Contrasting determinants for the introduction and establishment success of exotic birds in Taiwan using decision trees models

Species of the data set

The four stages of the invasion process were defined in Duncan, Blackburn & Sol (2003) as transport, introduction, establishment, and spread. In this study, we focused on the introduction and establishment stages. For a species to reach the introduction stage, it must have passed the transport stage. Therefore, we selected all the exotic species which had been transported to Taiwan’s main island (not including surrounding islands, such as Lanyu and Kinmen Island) as documented in Shieh et al. (2006) which included the results of Chi (1995), Severinghaus (1999) and Lin (2004). Whether a transported species has passed the subsequent stages of the invasion process was based (1) on escaping records in the field (introduction success) and (2) breeding record in the field (establishment success). We followed the detailed methods of how to define introduction success and establishment success which were given in Su, Cassey & Blackburn (2016). However, Su, Cassey & Blackburn (2016) based their decision of establishment success on the respective species having been recorded to be breeding at least twice; instead, we based it on at least one record of fledglings actually having left the nest successfully.

In order to record all the escaping and breeding records of bird species up to 2015, we continuously (1) checked information from the Chinese Wild Bird Federation (http://www.bird.org.tw/) database which is the main collector of wild bird data in Taiwan, as well as other Taiwanese websites dedicated to natural history observations of birds, (2) remained in contact with local ornithologists, birdwatchers and bird societies, and (3) included any relevant publications (e.g., Walther, 2011; Walther, 2014 for red-whiskered bulbul, Pycnonotus jocosus; Fan et al., 2009 for white-rumped shama, Copsychus malabaricus, or Shieh, Lin & Liang, 2016 for Asian glossy starling, Aplonis panayensis). Most of this updated information was published recently in a project report for the Taiwan Forestry Bureau (Liang & Shieh, 2016).

Despite following the methods as described in Su, Cassey & Blackburn (2016), we independently collected all the data used in this analysis beginning in 2004 and ending in 2015. Our dataset thus contains 283 full species (although we entered subspecies in our dataset, for this analysis, we only used full species), which were transported to Taiwan (see above). Of these 283 species, 95 species escaped to the field successfully (introduction success). Of these 95 species, 36 species reproduced in the field of Taiwan successfully (establishment success) (see Table S1 for species list).

Variables

We collected 22 variables for each species, including two nominal ones (order and family), six ordinal ones (latitude overlap with Taiwan: 0–2, migration pattern: 0–3, nesting location: 0–3, feeding: 1–3, diet: 1–6, and habitat: 0–6), three binary ones (hole nest, Taiwan genus_resident, dichromatism), and 11 interval ones (clutch size: clutch, maximum clutch size: Mclutch, incubation days: incubation, minimum incubation days: Minincub, body length: length, maximum body length: Mlength, body mass: Mass, maximum body mass: Mmass, the number of invaded countries: Invcountry_Max, distribution range (km²): Range, the number of subspecies: subspecies) (see supplementary file Table S2 for code descriptions of variables). The variable Taiwan genus_resident was based on the information in Hsiao & Li (2014). For the other variables, we gathered species information from the books of Del Hoyo et al. (1992–2011), Dunning Jr (1993), and internet databases of IUCN (http://www.iucn.org) and BirdLife International Datazone (http://datazone.birdlife.org) (see Table S1 for associated information of each species and Table S2 for code descriptions of variables). When we collected the values for reproduction and body size for each species, we usually found a given range instead of fixed values in the references. In order to account for the maximum adaptation and reproduction potential in the invasion process, we used maximum values such as maximum body mass or minimum values such as minimum incubation days in addition to averaged values. To determine the number of invaded countries (Invcountry_Max), we counted the total (or maximum) number of countries in which occurrences of introduced populations of each respective species were reported.

Decision trees models and variable treatments

To investigate the possible effects of nominal variables (family and order) and ordinal variables on the performance of the decision tree models, we conducted three variable treatments for modeling: (I) including all variables, (II) excluding nominal variables, and (III) excluding nominal variables and replacing ordinal values with binary ones; e.g., changing habitat values of 0–4 to 0 (natural habitats) and habitat values of 5–6 to 1 (artificial habitats).

For each variable treatment, we used five decision tree models (DT_no bagging, DT_bagging 90%, DT_bagging 100%, gradient boosting, and HP forest) to predict the outcomes of introduction and establishment, respectively. Modeling processes and comparisons of model performance were implemented using SAS Enterprise Miner 13.1 (for diagrams of process flow, see supplementary files Figs. S1 & S2). Because of the small data set, no data partition was implemented; that is, all data were used as training data. Instead, other methods, such as bagging and cross validation, which have been suggested for the use with small data sets (SAS Institute Inc., 2013), were used in the present study.

DT_no bagging is the traditional classification tree method by constructing a layered tree model with the following settings: splitting rule = Gini, cross validation with 10 subsets and 100 repeats. The DT_bagging 90% and DT_bagging 100% methods used the same setting of splitting rule and cross validation as the DT_no bagging method but with bagging 90% or 100% of the data set for 50 times, respectively. Gradient boosting is a boosting method that resamples the data set to produce a series of decision trees which together form a single predictive model which has been found to be less prone to overfitting the data than a single decision tree (Georges, 2008). HP Forest is the random forest method which builds many parallel trees forming a forest; a tree in the forest is a sample without replacement from all the available observations, and the input variables that are considered for splitting a node are randomly selected from all the available inputs (Hall et al., 2014).

We calculated five performance measures to compare models, namely, the area under the receiver operating characteristic curve (AUROC), the specificity which measures the fraction of negative events that were correctly labeled, the precision which measures the fraction of positively labeled outcomes that were correctly labeled, the recall which measures the fraction of positive events that were correctly labeled, and the accuracy which measures the fraction of all events that were correctly labeled (accuracy =1 − misclassification rate) (Söhngen, Chang & Schomburg, 2011). These five performance measures have the same range (0–1), and we gave each measure equal weight in evaluating the model performance in accordance with Chen, Peng & Yang (2015). The higher the values of these five performance measures are, the better the model performs; therefore, we summed up the five values (from hereupon called the “total score”) and chose the model with the highest sum as our final optimal model. We then compared the relative importance of each of the variables in the optimal introduction model and establishment model.

For illustrative purposes, we chose the visual output of the resulting trees of DT_no bagging of variable treatment I for our figures (Figs. 1 and 2). Such visual outputs are not possible for the other four methods (namely, DT_bagging 90%, DT_ bagging 100%, gradient boosting, and HP forest).

We used the decision tree models described above to build various versions of two kinds of models: (1) introduction success prediction models and (2) establishment success prediction models. However, for brevity’s sake, from hereupon we will call them introduction models and establishment models, respectively.

Model comparisons and variable treatment comparisons

Our results showed that for the complete data set of 283 transported species or for the data set of 95 introduced species, the gradient boosting method performed better than the other four decision tree methods. While we calculated five performance measures, the only other study which used the decision tree method on a bird data set was Keller, Kocev & Džeroski (2011) who calculated only AUROC and accuracy values. Considering AUROC values first, the AUROC values of gradient boosting of our study were over 0.919 in the introduction models and over 0.971 in the establishment models; thus, they were all higher than our values for the random forests method. This is in contrast to the results of Keller, Kocev & Džeroski (2011) who found that, based on the AUROC values, random forests performed better than gradient boosting for both their New Zealand and Australia bird data sets. Specifically, AUROC values for gradient boosting for their New Zealand (79 species with 11 traits) and Australia (52 species with 11 traits) data sets were 0.682 and 0.681, respectively, whereas AUROC values for random forests were 0.731 and 0.745, respectively. Pearce & Ferrier (2000) suggested that AUROC values between 0.7 and 0.9 indicate a reasonable discrimination ability of models, and values higher than 0.9 indicate a very good discrimination ability of models. The higher AUROC values of our study might have resulted from the inclusion of more variables (up to 22 variables) rather than larger samples used for analysis. In our study, both the introduction model and establishment model used 22 variables, and we found higher AUROC values (0.971–0.985) in the smaller data set (namely, the establishment model with 95 species) than in the larger data set (namely, the introduction model with 283 species) (AUROC values 0.919–0.936). We therefore suggest that even a small data set (less than 100 species) with up to 22 variables can achieve a prediction model of good performance using the gradient boosting method.

Comparing the performances of variable treatment I with variable treatments II and III, we found little difference on model performance. Treatment II excluded nominal variables, and treatment III changed ordinal variables of species traits into binary variables, but neither one of these changes really had much discernable influence on overall performance. Our results therefore provide evidence to support the use of ordinal variables of species traits, and that there is no need to convert ordinal variables of species traits to binary ones for their use in decision tree models.

Predictors of introduction and establishment success in exotic birds

Perhaps the most interesting and novel result of our study is that, for both the establishment model and introduction model, the number of previously invaded countries was the most important or second most important determinant in all the models. Therefore, our study suggests that future success for introduction and establishment of birds can be gauged by simply looking at previous success in invading other countries or regions. Future studies should include this variable to confirm our supposition because it might be a very simple and straightforward way to predict the potential invasion success of a species: if it has been successful before, it will probably be successful again. While this variable could not have been established a few decades ago, we now have a global track record of successful species invasions, and we might therefore be able to use it to better predict future local or regional invasions. Furthermore, global studies could investigate what species traits and other relevant factors, e.g., local ecological factors, are related to the number of successfully invaded countries; or, given the differential size of countries, the actual area invaded.

Another important determinant was family. While family was the most important variable in the optimal introduction model, it dropped to being only the seventh most important variable in the optimal establishment model. In other words, family was an important determinant of introduction but not establishment in Taiwan. Our results thus differ from those of a global study which found that bird family was also a good predictor for establishment success (Lockwood, 1999). The discrepancy between this study and our study could result from the fact that exotic birds in Taiwan are primarily introduced for aesthetic reasons but not for hunting (Shieh et al., 2006; Su, Cassey & Blackburn, 2016), while the global data set included many hunted species.

Several species traits were also chosen as important determinants for the introduction and establishment models. For the optimal introduction model, the top three selected species traits were maximum body mass (Mmass), latitude overlap with Taiwan (Overlap), and distribution range (Range). Among these three variables, maximum body mass was ranked the most important, and it also had a relative importance greater than that of two other closely related measures, specifically, the averaged body mass (Mass) and body length (Length). One possible explanation is that birds are usually heavier in captivity under well fed condition. Our data set contained primarily pet species (and not game species, which are prevalent in many other studies), and the body mass of pet birds might be higher than the average body mass of their wild congeners and therefore closer to the maximum attainable body mass. In order to consider the representability and the maximum adaptation potential in the invasion process, we therefore suggest that including maximum body mass may be important in order not to miss a potentially important determinant for the invasion success of exotic pet birds in particular. For example, Su, Cassey & Blackburn (2016) did not find that body mass had any influence on introduction success. However, they only used averaged body mass, and perhaps their result would have been different if they had also included maximum body mass. Furthermore, Cassey’s (2001) global study found that averaged body mass was significantly correlated with introduction success which further supports the role of some measure of body mass being an important determinant of introduction success.

Finally, several species traits related to reproduction were also important, such as minimum incubation days (Minincub), clutch size (Clutch), dichromatism, and nesting location (Nesting); however, these determinants were more important in establishment success than in introduction success. Furthermore, given that some top ranking variables were associated with maximum or minimum values of species traits, we suggest that in addition to averaged values, reproductive potential represented by minimum and maximum values of species traits should be considered in prediction models of invasion studies.

We conclude that decision tree models are efficient for the analysis of small data sets with mixed types of variables, including nominal, ordinal and interval variables, in predicting the invasion success of exotic birds. Our results further demonstrate that the most important determinants in predicting introduction success of exotic birds in Taiwan were the bird family, the number of invaded countries, and variables associated with environmental adaptation, whereas the most important determinants in predicting establishment success were the number of invaded countries and variables associated with reproduction.