Disentangling niche and neutral influences on community assembly: assessing the performance of community phylogenetic structure tests

Transcription

1 Ecology Letters, (2009) 12: doi: /j x LETTER Disentangling niche and neutral influences on community assembly: assessing the performance of community phylogenetic structure tests Steven W. Kembel*, Department of Integrative Biology, University of California, Berkeley, CA 94720, USA *Correspondence: Center for Ecology and Evolutionary Biology, University of Oregon, Eugene, OR 97403, USA. Abstract Patterns of phylogenetic relatedness within communities have been widely used to infer the importance of different ecological and evolutionary processes during community assembly, but little is known about the relative ability of community phylogenetics methods and null models to detect the signature of processes such as dispersal, competition and filtering under different models of trait evolution. Using a metacommunity simulation incorporating quantitative models of trait evolution and community assembly, I assessed the performance of different tests that have been used to measure community phylogenetic structure. All tests were sensitive to the relative phylogenetic signal in species metacommunity abundances and traits; methods that were most sensitive to the effects of niche-based processes on community structure were also more likely to find non-random patterns of community phylogenetic structure under dispersal assembly. When used with a null model that maintained species occurrence frequency in random communities, several metrics could detect niche-based assembly when there was strong phylogenetic signal in species traits, when multiple traits were involved in community assembly, and in the presence of environmental heterogeneity. Interpretations of the causes of community phylogenetic structure should be modified to account for the influence of dispersal. Keywords Community phylogenetic structure, dispersal, filtering, limiting similarity, metacommunity, neutral, niche, null model. Ecology Letters (2009) 12: INTRODUCTION The increasing availability of phylogenetic data for entire species assemblages has driven a proliferation of studies examining the phylogenetic diversity and phylogenetic structure of ecological communities at a range of spatial, temporal and taxonomic scales (Webb et al. 2002; Vamosi et al. 2009). Despite the increasing number of studies measuring patterns of community phylogenetic structure and using these patterns to infer the importance of different ecological and evolutionary processes (Emerson & Gillespie 2008), there have been few quantitative comparisons of the ability of different methods and null models to detect the signature of both species-neutral and niche-based assembly processes under different scenarios of trait evolution and community assembly. In the present study, I address this problem by comparing the ability of different community phylogenetic structure methods and null models to detect the signature of neutral and niche-based processes using simulated datasets created under known scenarios of trait evolution and community assembly. Many different methods have been used to quantify community phylogenetic structure, but most studies have used the same terminology of phylogenetic clustering and evenness to describe patterns of phylogenetic relatedness in ecological communities (Webb et al. 2002). Patterns of phylogenetic clustering and evenness have been measured to draw conclusions about different community assembly processes, and in particular to infer the relative importance of niche-based vs. neutral processes in structuring ecological communities. Based on a conceptual framework linking phylogenetic niche conservatism (Wiens & Graham 2005) with different assembly processes to predict patterns of phylogenetic relatedness (Webb et al. 2002), many studies

2 950 S. W. Kembel Letter have interpreted patterns of phylogenetic clustering as evidence for environmental filtering of phylogenetically conserved traits (Webb 2000; Horner-Devine & Bohannan 2006; Vamosi & Vamosi 2007), and phylogenetic evenness as evidence either for limiting similarity due to competition between close relatives when niches are phylogenetically conserved (Lovette & Hochachka 2006; Slingsby & Verboom 2006) or environmental filtering of convergently evolved traits (Cavender-Bares et al. 2004a). At the same time, random patterns of community phylogenetic structure have been interpreted as evidence for the importance of species-neutral processes during community assembly (Kembel & Hubbell 2006; Swenson et al. 2006). Despite the proliferation of empirical studies of community phylogenetic structure (Emerson & Gillespie 2008; Vamosi et al. 2009), very little is known about the performance of the different tests that have been used to infer the relative importance of community assembly processes based on patterns of community phylogenetic structure (Cavender-Bares et al. 2009). To date, inferences of process from pattern using these statistical methods has been largely based on several related assumptions: that limiting similarity or environmental filtering of traits explain non-random patterns of phylogenetic diversity, that phylogenetic niche and trait conservatism are widespread, and that patterns of community phylogenetic structure can be explained by the interaction between trait evolution and niche-based community assembly. While many studies have focused on limiting similarity and environmental filtering as competing explanations for community phylogenetic structure (Vamosi et al. 2009), a variety of other niche-based and species-neutral processes such as facilitation and dispersal may structure ecological communities, and while their potential role in generating phylogenetic community structure is increasingly being acknowledged (Emerson & Gillespie 2008), the role of these processes have not been widely incorporated into conceptual models of the processes that give rise to community phylogenetic structure (Cavender-Bares et al. 2009). In the absence of detailed information on the niches and traits of species, many studies of community phylogenetic structure have used phylogenetic relatedness as a surrogate measure for the ecological similarity of species, under an assumption of phylogenetic niche conservatism (Prinzing et al. 2001; Wiens & Graham 2005; Donoghue 2008). However, studies that measure phylogenetic distributions of traits directly have found a great deal of variation in the ecological similarity of relatives, including numerous examples of random or convergent patterns of niche and trait evolution (Losos et al. 2003; Cavender-Bares et al. 2004a), suggesting that an assumption of niche conservatism may not always be justified. More generally, the concepts of phylogenetic niche conservatism and convergence do not necessarily correspond to a testable evolutionary model (Losos 2008; Wiens 2008), and this may have prevented quantitative links between models of trait evolution and community assembly. Existing community phylogenetic structure tests incorporate both measures of patterns of relatedness in ecological communities and null model testing of observed patterns (Vamosi et al. 2009). While the performance of different ecological null models has been evaluated statistically in the context of studies of species co-occurrence patterns (Gotelli 2000), our understanding of the quantitative performance of these null models when used with measures of community phylogenetic structure is fragmentary, and the performance of community phylogenetic structure metrics and null models under a quantitative models of trait evolution and community assembly is essentially unknown. Kraft et al. (2007) evaluated the Type I and Type II error rates of different community phylogenetic structure metrics, but analysed individual communities assembled directly from a larger species pool with no dispersal limitation and assuming perfect knowledge of the species pool from which communities were assembled. Kembel & Hubbell (2006) and Hardy (2008) found that patterns of species occurrence across multiple local communities could bias measures of community phylogenetic structure and suggested that analysis of multiple local communities is required to account for the effects of neutral processes, but these studies only evaluated the Type I error rate of different methods since they analysed locally neutral datasets. To date no study has quantified the statistical performance of different community phylogenetic structure tests by analysing multiple local communities generated by quantitative models of trait evolution and ecological community assembly. Quantitative simulations will be essential if we wish to make robust inferences of process based on patterns of phylogenetic diversity. In the present study I conduct such a simulation to compare the relative power of different metrics and null models of community phylogenetic structure to detect niche-based community assembly mechanisms in communities assembled via both neutral dispersal from a metacommunity and niche-based filtering and limiting similarity. METHODS Overview of simulations To compare the performance of different metrics and null models of community phylogenetic structure, I simulated metacommunities of species for which patterns of trait evolution and community assembly into local communities were known. I generated numerous metacommunities of

3 Letter Measuring phylodiversity to detect niche and neutral processes 951 varying sizes by simulating the phylogenetic relationships among species in the metacommunity and evolving species traits and abundances on each phylogeny. I then assembled local communities from these metacommunities, while varying parameters such as local community size and the strength of environmental filtering or limiting similarity. This is similar to the general approach used by previous simulations (Raup et al. 1973; Colwell & Winkler 1984; Kraft et al. 2007), but with explicit simulation of abundance and trait evolution across multiple local communities assembled from a metacommunity. This approach creates local community data sets generated with a known set of processes, in which species differ realistically in their relative frequency of occurrence across samples (Fig. 1). Trait evolution Phylogenetic relationships among the species in each metacommunity were modelled as a stochastic Hey process with exponential waiting times to lineage coalescence using the APE package (Paradis et al. 2004) in the R statistical language (R Development Core Team 2008). This resulted in ultrametric phylogenetic trees with realistic branch lengths linking all species in a metacommunity. Given this phylogenetic tree linking taxa in each metacommunity, I then evolved traits and abundances on the tree in one of three ways to generate differing amounts of phylogenetic signal, with phylogenetic signal quantified using the K statistic (Blomberg et al. 2003). The first method (ÔrandomÕ) corresponded to a random or convergent pattern of trait evolution, where traits evolved on the tree under a Brownian motion model of evolution were shuffled across the tips of the phylogeny, producing traits and abundances with K = 0.25 ± 0.13 (mean ± SD). The second method (ÔBrownianÕ) corresponded to a moderately phylogenetically conserved pattern of trait evolution, with trait evolution via Brownian motion leading to some phylogenetic signal and close relatives tending to resemble one another. This resulted in traits and abundances with K = 1.00 ± The third method (ÔconservedÕ) corresponded to a strongly phylogenetically conserved pattern of trait evolution, with trait evolution via a decelerating model of trait evolution [accelerating-decelerating (ACDC) model with parameter g = 1.05 (Blomberg et al. 2003)] resulting in a strong phylogenetic signal (K = 3.55 ± 1.1). Thus, the range of phylogenetic signal in the simulated data was comparable with the range observed in many empirical data sets (Blomberg et al. 2003). Metacommunity abundances were exponentially transformed after evolution, to create a log-normal distribution of species metacommunity abundances that would give rise to a distribution of occurrence frequencies similar to those observed in real communities (Preston 1948). Multidimensional niches were simulated by evolving 1, 5 or 10 traits independently for each metacommunity. Community assembly For each metacommunity, multiple local communities of n species were first assembled through a process of dispersal from the metacommunity into each local community, with probability of a species dispersing to a local community proportional to metacommunity abundance. The abundance Metacommunity (n =100) # species Trait Abundance Figure 1 Simulation of local community assembly from a metacommunity via dispersal, filtering and limiting similarity. In this example, species in the metacommunity (n = 100 species) have abundances and traits that evolve independently via a Brownian motion model of trait evolution. Multiple local communities are assembled from the metacommunity with colonization probability weighted by metacommunity abundance. Species in local communities experience trait-based limiting similarity leading to local extinction of species that are too similar to co-occurring species. # species Occurrence frequency Dispersal (abundance-weighted colonization probability) Local community 1 (n = 40) # species Trait 20 local communities Limiting similarity (trait-mediated extinction probability) # species Local community 1 (n = 20) Occurrence frequency # species Trait 20 local communities

4 952 S. W. Kembel Letter of species in the metacommunity affected only their probability of colonizing local communities via dispersal, not their interactions or probability of survival after arrival, and thus the dispersal stage of community assembly was species neutral. After arrival in a local community, species could experience trait-based limiting similarity or filtering within the local community. In case of limiting similarity, I used the GAUSE algorithm of Colwell & Winkler (1984), whereby to generate a local community with a proportion PKILL of the n species in the community eliminated through competition, the species whose pairwise Euclidean trait distance to other species in a local community was smallest were sequentially eliminated until PKILL n species were removed. In case of filtering, a random trait value was selected as the optimum across all local communities, and the PKILL n species whose trait values were most distant from the optimum were eliminated from each local community. Filtering in multiple habitats was simulated by assigning local communities randomly to different habitats and selecting a different optimal trait value in each habitat. To evaluate the influence of phylogenetic signal in metacommunity abundance and traits, and of metacommunity and local community size on community phylogenetic structure, I evolved 100 metacommunities for each combination of simulation parameters listed in Table 1. I then assembled 20 local communities from each metacommunity under three different scenarios: dispersal, dispersal plus limiting similarity, or dispersal plus filtering. Under dispersal assembly, each local community was colonized by n species from the metacommunity with probability proportional to species metacommunity abundances. For dispersal plus limiting similarity or filtering, each local community was colonized by 2n species from the metacommunity and these species then experienced either limiting similarity or filtering to create local communities with n species remaining (PKILL = 0.5). Metrics of community phylogenetic structure Numerous metrics of community phylogenetic structure have been proposed, but most can be classified into one of two categories: methods that measure the relatedness of species occurring together in samples, and methods that measure the concordance between the co-occurrence and phylogenetic relatedness of species pairs. I will refer to these categories as sample-based and distance-based metrics respectively. An exhaustive review of these different measures and null models is outside the scope of this article, and I focus on the most commonly used measures of community phylogenetic structure and null models (see Hardy & Senterre (2007), Helmus et al. (2007), Hardy (2008) and Vamosi et al. (2009) for a comparison of terminology, Abundance signal Trait signal No. traits No. habitats Metacommunity richness Local community richness Table 1 Simulation parameters used to generate metacommunities Random Random Random Brownian Random Conserved Brownian Random Brownian Brownian Brownian Conserved Conserved Random Conserved Brownian Conserved Conserved Random Random Random Brownian Random Conserved Random Random Random Brownian Random Conserved Random Random Random Brownian Random Random Random Brownian Random Brownian For each simulation, 100 metacommunities were evolved for each combination of parameters, and then local communities were assembled from each metacommunity via dispersal, dispersal followed by limiting similarity, and dispersal followed by filtering.

5 Letter Measuring phylodiversity to detect niche and neutral processes 953 methods and heuristic interpretations of the metrics and null models used by different studies). I evaluated two of the most commonly used measures of relatedness within samples. The first is the previously mentioned mean pairwise phylogenetic distance (MPD) separating all pairs of species within a sample (Webb et al. 2002). The second is the mean phylogenetic distance to each speciesõ closest relative in a sample, which has been referred to as mean nearest phylogenetic taxon distance (MNTD; Kraft et al. 2007) or mean nearest phylogenetic neighbour distance (MNND; Webb et al. 2002). For both of these metrics, I calculated a standardized effect size (SES; Gurevitch et al. 1992) of the metric within each local community, based on a comparision of observed MPD MNTD values with the distribution of MPD MNTD generated by some null model: SES Metric ¼ Metric Observed MeanðMetric Randomized Þ SDðMetric Randomized Þ ð1þ SES MPD and SES MNTD are equivalent to )1 times the net relatedness index (NRI ) and nearest taxon index (NTI ) respectively (Webb et al. 2002). Tests for phylogenetic clustering and evenness using these metrics have generally tested whether the mean SES across all samples differs from the mean SES value of zero expected if the average phylogenetic structure of local communities is the same as the null communities, or tested whether the number of samples with statistically significant phylogenetic structure is greater than expected. In both cases, SES values higher than zero indicate phylogenetic evenness or overdispersion (species more distantly related than expected), and SES values lower than zero indicate phylogenetic clustering (species more closely related than expected). Distance-based methods compare measures of the co-occurrence of pairs of species with the phylogenetic distances separating those species. The most commonly used metric of co-occurrence has been a measure of proportional similarity (Schoener 1970): C ij ¼ 1 0:5 X p ik p jk ð2þ k where p ij represents the proportional abundance of species i in sample k (abundance in sample k divided by total abundance in that sample). Hardy (2008) proposed an alternative measure of species co-occurrence equivalent to a standardized version of the checkerboard score (Stone & Roberts 1990): DO ij ¼ P ij P i P j ð3þ P i P j where Pi is the proportion of sites where species i occurs and P ij is the proportion of sites at which species i and j co-occur. Pairwise co-occurrence measures were compared with the phylogenetic distances separating species using a Pearson correlation (phylogenetic distance vs. C ij : R PD-CO ; phylogenetic distance vs. DO ij : R PD-DO ; Hardy 2008), with positive correlations between co-occurrence and phylogenetic distance indicating phylogenetic evenness, and negative correlations indicating phylogenetic clustering (Cavender-Bares et al. 2004a). The strength of the observed correlation was compared with the strength of correlations in random data sets generated with a null model to assess the statistical significance of the result. Null models I evaluated several null models that have been commonly used with sample-based and distance-based metrics of community phylogenetic structure. Null model 1: taxa shuffle This null model involves shuffling taxa labels across the tips of the phylogenetic tree to randomize phylogenetic relationships among species, given a tree topology and branch lengths. When used with distance-based metrics, this null model is equivalent to a Mantel test of the correlation between co-occurrence and phylogenetic distance (Legendre & Legendre 1998). Taxa labels were shuffled among species occurring in local communities (the local pool). Null model 2: samples become random draws from local pool Null model 3: samples become random draws from regional pool If the species pool from which local communities are thought to be assembled is known, null communities can be generated by drawing species from the pool. The species richness of each null community is maintained but species are drawn at random with equal probability from the pool. The pool of potential colonists may simply represent the pool of species encountered across all samples included in a study (the local pool; null model 2), or it may be assembled from some larger spatial scale such as a regional species list (the regional pool; null model 3). Both null model 1 (shuffling taxa labels across the tips of the phylogeny) and null models 2 and 3 will cause the expected phylogenetic distance between any two species to converge on the same expected value over many randomizations: the MPD across all species pairs in the phylogeny or pool. These null models produced essentially identical results and I present combined results for null model 1 (tip shuffling across phylogeny of species encountered in local communities) and null model 2 (local communities become random draws from pool of species encountered in local communities).

6 954 S. W. Kembel Letter Null model 4: randomize co-occurrence maintaining sample richness and species occurrence frequency Species tend to differ in their frequency of occurrence, and differences in species frequencies have been shown to affect measures of species co-occurrence (Gotelli 2000) as well as measures of community phylogenetic structure (Kembel & Hubbell 2006; Hardy 2008). Null model 4 was implemented using the independent swap algorithm (Gotelli & Entsminger 2003) to generate null communities by randomizing species co-occurrences 1000 times per randomization while maintaining both sample richness and species occurrence frequency. Null model 5: shuffle species occurrences among samples As measures of species co-occurrence across samples are not as sensitive to sample richness as sample-based metrics of community phylogenetic structure, several studies using distance-based metrics have also included a null model where null communities are generated by shuffling occurrences of each species across all samples, which maintains species frequencies but not sample richness (Cavender- Bares et al. 2004a). Statistical analyses I analysed all simulated datasets with all combinations of metrics and null models using 999 randomizations for each analysis, and summarized the results both in terms of the mean value of the metrics (mean SES or mean correlation between co-occurrence and phylogenetic distance across metacommunities), and in terms of the statistical power of the metrics to detect assembly processes. For sample-based metrics, power was defined both as the proportion of times the mean value of SES across local communities assembled from a metacommunity was significantly different from zero according to a one-sample t-test (a = 0.05), and as the average proportion of local communities which exhibited non-random phylogenetic structure (observed phylogenetic distance in lower or upper 2.5% of distribution of null community phylogenetic distances; a = 0.05). For distancebased metrics, power was defined as the proportion of times non-random phylogenetic structure was detected (correlation in top 2.5% or bottom 2.5% of null distribution; a = 0.05). In all cases, I then tested whether the power of each metric and null-model was significantly greater than the expected value of 0.05 using a one-tailed binomial test. Following a widely-used framework of predicted community phylogenetic structure under different trait evolution and community assembly processes (cf. table 1 in Webb et al. 2002), Type I error was defined as the tendency for a method to find significant phylogenetic structure when communities were assembled by species-neutral dispersal (power under dispersal assembly), and Type II error was defined as the tendency for a method to find non-significant phylogenetic structure even when communities were assembled via trait-based limiting similarity or filtering (1 ) power under niche-based assembly). All analyses were carried out using PHYLOCOM version 3.41 software (Webb et al. 2008) and the PICANTE version 0.4 package (Kembel et al. 2008). RESULTS Simulation and test results are summarized graphically (Figs 2 and 3) using plots of the power of tests under neutral dispersal assembly from a given metacommunity (x-axis) vs. their power under niche-based assembly from the same metacommunity (y-axis). In each plot, dashed lines indicate the 0.05 significance level (a = 0.05), and the dotted line indicates the power above which tests would be considered significant according to a binomial test of power vs. the expected value of The ideal test to detect nichebased community assembly would have high power under niche-based assembly and low power under neutral assembly (the upper left corner of the plot). Tests falling in the shaded region of Fig. 3 would yield significant results under niche-based assembly and random results under dispersal assembly, and could potentially be used to detect nichebased community assembly. Tests outside this shaded region closer to the 1 : 1 line would not be recommended since they would be unable to distinguish patterns generated by niche-based and neutral community assembly. Many tests performed poorly in the framework predicting random patterns of phylogenetic structure under neutral community assembly and non-random patterns under nichebased assembly (Figs 2 and 3; Tables S1 and S2). In general, many tests performed better with less phylogenetic signal in abundances, more phylogenetic signal in traits, higher numbers of traits involved in niche-based community assembly, and in the presence of environmental heterogeneity. Many tests that had high statistical power to detect non-random patterns of community phylogenetic structure under niche-based community assembly also had poor Type I error rates (high statistical power under neutral dispersal assembly; Fig. 2), while tests with good type I error rates often had relatively low power to detect niche-based community assembly (Fig. 2). When both abundance evolution and trait evolution were phylogenetically conserved (ÔBrownianÕ or ÔconservedÕ abundance evolution and ÔconservedÕ trait evolution), most metrics used with null models 4 or 5 performed well when detecting the signal of competition. Null models 1 3 performed poorly at detecting competition, and all null models and metrics had relatively low power to detect the signal of filtering (Fig. 2), except when filtering was

7 Letter Measuring phylodiversity to detect niche and neutral processes 955 Abundance signal Random Brownian Conserved Random Trait signal Brownian MNTD/P/3 Conserved MNTD/P/3 MNTD/P/3 MPD/M/4 Figure 2 Plot of average power of phylogenetic community structure metrics and null models for 20 local communities of 10 species assembled via dispersal and dispersal followed by (a) filtering or (b) limiting similarity of a single trait from 100 simulations of a 50 species metacommunity. The solid line indicates a 1 : 1 relationship between power under dispersal assembly alone vs. power under dispersal plus trait-based assembly processes. Dashed lines indicate a = 0.05 significance level, dotted lines indicate threshold for power to be significantly greater than expected value of 0.05 (a = 0.05; one-tailed binomial test, n = 100 metacommunities). Metrics and null models are labelled in the format metric power null model (metric: MPD = SES MPD, MNTD = SES MNTD,CO=R PD-CO,DO=R PD-DO ; power: M = proportion of metacommunities for which the mean value of the standardized effect size was significantly different than zero, P = proportion of metacommunities (R PD-CO R PD-DO ) or average proportion of local communities (SES MPD SES MNTD ) for which there was statistically significant phylogenetic structure relative to a given null model; null model: 1 = taxa shuffle, 2 = samples become draws from local pool, 3 = samples become draws from regional pool, 4 = randomize co-occurrence maintaining sample richness and species occurrence frequency, 5 = shuffle species occurrences across samples). See Methods for further details of metrics, null models and simulation parameters. operating on multiple phylogenetically conserved traits (Fig. 3). When trait evolution was phylogenetically conserved (ÔconservedÕ or ÔBrownianÕ with multiple traits) and with no phylogenetic signal in abundance, several methods performed well, resulting in high power and significant results when competition or filtering were operating, but nonsignificant results otherwise (Fig. 2; Table S2). When

8 Trait signal 956 S. W. Kembel Letter Random Abundance signal Random Brownian Conserved MNTD/M/1 2 Brownian MPD/M/4 CO/P/5 Conserved CO/P/1245 MPD/M/4 MNTD/M/4 DO/P/1245 MNTD/P/3 MNTD/P/1 2 MNTD/P/4 MPD/P/4 CO/P/4 5 MNTD/M/4 MPD/M/123 MPD/M/4 DO/P/1245 MNTD/P/3 MNTD/P/1 2 MNTD/P/4 MPD/P/4 CO/P/5 MNTD/M/4 DO/P/125 MPD/M/4 CO/P/4 DO/P/4 MNTD/P/1 2 MNTD/P/3 MNTD/P/4 MPD/P/4 Figure 2 continued competition was operating, these included nearly all metrics and null models except the measures based on the proportion of significant MPD MNTD across local communities. When filtering was operating on multiple traits, only tests based on the mean value of standardized MPD MNTD plus null model 4 had power to detect filtering (Fig. 3). When filtering occurred in multiple habitats, the power of tests using null model 4 to detect filtering increased (Table S2). In general, tests based on the mean value of samplebased metrics departing from zero were more powerful than tests of the proportion of significant samples, and more powerful than the distance-based tests (Fig. 2; Table S2). But the mean value of the sample-based metrics also had the worst performance in terms of detecting patterns of non-random community phylogenetic structure under dispersal assembly. Similarly, the use of MNTD as a metric of community phylogenetic structure was less prone to Type I error than MPD when communities were assembled at random, but also resulted in less power to detect limiting similarity and filtering when they were operating. The power of all tests increased with weaker phylogenetic signal in abundances, stronger phylogenetic signal in traits, smaller metacommunity size, larger local community size, higher ratios of local community to metacommunity size, when multiple

9 Letter Measuring phylodiversity to detect niche and neutral processes 957 Random trait signal Brownian trait signal MPD/M/12 MNTD/M/12 MNTD/M/12 MPD/M/12 MNTD/M/4 MPD/M/4 CO/P/12 Figure 3 Plot of average power of phylogenetic community structure metrics and null models for 20 local communities of 10 species assembled via dispersal and dispersal followed by filtering of five traits from 100 simulations of a 50 species metacommunity. The solid line indicates a 1 : 1 relationship between power under dispersal assembly alone vs. power under dispersal plus trait-based filtering assembly. Dashed lines indicate a = 0.05 significance level, dotted lines indicate threshold for power to be significantly greater than expected value of 0.05 (a = 0.05; one-tailed binomial test, n = 100 metacommunities). The shaded region of the plot indicates tests with desirable statistical properties (high power to detect niche-based assembly, low power to detect dispersal assembly). Metrics and null models are labelled in the format metric power null model (see Fig. 2 caption for details). See Methods for further details of metrics, null models and simulation parameters. traits were involved in community assembly, and in the presence of habitat heterogeneity (Table S2). Under neutral dispersal assembly, when there was phylogenetic signal in species frequency many tests were too statistically liberal, finding significant phylogenetic clustering regardless of the community assembly process (Fig. 2). Even when there was no phylogenetic signal in abundance, null models that did not maintain observed frequencies resulted in statistically significant results when analysing neutrally assembled communities. In terms of their sensitivity to phylogenetic signal in abundance, distancebased methods performed better than sample-based metrics, MNTD performed better than MPD, and null model 4 performed better than other nulls (Fig. 2). With dispersal assembly followed by trait filtering within communities and random phylogenetic signal in abundances, mean values of all metrics were in the predicted direction of phylogenetic clustering (Webb et al. 2002; Table S1). With dispersal assembly plus limiting similarity within communities, if the phylogenetic signal of frequency was random most metrics could detect limiting similarity via a pattern of phylogenetic evenness. However, when there was phylogenetic signal in abundance, many methods suffered from a lack of power, or found patterns of phylogenetic clustering when random structure or phylogenetic evenness was the predicted outcome of a given combination of assembly processes and trait evolution (Webb et al. 2002; Table S2). Phylogenetic evenness was only detected when there was no phylogenetic signal in species abundance, phylogenetic signal in traits, and limiting similarity operating during community assembly. DISCUSSION Using and interpreting community phylogenetic structure tests The ability of community phylogenetic structure tests to detect niche-based assembly processes depended on both the details of trait evolution and the relative importance of different assembly processes. Studies of community phylogenetic structure that wish to detect trait-based community assembly processes should first quantify the phylogenetic signal in species occurrence frequencies (Hardy 2008) and in ecologically relevant traits that are hypothesized to play a role in community assembly, given the sensitivity of all tests to phylogenetic signal in abundances and traits. The increase in power to detect filtering when local communities were environmentally heterogeneous suggests that environmental covariates may be important predictors of community phylogenetic structure (Cavender-Bares et al. 2009). In general, community phylogenetic structure tests using null models that randomized the phylogeny or did not maintain species frequencies when randomizing the community matrix (null models 1, 2 and 3) were unreliable and would not be recommended to detect niche-based community assembly, even in the absence of phylogenetic signal in abundances. Several tests using null model 4 (independent

10 958 S. W. Kembel Letter swap) combined good Type I error rates with power to detect niche-based assembly processes. Tests based on mean values of SES MPD or SES MNTD plus null model 4 had the best overall ability to detect filtering when multiple traits were involved in community assembly, while R PD-CO and R PD-DO plus null models 4 or 5 were best able to detect the effects of limiting similarity on community phylogenetic structure. Linking pattern and process in ecophylogenetics The results of this study suggest that the patterns of phylogenetic clustering that have been observed by many empirical studies (Emerson & Gillespie 2008; Vamosi et al. 2009) do not actually provide strong evidence for widespread niche-based environmental filtering of conserved traits, since many of these studies used null models that did not maintain species frequencies, and thus the observed pattern of phylogenetic clustering could also be explained by species-neutral dispersal assembly of these communities. Given the relatively small set of parameter space explored by this study, and the fact that multiple processes could produce similar patterns of phylogenetic relatedness, tests of community phylogenetic structure should be seen as a starting point to suggest further experimental tests and model parameterization, rather than strong evidence for the prevalence of one process or another. This study provides further evidence that neutral models are not null models in terms of their predictions about community structure and evolution (Gotelli & McGill 2006). Non-random patterns of community phylogenetic structure were observed even when communities were assembled purely by neutral dispersal from a metacommunity. This result agrees with previous findings (Hardy 2008) that variation in species occurrence frequency and environmental heterogeneity influence measures of relatedness across multiple local communities, which explains the difference between my findings and those of Kraft et al. (2007) who used very similar methods to assemble communities, but only considered a single local community assembled from a larger pool of species directly without dispersal limitation. The present study is not an exhaustive examination of the huge potential parameter space for models of trait evolution and community assembly, but was meant to test an extreme case where a few processes were operating at high intensity. I did not simulate or analyse within-community abundance variation, tested a subset of the methods and null models that have been proposed, examined only a single intensity of niche-based assembly, and simulated relatively small metacommunities and local communities, but even with these limitations, there was a consistent trend in the performance of different tests. Environmental heterogeneity among local communities and variation in the spatial scale used to define communities and species pools would be likely to affect the power of many statistical tests of community phylogenetic structure (Kraft et al. 2007; Hardy 2008, Cavender-Bares et al. 2009), and future studies should examine the influence of these and other factors on the performance of different tests. Several authors have suggested that the concepts of conserved vs. convergent evolution be replaced with measures of phylogenetic signal (Revell et al. 2008), the similarity of close relatives relative to that expected under a quantitative models of trait evolution (Blomberg & Garland 2002; Blomberg et al. 2003), but there remains a great deal of heterogeneity in how studies have measured and defined niche conservatism (Losos 2008). This and other studies have shown that traits that evolve via drift or a random walk can give rise to very non-random looking patterns of trait similarity, abundance distributions on a phylogeny, and community structure (Bell 2000; Hubbell 2001). More generally, there are many ways that traits and niches of species may evolve, from niche-filling models (Price 1997) to random walks or stabilizing selection (Estes & Arnold 2007), or randomly with respect to phylogenetic relationships if evolution is rapid or during adaptive radiations into novel habitats (Losos et al. 2003). Quantification of the phylogenetic signal in ecological traits and niches in a model-based framework, along with further studies to understand patterns of phylogenetic community structure generated by different processes, will be required to understand the prevalence of different modes of evolution and what this implies for the evolution of species niches and habitat associations (Donoghue 2008; Losos 2008). Multiple processes are likely to be operating at the same time to structure ecological communities (Emerson & Gillespie 2008), and there is a need for ecophylogenetics studies to move beyond either-or tests for different processes towards using models to directly estimate parameters of interest such as dispersal rates or the intensity of species interactions (Jabot & Chave 2009). Most ecological null models used to test community assembly processes are based on rearranging a species co-occurrence matrix, but these null models may not be able to easily differentiate among the numerous processes that can give rise to local community structure (Ulrich 2004), which include dispersal as well as niche-based assembly processes such as limiting similarity, facilitation, or environmental filtering (Emerson & Gillespie 2008; Vamosi et al. 2009). This disconnect between traditional ecological null models and the ecological and evolutionary processes that many studies are trying to detect suggest a need to develop new strategies for quantifying community phylogenetic structure and comparing observed patterns to those expected under different models of evolution and community assembly.

11 Letter Measuring phylodiversity to detect niche and neutral processes 959 Models similar to existing biogeographic methods (Ree et al. 2005) could be applied to studies of community structure to directly estimate the relative importance of processes such as colonization, local extinction, dispersal and niche-based local interactions on community structure. Similarly, models of the effects of neutral and niche-based processes on species abundance distributions (Jabot et al. 2008) could potentially be extended to measures of community phylogenetic structure and phylogenetic patterns of species abundance, given the phylogenetic predictions inherent in neutral theory (Hubbell 2001). Direct estimation of the strength of fundamental processes and comparison of model performance, as well as continued simulation studies to directly evaluate the ability of different methods to measure ecological and evolutionary processes, will be required to improve our understanding of the processes responsible for patterns of community phylogenetic structure. ACKNOWLEDGEMENTS The author thanks the National Sciences and Engineering Research Council (NSERC) of Canada for support. Comments from David Ackerly, Jeannine Cavender-Bares, Jerome Chave, Franck Jabot, Nathan Kraft and three anonymous reviewers improved the quality of this manuscript. REFERENCES Bell, G. (2000). The distribution of abundance in neutral communities. Am. Nat., 155, Blomberg, S.P. & Garland, T. (2002). Tempo and mode in evolution: phylogenetic inertia, adaptation and comparative methods. J. Evol. Biol., 15, Blomberg, S.P., Garland, T. & Ives, A.R. (2003). Testing for phylogenetic signal in comparative data: behavioral traits are more labile. Evolution, 57, Cavender-Bares, J., Ackerly, D.D., Baum, D.A. & Bazzaz, F.A. (2004a). Phylogenetic overdispersion in Floridian oak communities. Am. Nat., 163, Cavender-Bares, J., Kozak, K., Fine, P.V.A. & Kembel, S.W. (2009). The merging of community ecology and phylogenetic biology. Ecol. Lett., 12, Colwell, R.K. & Winkler, D.W. (1984). A null model for null models in biogeography. In: Ecological Communities: Conceptual Issues and the Evidence (eds Strong, D.R., Simberloff, D., Abele, L.G. & Thistle, A.B.). Princeton University Press, Princeton, pp Donoghue, M.J. (2008). A phylogenetic perspective on the distribution of plant diversity. Proc. Natl Acad. Sci., 105, Emerson, B.C. & Gillespie, R.G. (2008). Phylogenetic analysis of community assembly and structure over space and time. Trends Ecol. Evol., 23, Estes, S. & Arnold, S.J. (2007). Resolving the paradox of stasis: models with stabilizing selection explain evolutionary divergence on all timescales. Am. Nat., 169, Gotelli, N.J. (2000). Null model analysis of species co-occurrence patterns. Ecology, 81, Gotelli, N.J. & Entsminger, G.L. (2003). Swap algorithms in null model analysis. Ecology, 84, Gotelli, N.J. & McGill, B.J. (2006). Null versus neutral models: whatõs the difference? Ecography, 29, Gurevitch, J., Morrow, L.L., Wallace, A. & Walsh, J.S. (1992). A meta-analysis of competition in field experiments. Am. Nat., 140, 539. Hardy, O.J. (2008). Testing the spatial phylogenetic structure of local communities: statistical performances of different null models and test statistics on a locally neutral community. J. Ecol., 96, Hardy, O.J. & Senterre, B. (2007). Characterizing the phylogenetic structure of communities by an additive partitioning of phylogenetic diversity. J. Ecol., 95, Helmus, M.R., Bland, T.J., Williams, C.K. & Ives, A.R. (2007). Phylogenetic measures of biodiversity. Am. Nat., 169, E68 E83. Horner-Devine, M.C. & Bohannan, B.J. (2006). Phylogenetic clustering and overdispersion in bacterial communities. Ecology, 87, S100 S108. Hubbell, S.P. (2001). The Unified Neutral Theory of Biogeography and Biodiversity. Princeton University Press, Princeton, NJ. Jabot, F. & Chave, J. (2009). Inferring the parameters of the neutral theory of biodiversity using phylogenetic information, and implications for tropical forests. Ecol. Lett., 12, Jabot, F., Etienne, R.S. & Chave, J. (2008). Reconciling neutral community models and environmental filtering: theory and an empirical test. Oikos, 117, Kembel, S.W. & Hubbell, S.P. (2006). The phylogenetic structure of a Neotropical forest tree community. Ecology, 87, S86 S99. Kembel, S., Ackerly, D., Blomberg, S., Cowan, P., Helmus, M. & Webb, C. (2008). picante: Tools for Integrating Phylogenies and Ecology, Version Software package available from r-forge.r-project.org. Last accessed 1 January Kraft, N.J., Cornwell, W.K., Webb, C.O. & Ackerly, D.D. (2007). Trait evolution, community assembly, and the phylogenetic structure of ecological communities. Am. Nat., 170, Legendre, P. & Legendre, L. (1998). Numerical Ecology, 2nd English edn ed. Elsevier, New York. Losos, J.B. (2008). Phylogenetic niche conservatism, phylogenetic signal and the relationship between phylogenetic relatedness and ecological similarity among species. Ecol. Lett., 11, Losos, J.B., Leal, M., Glor, R.E., de Queiroz, K., Hertz, P.E., Schettino, L.R. et al. (2003). Niche lability in the evolution of a Caribbean lizard community. Nature, 424, Lovette, I.J. & Hochachka, W.M. (2006). Simultaneous effects of phylogenetic niche conservatism and competition on avian community structure. Ecology, 87, S14 S28. Paradis, E., Claude, J. & Strimmer, K. (2004). APE: Analyses of phylogenetics and evolution in R language. Bioinformatics, 20, Preston, F.W. (1948). The commonness, and rarity, of species. Ecology, 29, Price, T. (1997). Correlated evolution and independent contrasts. Philos. Trans. R. Soc. B, 352, Prinzing, A., Durka, W., Klotz, S. & Brandl, R. (2001). The niche of higher plants: evidence for phylogenetic conservatism. Proc. R. Soc. Lond. B Biol. Sci., 268,

12 960 S. W. Kembel Letter R Development Core Team (2008). R: A language and environment for statistical computing, R Foundation for Statistical Computing, Vienna, Austria. Version 2.6. From R-project.org, last accessed on 1 January Raup, D.M., Gould, S.J., Schopf, T.J.M. & Simberloff, D.S. (1973). Stochastic models of phylogeny and the evolution of diversity. J. Geol., 81, Ree, R.H., Moore, B.R., Webb, C.O. & Donoghue, M.J. (2005). A likelihood framework for inferring the evolution of geographic range on phylogenetic trees. Evolution, 59, Revell, L.J., Harmon, L. & Collar, D. (2008). Phylogenetic signal, evolutionary process, and rate. Syst. Biol., 57, Schoener, T.W. (1970). Nonsynchronous spatial overlap of lizards in patchy habitats. Ecology, 51, Slingsby, J.A. & Verboom, G.A. (2006). Phylogenetic relatedness limits co-occurrence at fine spatial scales: evidence from the schoenoid sedges (Cyperaceae: Schoeneae) of the Cape Floristic Region, South Africa. Am. Nat., 168, Stone, L. & Roberts, A. (1990). The checkerboard score and species distributions. Oecologia, 85, Swenson, N.G., Enquist, B.J., Pither, J., Thompson, J. & Zimmerman, J.K. (2006). The problem and promise of scale dependency in community phylogenetics. Ecology, 87, Ulrich, W. (2004). Species co-occurrences and neutral models: reassessing J.M. DiamondÕs assembly rules. Oikos, 107, Vamosi, J.C. & Vamosi, S.M. (2007). Body size, rarity, and phylogenetic community structure: insights from diving beetle assemblages of Alberta. Divers. Distrib., 13, Vamosi, S.M., Heard, S.B., Vamosi, J.C. & Webb, C.O. (2009). Emerging patterns in the comparative analysis of phylogenetic community structure. Mol. Ecol., 18, Webb, C.O. (2000). Exploring the phylogenetic structure of ecological communities: an example for rain forest trees. Am. Nat., 156, Webb, C.O., Ackerly, D.D., McPeek, M.A. & Donoghue, M.J. (2002). Phylogenies and community ecology. Annu. Rev. Ecol. Syst., 33, Webb, C.O., Ackerly, D.D. & Kembel, S.W. (2008). Phylocom: software for the analysis of phylogenetic community structure and trait evolution. Bioinformatics, 24, Wiens, J.J. (2008). Commentary: niche conservatism déjà vu. Ecol. Lett., 11, Wiens, J.J. & Graham, C.H. (2005). Niche conservatism: Integrating evolution, ecology, and conservation biology. Ann. Rev. Ecol. Evol. Syst., 36, SUPPORTING INFORMATION Additional Supporting Information may be found in the online version of this article: Table S1 Mean value of various metrics of community phylogenetic structure across 100 metacommunities, under different metacommunity simulation parameters, local community assembly processes and null models. Table S2 Statistical power of various metrics of community phylogenetic structure under different metacommunity simulation parameters, community assembly processes, and null models. As a service to our authors and readers, this journal provides supporting information supplied by the authors. Such materials are peer-reviewed and may be re-organized for online delivery, but are not copy-edited or typeset. Technical support issues arising from supporting information (other than missing files) should be addressed to the authors. Editor, Jerome Chave Manuscript received 7 January 2009 First decision made 13 February 2009 Second decision made 4 May 2009 Manuscript accepted 25 June 2009