1. Concern over the impact of invasive species has led to the development of risk assessment methodologies to identify potential invaders and prevent future ecological and economic problems. However, developing a risk assessment tool is challenging because of the difficulties of accurately predicting the outcome of species introductions.
2. In this study, we develop a global risk assessment for birds. We integrate two approaches, generalized linear mixed models (GLMM) and hierarchical tree models, to help identify those introductions with the highest risk of establishment success.
3. Past work has shown that the number of individuals released is the main factor influencing establishment success in animals, a conclusion that was supported in our analyses. Establishment success was also higher for species with broader ecological niches and larger brains relative to body size. These features should increase the likelihood of finding an appropriate niche in the region of introduction.
4. The GLMM and tree model predicted the probability of establishment success of birds in Europe and Australia with high accuracy (over 80% of introductions correctly classified). This highlights that establishment risk can be reasonably assessed with information on general habitat use, brain size and the size of the founder population. When compared with an alternative risk assessment tool based on a qualitative ranking, our quantitative approaches achieved higher accuracy with less information.
5.Synthesis and applications. Quantitative risk assessments based on traits related to establishment success are difficult but feasible, providing a useful tool for guiding preventive polices aimed at mitigating the impact of invasive species.
Concern over the impact of invasive species on biodiversity loss and homogenization has spurred governments worldwide to increase investment in prevention and control measures (Leung et al. 2005). Once a non-indigenous species is established in a new region it is extremely difficult and costly to eradicate or control (Bergman, Chandler & Locklear 2000; Fagerstone 2003). Recent attention has therefore switched to developing systems for preventing the establishment or the spread of the invasive species (Hulme et al. 2008). One possible strategy is to develop risk assessment protocols that identify high-risk situations for species that may become established and have a negative impact in a new region. Risk assessments for alien species seek to simplify the decision-making process by identifying those characteristics that are linked to invasion success and using them to make predictions about the outcome of future introductions (Kolar & Lodge 2002). An effective risk assessment method would be highly valuable in guiding and designing environmental policies; for instance, it can inform decisions about the planned introduction of alien species, and help establish priorities in eradication and control programmes. The goal of our study is to develop a global risk assessment for birds, a group of organisms that has become a model for the study of invasion biology (Duncan, Blackburn & Sol 2003).
Recent attempts to develop risk assessment protocols have progressed in two distinct directions. The first uses a ranking approach that converts responses to questions related to species invasiveness and expected impacts to a score whose value determines the overall risk posed by the invader (hereafter, ranking systems). These are non-statistical, qualitative methods based on expert knowledge, and they require a large amount of information in order to answer the questions. Moreover, the nature of the questions and their weight on the final score are arbitrary and not always based on rigorous scientific knowledge. The second approach uses advanced statistical techniques to predict the probability that a species can establish itself or spread based on their intrinsic features (hereafter, quantitative systems). The work reported here utilizes the statistical approach, which a priori should be more powerful than qualitative methods.
We built two types of quantitative risk assessment, one based on generalized linear mixed models (GLMMs) and the other on hierachical decision trees. The GLMM approach is an extension of Generalized Linear Models that has been used previously to identify the predictors of establishment success. It allows problems associated with phylogenetic and spatial non-independence among introductions to be corrected (Blackburn & Duncan 2001a; Sol, Vilà & Kühn 2008). GLMMs can also be useful in making predictions, but to date there have been no attempts to use this methodology as a risk assessment tool for biological invasions. Hierarchical decision trees, on the other hand, use a likelihood approach to split a response variable (e.g. success or failure in an introduction) into increasingly homogenous subsets by binary recursive partitioning on a set of predictor variables (Clark & Pregibon 1992). Our risk assessment was based on a particular type of hierarchical decision trees called regression trees. They allow quantitative estimations of the risks, are able to handle complex interactions between variables and present the results in a graphical format that is easy to understand even for non-experts. A similar approach has been used successfully to predict the invasive potential of freshwater fish (Kolar & Lodge 2002) and this provides the foundation of our second approach.
Developing risk assessment protocols is not an easy task. Because the invasion process is generally seen as highly idiosyncratic, there is a widespread perception that it is impossible to predict which species will invade in any given situation (Ehrlich 1989; Kolar & Lodge 2001). In birds, for example, a large number of species-traits have been proposed to influence establishment success (Newsome & Noble 1986; Blackburn & Duncan 2001a,b; Kolar & Lodge 2001; Cassey 2002a,b; Cassey et al. 2004; Jeschke & Strayer 2006) (see Table 1), but very few of these features have been supported firmly by empirical evidence (Duncan et al. 2003). This could indicate that only a few traits are relevant in determining the outcome of an introduction. If so, we can ask whether these few traits are sufficient to produce reliable risk assessment tools. Alternatively, the lack of success in identifying the features of successful invaders could reflect the low power of the tests used due to insufficient sample sizes or the failure to deal with biases and confounding effects associated with historical introductions (Blackburn & Duncan 2001a;Sol et al. 2008). In this case, the solution is to re-analyse whether these features influence establishment success with a larger sample of introductions and using approaches that allow controlling for possible biases and confounding effects. To develop our quantitative risk assessments, we employ a global data base documenting the outcome of 832 introductions of 202 avian species to new locations, 311 of which were successful. With this information, we test the predictive power of 17 traits that have been either demonstrated or suggested to be associated with establishment success (Table 1), using both the GLMM and tree regression approaches. We validate our risk assessment models with the subset of species introduced to Europe and Australia. Finally, we compare the performance of both methods with a ranking system proposed for vertebrates in Australia (Bomford 2003).
Table 1. Hypotheses proposed relating species-level traits with establishment success (adapted from Sol 2007)
*Information on relative brain size was available for 157 of the 202 introduced species (77·7%) that were considered; for the remaining species, brain residuals were estimated by using the average brain residual of the genus, which predicts 91% of the variance at the species level (Sol et al. 2002).
Generalist species should be better invaders than specialists because they are more likely to find appropriate resources in a new environment (Mayr 1965; Brown 1989)
Sedentary Nomadic Local movements Partial migrant Total migrant
Sexually selected species may be more vulnerable to extinction risks and therefore have lower introduction success when compared with non-sexually selected species (Sorci et al. 1998; McLain et al. 1999)
Species with life histories that increase population survival are expected to have a higher chance of invasion success because these species may attain a large population size faster (Moulton & Pimm 1986)
The information on avian introductions comes from a compiled global data base (Sol et al. 2005), reported mainly in Long (1981) and Lever (1987), which we updated with information from Lever (2005). An introduction event was defined as the human-driven accidental or deliberate release of a species to a new location, outside its area of natural distribution. All introductions of a given species to the same place within a period of 10 years were considered as a sole introduction event (Cassey et al. 2004). From this global data base, we restricted the analyses to introduction events that included information on propagule size, as this is a major determinant of introduction success in birds (Lockwood, Cassey & Blackburn 2005). Data on introduction effort were available for 832 introduction events (i.e. 25% of all introduction events reported in the literature), comprising 202 species from 36 families, 311 of which were successful (see Table S1, Supporting Information). The outcome of each introduction was scored as ‘established’ if the species succeeded in establishing itself in the new region, or ‘extinct’ if it did not. A species was considered established when it had developed a self-sustaining population, defined as a population that persisted without declining for at least 20 years after release. Consequently, introductions carried out after 1987 were excluded from the analysis. We also excluded recorded introductions with unknown or uncertain outcome, failures associated with human eradication and natural colonizations.
Explanatory and confounding variables
Those characteristics hypothesized to influence establishment can be grouped into three categories (Duncan et al. 2003; Sol 2007): (i) traits that pre-adapt species to the new environment, (ii) traits that favour population increase from a low level and (iii) traits that constrain establishment success (Table 1). We extracted data from the literature (see Appendix S1, Supporting Information) for all major traits that have been hypothesized to influence establishment success in birds (Table 1).
In addition, we quantified several factors that could inflate or obscure the predicted link between species-level traits and establishment success: (i) propagule size, measured as the minimum number of individuals released (log-transformed) (Cassey et al. 2004); (ii) locations where species were introduced, assigned to the six biomes (Australasia, Ethiopian, Nearctic, Neotropical, Oriental and Palaearctic) (Blackburn & Duncan 2001a; Cox 2001; Cassey 2003); (iii) latitudinal difference, measured as the difference between the latitudinal mid-point of the native range of the species and the latitude of introduction calculated without reference to the hemisphere (Blackburn & Duncan 2001a); (iv) island/mainland location, whether the introduction was to a mainland or island location, and in that case which type of island (continental or oceanic) (Cassey 2003); and (v) intra-/inter-regional location, whether the introduction site and the native range were in the same biome, noted as ‘intra-regional’, or in different biomes, noted as ‘inter-regional’(Blackburn & Duncan 2001a).
Following previous studies (Blackburn & Duncan 2001a; Cassey et al. 2004; Sol et al. 2005), we used GLMM to identify the traits related to establishment success. GLMMs allow taxonomic and regional variables to be specified as random factors in the model, effectively dealing with the fact that the species introduced and the regions of introduction are a non-random sample of all species and regions available (Blackburn & Duncan 2001b). The inclusion of taxonomy and region as random effects in the model also deals with pseudoreplication that arises from introductions of the same species in different regions or different species in a same region. Because the taxonomy was incorporated in the model with a hierarchical structure (i.e. species nested within genera, genera nested within families and families nested within orders), the model also helps correct for the phylogenetic effects described between these taxonomic categories. The models we used assume a common positive correlation between introduction outcomes within the same taxa or region, but a zero correlation between outcomes involving different taxa or regions.
The GLMMs were fit using the GLIMMIX procedure (Littell et al. 1996) in SAS (SAS Institute Inc. 2002–03). The outcome of each introduction (success or failure) was modelled by specifying a binomial error distribution and a logit link function (Blackburn & Duncan 2001b; Sol et al. 2005), with the explanatory variables coded as fixed effects and region, species and higher taxonomic levels coded as random effects. The minimum adequate model (MAM) was obtained by removing from the full model those fixed effect variables that did not lead to an improvement. At each step, the less significant variable was removed until the model retained only significant predictors. We investigated the significance of alternative models by adding the previous variable removed from the model. The MAM was the one that only retained variables with a significant effect on establishment success. Standardized coefficients of the fixed effects (Kramer 2005) are reported to help assess the relative importance of each explanatory variable.
The MAM was used to produce predictive risk assessment models with GLMMs with the OUTPUT statement of the GLIMMIX procedure. To assess the accuracy of our models in predicting the success of introductions, we used a cross-validation approach, partitioning our sample of introductions into subsets such that the analysis was initially performed on a single subset that generates the predictions, while the other subsets were used in confirming the model (observed values). We ran a series of 50 10-fold cross-validations to assess the accuracy of our GLMM risk assessment. The overall predictive accuracy of the model was evaluated comparing the predicted and observed values. As the GLMM model delivered specific predicted probabilities for each introduction event, we estimated the predicted values as follows. First, we classified each predicted value in 10 probability categories (0–10%, 10–20%, 20–30% and so on). We then calculated the expected outcomes for each category as the product of the total observed outcomes and the respective probability of establishment. Finally, we compared the predicted values with the observed outcomes of the introduction events to obtain the overall predictive accuracy.
Tree regressions were performed with the statistical package r 2.4.1 (R Development Core Team 2005). The library MVPART (MultiVariate PARTitioning), developed by Therneau et al. (2007), was used to perform the regression tree. Trees are modelling techniques that aim to explain variation of a single response variable by one or more explanatory variables. The tree is constructed by repeatedly splitting the data so that at any node, the split which maximally distinguishes the response variable in the left and the right branches is selected. The splitting is then applied to each group separately and the procedure continues until nodes are homogenous or the data are too sparse (Crawley 2002).
To determine the robustness and the number of variables to be included in the model, we ran a series of 50 10-fold cross-validations and selected the most frequently occurring tree size using the 1-SE rule (De’ath & Fabricius 2000). Agglomeration of nodes was considered appropriate if some node did not have biological meaning and the change did not decrease the power of the model.
To demonstrate the practical utility of the GLMM and tree regression approaches, we developed separate models for Europe and Australia based on information from all the introduction events of our data base but excluding those introductions that took place in their respective continent. The overall predictive accuracy of these models was then evaluated using the observed introduction outcomes of the subset of introductions carried out in the particular continent compared with those predicted by the models.
We also compared the performances of our quantitative models in predicting the outcome of introductions to Europe and Australia with the performance of an alternative ranking system model, the risk assessment developed by Bomford (2003) for Australian vertebrates. Bomford’s risk assessment also includes questions about the impact of invaders, not considered in our study; consequently, to facilitate the comparison among approaches, we only used those questions that were related to establishment success (see Appendix S2). As our models classify introduction events according to the probability of establishment success while the Bomford’s system classifies the introduced species in categories of risk, we had to adapt our data in order to make the models comparable. From our global data base we selected those introduction events that had been carried out to Europe or Australia. For the evaluation we compiled all same species introduction events in a single entry and, as outcome, we recorded the final result of all their introductions to the region. We then used the information already collected in our data base to answer the questionnaire according to Bomford’s system criteria. We used the information on climate matching published in Bomford (2003), but for Europe that information was not available. Following Bomford’s criteria, when the required data were unavailable, the maximum score was given to that question. We then calculated the establishment risk score and used it to assign the species to a risk category of establishing a wild population (Bomford 2003). After classifying the species we compared the predictions with the known outcome of the introductions to both regions. To be able to calculate the overall predictive accuracy of this ranking system, we assigned an establishment probability to each risk category as follows: low risk, 0·25; moderate risk, 0·50; high risk, 0·75; extreme risk, 1·00.
As in previous analyses (Blackburn & Duncan 2001a; Sol, Timmermans & Lefebvre 2002), the GLMM approach revealed that most variation in establishment success is found at the species level (for the complete set of random effect solutions, see random effects in Table S2). This suggests low phylogenetic autocorrelation in establishment success. The MAM retained four variables: habitat generalism, propagule size (log-transformed), relative brain size and island/mainland introduction location (Table 2). Establishment success was higher for species that occupy a higher diversity of habitats in their native ranges, have bigger brains in relation to their body size, are introduced in higher numbers and are released onto islands. The standardized coefficients indicated that propagule size has the most influence on the probability of establishment, followed by relative brain size and habitat generalism, while island/mainland introduction location is the least important factor. This model showed high consistency: a 10-fold cross-validation estimated 91·50% of overall predictive accuracy.
Table 2. Fixed and random effects in a minimum adequate generalized linear mixed model explaining variation in bird establishment success while controlling for biome of introduction and taxonomic levels
Type III, F
−14·544 to −6·6150
Relative brain size
Families within orders
Genus within families
Species within genus
Biome of introduction
Regression tree approach
The resulting regression tree had four terminal nodes, and only included three variables: habitat generalism, relative brain size and propagule size (Fig. 1). Big-brained species with generalist habits had higher probabilities of establishment (predicted probability of establishment: 0·63) than specialist species (0·235). Within the generalist species with smaller brains, the probability of establishment was associated with propagule pressure: species released in larger numbers had a higher probability of success (0·595) than those released in smaller numbers (0·276). As in the GLMM model, the consistency of the regression tree model was high: a 10-fold cross-validation estimated 91·50% of overall predictive accuracy.
Comparing risk assessment approaches
Do our risk assessment tools accurately predict the establishment of birds introduced into Europe and Australia? To answer this question, we developed separate models for Europe and Australia based on information from all the introduction events of our data base but excluding those introductions that took place in their respective continent. For Europe, we used the models to classify the 62 introduction events belonging to 29 species. The tree model classified the introductions with 80·65% accuracy: only 12 of 62 introductions were misclassified (seven false negatives and five false positives). The GLMM predictive model classified the same subset of introductions with 74·19% accuracy, only failing to classify correctly 16 events (seven false positives and nine false negatives). For Australia, we analysed 80 introductions events belonging to 36 species. The tree model classified the introduction with 100% accuracy, while the GLMM applied to the same subset predicted the outcome of these introductions with 87·5% accuracy, only yielding four false negatives and six false positives.
Finally, we asked whether our quantitative risk assessments have a higher predictive power than Bomford’s (2003) ranking system, when applied to Europe and Australia. The ranking system correctly classified 87·07% of introductions carried out in Europe, with only four of twenty-nine species misclassified. However, the ranking system applied to the Australian species subset correctly classified the introduced birds with only 67·93% of accuracy: 28 of 29 species were classified by the model as ‘Extreme’ or ‘High’ risk, while in fact 12 of these high-risk potential species did not become established in Australia.
A common belief among students of invasion biology is that risk assessment tools are of little use given the idiosyncratic nature of the invasion process, which makes it difficult to accurately predict the outcome of introductions (Ehrlich 1989; Kolar & Lodge 2001). At first sight, our results appear to confirm this perception. Despite analysing more than 800 introduction events, the largest data set ever used in any recent analysis of bird introductions, and many traits reported as affecting establishment success, we found that only four of the twenty-one variables considered were significantly associated with establishment success. Of these variables, only two were species-level traits. These were habitat generalism and relative brain size, traits that previous analyses have related to the ability of the introduced species to find an appropriate ecological niche in the region of introduction (Cassey et al. 2004; Sol et al. 2005). The other two variables significantly associated with establishment success were propagule pressure, an event-level factor that has been previously considered as the most important predictor of establishment success (Green 1997; Forsyth & Duncan 2001; Cassey et al. 2004; Lockwood et al. 2005), and whether the introduction was on an island or on the mainland. Islands are species-poor and are traditionally perceived as easy to invade (Williamson 1996; Sax & Brown 2000).
The limited success in identifying traits that pre-adapt species to novel environments is not surprising because we should expect that adaptations that have arisen under certain environmental circumstances do not generally function so well when the circumstances change (see Sol 2007). More surprising is the lack of success in identifying traits that constraint establishment or life-history traits that influence population dynamics (see Duncan et al. 2003; Sol 2007). If the vast majority of species’ traits are not useful as a tool to predict the outcome of bird introductions, can we still assess the risk of establishment with enough accuracy to be used for guiding conservation policies? Previous work suggests that it is possible to build reliable risk assessment protocols with only a few variables. The regression tree-based risk assessment developed by Kolar & Lodge (2001) used four predictors (minimum temperature threshold, diet breadth and two measures of relative growth) to predict establishment success for alien fishes introduced in US Great Lakes with 94% accuracy. Likewise, the regression tree developed by Caley & Kuhnert (2006) based on the Australian Weed Risk Assessment (Pheloung, Williams & Halloy 1999) used only four questions, all them surrogates of introduction effort, to classify the species according to the probability of being a weed; this predictive model correctly classified the species with a 93·6% accuracy. Our results confirm that risk assessment tools able to predict the outcome of new introductions with high accuracy can be produced with just a few predictors. The risk assessments based on GLMM and tree regression approaches only used four and three variables, respectively, but correctly classified the outcome of a high fraction (more than 80%) of birds introduced into Europe and Australia.
A main advantage of tree models over other approaches to invasive species risk assessment is that it delivers the results in a graph that is easy to understand even for non-experts. In addition, the regression tree organizes the predictors in a hierarchical way, giving different weight according to their power to classify introductions, and is able to deal with missing information for the factors through the use of surrogate variables. Tree models also have some limitations, however. For example, the approach assumes that introduction events are statistically independent, but this is unlikely to be true (Duncan et al. 2003; Sol et al. 2008a). Introduction outcomes are likely to be correlated because the same species were introduced to many locations, and because most locations were subject to several introductions (Blackburn & Duncan 2001a,b). Moreover, there is evidence that the identity of the species introduced by humans is non-random. In birds, most of the species chosen for introduction come from temperate regions, and hence it is expected that traits characteristic of the taxa in these regions are over-represented (Duncan et al. 2003). Without taking into account these issues, it is difficult to draw firm conclusions about the species’ traits that facilitate or limit establishment success. The GLMM approach overcomes these problems by implementing as random effects variables that code for the clustering of introduction events within species, higher taxa and biogeographical region of introduction (Blackburn & Duncan 2001a,b). Thus, the GLMM approach is more robust in identifying the factors that actually affect the success of introductions. Despite the differences between the tree model and GLMM approaches, however both produced risk assessments with similar accuracy levels. When comparing the pool of misclassified species, we found few differences in the misclassified introductions between the two approaches. We suggest that the best strategy when developing risk assessment tools would be to exploit the respective strengths of both methodologies and combine them to identify potentially risky species with higher exactitude.
Risk assessments based on tree models and GLMMs can be criticized on the grounds that they assign equal costs to false positives and false negatives (Caley & Kuhnert 2006). This issue is particularly relevant because the risks of importing a pest are higher than the risks of excluding an innocuous species (Smith, Lonsdale & Fortune 1999). In the present study unsuccessful introductions were predicted by the tree model better than successful introductions. The implication is that at least for some species (i.e. those characterized by being habitat specialists and small-brained), we have some certainty that they are unlikely to become established when introduced. These species should be the ones for which international trade can be allowed, while the species for which the predictions are less reliable merit a cautionary approach, i.e. prohibiting any importation and release of these species, and, where a release has already taken place, carrying out monitoring and eradication programmes to avoid increases in numbers and the spread of the invader.
A comparison between our quantitative risk assessments and the alternative semi-quantitative ranking system of Bomford (2003), found that quantitative models performed better when all the required information was available to answer the ranking system’s questionnaire (the Australian case), while the ranking system was better for the European subset when one question was overscored due to lack of key data. Semi-qualitative ranking systems are more arbitrary than quantitative approaches, but also require more information and are sensitive to a lack of key information. This is because the ranking system handles missing data by scoring them with the highest value of the ranking. This decision is justified from a conservation point of view, yet the consequence is that the method overestimates the total risk of the species. Thus, an advantage of quantitative risk assessments is that they deal with missing data through the use of surrogate factors when the information for a determinate split is unknown. Despite the advantages of quantitative risk assessments, ranking systems can also achieve an acceptable level of accuracy and can still be a useful tool for some organisms for which quantitative predictive models are unreliable.
Much of the focus of current research in invasion biology is on the development of screening methodologies to limit the introduction of noxious species. Recent studies have shown that the implementation of risk assessments with 80% accuracy, such as the one presented here, is economically beneficial for planning medium-term policies (Keller, Lodge & Finnoff 2007). In fact, recently developed risk assessments for exotic species importation are, at present, producing net economic benefits in those regions where they are implemented (World Trade Organization 2005). The use of risk assessments as preventive tools to guide future environmental policies provides a means to reduce the damage caused by invasive species thereby contributing to the conservation of biodiversity and ecosystem function.
We thank Montse Vilà, Sven Bacher, Richard Duncan, Dan Blumstein, Mark Cadotte, Esteban Fernandez-Juricic, Gillian Kerby, Francisco Lloret, Josep Piñol, Bernat Claramunt and two anonymous referees for statistical advice and/or useful comments on early versions of the paper. M.V. was supported by a scholarship of the General Direction of Research of the Department of Universities, Innovation and Enterprise of the Generalitat de Catalunya and the European Social Fund (2005SGR00090). The project was funded by a ‘Proyecto de Investigación’ (ref. CGL2007-66257/BOS) to D.S., a project Consolider (CSD 2008-00040) and European 6th integrated project ALARM grant (GOCE-CT-2003-506675).