Plant traits and extinction in urban areas: a meta-analysis of 11 cities


Richard P. Duncan, Bio-Protection Research Centre, PO Box 84, Lincoln University, Christchurch 7674, New Zealand. E-mail:


Aim  Urban environments around the world share many features in common, including the local extinction of native plant species. We tested the hypothesis that similarity in environmental conditions among urban areas should select for plant species with a particular suite of traits suited to those conditions, and lead to the selective extinction of species lacking those traits.

Location  Eleven cities with data on the plant species that persisted and those that went locally extinct within at least the last 100 years following urbanization.

Methods  We compiled data on 11 plant traits for 8269 native species in the 11 cities and used hierarchical logistic regression models to identify the degree to which traits could distinguish species that persisted from those that went locally extinct in each city. The trait effects from each city were then combined in a meta-analysis.

Results  The cities fell into two groups: those with relatively low rates of extinction (less than 0.05% species per year – Adelaide, Hong Kong, Los Angeles, San Diego and San Francisco), for which no traits reliably predicted the pattern of extinction, and those with higher rates of extinction (> 0.08% species per year – Auckland, Chicago, Melbourne, New York, Singapore and Worcester, MA), where short-statured, small-seeded plants were more likely to go extinct.

Main conclusions  Our analysis reveals patterns in trait selectivity consistent with local studies, suggesting some consistency in trait selection by urbanization. Overall, however, few traits reliably predicted the pattern of plant extinction across cities, making it difficult to identify a priori the extinction-prone species most likely to be affected by urban expansion.


Humans often monopolize resources and transform landscapes for their own purposes, which can lead to changes in abundance or the local extinction of native plant species (Vitousek et al., 1997). Understanding these changes will help ecologists determine how species are likely to respond to ongoing human impacts and identify ways to mitigate adverse effects. Identifying predictable relationships between plant traits and environmental conditions or disturbances is a promising approach for understanding how plant communities change in response to human land-use modification (Westoby & Wright, 2006; Díaz et al., 2007). Particular environments tend to favour the persistence of plant species with particular trait combinations that appear well suited to the conditions. A marked change in environmental conditions may alter the suite of suitable traits, causing the loss of some species while favouring the persistence of others. The nature of these changes may be predictable given consistent trait–environment associations, which define how changes in environmental conditions should select for or against species with certain traits.

Human impacts are nowhere more evident than in urban environments, which are often radically transformed in order to meet human needs. The intensive development associated with urbanization has been linked to declines and local extinctions of native plant species (Duncan & Young, 2000; Preston, 2000; Williams et al., 2005; Hahs et al., 2009), although not all native species are adversely affected. Urban environments share many features in common because they are designed to perform standard functions to meet human needs (Grimm et al., 2008). Such an environment, common to cities around the world, might be expected to select for species with a similar suite of traits favouring persistence in highly disturbed and human-modified habitats. Indeed the process of urbanization has been conceptualized as a series of filters acting on an existing species pool and selectively removing those species with traits unfavourable for persistence in this new environment (Williams et al., 2009). The distinctive nature of urban habitats and the similarity in urban environments around the world suggests that the native floras of cities should converge toward a subset of species sharing a common suite of traits, with the selective extinction of species lacking those traits (McKinney, 2006; Schwartz et al., 2006; Knapp et al., 2008).

Our aim in this study is to test this hypothesis. To do this, we obtained comprehensive data on changes, over at least the last 100 years, in the native flora of 11 cities from North America, Asia and Australasia (Table 1). All of these cities were founded relatively recently (all after ad 1600, with all but two founded after ad 1800) and had initial floristic surveys documenting the composition of the flora around the time that extensive urban development commenced. For each city we could extract a list of the native plant species initially present in the area and then classify each species as having persisted or having gone locally extinct from the area following at least 100 years of urbanization. Given that cities share many environmental features in common, regardless of where they are, we predicted that a shift to an urban environment should select for species with a predictable suite of traits (see, for example, Knapp et al., 2008). A consequence of this selection pressure being similar across urban areas is that we expected the traits that reliably distinguish locally extinct species from those that persisted to be consistent across the 11 cities.

Table 1.  Location (latitude and longitude), city foundation date, data sources, the number of native species that became locally extinct in the study period, the number of native species that persisted (extant), the length of the study period and the rate of species extinctions (percentage going extinct per year) for 11 cities.
CityLatitudeLongitudeCity foundation date (year ad)Locally extinct nativesExtant nativesStudy length (years)Percentage extinct per yearData source
Adelaide−34.9138.61836819641660.047 Tait et al. (2005)
Auckland−36.9174.81840802931140.188 Duncan & Young (2000), Esler (1991)
Hong Kong22.3114.218425615901630.021 Corlett et al. (2000)
Los Angeles34.1−118.31850186551450.018M.W.S., unpublished data
Melbourne−37.8145.018351178851400.083A.K.H. & N.S.G.W., unpublished data
New York40.7−74.0162440111592070.124 Moore et al. (2004)
San Diego32.8−117.2185077021450.007M.W.S., unpublished data
San Francisco37.8−122.41850196951450.018M.W.S., unpublished data
Singapore1.4103.818195981578 c. 1000.275 Chong et al. (2009)
Worcester, MA52.2−2.21722174644 c. 1000.213 Bertin (2002)


City flora data

We obtained data on changes in the native flora over at least the last 100 years for 11 cities from North America, Asia and Australasia: Adelaide, Auckland, Chicago, Hong Kong, Los Angeles, Melbourne, New York, San Diego, San Francisco, Singapore and Worcester (MA, USA). The cities were a subset of those analysed in Hahs et al. (2009), chosen for two reasons: (1) they were cities for which we had complete lists of the native species that were present in an area based on initial floristic surveys, with each species classified as having persisted or having gone locally extinct based on contemporary surveys, and (2) we excluded older cities founded prior to ad 1600 because cities with a long history of urbanization may already have lost vulnerable species prior to the initial floristic surveys, which occurred mostly in the 19th century (Hahs et al., 2009). Data on changes in the native flora for each area were obtained from a variety of sources including published papers and unpublished analyses of herbarium and agency databases (Table 1). The survey area for each city varied, but typically encompassed the heavily urbanized areas of central cities, surrounding suburban neighbourhoods and adjacent semi-urban environments. In the case of Hong Kong, the survey area was all of Hong Kong territory. To ensure consistency, we combined the species lists for all cities and then carefully checked taxonomy, correcting synonyms and ensuring that the family-level taxonomy was consistent and matched that of the Angiosperm Phylogeny Group (Angiosperm Phylogeny Group, 2003). Subspecies and varieties were excluded by merging them into species, so that local extinction refers only to species level.

Habitat affinity and trait data

For each species we obtained data on habitat affinity and plant traits. Due to the size of the dataset (8269 species) our choice of which traits to include depended on data availability, although we selected traits that have previously been shown to correlate with extinction risk in plants (plant height, seed mass, growth form, dispersal mode, longevity; Leach & Givnish, 1996; Turner et al., 1996; Duncan & Young, 2000; Williams et al., 2005; Fréville et al., 2007) and traits hypothesized to be selected for or against in urban environments (Godefroid, 2001; Thompson & McCarthy, 2008; Williams et al., 2009), including habitat affinity (due to selective habitat transformation associated with urbanization), nutrient uptake strategy and pollination system (with species having specialized systems being more vulnerable to extinction), photosynthetic pathway (with C4 plants potentially profiting from warmer temperatures due to urban heat island effects) and spinescence (with humans favouring non-spiny species in urban areas).

We determined the habitat affinity of each species using the habitat categories of forest, aquatic, riparian, grassland, coastal dunes and scrubland by consulting floras, ecological texts and expert ecologists. Species were assigned to a habitat category if they were commonly found in that habitat, and species could be assigned to more than one category.

Using published floras, trait databases, expert knowledge and other published sources, we obtained data on the growth form, clonality, dispersal mode, nutrient uptake strategy, spinescence, pollination system, photosynthetic pathway, plant height and seed mass for each species (see Table S1 in Supporting Information). Traits were initially classified following the protocols outlined in Cornelissen et al. (2003), and subsequently modified by combining classes when there were trait categories with few representatives. Modifications to Cornelissen et al. (2003) included reducing growth form from 20 to nine classes, nutrient uptake strategy to four classes and dispersal mechanism to five (see Table S1 for a description of the trait classes). Pollination system was not considered in Cornelissen et al. (2003); we defined it as having three classes, namely abiotic (i.e. wind, water), specialized biotic (one or a few specialist animal pollinators) and unspecialized biotic (pollinated by many animal species).

Due to the size of the dataset and the large number of species for which no formal trait screening has occurred, we made several generalizations. Photosynthetic pathway was assumed to be C3 unless the plant was from a family known to be otherwise, and we combined CAM (crassulacean acid metabolism) and C4 photosynthetic pathways into one category. Although some species can exhibit multiple photosynthetic pathways under different environmental conditions (Kluge, 1977; Sayre & Kennedy, 1977) we assigned only one, and used C3 by default if several were coded. Species in the families Cyperaceae and Poaceae and all ferns and gymnosperms were assigned as wind pollinated, and species in the family Orchidaceae were assigned to specialized biotic pollination. Where no data were available we assumed the nutrient uptake strategy for species in the families Orchidaceae and Ericaceae to be mycorrhizal, and species in the family Fabaceae to be nitrogen fixers.

Plant heights (in metres) were obtained from floras and other published and unpublished descriptions, with all floating and submerged aquatic plants given a nominal height of 0.01 m. Seed mass (in grams) was obtained from the Kew Seed Information Database (SID; Liu et al., 2008) and USDA PLANTS (USDA, 2008) databases and published plant trait studies. For ferns, spore mass was used to represent seed mass. Where there was no measurement in the databases, seed mass for ferns and orchids was assigned 0.001 mg. When we obtained multiple values for any species we used the mean of the values for plant height and the median of the values for seed mass.

Our final dataset comprised 8269 native species with 11,936 species by city occurrence records. Given the size of our dataset, we were unable to compile complete trait data for all species and our dataset contained missing values: of the 90,959 species by trait combinations 11,762 were missing (about 13%; see Table S1). We describe how we dealt with these missing values in our analysis below.

Analysis of extinction in individual cities

Figure 1 summarizes the three stages of our analysis, which we describe here in more detail. First, we analysed the data for each city separately using hierarchical logistic regression models to identify attributes that distinguished species that persisted in each urban area from those that became locally extinct. The explanatory variables used to discriminate between extant and locally extinct species were the plant traits (growth form, clonality, dispersal mode, nutrient uptake strategy, spinescence, pollination system, photosynthetic pathway, plant height and seed mass), the habitat affinities of the species, and plant family. The analysis for each city was based on a model with the form:

Figure 1.

Diagram summarizing the three main stages in our analysis of the native plant extinction data across 11 cities. See Methods for more details.


where pi,j is the probability that species j became extinct in city i, ai is the intercept term, bi,k is the coefficient estimating the effect of trait or habitat k in city i on the probability of extinction, xi,j,k is the value of trait or habitat k for species j in city i, and inline imageis the coefficient estimating the effect of belonging to taxonomic family φj in city i on the probability of extinction. Categorical trait variables were included by coding them as dummy variables and choosing one of the classes as a reference class with the coefficient set to zero. Continuous trait variables (height and seed mass) were log transformed and then standardized by subtracting the mean and dividing by two standard deviations, to ensure the regression coefficients were comparable with those for the categorical variables (Gelman & Hill, 2007). The regression coefficients that describe the effect of each trait, habitat and plant family on the probability of extinction are all conditional on other variables in the model, and thus describe the effect of a particular trait, habitat or family having accounted for the other effects.

We used a Bayesian framework to accommodate the missing data and the hierarchical structure of the analysis (Gelman & Hill, 2007). The regression coefficients for the continuous variables plant height and seed mass were treated as fixed effects. The regression coefficients for the categorical variables bi,k were treated as random effects to help stabilize the parameter estimates when data were sparse, and family inline imagewas also treated as a random effect. The regression coefficients describing the effect of each family were assumed to be drawn from a common normal distribution with a mean of zero and a standard deviation (σf,i) that was estimated from the data. The regression coefficients for each categorical variable were assumed to be drawn from common normal distributions with a mean of zero and a standard deviation (σb,i) that was estimated from the data. Thus:




where N(a, b) indicates a variable that is normally distributed with mean a and standard deviation b. Data were not available for all traits for all species. Deleting species with missing data from the analysis would reduce the number of data points substantially, and probably bias the results due to the selective removal of species that were less well known. To accommodate missing data, we modelled traits as a function of plant family, thereby using data on traits from closely related species to model the missing values. Missing continuous trait data (height and seed mass) were modelled thus:


where inline imageis the mean value of trait k of species in family φj in city i. The standard deviation inline imagedefines the variation in trait k among species within family φj. Missing categorical data were modelled as being drawn from a multi-nomial distribution with the probability of occurrence in each trait category determined by the distribution of occurrences in the family to which the species belonged. We used the trait and family data for all species in all urban areas to model missing data for species in each urban area.

The Bayesian logistic regression models were fitted using Markov chain Monte Carlo (MCMC) methods as implemented in the Open BUGS software (Thomas et al., 2006), called from R vs. 2.7.1 (R Development Core Team, 2004). We specified non-informative prior distributions to allow the data to drive parameter estimation: variance terms were assigned broad uniform priors (0–100) on the standard deviations following Gelman (2006), while the remaining parameters were assigned normal prior distributions with mean 0 and variance 1000. Each model was run for 20,000 iterations with a burn-in of 5000 iterations, which was sufficient to achieve convergence.

We calculated the area under the receiver operating characteristic curve (AUC) as a measure of the predictive accuracy of each model. AUC measures the likelihood that an extinct species will have a higher predicted probability of extinction from the model than an extant species. An AUC value of 0.5 indicates that a model has no ability to discriminate between these classes (it performs no better than chance), while AUC values closer to 1 indicate models that more accurately assign probabilities that correctly discriminate the classes. Hosmer & Lemeshow (2000) suggest interpreting AUC values as follows: 0.7 ≤ AUC < 0.8 = acceptable; 0.8 ≤ AUC < 0.9 = excellent; 0.9 ≤ AUC = outstanding.


To determine whether particular traits had a consistent effect on the probability of extinction across cities, we used meta-analysis to summarize trait effects by modelling the regression coefficients of each city as random effects drawn from a common distribution for a particular trait. Thus


where bk is the mean effect of trait k across cities, the standard deviation σb,k defines the variation in the effect among cities, and the standard deviation si,b,k defines the uncertainty in the estimate of bi,k (the standard deviation of the posterior distribution of the parameter, equivalent to the standard error). By including si,b,k in the meta-analysis, more weight is placed on cities with more precise estimates of the regression coefficients. This was important because some cities had few recorded extinctions, with a small number of extinct species meaning that traits were sometimes estimated to have large effects (because the few extinct species possessed those traits) but high uncertainty. Weighting the regression coefficients ensured that anomalously large but imprecise estimates did not have an undue influence on the results. Indeed, removing the cities with very low extinction rates (San Diego, Los Angeles and San Francisco) had no discernible effect on the results of the meta-analysis. The meta-analysis was fitted using Open BUGS, called from R v.2.7.1, specifying non-informative prior distributions and running the model for 10,000 iterations with a burn-in of 10,000 iterations, which ensured convergence.

Variable importance

The meta-analysis tests whether each trait shows evidence for a consistent effect on extinction probability across all cities in this study. Evidence for this would provide us with the greatest generality; traits showing consistent effects are potentially useful predictors of extinction in other urban areas, and may reflect extinction processes common to all urban environments. Lack of evidence for a consistent effect could arise for several reasons, of which we note two.

First, none of the traits we considered may be useful predictors of extinction anywhere, and the lack of generality simply reflects this. Under these circumstances, we expect the logistic regression models for individual cities to fit the data poorly and for each model to have low explanatory power. We evaluated this using the AUC scores calculated for each model.

Second, if the models do have useful explanatory power then lack of generality could arise because: (1) outcomes are idiosyncratic, with different attributes predicting extinction in different urban areas, or (2) some traits do consistently predict extinction, but the direction of their effect varies among urban areas, which cancels out in the meta-analysis. This could arise, for example, if height was consistently associated with extinction probability, but tall plants were prone to extinction in some areas and short plants in others. We can distinguish between these outcomes by identifying which traits are the most important determinants of extinction in each city, which may further allow us to identify clusters of cities that share extinction-predicting traits in common.

To do this, we used the variable selection approach proposed by Kuo & Mallick (1998) in which the logistic regression models for each city were expanded by assigning each explanatory variable a latent indicator variable wi,k, taking the value 1 if the variable k in city i is included in the model and 0 if it is not. The indicator variables were given prior distributions:


which equate to a uniform prior on the number of traits to include in a model (Ley & Steel, 2007). At the completion of the MCMC iterations we obtained a distribution of values (either 0 or 1) for each wi,k, the mean of which can be interpreted as the probability that variable k would be included in the most probable model for city i as defined by all possible variable combinations. Variables that are consistently selected (mean wi,k close to 1) can be thought of as more important in explaining extinction outcomes in a particular city than variables that are rarely selected (mean wi,k close to 0).

While vague priors were used to allow the data to drive parameter estimation in model fitting, vague priors are usually not uninformative when it comes to assessing model probabilities (Link & Barker, 2006). To deal with this issue, we used the approach suggested by Aitkin (1991), and set the priors for each coefficient to the posterior distributions obtained from the initial logistic regression models, assuming that these followed a normal distribution. The logistic regression models with latent indicator variables were fitted using Open BUGS, called from R v.2.7.1, and run for 20,000 iterations after a burn-in of 5000 iterations, which achieved convergence.


Extinction rates

The proportion of native species that became locally extinct in the study period varied substantially among the 11 cities, ranging from less than 1% in San Diego to nearly 28% in Singapore (Table 1). Expressing these as a rate (percentage of species going extinct per year) revealed two distinct groups: six cities (Auckland, Chicago, Melbourne, New York, Singapore and Worcester) had extinction rates that exceeded 0.08% species per year, while five cities had substantially lower extinction rates (all less than 0.05% species per year; Adelaide, Hong Kong, Los Angeles, San Diego and San Francisco).


Figure 2 shows the mean effect sizes for traits from the meta-analysis, with effect sizes of 0.2, 0.4 and 0.7 corresponding to approximately 20%, 50% and 100% increases in the odds of extinction (or persistence).

Figure 2.

Posterior distributions of mean effect sizes, and associated 95% credible intervals, for traits predicting the local extinction of native plant species across 11 cities. For continuous trait variables (height and seed mass) and for habitat, positive values indicate larger values of the trait or that presence in that habitat was associated with a higher probability of extinction. For categorical variables, one category was chosen as a reference class (shown in brackets for each trait), with positive values for other categories indicating a higher probability of extinction relative to the reference class.

Plant height was strongly and consistently associated with extinction probability across the 11 cities, with shorter plants more likely to become extinct. The mean effect size for height (−0.64) implies that a height increase of two standard deviations on the log scale (i.e. shifting from the mean to the upper 95th quantile) would reduce the odds of local extinction by about 80%.

Several other traits had large effects (> 0.3), with 95% confidence intervals for the posterior distributions only slightly overlapping zero – species with lighter seeds and those with affinity for forests or riparian habitats appeared more likely to go extinct across cities. Other traits, including spinescence and affinity with aquatic habitats, had potentially large effects but greater uncertainty.

Along with trait variables, we included plant family as a random effect in the model for each city, allowing us to identify any families consistently associated with extinction. We did this using meta-analysis in the same way as for plant traits, modelling the family coefficients in each city as random effects drawn from a common distribution for each family. Across cities, only one family showed a strong effect: species in the family Orchidaceae appeared consistently more likely to go extinct (mean effect size = 0.26 with 95% credible intervals −0.10 to 0.63).

Having accounted for the differential susceptibility of species to extinction due to trait variation, we can compare extinction probabilities across cities using the model intercept terms. This comparison controls for differences due to trait composition: some urban areas may have high extinction rates because they contain a high proportion of species with traits that predispose them to extinction. To account for this, Figure 3 shows the estimated probability of extinction for a species with trait values fixed at the reference class for each categorical trait, and having mean height and seed mass. The distinction between the two groups of cities with high and low extinction rates is again clear.

Figure 3.

Estimated probability of local extinction (and associated 95% credible intervals) for a plant species with traits fixed at the reference class for each categorical trait (see Fig. 2), and having mean height and seed mass, for each of the 11 cities.

Model fit and variable importance

The general lack of traits consistently associated with extinction across cities may have arisen because the traits we chose were not those strongly selected for or against in urban environments, an outcome that might be reflected in poor predictive performance of the models. However, AUC values were all > 0.69, with seven models ≥ 0.8 (Table 2), which suggests acceptable to excellent model fits (Hosmer & Lemeshow, 2000). This implies that at least some traits had good predictive ability. The high AUC values for the three cities with few extinctions (Los Angeles, San Diego and San Francisco) most likely resulted from overfitting: these models included a large number of parameters relative to the number of species that became extinct. Regression coefficients for these models had a correspondingly high degree of uncertainty and were down-weighted in the meta-analysis (see Methods).

Table 2.  AUC (area under the receiver operating characteristic curve) values for models predicting local native plant species extinction using trait data, for each of 11 cities.
Hong Kong0.90
Los Angeles0.90
New York0.70
San Diego0.99
San Francisco0.86
Worcester, MA0.72

Given that trait models for each city have reasonable predictive ability, the lack of consistent trait effects is more likely to be due to inconsistent selection of traits that increase the probability of extinction across cities. This suggests that different sets of traits are associated with extinction and persistence in different urban areas, and is supported by the considerable variation in the probability that a given trait would be included in the most probable model for each city (Table 3). Habitat, for example, was highly likely to be included in the most probable model for Auckland, Melbourne and Singapore, suggesting that in these urban areas habitat affinity was closely associated with extinction outcomes but had a very low probability of inclusion in other urban areas, implying it was of little importance elsewhere.

Table 3.  The probability that a trait was included in the most probable model predicting local native plant species extinction, for each of 11 cities.
TraitAdelaideAucklandChicagoHong KongLos AngelesMelbourneNew YorkSan DiegoSan FranciscoSingaporeWorcester, MA
Growth form0.950.770.270.410.850.330.200.850.601.000.25
Clonal spread0.990.380.
Nutrient uptake0.320.460.290.491.000.310.200.971.000.640.16
Photosynthetic path0.
Seed mass0.210.350.760.340.000.800.

To summarize the variation in trait importance across cities, we constructed a distance matrix of Euclidean distances between cities in terms of the probabilities of trait inclusion from Table 3, excluding habitat, which distinguished only Auckland, Melbourne and Singapore from the rest. We then used non-metric multidimensional scaling (as implemented in the package isoMDS; R Development Core Team, 2004) to display the similarity among cities as two ordination axes (Fig. 4). Cities located close together in ordination space would therefore tend to have similar traits identified as important in determining extinction outcomes.

Figure 4.

Ordination of the trait by city data in Table 3 using non-metric multi-dimensional scaling. Cities located close together in this ordination space tend to have similar traits identified as important in determining extinction outcomes relative to cities located further apart. The ordination stress value was 8.9.

The distribution of cities in ordination space was related to their extinction rates (Table 1, Fig. 3); the six cities with high rates of extinction (Chicago, Worcester, Auckland, Singapore, New York and Melbourne) had low axis 1 and high axis 2 scores, and occupied the lower right of the ordination diagram (Fig. 4). Correlations of ordination scores with the probabilities of trait inclusion revealed that cities with low axis 1 and high axis 2 scores tended to include height and seed mass in the most probable model, with low importance on most other traits, while the remaining cities showed the opposite (Table 4).

Table 4.  Pearson correlations (r) between the probability a trait was included in the most probable model for a city (from Table 3) and the axis 1 and axis 2 ordination scores for each city (from Fig. 4). A strong positive (or negative) correlation between a trait and an ordination axis score indicates that cities having high values on that ordination axis are more (or less) likely to include that trait in the most probable extinction model.
TraitAxis 1Axis 2
Growth form0.820.24
Clonal spread0.720.16
Nutrient uptake0.900.09
Photosynthetic path0.870.34
Seed mass−0.420.23

Given these groupings, we repeated the meta-analysis, analysing cities with a high extinction rate (Chicago, Worcester, Auckland, Singapore, New York and Melbourne) separately from those with a low extinction rate (Adelaide, Hong Kong, San Francisco, Los Angeles and San Diego). Consistent with the ordination axis correlations, both seed mass and plant height were strongly associated with extinction in cities with high extinction rates, with small-seeded, short plants being consistently more likely to become extinct (Fig. S1), and the remaining trait effects similar to those in the full meta-analysis (Fig. 2). In contrast, cities with low extinction rates showed no consistent trait effects, with many having high uncertainty and all with 95% credible intervals that substantially overlapped zero (Fig. S2).


Cities in our dataset fell into two groups: those with low rates of plant extinction for which no traits consistently predicted extinction outcomes, and those with higher rates of extinction for which two traits, height and seed mass, were consistent predictors. Cities with high and low rates of extinction appeared to differ in their history of modification prior to urban development. With the exception of Chicago and Melbourne, which were originally a mosaic of woodlands and grasslands, cities with high rates of extinction were naturally regions of closed forest that was cleared, usually for agriculture, prior to urban development. In contrast, cities with lower rates of extinction were in areas not subject to intensive agricultural conversion prior to urbanization (Adelaide, Los Angeles, San Diego and San Francisco), and thus which may have retained more natural remnants within developing urban areas, or had a study region that encompassed a large area of remnant vegetation not yet affected by urbanization (Hong Kong). Such patterns are consistent across a larger sample of cities, where the history of development accounts for much of the variation in rates of plant extinction among cities (Hahs et al., 2009).

Among cities with higher rates of extinction, urbanization resulted in the loss of shorter plants with lighter seeds, implying that urban environments favour taller, larger-seeded species. This result is consistent with other studies showing that shorter plants have higher extinction rates or are less likely to occur in cities, including New York (DeCandido, 2004), southern England (Preston, 2000) and Sheffield and Birmingham (Thompson & McCarthy, 2008). This may be surprising given that cities are generally regarded as disturbed environments that should favour species with a ruderal life history, which are often shorter plants. However, areas of taller vegetation may be more likely to be preserved as remnant native patches in urban settings because they are attractive to humans, while other habitats may be more readily exploited (grasslands) or favoured for conversion (wetlands). Preston (2000) found that the species more likely to go extinct with urban expansion were short plants of open habitats. Indeed, while habitat destruction is a feature of urbanization, remnant native patches may experience a lower frequency of disturbance in cities than in surrounding agricultural landscapes, and in the absence of repeated disturbance, taller herbaceous species may have a competitive advantage over shorter species (Leach & Givnish, 1996). In addition, taller plants are likely to live longer, meaning that even if populations become non-viable at the same rate, those with taller, longer-lived individuals will persist for longer in the landscape (Duncan & Young, 2000). This would imply that some tall, longer-lived species may be persisting as relict populations – an urban extinction debt that is yet to be realized (Hahs et al., 2009).

The tendency to lose species in the family Orchidaceae was the only other consistent pattern in the data. Orchids tend to be small-seeded, short plants with the additional feature that they have often highly specialized pollination systems. The tendency for these species to be lost from urban areas having accounted for plant size and seed mass hints that species with highly specialized mutualisms may be at greater risk of extinction.

Apart from height and seed mass, we found little evidence of a consistent shift in the composition of urban floras towards species with a predictable subset of traits. This might appear surprising given that urban areas share many environmental features and that urbanization should impose a consistent set of filters on the vegetation (Grimm et al., 2008; Williams et al., 2009). Nevertheless, there may be several reasons for this disparity. First, additional traits that have proved to be related to urban–rural gradients in other studies (e.g. Ellenberg numbers and specific leaf area; Chocholouskova & Pysek, 2003; Lososova et al., 2006; Knapp et al., 2008; Thompson & McCarthy, 2008) were unavailable for our analysis. We also lacked data for most cities on the initial abundance of species, which was the strongest determinant of extinction probability in Auckland (Duncan & Young, 2000) and fragmented Victorian woodlands (Sutton & Morgan, 2009), with initially small populations being much more likely to go extinct. This factor clearly has the potential to confound outcomes.

Second, habitat destruction must be a key driver of extinction in urban areas and may cause plant extinctions irrespective of species traits (Williams et al., 2009). The lack of consistent trait selectivity in urban environments may arise because extinction is more about being in the wrong place than possessing a particular suite of traits. In such circumstances, being initially abundant and widely distributed is likely to provide some buffering against extinction. Nevertheless, exceptions could arise when particular habitats are preferentially destroyed (Seabloom et al., 2002) and those habitats have selected for species with a particular suite of traits. The clearance of large areas of mangroves in Singapore, for example, led to the extinction of many species of epiphytic orchids that specialized in those habitats (R.T.C., unpublished data).

Finally, the data from urban areas with relatively low extinction rates in our study will have less statistical power to identify attributes that reliably distinguish extinct from extant species. Indeed, the regression coefficients in models for urban areas with the lowest extinction rates had large uncertainties, and no traits showed strong and consistent effects. In urban areas where relatively few species have yet become extinct, chance and idiosyncratic factors may obscure any underlying patterns in trait selectivity.

Yet there are reasons not to be gloomy about these results. Other global analyses of trait responses to disturbances have highlighted the importance of context: plant trait responses to grazing depend upon the evolutionary history of grazing and site productivity (Díaz et al., 2007); fire responses depend upon the nature of the fire regime (Pausas et al., 2004); resprouting responses depend upon the intensity of biomass removal and site productivity (Vesk & Westoby, 2004). In the case of urbanization, variation in extinction rates across cities is at least partly a function of their history of development (Hahs et al., 2009) but could also vary as a function of features such as regional climate. Given that urban areas converge on similar environments, the response of the biota may depend on the degree to which new urban environments differ from the regional conditions; the flora of arid regions, for example, may be better adapted to persisting in urban environments than those of wet tropical forests. Such context dependence would parallel the findings for relationships between plant traits and rarity, with the traits that predict rarity varying substantially among locations (Murray et al., 2002).

Moreover, the strength of trait signals will depend on the nature of the environmental driver. One might expect drivers with simple, acute effects (e.g. fire or logging) to have stronger trait signals than more complex, interacting drivers (e.g. rangeland grazing, urbanization). Global analysis of plant traits related to grazing responses has resulted in only a few clear patterns, with consistent differences between annual and perennial life histories and between some growth forms (Díaz et al., 2007). The complexity of the environmental changes associated with urbanization might obscure simple trait responses because there are multiple drivers affecting multiple traits, which may not be independent.

Identifying predictable relationships between plant traits and environmental conditions provides a promising framework for understanding how vegetation responds to environmental change in a variety of ecosystems (Westoby & Wright, 2006; Díaz et al., 2007). Despite the fact that urbanization involves a set of filters that we expect to select for a consistent suite of traits (Williams et al., 2009), we find only a weak tendency for the traits of native vegetation in urban areas to converge, at the global scale analysed here. Such a lack of evidence for trait selectivity may be a consequence of relatively low extinction rates in some cities, with the possibility of a yet to be realized extinction debt (Hahs et al., 2009), coupled with the idiosyncratic loss of species due to indiscriminate habitat destruction. Regardless of this, our results suggest that it will be difficult to reliably identify those species most prone to extinction in the face of increasing rates of urban expansion.


This paper is dedicated to our colleague Steven Clemants, who died during its preparation; we owe him a great deal. We thank Catherine Tait for providing the Adelaide data, Monique Hallet for helping to compile the Melbourne and Adelaide trait data, and Andrew Hipp and Myla Aronson for help with the Chicago data. This is a product of the Urbanization and Plant Functional Traits working group of the ARC-NZ Network for Vegetation Function. We acknowledge support of the Australian Research Council, The Commonwealth Environment Research Facility, and hosts of the workshops at which this work was conducted (the Australian Research Centre for Urban Ecology, the School of Botany and the Department of Resource Management and Geography, all at the University of Melbourne). Thanks to Brad Murray, Sandra Lavorel and an anonymous referee for helpful comments.


This work was carried out as part of the Urbanization and Plant Functional Traits working group of the ARC–NZ Network for Vegetation Function. All authors were involved in study design, and in compiling and checking the data. R.P.D was responsible for data analysis. R.P.D, N.S.G.W, P.A.V and M.A.M wrote the first draft of the manuscript, with all authors contributing to subsequent versions.

Editor: Brad Murray