Invasive predators and global biodiversity loss

For all threatened species in the taxonomic classes Aves, Mammalia, and Reptilia, we downloaded data on taxonomy and conservation status from the International Union for Conservation of Nature and Natural Resources (IUCN) Red List in December 2014 (version 2014.3) using the inbuilt search and export functions (n = 3,745 species) (Dataset S2). We did not assess amphibians here because our preliminary research indicated that the invasive predators impacting them are mostly nonmammalian (e.g., snakes, fish, crayfish, and other amphibians). Threatened species were those listed as vulnerable, endangered, critically endangered, extinct, or extinct in the wild. We then used a custom R script (Dataset S3) to download additional Red List information on each species’ range and major threats.

We filtered this database (n = 3,745 species) in Microsoft Access by searching the “major threats” section for any of the following keywords: predator*, predation, cat, cats, fox*, dog, dogs, rat, rats, rodent*, Rattus, mouse, mice, stoat*, mongoose*, pig, pigs, mink, ferret*, weasel*, mustelid*, possum*, macaque*, coati*, and civet*. These predators were chosen based on consultation of the Global Invasive Species Database (43) and Long (16). This search returned 771 records, which we inspected to determine whether invasive alien predators were identified as a known or likely threat to each species (n = 703 species identified as negatively impacted by invasive predators). We cross-checked this list against previous reviews (1, 18, 20, 44–48) and added 35 additional threatened species recorded as being negatively affected by invasive predators, but not revealed in our Red List search. Given the small number of additional species identified and the broad geographic coverage of the previous studies used for cross-checking, we do not consider that this exercise brings any systematic bias to our analyses.

For each of the 738 study species, we recorded information on taxonomic classification (class, order, family), Red List status, insularity (insular endemic or found on continents also), and region (Fig. S3 and Table S2). Information on species distributions was sourced primarily from the Red List although other sources were consulted in a small number of cases. For the analyses, we included in the extinct category four species classed as extinct in the wild.

To find information on the impact of invasive predators on each of the study species, we initially searched the Red List and Scopus database for relevant material using species names and synonyms, followed by consultation of primary and gray literature cited therein. We defined impact as any inference that an invasive predator had caused a decline in the abundance or distribution of a species. In most cases, predation was inferred as the primary mechanism of predator impacts although competition, disease transmission, and habitat disturbance were also cited in some cases. For accounts that referred only to “introduced/invasive predators” and not a specific species, we assigned the impact to a generic predator group. We took any reference to “domestic predators/carnivores/pets” to mean cats (F. catus) and dogs (C. familiaris). We did not distinguish the impacts of individual rodent species because many accounts did not provide sufficient information to allow discrimination of individual species effects and because the relative impacts of the different rodent species have been reviewed elsewhere (18–20, 49, 50).

Given the difficulties in attributing causation in species declines and extinctions, most inferences regarding the impact of invasive predators were based on observational evidence, rather than experimental data. For this reason, we used a similar approach to that of previous studies (1, 19) and coded the degree of predator impacts as follows: mixed (0.25, when the predator was a secondary cause of species decline); high (0.75, when the predator was a primary cause of species decline); and strong (1.0, when the extinction of the species was attributed to the predator). Unlike previous studies (1, 19), however, we did not include a “nil impact” level (e.g., 0.01) because such information is not systematically reported in the literature. Other threats may have contributed to the species’ declines/extinctions although assessing their relative importance was beyond the scope of this study. We assessed species across their entire geographic ranges and thus did not code predator impacts for individual populations (e.g., multiple islands). This exercise was conducted between March and September 2015, and it revealed 1,381 individual predator-threatened species cases, plus an additional 58 cases where the predator species were not named. The 996 references supporting the rankings are listed in Dataset S4.

We first summarized numbers of extinct and threatened species impacted by invasive predators, based on taxonomic classes and geographic regions where they occur, or occurred. We then used metaanalysis in the metafor package version 1.9-6 in R version 3.1.2 (51, 52) to analyze these trends based on three categorical variables: (i) taxonomic class model (levels: Aves, Mammalia, Reptilia); (ii) insularity model [levels: insular endemic, or continental (either wholly or partially)]; and (iii) predator model [levels: rodent (Rodentia), cat, dog, red fox (V. vulpes), stoat (M. erminea), small Indian mongoose (H. auropunctatus), and pig (S. scrofa)].

For the predator model, we excluded 19 predator species that impacted fewer than 15 threatened species each (range 30–430 threatened species impacted by each of the seven remaining predators). We conducted separate tests for each of these variables using the restricted maximum-likelihood estimator. We pooled impacts across all predators for the taxonomic class and insularity models; if a threatened species was impacted by multiple predators, we used the highest impact and its associated weight. For example, if a bird species was impacted by both cats (impact = 0.75, weight = 10) and rodents (impact = 0.25, weight = 100), we used the former pair of values for the pooled category, which means that the models estimate the strongest predator impacts across taxonomic classes and insularity. To examine individual responses of the three taxonomic classes, we conducted separate analyses for birds, mammals, and reptiles across insular endemism and predators. The response variable was the impact rankings described above, such that higher effect sizes represented greater predator impacts. We inferred “significant” effects where the 90% confidence intervals of the different predictor variable levels did not overlap. Data used in the analyses are available as Dataset S1 (see also Table S3).

Metaanalysis traditionally weights effect sizes based on each study’s sample variance and/or size. However, these data do not exist for our database because each case consists of a predator × threatened species combination that is assigned a categorical level of impact. Instead, we used a weighting system similar to that of Jones et al. (19) and Medina et al. (1) that weights individual cases based on the type and strength of evidence provided in each case. Assigned weights were as follows: 1 (lowest: no evidence provided apart from stating that the predator is thought to be a cause of species decline or extinction), 10 (single line of correlative evidence), 100 (multiple lines of correlative evidence), or 1,000 (highest: experimental evidence in a before–after and/or control–impact design). We used the inverse of the weights as the variance component in the metaanalysis. Examples of correlative evidence included artificial nest experiments, correlation between species decline and predator introduction, absence of a species from parts of its historical range now inhabited by predators, monitoring of predation events, and analysis of predator diet. Examples of experimental evidence included monitoring of population parameters in response to predator removal, and comparison of islands with and without predators. The weights were assigned during the impact ranking exercise described above. We conducted a fail-safe analysis to determine the number of cases showing no effect that would be needed to eliminate a significant overall effect size (SI Text). We also conducted a sensitivity analysis to determine how the selection of impact values and the use of weights influenced the results (SI Text and Figs. S4 and S5).

We used χ² analyses to determine (i) whether the proportion of impacted species that were insular endemics varied among taxonomic classes and (ii) whether the proportion of impacted species in each taxonomic class differed among predators. We restricted the second analysis to those seven predators included in the predator model described above. Significant effects were inferred at the 0.05 level.

We used evolutionary distinctiveness (ED) scores to examine whether invasive predators have had a disproportionate impact on evolutionarily distinct species. ED scores were calculated based on the “fair proportion” metric: i.e., the weighted sum of branch lengths along phylogenetic tree roots to tips, with weights based on the number of tips sharing that branch (see refs. 53–55 for detailed descriptions). This analysis was restricted to extant birds (53) and mammals (54, 55) because data limitations currently prevent ED scores being calculated for reptiles and extinct taxa from all classes. We used general linear models to compare the ED scores of the impacted species against threatened species for which invasive predators were not identified as a threat (“nonimpacted” species hereafter). We used a gamma error distribution because the data were positive, continuous, and skewed. Significant effects were inferred at the 0.05 level. Taxonomic differences between the Red List version 2014.3 and the source databases (53, 55) are detailed in Table S4. Because ED scores were not available for extinct species, the values presented here are likely to be an underestimate of the true effect sizes.

Dataset S1 contains a CSV file with information on species taxonomy, endemicity, regional occurrence, and the predator impact rankings that were used in the metaanalysis (Table S3); there is also a PDF file (Dataset S4) listing the 996 references that support the impact rankings.

Dataset S2 is a CSV file containing information on the taxonomy and conservation status of 3,745 birds, mammals, and reptiles classified as threatened in the IUCN Red List version 2014.3 (Data Collation).

We conducted a fail-safe analysis (56) to show that 3,203,182 cases showing an effect size of zero would be needed to eliminate a significant overall effect size. We also conducted a sensitivity analysis to determine how the selection of impact values (0.25, 0.75, 1) and the use of weights (1, 10, 100, 1,000) (Methods) influenced the results. We reran the taxonomic class and insularity models five times and changed a different variable each time, as follows: (i) an unweighted model; (ii) 0.25 decreased to 0.1; (iii) 0.25 increased to 0.4; (iv) 0.75 decreased to 0.6; and (v) 0.75 increased to 0.9. We plotted the effect sizes and confidence intervals (Figs. S4 and S5) to identify any qualitative changes in the results relative to the original results (i.e., the pattern changed across taxonomic classes or insular endemicity). We did not rerun the predator model because the results are unlikely to be informative, given the large number of models involved (5 models × 7 predators × 3 taxonomic classes = 105 sets of parameter estimates). The unweighted analysis revealed no systematic qualitative changes, except that the effect size decreased for continental birds, although the confidence intervals were wide and overlapped with those for insular species (Fig. S4). Varying the impact rankings likewise revealed no systematic qualitative differences relative to the original results, except that effect sizes for birds and mammals were more greatly influenced by changes in the “high” than the “mixed” level, whereas the opposite was true for reptiles (Fig. S5). This result is to be expected, given that reptiles on average had lower impact scores, but it does not change the interpretation of the results.