dissimilarity of two patches of vegetation from their distance apart (Mistral et al. 2000). The strength of this relation is a measure of the community's spatial structure, and logic suggests that so long as the focus (e.g. quadrat size) is smaller than the grain of micro‐environmental variation, predictability should increase with time from disturbance, as species composition responds to variation in micro‐environmental conditions. However, there is only limited and contradictory evidence on this (e.g. Seabloom et al. 2005; Erfanzadeh et al. 2010).

A spatial autocorrelation fit can also be extrapolated backwards to zero distance (the nugget), at which all environmental dissimilarity and dispersal limitation have been removed, giving an estimate of the point randomness of the community (Fortin & Dale 2005; Lawrence Lodge et al. 2007). We argued above that this should be higher in more disturbed communities.

Thus, the spatial community assembly hypotheses are that:

• 1

  Spatial heterogeneity: will be high in the most disturbed community and in undisturbed communities, but lower in those with intermediate disturbance.

• 2

  Spatial autocorrelation strength: will be greater in the least disturbed communities.

• 3

  Spatial autocorrelation nugget: will be smaller in the least disturbed communities.

Abundance‐pattern assembly rules

Community re‐assembly after a disturbance involves adjustment of the abundances of the species, and the most basic aspect of this is evenness (Bartha et al. 2001). There is little theory to guide us on the response here, but the most rigorous tests to date have shown evenness to increase through succession, at least in herbaceous/shrubby communities (Wilson et al. 1996).

To analyse abundances more closely, we examine relative abundance distributions (RAD), the ‘dominance/diversity’ relations of early workers. These have also been related to succession, with the suggestion from mathematical modelling that a General Lognormal distribution will be seen in equilibrium, late‐successional communities but not in disturbed ones (Stenseth 1979; Ugland & Gray 1982). However, this remains largely unexplored, especially for plant communities.

Ecologists have distinguished between ecological/founder control, whereby composition depends on the chance order of species arrival, versus evolutionary/competition control, whereby it depends on the (evolved) competitive abilities of the species. This will be reflected in whether the ranks of an RAD are occupied by the same species in different patches/quadrats, for example, whether the most abundant species is always has the same identity (Watkins & Wilson 1994). Just after a disturbance, the ranks should depend largely on chance dispersal (i.e. ecological/founder control), giving low consistency between patches/quadrats. As the community assembles, rank consistency should increase as the relative competitive abilities of species determine their presence and abundance (evolutionary/competition control).

Thus, the hypotheses for abundance‐pattern assembly rules are that:

• 4

  Evenness: will be highest in the undisturbed community.

• 5

  Relative abundance distribution: the undisturbed communities will show a lognormal RAD; communities with less disturbance and especially the most disturbed will show geometric, Zipf‐Mandelbrot or possibly broken stick RADs.

• 6

  Rank consistency: will be higher the less disturbed the community, that is, there should be more consistency in the identity of species in the ranks of the RAD.

Niche‐based assembly rules

The concept of the niche (Hutchinson 1957) implies that, in the longer term, limiting similarity (MacArthur & Levins 1967) will restrict the number of species to the number of distinct niches: ‘niche limitation’. The identity of the species may be different in different patches of a community because of chance colonization and subtle micro‐environmental variation, but there should be a trend towards less variation in species richness, that is, greater constancy in species richness, than if the species occurred independently of each other (Wilson et al. 1987). The appropriate null model for such independent occurrence of species fixes the number of occurrences of each species at that observed in the real data, but assigns those occurrences at random to sub‐quadrats, independent of the assignment of other species (Wilson et al. 1987).

Likewise, it should be easier for a species to invade or to avoid exclusion if it is in a different guild from the majority of the species already present so long as the guilds represent the limitations to co‐existence. This would give a more constant representation of different guilds than under a null model which randomizes occurrences while retaining species frequencies and quadrat richness: the assembly rule of guild proportionality (Wilson 1989). We here use a priori guilds, based on morphology/taxonomy.

However, after disturbance there might be dispersal limitation, empty niches (Hutchinson 1959) and incomplete competitive exclusion, so these rules would not apply. Thus, the niche‐based assembly rule hypotheses are that:

• 7

  Constancy in species richness: in less‐disturbed communities, species richness will show lower variance between sub‐quadrats than expected at random.

• 8

  Guild proportionality: in less‐disturbed communities, the proportions of species from each guild will show lower variance between sub‐quadrats than expected at random.

Purpose

In spite of all this theory, there is actually little evidence that there is more spatial disorder or weaker assembly rules after disturbance. Wilson et al. (2000) did find that in two semi‐arid grassland sites, the one which had been more disturbed gave no evidence that species constrained each other's biomass, whereas the less‐disturbed site did. Sanders et al. (2003) found that disturbance of an ant community in the form of invasion by an exotic species reduced the chequerboarding pattern, which they saw as breakdown in assembly rules. However, there seem to have been no direct comparisons of assembly rules under various degrees of disturbance, or of the differences in spatial organization that would be expected to follow.

We examined the relation between disturbance, spatial community assembly and assembly rules using ski‐related disturbance on two southern New Zealand mountains (a minimal form of replication). Disturbance of ecological communities generally comprises either repeated removal of vegetation (usually mowing or grazing) or mechanical disturbance exposing bare soil (e.g. by burrowing animals). Ski‐slope making combines the two. Mechanical soil disturbance is one of the major factors involved in human perturbation of plant communities, so determining its effect on assembly rules during ski‐run formation is a significant case study of human impacts. To summarize, we examined the eight hypotheses listed above, under the following categories:

Spatial community assembly:

• 1

   spatial heterogeneity,

• 2

   spatial autocorrelation strength,

• 3

   spatial autocorrelation nugget,

Abundance‐pattern assembly rules:

• 4

   evenness,

• 5

   relative abundance distribution,

• 6

   rank consistency,

Niche‐based assembly rules:

• 7

   constancy in species richness, and

• 8

   guild proportionality.

Sites

Two areas with extensive ski activity were chosen on nearby mountain ranges in southern New Zealand (Fig. 1). The Cardrona area (45°52′S, 168°56′E) was at 1560–1640 m a.s.l., with 2.6 m per year of snowfall, and Coronet Peak (45°56′S 168°43′E, referred to as ‘Coronet’ below) at 1180–1200 m a.s.l., with extensive snowmaking. Within each area, three sites were chosen:

Sampling

At each of the three sites in each area, six 1 × 1 m quadrats were placed by restricted randomization. Each quadrat was divided into a grid of 100 sub‐quadrats, each 10 × 10 cm, in which the shoot presence/absence of all vascular plant species was recorded. In the first row of 10 sub‐quadrats in each quadrat, all above‐ground plant material in the volume defined by each sub‐quadrat was collected, sorted by species, and the biomass determined.

Analyses: spatial community assembly

To measure spatial heterogeneity in the community at a site, the mean dissimilarity was calculated between all possible pairs of 10 × 10 cm sub‐quadrats within each quadrat. Within each of the six quadrats, 4950 pairs were available for presence/absence, 45 for biomass. The dissimilarity measure for the grid of presence/absence records was Jaccard (1.0 – Jaccard similarity: Jongman et al. 1987), and for the line of biomass records proportional dissimilarity (PD: Jongman et al. 1987). Jaccard and PD both give 0.0 for identical species composition and 1.0 for maximum dissimilarity, and neither is affected by species absent from both samples being compared. These values were used in a two‐way disturbance × area analysis of variance, with the six quadrats within each site used as replicates.

For spatial autocorrelation within a quadrat, the same dissimilarities were compared with the distance apart of the two sub‐quadrats by fitting the asymptotic relation (cf. Lawrence Lodge et al. 2007):

The dissimilarities and distances within each of the six quadrats were accumulated for each site in each area. The goodness of fit (% of the variance dissimilarity explained by distance using this formula) for each quadrat was arcsine transformed and used in analysis of variance (an alternative logit transformation gave the same pattern of significances). Transformation removed minor heteroscedasticity, but the means and analysis results were very similar to those on untransformed data and untransformed means are presented. The intercept is termed the ‘nugget’ in spatial autocorrelation, has been seen as an estimate of community randomness (Fortin & Dale 2005). The significance of the relation was examined by a randomization test in which the species composition of a sub‐quadrat was kept as a unit, but randomized 2000 times with respect to the position of the sub‐quadrat within its quadrat (Lawrence Lodge et al. 2007).

Analyses: abundance‐pattern assembly rules

Evenness in species biomass was measured in each of the 10 biomass sub‐quadrats of a quadrat using the Camargo index E′, selected for its ability to meet the 15 requirements and features evaluated by Smith and Wilson (1996). The means over the sub‐quadrats were used in analysis of variance.

To characterize the RAD, four models of RAD (Broken Stick, Geometric, General Lognormal and Zipf‐Mandelbrot) were fitted to each of the 10 × 10 cm biomass sub‐quadrats by the method of Wilson (1991). In order to allow for any tendency for all models to fit some observed distributions poorly, that is, to compare models on a standard basis, the sum of squares of deviations from the best‐fitting model was subtracted from all four models, separately for each sub‐quadrat (Wilson et al. 1996), and quadrat means calculated over these sub‐quadrat values.

Rank consistency across the ten biomass sub‐quadrats of a quadrat was determined with index C_r (Watkins & Wilson 1994), which ranges from +1 when the ranks of the species (1 if the species is the most abundant, 2 if it is the second‐most abundant, etc.) are identical between replicate sub‐quadrats and −1 when they are as different as they could be. The null model used randomly permutes the non‐zero species abundances within a sub‐quadrat. The observed value of C_r was compared with random expectation using an IV index:

Analyses: niche‐based assembly rules

For these analyses, constancy in species richness and guild proportionality, the test statistic (see below) from observed data was compared with those from 5000 randomizations under the appropriate null model (see below), again using the IV index (see above). Observed, expected and IV values were calculated separately for each sub‐quadrat. The significance (P) of IV was calculated as the proportion of randomizations giving an IV value equal to, or more extreme than, that observed, adjusted to a two‐tailed test (previous work has shown departures from the null model in either direction, and this is expected ecologically).

For constancy in species richness, the test statistic was the variance in richness across the 100 10 × 10 cm sub‐quadrats of a quadrat. The null model comprised the number of occurrences of each species in the quadrat being fixed at that observed, but allocated to sub‐quadrats at random, separately for each species (Wilson et al. 1987). An overall null model was used in which species occurrences were randomized over all 100 sub‐quadrats in a quadrat, and also a patch model in which the null model for each target sub‐quadrat was based on a 3 × 3 sub‐grid of sub‐quadrats around the target (Watkins & Wilson 1992).

For guild proportionality, the guilds used were forbs versus all other species (there was no separation among the latter because of the small number of species involved). The test statistic was the variance in the proportion of species in the sub‐quadrat that were forbs, that is, the variance in (number of forb species)/(total number of species) (Wilson & Watkins 1994). The guild proportions of the non‐forb guild were not analysed because they are the complement of the forb guild. The null models, overall and patch, comprised holding the species frequencies in each quadrat (with a site model) or patch (with a patch model: Watkins & Wilson 1992) and the richness of each sub‐quadrat at those observed, but with occurrences randomized within these constraints (Wilson 1989; using a swapping method of the type validated by Miklós & Podani 2004).

Spatial community assembly

Spatial heterogeneity between sub‐quadrats using species presence/absence (Jaccard index; Fig. 2a) was significantly higher at Coronet than at Cardrona (F = 22.2, P < 0.001), but more notably the responses to disturbance were quite different between the two areas (area × disturbance interaction F = 21.7, P < 0.001). At Cardrona, the lowest presence/absence heterogeneity was seen in the Ski‐run, but at Coronet heterogeneity was highest there (Fig. 2a).

When heterogeneity was weighted by species biomasses (abundance, index PD; Fig. 2b) the two areas differed in the opposite way from presence/absence, within‐quadrat heterogeneity being higher at Cardrona (F = 16.1, P < 0.001). Response to disturbance again differed between the two areas, and in ways different from that in presence/absence: at Cardronathe lowest heterogeneity was in the Adjacent, but at Coronet it was highest there (interaction F = 3.54, P = 0.043).

In summary, the response of heterogeneity to disturbance was inconsistent between the two areas, and was also dependent on whether heterogeneity was weighted by species biomasses.

Spatial autocorrelation within quadrats was significant in the Adjacent and Undisturbed sites in both areas (P < 0.01, by randomization test), but not in the Ski‐runs except at Coronet in terms of biomass (index PD, P = 0.028). The strength of spatial autocorrelation is indicated by examining the variation in dissimilarities between quadrats in their species composition, and calculating the percentage of it that is explained by distance. This percentage differed significantly between disturbance regimes, for presence/absence (Jaccard; Fig. 3a, F = 5.16, P = 0.012) and for abundance (PD; Fig. 3b, F = 3.83, P = 0.033). This difference between regimes was consistent between the two areas sampled (area comparison P with Jaccard 0.84, with PD 0.52; area × disturbance interaction P with Jaccard 0.20, with PD 0.66). In particular, spatial autocorrelation was stronger (by Duncan's test) in the Undisturbed than in the Ski‐run, with the Adjacent generally being intermediate (Fig. 3). Thus, spatial autocorrelation in the areas sampled was more pronounced in the less‐disturbed sites, as hypothesized.

Abundance‐pattern assembly rules

Evenness at 10 × 10 cm was significantly higher in the Ski‐run sites than in the Adjacent or Undisturbed ones (Fig. 4a, P < 0.022 for the disturbance effect). The two areas showed similar, though not identical, effects (the area × disturbance interaction was non‐significant, P = 0.076), and this was not scale‐dependent, for very similar patterns were seen lumping adjacent quadrats to 20 × 10 cm and 100 × 10 cm scales (results not shown).

Of the RAD models, the Zipf‐Mandelbrot gave the best fit (i.e. it had the lowest sum of squares deviance) in every case except Undisturbed at Cardrona, for which it was second‐best to General Lognormal (Table 1). The Zipf‐Mandelbrot fit was especially close for the two Ski‐run sites. Broken Stick did not give the best fit for any individual quadrat, or for any site (the Broken Stick model implies very high evenness). For most sites, the second‐best fit was Geometric or General Lognormal; in fact, overall the fits of these two models were very close (Geometric average = 0.571, General Lognormal = 0.584), and there was no tendency for a particular model (Broken Stick, Geometric, General Lognormal and Zipf‐Mandelbrot) to fit better in one of the three disturbance regimes.

As expected, some rank consistency was always present, as indicated by IV values being positive (there should be some consistency among the 10 sub‐quadrats of a quadrat in which species have high biomass and which have low biomass). It was significantly greater than the null model in 34 out of the 36 quadrats (3 disturbance regimes × 2 areas × 6 replicates). There was no overall difference between disturbance states in this consistency in rank (P = 0.66), but the response to disturbance at the two areas was quite different: at Cardrona the Ski‐run had the lowest rank consistency and at Coronet it had the highest (Fig. 4b, area × disturbance interaction P = 0.019).

Niche‐based assembly rules

Constancy in species richness was not generally apparent, with only sporadic departure of variance in richness from that expected under the null model and with some departures in each direction (Table 2). In the Coronet Adjacent sites using an overall model, richness was significantly more constant than expected (i.e. IV < 0), as would be expected from niche limitation. However, once environmental heterogeneity and spatial autocorrelation effects were excluded by using a patch model, the effect, though still present (IV = −2.5), was no longer significant. The only effect remaining significant under a patch model was high variance in richness in the Cardrona Adjacent sites, suggesting environmental variation rather than an assembly rule (IV = +6.1).

Guild proportionality was strong and highly significant (i.e. lower variation in the proportional representation of the forb guild than expected under a null model, IV < 0.0) in the Ski‐run sites in both areas (Table 3). However, these effects decreased and became non‐significant using a patch model.

Trends in heterogeneity

Greig‐Smith (1952) and Gitay and Wilson (1995) suggested that there would be three phases in succession after a disturbance: (i) Pioneer, with moderate heterogeneity resulting from random initial colonization; (ii) Building, with reduced heterogeneity as species spread; and (iii) Mature, with strong heterogeneity arising from micro‐habitat sorting and the operation of assembly rules. Gitay and Wilson (1995) found support for this model in the recovery of New Zealand tussock grasslands from fire. The observation of Glenn et al. (1992) also matches the model: in prairie sites that had not been experimentally disturbed by grazing and/or burning (i.e. the Mature phase) there was higher heterogeneity. Abundance‐based heterogeneity at Cardrona (Fig. 2b) exactly matched the theory, higher in the Ski‐run (Pioneer phase) than in the Adjacent vegetation (Building phase), then higher again in the Undisturbed (Mature phase). However, the opposite pattern was seen at Coronet. Apart from the higher heterogeneity in the Ski‐run at Coronet, presence/absence heterogeneity did not fit the theory either (Fig. 2a). The differences between the two sites cannot be explained in terms of known disturbance history, as it is clear that both were originally denuded to bare soil and then colonized from sown and volunteer seed. Some reports in the literature conflict with the theory too. Seabloom et al. (2005) studied experimental grasslands in California over the 4 years after sowing or burning. This timespan surely represents the span of the Pioneer to the Building phases, when heterogeneity should decrease by the theory, but instead it increased. Ruprecht et al. (2007) found in a ‘chronosequence’ of old fields in Romania that there were no trends in the proportion of statistically significant species associations, contrary to the theory. Lepori and Malmqvist (2009), investigating freshwater invertebrates, found the highest heterogeneity under the least disturbed conditions and the lowest under intermediate disturbance, in conformance with theory. It seems that changes in community heterogeneity after disturbance are not simple: the pattern varies between areas, as it did here, and different measures of species composition need to be examined to check that they are consistent. Much of the evidence is from space‐for‐time substitution studies, and more studies of temporal changes within individual sites are needed.

Biomass heterogeneity (Fig. 2b) comprises two aspects: which species are present (Fig. 2a) and whether they are present in consistent amounts (Fig. 4b). The latter aspect is rank consistency, or rather the inverse of it because the greater the rank consistency the lower the abundance heterogeneity. The results for rank consistency (Fig. 4b) do indeed closely mirror those for abundance heterogeneity (Fig. 2b). The unexpected Coronet result, of high rank consistency in the ski‐run, best matches the observation of Collins et al. (2008), of higher rank consistency with higher burning disturbance in prairie. For the ski‐run, this may have been because the species present were predominantly exotic, with efficient dispersal, but it is not clear whether this applies to the burnt prairie.

Spatial differences are not just a question of the amount of heterogeneity, but also of the consistency of spatial trends in that heterogeneity, that is, whether the dissimilarity in community composition increases steadily with distance apart. This is spatial autocorrelation. The study of spatial autocorrelation at the community level is relatively new (Kleb & Wilson 1999; Nekola & White 1999; Mistral et al. 2000) and little is known of its change through succession. We would expect the strength of spatial autocorrelation to increase through succession as the community becomes organized, and it generally did (Fig. 3). In the only comparable study, Lu et al. (2009) reported for four Chinese wetlands stronger spatial autocorrelation in those with lower disturbance, indicated by a lower agricultural ditch density in the surrounding lands. This parallels our result, and may be a general rule in succession, but observations over time are needed.

In summary, the prediction of Fowler (1990) that disturbance would increase the role of ‘disorderliness’/random‐events receives only limited support in terms of pure heterogeneity, but it is strongly supported by spatial autocorrelation: the strength of the fitted dissimilarity/distance relation increased as the amount of disturbance decreased. Although the amount of heterogeneity cannot be predicted, its spatial distribution can.

Trends in abundance patterns

The evenness of a community is a very obvious result of species interactions and assembly: the simplest abundance‐based assembly rule. It represents species dominance, that is, heterogeneity among the species. There is no theory on its response to disturbance, but many empirical studies. Mackey and Currie (2001) found that in 50% of published studies there was no significant trend, but of the remainder the majority indicated that disturbance reduces evenness. However, almost all such studies, those reviewed by Mackey and Currie and those published since, are flawed in that: (i) the measure of plant abundance has usually been subjective, or even not stated; or (ii) an evenness index such as J′ has been used that reflects species richness as well as evenness (Smith & Wilson 1996) and usually species richness did change with disturbance in the studies Mackey and Currie reviewed. In a study of two successional sequences using biomass and the richness‐independent index E′, Wilson et al. (1996) showed evenness to increase through succession, but the present work with the same methods found the opposite pattern: evenness was lower in the less‐disturbed communities (Fig. 4a). No generalization is yet possible. Evenness will depend on the relative competitive ability of the species and the opportunity for niche complementarity, so successional trends in evenness as in the Wilson et al. (1996) study, or differences between sites with difference disturbance histories as here, may be individualistic to each site with its unique assemblage of species.

A more subtle abundance‐based assembly rule is the RAD. Community models have predicted that a lognormal RAD will be found at equilibrium but not in disturbed communities (Stenseth 1979; Ugland & Gray 1982). Similarly, in the theoretical study of Wissel and Maier (1992) only those models that included competition, such as would predominate later in succession, gave lognormal distributions. Some animal ecologists have claimed to find confirmatory evidence that a lognormal distribution is typical of, and restricted to, equilibrium communities (e.g. Hamer et al. 1997), but others have failed to see a general pattern (Nummelin 1998). Evidence for plant communities is very sparse. Wilson et al. (1996) could find no consistent pattern of change in RAD model fit through succession. In the present results, the General Lognormal did give the best fit to the Undisturbed vegetation at Cardrona, conforming to the theory, but for the Undisturbed at Coronet it was inferior to both the Zipf‐Mandelbrot and the Geometric. Insight into the cause of abundance distributions can also be found via rank consistency since the frequent comments that disturbance will cause founder events to predominate and prevent communities from self‐organizing (Fowler 1990; Drake et al. 1999) imply that, in the absence of disturbance, evolutionary/competition control will be present, giving higher rank consistency (Watkins & Wilson 1992). This prediction was borne out at Cardrona, but the opposite was true at Coronet (Fig. 4b). Again, any generalizations would be premature.

The conclusion on abundance‐based assembly rules must be that there are no trends consistent between sites, or that conform to theoretical predictions.

Niche‐based assembly rules

Assembly rules such as constancy in species richness and guild proportionality suggest that different patches of a community will converge in structure, even if they differ in species composition. The rules will be only trends because of the noise of incomplete competitive exclusion, overlap between niches, empty niches, etc., but they may still be detectable when compared to a null model in which no niche constraint is present (Wilson et al. 1987; Wilson 1989). We started with the widely held concept that this would be true only for equilibrium communities (Bartha et al. 1995; Belyea & Lancaster 1999; Cornell 1999; Drake et al. 1999). Part of the reasoning has been that after a disturbance competitive intensity is reduced (Bazzaz 1987; Wilson & Tilman 1993). Violle et al. (2010) have questioned this, but the logic remains that time is required for species to disperse, exclusion to occur, and abundances to adjust.

Only a little evidence has been produced to support this. Wilson et al. (1987) and Wilson and Gitay (1995) demonstrated that the least disturbed of four dune slacks was the one showing evidence of niche limitation and guild proportionality. In the present work the only significant indication of constant richness (negative IV) was in the Adjacent sites at Coronet (Table 2), while guild proportionality was evident in both Ski‐run sites (Table 3): the opposite position in the disturbed–undisturbed spectrum from that expected by the logic of Belyea and Lancaster (1999) and others. One possible explanation for the guild proportionality results is that mutualisms are important early in the ski‐runs, which are in an early successional stage after disturbance, as after disturbance in other alpine areas (Choler et al. 2001), giving the relative constancy in guild proportions seen in the ski‐runs (Tables 2,3). In all three cases the effects became small and non‐significant using a patch model, so the cause may be environmental micro‐heterogeneity (Wilson 1999). Most other studies of constancy in richness and guild proportionality have not used a patch model, and their results may likewise reflect habitat sorting. However, we must bear in mind that our effects might be real, but not demonstrable with a patch model, as basing the null for a target quadrat on only the nine surrounding quadrats gives a conservative test.

Conclusions

The hypothesis with which we started was stated rather explicitly by Fowler (1990), that disturbance produces disorderliness in plant communities: it weakens the effectiveness of both the processes that promote species coexistence and also those that cause competitive exclusion, so that random events are more important. This affects the frequency with which the outcome of competition (winner vs. loser) can be identified. Our results come close to disproving this as a general phenomenon. Higher evenness was seen in disturbed site at Coronet, but this is contrary to trends observed elsewhere (Wilson et al. 1996), and there was no other abundance‐based assembly rule. The only niche‐based assembly rules might have been caused by environmental differences between 10 × 10 cm patches since they disappear with a patch model. Spatially, the lack of consistency between whether heterogeneity was higher or lower with disturbance reflects mixed reports in the literature, and indeed mixed prediction from theory. The one consistent pattern is stronger spatial autocorrelation in undisturbed sites, supporting the concept of the Mature phase in the Greig‐Smith/Gitay and Wilson model of community assembly: decreased randomness but increased patchiness, reflecting not only micro‐habitat sorting but also interspecific interactions of the type that would cause assembly rules.

In the Cardrona area, more consistent support can be seen for the arguments of Bartha et al. (1995), Belyea and Lancaster (1999) and Drake et al. (1999). There is low initial (Pioneer phase) heterogeneity in species presence since any species can arrive, but high biomass patchiness due to founder dominance (Fig. 2). Then, after a homogeneous intermediate Building phase, the final Mature phase has high heterogeneity in both presence and biomass due to habitat and competitive sorting (Fig. 2). This is accompanied by steadily increasing spatial autocorrelation, as expected (Fig. 3), although unexpectedly the nuggets were lower in the ski‐runs (Fig. 5). There is little theory to guide us on the development of evenness, but the evidence so far suggests that it will increase through succession, and at least the opposite trend is not seen at Cardrona (Fig. 4a). Several models suggest that equilibrium communities should have a general lognormal RAD, and this is seen at Cardrona (Table 1). The expected shift from founder control to dominance control, giving an increase in rank consistency from disturbed to undisturbed, is seen there (Fig. 4b). The analyses of constancy in richness (Table 2) and guild proportionality (Table 3) are dominated by the effects of patchiness. Neither show unequivocally the assembly rules predicted, but this is unsurprising because such rules are difficult to demonstrate (Wilson et al. 1987; Wilson 1989, 1999). The evidence for Cardrona points to the processes developing that lead to assembly rules, even if the end results are obscured.

However, the absence of many of these patterns at Coronet reminds us that nature does not always neatly follow our theories. More empirical studies are needed in order to establish general trends.