To obtain a comprehensive set of studies for our meta-analysis, we searched online databases of Web of Science, Biosis Previews and Biological Abstracts using combinations of the following keywords: introduced; alien; feral; predator; predation; experiment; manipulation; removal; reduction; control; effect; and impact. We also used the bibliographies of earlier reviews (Côté & Sutherland 1997
; Newton 1998
) and of papers already retrieved. Data searches ended in January 2006. A preliminary search yielded 159 articles, from which 84 were discarded as they did not meet our criteria (see below). This left 45 replicated studies (at least two control and two treatment plots or a before-and-after design) and 30 unreplicated studies (only one sample for either treatment or control or both; 30 studies describing 35 experiments) that were included in the final dataset (see Appendix S1 in the electronic supplementary material). Most of these were published in international scientific journals on ecology, conservation and wildlife, and we have also included book chapters and one unpublished Ph.D. thesis. Articles were published between 1939 and 2006, with most originating from the last 10 years.
We selected publications that described the effect of reduction or addition of avian or mammalian predators on avian or mammalian prey, excluding livestock and other non-native prey species. Studies that had removed both native and introduced predators were excluded, if the effects of these predator groups could not be separated. Acceptable prey responses to predator manipulations were classified as either population size or reproductive responses. Population size responses included those measured directly, as density, minimum numbers known to be alive, numbers of breeding pairs (as an index of population size), rate of increase or survival; and catch-per-unit effort indices such as the number of animals per area, trapline or transect. Reproductive responses included numbers of juveniles or broods produced, numbers of females with young, nesting success, survival of young and mean recruitment. Per capita measures, such as brood size per hen, number of juveniles per hen, number of broods per pair, number of fledglings/ducklings per pair, number of chicks fledged per pair, number of fawns/100 does, etc., were not included. The studies also had to have been run for long enough (one prey generation or more) for a prey demographic response to be possible. The studies measuring other parameters or using other units than those described were omitted. No authors were contacted to obtain missing data.
Necessary data (sample sizes, means of controls and treatments and their standard deviations/standard errors/confidence limits) were extracted from the text, tables or figures of the articles. In cases where error bars were not symmetrical about means, variances were calculated conservatively using the longest bars provided and, if no variances were given, these were calculated from raw results. Where possible, we used data taken at the end of experiments, but otherwise used an arithmetic mean of responses over the course of the studies. In cyclically fluctuating prey species, such as some small mammals, data were taken for consistency from the peak phase of the cycle. Reproductive responses were taken separately for each year, and an arithmetic mean calculated across the duration of the study to obtain one effect size per study.
Publications were scored also for the type of prey (e.g. rodent, ungulate, marsupial), prey class (bird/mammal), origin of predator (native/introduced), predator class (bird/mammal/both), the method of manipulation (addition/removal), location (mainland/island), habitat, spatial and temporal scales of the experiment (manipulation area and manipulation time), the continent in which the study was conducted and whether it was conducted in an exclosure or open terrain (manipulation type). We also recorded the mean weight of prey and predator species to calculate a predator/prey weight ratio for each study. Predators were considered to be introduced or native based on definitions provided in each study and confirmed using Long (2003)
. ‘Predator addition’ means either release of predators into experimental areas or attraction of predators (e.g. attraction of raptors with perch sites). ‘Predator removal’ refers to either exclosure experiments or manipulations, where predators were killed or relocated. Two persons (M.N. and P.S.) were responsible for data collection.
The 45 replicated experiments (table S1 in electronic supplementary material) were examined for publication bias using the normal quantile plot method (Wang & Bushman 1998
), and no evidence of publication bias was found (figure S1 in electronic supplementary material). This analysis was not possible for unreplicated studies and, therefore, the publication of such studies may have been biased towards large positive effects of predator removal on prey. However, it is rather unlikely that the results, whether significant or not, of very expensive, long-lasting predator manipulation experiments would remain unpublished, strongly reducing the likelihood for the file-drawer problem (Rosenthal 1979
) particularly in this meta-analysis.
For each replicated study, we calculated the standardized effect size as Hedges' d
v. 2.1 (Rosenberg et al. 2000
). There are also other metrics available for this type of primary data (means, variances and sample sizes), such as the log response ratio lnR
(Rosenberg et al. 2000
), but we chose d
because our data were not suitable for use of the response ratio (e.g. in some studies, the control group value was zero; Hedges et al. 1999
). Positive values of d
indicate that the predator treatment had a positive effect on prey species, zero means that there was no difference between treatment and control, and negative values signify a greater response in controls. For studies that reported the responses of multiple prey species to predator manipulation, we used the mean effect size across all species to retain independence. In one study, predators had been both added and removed; also here a mean effect for the whole study was calculated from the effect sizes of both treatments.
Our first prediction was that introduced predators should have more pronounced effects than native predators on the population sizes and reproductive outputs of their prey. To test this prediction with the 45 replicated experiments, we carried out a categorical summary analysis using the homogeneity statistic, Q
, in MetaWin
v. 2.1. As with variance in ANOVA, the total heterogeneity QT
can be partitioned into QM
, the variation explained by the model, and QE
, the residual error variance (Rosenberg et al. 2000
). Continuous summary analysis (weighted linear regression) was used to determine whether d
was affected by the spatial or temporal scale of the studies. We used random effects models and conducted resampling tests with 4999 iterations. Bias-corrected confidence intervals were used to evaluate the probability at 0.05. All tests were two-tailed.
To expand the coverage of research that has evaluated the impacts of predation, we conducted a similar analysis on the unreplicated predator removal experiments. Altogether, 34 unreplicated studies fulfilled the criteria, but two were excluded as they reported earlier stages of experiments that were represented in the analysis by later, more inclusive papers. In two cases, different aspects of the same experiment were reported in separate papers, which were then combined to gain one effect size. One study consisted of six experiments at different locations, which were therefore treated as independent studies in the dataset. Hence, the final dataset has 35 rows (table S2 in electronic supplementary material).
We classified the traits of the unreplicated experimental systems as described previously and defined the effect size as Xe/Xc, where Xe and Xc are the treatment and control prey responses, respectively. A ratio over 1 means that predator manipulation had a positive effect on the prey species, while a ratio up to 1 means that manipulation did not affect the prey species or the effects were negative. These unreplicated data cannot be analysed using typical meta-analysis approaches; therefore, we tested for differences in effect size in the study traits using Student's t-test with the Satterthwaite option for heteroscedastic variances (procedure TTEST, SAS Statistical Package, v. 9.1; SAS Institute, Cary, NC, USA). Effect size was ln transformed to meet the assumptions of normality.
Finally, a generalized linear model was built in order to further test our first and second predictions (i.e. that alien predators would have more impact than native predators on prey populations and that predation impacts would be greater on prey in island ecosystems compared with mainland ecosystems), and to explore possible interactions of the different explanatory variables. Neither MetaWin nor t-test allows the simultaneous analysis of multiple factors, and the sample sizes of replicated and unreplicated experiments alone were too small for such an analysis. Therefore, we pooled population size responses of the replicated and unreplicated experiments using Xe/Xc as the effect size measure. The model was fitted with a negative binomial distribution of the response and a log link function with the negative binomial GLM (glm.nb) procedure in the MASS library of S-Plus (v. 6, Insightful Corporation, Seattle, USA).
The main explanatory variables in the model were origin of predator (native versus alien), type of manipulation (open area versus predator exclosure), predator class (mammal, bird and both) and location (mainland versus island) together with their second-order interactions. Predator/prey weight ratio, manipulation area and duration of manipulation were included as continuous variables. Not all interactions of the classifying variables could be included, since they produced empty cells (singularities): origin of predator×predator class was removed because all introduced predators were mammals; origin of predator×manipulation type was removed because all except one study on introduced predators were conducted in open areas; and location×manipulation type was removed because there was only one study on islands using enclosures. Australia was classified as mainland in the analysis. The step Akaike information criterion (AIC) procedure in the MASS library of S-Plus was used for a stepwise model selection procedure, which selects the model with the lowest AIC value, starting with the global model. Owing to small sample size, AICc
was used (Burnham & Anderson 2000
). The support for each alternative model was evaluated by calculating: (i) AICc
where models with Δi
≤2 are considered to have substantial support and (ii) Akaike weights wi
, which describe the weight of evidence that model i
is the best model from the set of alternative models (Burnham & Anderson 2000