Biological control (biocontrol) can provide permanent cost-effective control of plant pests, but has variable success. The ability to predict the success of weed biocontrol should improve target prioritisation and the cost-benefit ratio of weed biocontrol.
We compiled a data base of the quantitative impacts of weed biocontrol programmes against 80 weed species and tested hypotheses regarding weed traits that contribute to weed biocontrol success using generalised additive models.
Modelling and cross-validation indicated that a model with three traits provided good ability to predict the responses of novel species in novel regions. Biocontrol impact varied according to whether a weed was reported to be a major weed in its native range, mode of reproduction (sexual or asexual) and ecosystem (aquatic or wetland versus terrestrial).
Biocontrol appears to be highly effective against weeds with the best combination of factors for success (aquatic, asexual species that are not major weeds in their native ranges), whereas most, but not all, programmes against ‘difficult targets’, which possess the worst combination of factors for success, have failed to result in a substantial impact. An additional analysis provides a preliminary indication that the success of pioneering programmes predicts the success of repeat programmes against the same target weed in other regions.
Synthesis and applications. Predictions generated by our model will assist weed prioritisation by improving the ability to predict the success of weed biocontrol. Nevertheless, prioritisation must also consider the importance of the candidate target weeds. Species that are predicted to be difficult targets could be targeted for biocontrol, provided that they are sufficiently important to offset the increased risk of failure against the greater benefits of successful control. Further investigation is needed to assess the ability of successful pioneering programmes to predict the success of repeat programmes in other locations, particularly in conjunction with plant traits.
Invasive plants can disrupt entire ecosystems and their economic impacts amount to many billions of dollars (Mack et al. 2000). Often, the number of weed (or potential weed) species is very large. For example, in New Zealand the 2430 naturalised, alien, vascular plant species (1780 fully naturalised and 650 casual; Howell & Sawyer 2006) outnumber the 2414 native vascular plants (De Lange & Rolfe 2010). Given high diversity of invasive plant species, the limited resources for tackling weed invasions must be prioritised effectively. Classical weed biocontrol tends to be a public, community-level activity carried out by institutions and public departments rather than private enterprise (Van Driesche & Hoddle 2002). The need to account for this public investment should demand robust decision-making processes, with the intention of selecting biocontrol targets that are not just important and have broad social support, but are also biologically and ecologically feasible. In practice, the means by which weed biocontrol targets are selected vary greatly between jurisdictions, and there have been few attempts to develop systems that select targets in a transparent and repeatable fashion.
Rationales for the prioritisation of weed control were proposed by Hiebert (1997), who advocated the development of decision-making tools to rank weeds according to their current impacts, future threat and the cost and feasibility of control. Many national schemes for setting weed management priorities, however, emphasise weed impacts (Thorp & Lynch 2000; Moran, Hoffmann & Zimmermann 2005) more than the cost or feasibility of control because the latter may be hard to estimate prior to the commencement of a control programme. This deficiency is particularly pertinent to classical biocontrol, which can have high development costs (Fowler, Syrett & Hill 2000) and does not always succeed; complete successes, where biocontrol is so dramatic that other control methods are no longer required, only account for approximately one-third of all completed programmes (McFadyen 1998). Approximately one in six programmes fail to have any impact (Hoffmann 1995; Fowler, Syrett & Hill 2000). Although much work has been directed towards improving the selection, testing and evaluation of weed biocontrol agents, there has been little underpinning research to identify criteria for selecting target weeds for which biocontrol might become an important management tool. In this paper, we consider various plant traits and ecological factors that have been proposed to influence weed biocontrol success as starting hypotheses, and test these hypotheses using observed data on biocontrol success.
McClay (1989) conducted a review of the 138 agricultural weeds in Alberta Canada and developed a scoring system, with the aim of selecting target weeds that were suitable for classical biocontrol. A revised system (Peschken & McClay 1995) ascribed scores according to the economic costs of the weed, the likelihood of ‘conflicts of interest’ (i.e. whether or not the target weed has some economic or environmental value that might result in opposition to a biocontrol programme), and to the levels of seven biological attributes, assumed to contribute to biocontrol success, namely (1) Intraspecific variation, which assumes that biocontrol success is more likely for weed populations of lower genetic variability. Evidence that intraspecific variation influences biocontrol success was presented by Burdon & Marshall (1981), who used the mode of reproduction of the target weed to infer genetic diversity (assuming that bottlenecks occurring during introduction are likely to result in genetically uniform populations of asexual species, while recombination of genes from parent plants should result in a range of different genotypes of sexually reproducing species). Sheppard & Chaboudez (1995) contradicted these claims, arguing that (i) the analysis was biased because obligate annuals are never agamospermous or clonal and annuals may be harder targets for biocontrol for reasons unrelated to their mode of reproduction; and (ii) biocontrol success was independent of mode of reproduction when they repeated Burdon & Marshall's (1981) analysis with updated data. Nevertheless, Charudattan (2005) subsequently noted that because host–pathogen interactions are often governed by single-gene differences, even limited genetic variability can cause a plant pathogen biocontrol agent to fail; (2) The native geographical range of the weed was included by Peschken & McClay (1995), as the biocontrol of native (indigenous) weeds in their natural range is considered in this study, and the likelihood of finding biocontrol agents for introduced weeds is assumed to be greater than for native weeds. Although native-weed biocontrol programmes have been conducted in the past, we discounted this factor from our analyses because, as Pemberton (2002) noted, it is impossible to limit biocontrol agents to situations where the target native weeds are problems and such programmes would no longer be sanctioned by most regulatory authorities (Barratt & Moeed 2005); (3) Relative abundance was included because it was considered that weeds which were not weedy (and so unlikely to be abundant) in the areas from which the biocontrol agents were imported, tend to be successfully controlled in the countries of introduction; (4) Success of previous biocontrol programmes was included because successful biocontrol programmes have often been repeated in different regions (Crawley 1989); (5) Number of known, promising candidate biocontrol agents was included by Peschken & McClay (1995), who assumed that prospects for successful biocontrol are better if surveys carried out in the native range identified promising biocontrol agents. We excluded this factor because a biocontrol programme must already be in process in order for the results of native range surveys to be known; (6) Habitat stability was considered important because most significant successes in classical biocontrol have occurred against weeds of uncultivated rangeland; (7) The number of economic, ornamental and native plant species in the same genus and/or tribe as the target weed was assumed to be negatively correlated to success because the more related a target weed is to crops, ornamental plants and native plants, the lower the chance of finding sufficiently specific and damaging biocontrol agents.
Little work has been performed to test McClay's (1989) and Peschken & McClay's (1995) hypotheses, or to assess their relative importance. Here, we use an extensive international compilation of the observed effectiveness of weed biocontrol to test, which factors best predict biocontrol programme success. We use modern regression techniques to model biocontrol success against plant traits and test the ability of this model to predict novel species in novel regions.
Materials and methods
Observed biocontrol efficacy
We collated an international data base of observed weed biocontrol success from a range of published sources, together with plant traits for each weed species. We searched Centre for Agricultural Bioscience (CAB) Abstracts, Google and regional reviews of weed biocontrol (Kelleher & Hulme 1984; Cameron et al. 1989; Olckers & Hill 1999; Mason & Huber 2002; Coombs et al. 2004; Munniappan, Reddy & Raman 2009; Moran, Hoffman & Hill 2011; Julien, McFadyen & Cullen 2012) and international symposia on biological control of weeds (http://www.invasive.org/proceedings/) for quantitative data regarding success of biocontrol against weeds. Impact data were reported in multiple ways (e.g. change in a weed's percentage cover; stems m−2 or biomass m−2), so we standardised data by creating an ‘impact index’ to allow comparison between weeds. Impact index is defined as the proportional reduction in weed density as a result of biocontrol (e.g. if biocontrol reduced a weed's density from 33 to 3·8 stems m−2, then the reduction in stem density = 33–3·8 = 29·2 and impact index = 29·2/33 = 0·885). Where there were several published reports for a particular weed species, we used the most recent provided that it updated previous studies (e.g. after the release of a new agent), or we calculated an average impact where the impact of the same biocontrol agent(s) was determined at different localities within the same region. Note ‘region’ generally refers to the country in which the biocontrol programme was conducted but, to allow for biogeographical separation, we considered Hawai'i distinct from continental USA and we combined Canada with continental USA (as North America). If no biocontrol agents were reported to have established, we assumed that the impact of biocontrol was zero. Where anecdotal reports indicated that the agents have a ‘negligible’ or ‘trivial’ impact on the target weed, we assigned a reduction in weed density of zero, even if quantitative data were lacking. Where reports indicated agents had an impact on weed populations but lacked quantitative data, we excluded these weed species from the data set.
We omitted ‘restricted’ programmes that only introduced agents, which attack plant reproductive structures to reduce the invasiveness of species that are valued (e.g. for timber, tannin and firewood production; Moseley et al. 2009), because these programmes were never intended to directly reduce weed density. We also omitted ‘reassociation’ weeds, where agents were released on weeds that are taxonomically related to the target weed, because these were opportunistic attempts at biocontrol and the reassociation weed species had not been surveyed for enemies in their native range.
Factors of biocontrol success
For each weed species, we determined a number of explanatory variables that have been proposed to determine the success of biocontrol efforts listed below:
Relative abundance (‘major weed in native range’, with two levels, Yes or No). We searched CAB Abstracts (using the plant name and ‘weed’ as search options), and we categorised a plant as a major weed in the native range if the CAB Abstracts search listed five or more references that described a plant as a weed in the native range (i.e. excluding articles reporting native range surveys for biocontrol agents). This tested the hypothesis (McClay 1989; Peschken & McClay 1995) that plants that are more abundant in the exotic range versus the native range should be easier to control because these species are more likely to be limited by natural enemies in the native range. We chose this approach because quantitative data regarding relative abundance in the native and introduced ranges are often lacking (Hierro, Maron & Callaway 2005). Our underlying assumption is that a plant must be abundant within its native range to be considered a weed.
Presence of potential valued non-target congeneric plants (henceforth ‘valued congener’), with two levels, Yes or No. We consulted floras and conducted Internet searches to determine whether native or valued exotic congeneric plants (e.g. crops) of the weed were present in the same region as the target weed. This tested the hypothesis that weeds with valued congeners should be harder to control, because of the greater potential for non-target attack constraining biocontrol agent selection.
Life cycle, with two levels (‘temperate annual’, Yes or No). This tested McClay (1989) hypothesis that annual plants are harder to control than biennial and perennial weeds, modified to allow for Sheppard & Chaboudez's (1995) observation that this generalisation was only true for temperate annuals.
Reproduction, with two levels: ‘asexual’ (i.e. apomictic species and species which only reproduce by vegetative means, e.g. Ageratina adenophora and Alternanthera philoxeroides, respectively; Burdon & Marshall 1981) and ‘sexual’ (i.e. species which are capable of sexual reproduction, including those that may also use vegetative reproduction, e.g. Cirsium arvense; Burdon & Marshall 1981).
Habitat stability, with two levels (‘weed of cultivated land or improved pasture’, Yes or No). This tested Peschken & McClay (1995) hypothesis that weeds that occur on cultivated land are harder to control than weeds of relatively undisturbed habitats. Our assumption being that cultivated land or improved pasture are subject to large-scale disturbance more frequently than other habitats, for example unimproved rangeland, where biocontrol has been applied.
Ecosystem, with two levels (‘Aquatic/wetland’ or ‘terrestrial’). This retested the hypothesis that aquatic weeds are easier targets than terrestrial weeds (Sheppard & Chaboudez 1995). Weeds were classified as aquatic/wetland if they were aquatic (emergent, floating or submerged) or wetland plants that are subject to regular seasonal flooding (e.g. Mimosa pigra; Paynter & Flanagan 2004).
Denoth, Frid & Myers (2002) showed that biocontrol success was correlated with the number of agents released, so we also recorded the number of agent species released to investigate this relationship. We also compiled a list of weeds where there were quantitative impact data from multiple regions, to test the hypothesis that biocontrol programme success in one region is a good predictor of success in other regions.
Our collated biocontrol data base included 80 weeds species in 34 plant families from 8 regions against which 232 agents had been released. Certain genera were represented by several weeds (e.g. Centaurea and Opuntia). Congeneric plants may share traits that are unrelated to their susceptibility to biocontrol, potentially biasing the analysis, so we averaged the impacts for congeneric species with identical traits. For example, for Centaurea, we calculated two impact indices: one for temperate annuals (Centaurea cyanus and Centaurea solstitialis) and one for perennials (Centaurea diffusa and Centaurea stoebe). Including such ‘averaged’ species reduced the data base to 69 species for subsequent analyses, of which 28 were classified as a major weed in native range; 26 had a valued congener; 12 were temperate annuals; 12 were asexual; 19 were weeds of cultivated land or improved pasture, and 11 were aquatic/wetland weeds. The eight regions in our data and the number of data base ‘species’ in each were as follows: North America (30). Australia (20), Republic of South Africa (7), Hawaii (6), New Zealand (3), Indonesia (1), Papua New Guinea (1) and Russia (1). If biocontrol data existed for a weed in more than one region, we used data from the first region that biocontrol was used against that particular weed for which data were available, so that regional cross-validation could be performed (see below).
We used Generalized Regression Analysis and Spatial Prediction (GRASP, Lehmann, Overton & Leathwick 2002) to model the efficacy of biocontrol against the plant factors. GRASP is a collection of scripts in Splus (Lehmann, Overton & Leathwick 2002) that facilitate the use of generalized additive models (GAMs). A GAM can include both categorical and continuous variables. We modelled the proportion reduction achieved for each species against the factors of biocontrol success, using a logistic link function. The Akaike information criteria (AIC; Burnham & Anderson 2000) were used for initial model selection.
A regional cross-validation was performed to test the stability of models and their ability to predict biocontrol success in new geographical regions. K-fold cross-validation and the correlation between observed and predicted were used to evaluate model performance. The region variable was used as a basis to subset the data for k-fold cross-validation. Regions were successively dropped from the data, and models fit on the remaining regions and predictions made for the missing region. The observed data from each region were compared to predictions from models fit on the data from other regions.
A backwards stepwise procedure was used to choose the model with the highest correlation between observed and predicted values under regional cross-validation. This model was used as the final model for predictive purposes. The final model was used to create a table that can be used to predict the expected proportion reduction achieved by biocontrol.
Finally, for the data set of weeds where there was quantitative impact data from multiple regions, we investigated the hypothesis that success in one region is a good predictor of success in other regions by plotting the impact of the pioneering programme against the impact of the repeat programmes (averaged if a programme was repeated in multiple regions). The data do not warrant statistical tests, so only indicative results are provided.
Across the 69 averaged ‘species’ in our analyses, the proportional reduction achieved using biocontrol varied from zero (no impact) to one (complete control), with a mean of 0·44. The raw relationships between the proportional reduction and the explanatory variables are shown in Fig. 1. The mean proportional reduction achieved was higher when the plant was not a major weed in its native range, and for species that were not temperate annuals, whereas the reduction was lower for species with sexual reproduction, for weeds of cultivated land or improved pasture, and in terrestrial species compared to aquatic or wetland species. The proportional reduction showed a generally positive correlation with the number of agents released, with a notable exception of Lantana camara in South Africa (where 19 agents were released and no reduction was achieved; Cilliers & Neser 1991). The only variable that showed little effect was whether the species had a valued congener. Of course, the explanatory variables are not entirely independent of each other, so rather than test their effects independently, we used GAM models to choose a set of explanatory variables to predict biocontrol success.
The validation and cross-validation performance of the GAM models during the backwards stepwise selection procedure is shown in Fig. 2. Starting with a full model with seven explanatory variables, the least important variable at each step is dropped from the model. For each step, the correlation between observed values and those predicted by the model are plotted for the model validation and the model cross-validation (by region). The validation performance decreases monotonically as variables are removed from the model, while cross-validation performance increases to a maximum when three variables remain in the model, and then declines rapidly. The model with the highest regional cross-validation was chosen as the best model because this will have the greatest ability to predict novel species in new regions. Although there are only seven variables, there are eight steps in selection procedure because the continuous variable ‘number of agent species released’ was tested with three and one degrees of freedom in the scatterplot smoother used in the GAM model.
The scatterplot of observed and predicted values under regional cross-validation is shown in Fig. 3. The model shows errors of both over- and under-prediction when predicting onto novel species and regions. Ten species for which the model predicts over 40% reduction were observed to have no impact of biocontrol (points on middle left: Carduus pycnocephalus, Halogeton glomeratus and Salsola tragus in North America, Cyperus rotundus, Myrica faya, Pluchea odorata and Schinus terebinthifolius in Hawai'i, Lantana camara in South Africa and Parkinsonia aculeata and Vachellia (Acacia) nilotica subsp. indica in Australia). Conversely, three species predicted to have less than 40% reduction were observed to have over 60% reduction because of biocontrol (points on bottom right of Fig. 3: Chondrilla juncea and Jacobaea vulgaris in Australia and Solanum elaeagnifolium in South Africa). Nevertheless, this regional cross-validation demonstrates useful ability to predict onto novel species in new regions. For the species with a predicted reduction because of biocontrol over 60%, all were observed to have at least a 40% reduction. Overall, the model explains 32% of the deviation in observed proportion reduction.
The relative importance of the explanatory variables in predicting the proportion reduction achieved by biocontrol is shown by their contributions to the model (Fig. 4). The alone contribution for an explanatory variable is the amount of deviance in the proportional reduction that is explained by a model with only that explanatory variable. Thus, the alone contribution provides a characterisation of the explanatory variable in predicting the proportional reduction when no other variables are important. The relative magnitudes of these alone contributions between variables accord well with the differences between groups seen in Fig. 1. The variable ‘major weed in native range’ has the largest alone contribution, followed by ‘habitat stability’. The variable ‘valued congener’ has a negligible alone contribution.
The drop contribution of each explanatory variable (Fig. 4) is the decrease in the total deviance explained by the final model when each variable is dropped from the model. For variables that are not in the final model, the drop contribution is not shown. The variable ‘major weed in native range’ had the highest drop contribution, followed by ‘reproduction’ and ‘ecosystem’.
The final GAM model for the proportion reduction because of biocontrol is shown in Fig. 5. This model accords to the pattern seen in Fig. 1; higher proportion reduction is predicted for species that are not a major weed in their native range, asexual species and aquatic species. The model contains only three, binary variables giving eight unique combinations of the traits. Table 1 contains the predicted values for the proportion reduction achieved for each of these eight trait combinations.
Table 1. Predictions of the proportional reduction achieved by biocontrol for each of the eight combinations of the predictor variables
Major weed in native range
Proportion reduction from biocontrol
A scatterplot of the impact of the ‘repeat’ biocontrol programmes against the impact of the pioneering programmes (Fig. 6) shows a perfect correspondence between success versus nonsuccess in the pioneering programme and success versus nonsuccess in the repeat programme(s). However, with only two programmes that failed, the result is not statistically significant. For programmes that succeeded, there is no relationship between the extent of weed reduction (i.e. success) in the pioneering programme and the repeat programme(s).
We consider our system, which predicts biocontrol impact using evidence-based criteria and quantitative data, to be an advance over previous prioritisation systems (McClay 1989; Peschken & McClay 1995; Palmer & Miller 1996; Syrett 2002).
Our modelling and cross-validation tested the ability of our model to predict the biocontrol success of novel species in novel regions. Our model predicts the success of biocontrol in novel species because it uses only information on the plants traits, not on the success of biocontrol elsewhere. Second, the regional cross-validation procedure used, where the model is fit on species from other regions and predicted onto the target species of a particular region, tests the ability of the model to predict onto new regions. Our model with the best regional cross-validation performance indicated that only three factors were significant in the best fit model (‘major weed in the native range’, ‘reproduction’ and ‘ecosystem’). Each of these factors has only two levels, which limits the model's ability to differentiate between candidate target weed species to eight possible combinations of the factors. Equally importantly, our results indicate that further discrimination is not possible, at least for our data. Success appears almost guaranteed against ‘good target’ weeds which possess the best combination of factors for successful biocontrol (aquatic, clonal species that are not major weeds in their native ranges), while most programmes against ‘difficult targets’ (i.e. weeds with the worst combination of factors for success) have failed to result in a measurable impact.
When tested individually, additional factors that were excluded from the best model were nevertheless significantly correlated to biocontrol success (life cycle and habitat stability; Fig 1). The most likely explanation for their omission from the model is that other variables with which they have low correlations are better explanatory variables. However, the importance of habitat stability may have been underestimated because of the way some data were collected. For example, the chrysomelid beetle Leptinotarsa texana had a major impact on the cropping weed Solanum elaeagnifolium in South Africa, but its impact was monitored on plots that were set aside for the duration of the impact study (Hoffmann, Moran & Impson 1998). To our knowledge, the effectiveness of L. texana under different crop management regimes has not been documented, and it is conceivable that the impact of this species may be diminished under typical cropping systems. Indeed, the ragweed leaf beetle Zygogramma suturalis initially suppressed ragweed Ambrosia artemisiifolia following introduction into an uncultivated site in Russia, but its population density drastically declined and its impact on the target weed was negligible after crop rotation practices were resumed (Reznik et al. 2008).
The presence of a valued congener did not influence impact. Regulation of biocontrol has, however, become increasingly risk-averse, largely due to increased recognition of the risks of non-target attack (Sheppard et al. 2003). Most examples of serious non-target attack in weed biocontrol involve plants that are closely related to the target weed (Pemberton 2000). Host-specificity testing is used to discard potential weed biocontrol agents that are likely to cause significant undesirable damage to non-target plants. In the past, little consideration was given to the potential for non-target attack on native plants and test-plant lists often comprised of economically important plants, which were not necessarily closely related to the target weed. For example, Fowler, Syrett & Hill (2000) reported that from 1943 to 1982, all weed biocontrol introductions in New Zealand relied upon overseas host-range testing and that native New Zealand test plants were invariably not examined. Consequently, there are few past examples of weed biocontrol programmes being abandoned because of an absence of demonstrably specific agents (Maw & Schroeder 1981). However, had permission to release agents been sought under today's regulations, it is likely that the predictable potential for non-target attack would have prevented the release of a number of successful agents (e.g. Chrysolina spp., Rhinocyllus conicus and Trichosirocalus horridus) that have subsequently been reported to feed on non-target plants (Pemberton 2000; Groenteman, Fowler & Sullivan 2011; Fowler et al. 2012), greatly reducing the likelihood of control of their target weeds. Thus, while we found that presence of a valued congener may not have influenced the success of past biocontrol programmes, there is good reason to suppose it should influence the success of contemporary programmes.
Variables not tested
Biocontrol has had a major impact against some ‘difficult’ target weeds, such as ragwort Jacobaea vulgaris (McEvoy, Cox & Coombs 1991; McLaren, Ireson & Kwong 2000). It is not clear to us whether this indicates that biocontrol can potentially succeed against any target weed, provided programmes are adequately resourced, or if other factors that we have not identified are important determinants of success. A lack of resources can play a role in biocontrol success (Fowler 2000); however, within our data, the numbers of agents released (a likely surrogate for how well a programme is resourced) had only a small and statistically insignificant additional ability to predict biocontrol success. Other factors can also influence biocontrol success, such as biotic resistance owing to predation or parasitism (Goeden & Louda 1976; Paynter et al. 2010) of biocontrol agents.
Recommendations for improving prediction
Predicting biocontrol impact may be enhanced by improved quantification of key plant attributes. For example, mode of reproduction is an unreliable indicator of genetic diversity; for instance, despite being apomictic (Burdon & Marshall 1981), the genetic diversity of Ageratina adenophora is high in China, where biocontrol is ineffective (Wan et al. 2010). The genetic diversity of a weed population can be directly estimated using molecular techniques (Amsellem et al. 2000), but this has rarely been performed. Moreover, some plants may not be perceived to be weedy or undesirable in their native ranges, even if they are abundant there. Rather than relying on the factor ‘weed in native range’ as a surrogate for abundance in the native range, it is preferable to quantify the relative abundance of a weed and its native and introduced ranges, and this has also rarely been attempted (Hinz & Schwarzlaender 2004; Hierro, Maron & Callaway 2005).
Our preliminary analysis of the predictive power of previous control programmes suggests that this effect deserves further investigation to improve predictive performance. It is possible that our model could be improved with the inclusion of information on the outcome of previous programmes in other regions. Note, the only predictive ability evident in Fig. 6 is the perfect correlation between success in the pioneering programme and repeat programmes, but our data base contained only two-failed biocontrol programmes; therefore, conclusions about this effect await further research. For those programmes with success in pioneer programmes followed by success in further programmes, there is no relationship between the proportional reduction in the pioneering programme and the repeat programmes evident from our data.
Recommendations for improving target prioritisation
An analysis of the return on investment of Australian weed biocontrol programmes indicated that a minor reduction in a highly important weed can result in greater economic benefits, compared to the complete control of a relatively minor weed (Page & Lacey 2006). Moreover, the impacts of some alien invasive species cannot be minimised without integrating biocontrol with other management practices (Moran, Hoffmann & Zimmermann 2005). There is a risk, therefore, that by emphasising ‘impact score’, our system will lead to greater focus on picking ‘winners’ and even less emphasis on the potential benefits of biocontrol when integrated with other options. The predictions generated by our model should not, therefore, be used slavishly to select the programmes that are most likely to succeed without due consideration of the importance of the candidate target weeds and the potential for biocontrol to play a role in integrated weed management (Paynter & Flanagan 2004). Species that are predicted to be difficult targets should be targeted for biocontrol, provided they are sufficiently important to ensure that the increased risk of failure is offset by the greater benefits of success. Our results from Fig. 6 indicate that the success of previous programmes on the same species may be a very useful predictor of biocontrol success, so this effect warrants better quantification and testing.
This work was funded by Land & Water Australia and FRST Contract no. C09X0905. We thank Simon Fowler and Lynley Hayes for helpful discussions.