Understanding dispersal rates of invading weed biocontrol agents

Authors


Correspondence author. E-mail: paynterq@landcareresearch.co.nz

Summary

1. Classical weed biological control programmes aim to rapidly establish biocontrol agent populations throughout the range of a weed. Release strategies, however, may often be suboptimal because the natural dispersal ability of a new agent is rarely known, potentially resulting in unnecessarily intensive release programmes for fast-dispersing agents or inadequate release programmes for slowly dispersing agents.

2. We reviewed published dispersal data for 66 arthropod and 11 fungal pathogen weed biocontrol agents. We tested hypotheses regarding agent characteristics that were predicted to affect dispersal and whether agents that dispersed rapidly were more successful than those which dispersed more slowly.

3. Dispersal rates varied by four orders of magnitude: the fastest agents dispersed several hundred kilometres per year and the slowest by only tens of metres per year. Approximately 30% of arthropod agents and four of 11 pathogen agents dispersed <1 km year−1, indicating that intensive redistribution is often required for rapid widespread establishment. Successful agents were equally likely to be fast or slow dispersers indicating that effort made redistributing slowly dispersing agents can often be beneficial.

4. Both pathogen and arthropod dispersal rates were positively correlated with voltinism. Arthropod dispersal also significantly varied according to fecundity, dispersal type (crawling or passive wind dispersal vs. flight); taxon, life-style, habitat and the diversity of parasitoids attacking the agent in the native range.

5. We developed dispersal models which explained up to 73% of the variance in arthropod agent dispersal rates and we validated these models by using them to predict the rate that invasive arthropod herbivores should invade.

6.Synthesis and applications. We conclude that knowledge of a few parameters (agent voltinism, parasitoid diversity in the native range, habitat, fecundity, taxon and life-style category), prior to introduction of a biocontrol agent, could be used to predict how fast it is likely to invade a new environment. This should assist optimization of release strategies by determining the geographic scale at which to release agents, according to the agents’ ability to rapidly close the gaps by natural dispersal.

Introduction

To minimize the delay between agent release and successful weed suppression, classical weed biological control programmes must rapidly establish biological control agent populations throughout the range of a weed. Conducting numerous small releases can potentially establish an agent across a wide geographical area more rapidly than making fewer, larger releases. Establishment rates decline with decreasing release size (i.e. the number of organisms released) but the optimal release size for maximizing the number of successful releases can be quantified experimentally (e.g. Memmott, Fowler & Hill 1998; Grevstad 1999; Memmott et al. 2005). The distance between release sites, however, cannot be optimized if the natural dispersal rate of an agent is unknown. Moreover, agent dispersal rates are highly variable, ranging from just tens of metres (e.g. Dactylopius opuntiae and Puccinia myrsiphylli; Foxcroft & Hoffmann 2000; Morin et al. 2006) to hundreds of kilometres within a year of their release (e.g. Cullen, Kable & Catt 1973; Edwards, Holtkamp & Adair 1999). Therefore, slowly dispersing agents may require numerous releases in relatively close proximity, while rapidly dispersing agents may only require a few widely dispersed releases, to achieve rapid widespread establishment. An ability to predict an agents’ rate of dispersal should enable biocontrol practitioners to better plan and allocate resources towards optimizing agent releases.

In contrast to weed biocontrol programmes, a fundamental strategy of invasive pest control is to limit dispersal (Krushelnycky, Loope & Joe 2004). Failed eradication attempts can be economic and environmental disasters (Myers et al. 2000), so an ability to predict an organism’s dispersal ability should assist decision-making for authorities responding to a new incursion of a pest species. Similar to weed biocontrol agents, there may be only limited information regarding the dispersal abilities of the invading organism. Retrospectively investigating factors that influence the dispersal of invasive pests can be problematic because precise introduction information is often unknown. For example, when the moth Uraba lugens Walker was discovered in New Zealand it was already so well established that it was considered ineradicable (Ross 2003), implying that it had either been present for several generations or that there were multiple introduction loci. Similarly, the presence of the hadda beetle Epilachna vigintioctopunctata (F.) in New Zealand was first recognized in January 2010, but the subsequent discovery of a misidentified specimen in a student’s insect collection indicated that it was already present in April 2006 (D. Jones, personal communication). Weed biocontrol programmes, therefore, make convenient experimental systems for studying the factors which influence dispersal, because the locations and dates of introduction are usually well documented. We, therefore, attempted to identify which measurable factors best explain or predict weed biocontrol agent dispersal rates. We also tested whether faster dispersing agents are also more successful at controlling their target weeds to determine whether effort spent establishing slowly dispersing agents is worthwhile.

Models (see Williamson 1996) predict that dispersal rate should be related to an organism’s diffusion coefficient (i.e. dispersal ability) and its intrinsic rate of increase (r). We hypothesized that fecundity and voltinism should influence dispersal. For example, voltinism is correlated with the set of life-history adaptations resulting from r- and k-selection (Shapiro 1975): Species living in ephemeral habitats would be expected to undergo selection for the following characteristics: ability to discover the habitat quickly; rapid reproduction, to exploit the habitat before competing species arrive; and dispersal in search of new habitats as the old one deteriorates. The overall effect is to raise the intrinsic rate of increase. Another factor that may influence an agent’s intrinsic rate of increase is enemy release: arthropod weed biocontrol agents commonly escape from specialist parasitoids when they are introduced into a novel environment and this may be critical for agent success (Paynter et al. 2010). Moreover, species which are heavily parasitized in their native range could be considered to occupy ephemeral habitats, where they undergo similar selection pressures to those noted by Shapiro (1975), not because the habitat is ‘used up’, but because specialist parasitoids eventually ‘catch up’ with their hosts and render the habitat unsuitable due to the high mortality rates they impose on their hosts. Reliable quantitative data regarding parasitism rates of weed biocontrol agents are rare in the literature, but the number of parasitoid species that attack weed biocontrol agents in the native range is often recorded (Paynter et al. 2010). We, therefore, hypothesized that biocontrol agent dispersal rates should be correlated to the number of parasitoid species that attack them in the native range (henceforth ‘parasitoid diversity’). We also hypothesized that variation in the dispersal ability of weed biocontrol agents might be explained by taxon, habitat, larval life-style category and dispersal type (for example, passive wind dispersal or active flight). To test these hypotheses, we reviewed the international literature regarding biocontrol agent dispersal.

Materials and methods

We used CAB Abstracts® and Web of Science (ISI) to generate a data base of publications that quantitatively assess weed biocontrol agent dispersal. Many publications reporting the establishment of weed biocontrol agents did not mention dispersal in the title or the key-words, but nevertheless quantified dispersal in the text. Therefore, we used the search terms ‘weed’ AND ‘biocontrol’ and ‘establishment’ as well as ‘weed’ AND ‘biocontrol’ and ‘dispersal’ as search terms. We excluded examples of accidentally or illegally introduced biocontrol agents (e.g. Syrett & Harman 1995; Morin, Auld & Smith 1996), due to uncertainties regarding the introduction dates and the number and location of introduction sites. Where several estimates of dispersal rate were given for a species, we calculated an average dispersal rate. We also reviewed published literature for records of fecundity, voltinism and native range parasitoid diversity. For some agents for which the voltinism was not explicitly stated, we estimated the potential number of generations per year, based on the published development times for each life cycle stage and the stated activity period of the organism. For example, we estimated that the fungal pathogen Entyloma ageratinae RW Barreto and HC Evans can complete c. 21·5 generations per year in New Zealand based on a the generation time of 7–10 days and that the climate is suitable for the fungus to proliferate for 6 months of the year (see Table S1, Supporting information).

Analysis

We compiled quantitative dispersal data for 66 arthropod and 11 fungal pathogen species (Tables S1 and S2, Supporting information). Voltinism data existed for all but one agent, and we found fecundity data and native range parasitoid diversity data for 56 and 42 of the 66 arthropod agents, respectively (Table S1, Supporting information).

For all analyses, dispersal data were standardized to km year−1, and then log (n + 1) transformed, to normalize the data. Arthropod and pathogen agents were analysed separately, using the general linear models option in genstat® 8.1 (Lawes Agricultural Trust).

Arthropod agents

Preliminary analyses of variance using the 66 species data set tested the hypotheses that the variable ‘dispersal rate’ varied according to the following predictors that were each fitted individually: taxon (declared as a factor, with seven categories corresponding to the arthropod orders: Acarina, Coleoptera, Diptera, Hemiptera, Hymenoptera, Lepidoptera, Thysanoptera); habitat (declared as a factor with two categories: aquatic/wetland vs. terrestrial); dispersal type (declared as a factor with two categories, flight vs. wind/crawling); life-style category (declared as a factor based on larval feeding niche with six categories: external defoliator or sap-sucker; gall-inducer; leaf/stem miner/borer/roller/tier; seed-feeder; underground root or rosette-borer). Linear regressions were performed to test the hypotheses that ‘dispersal rate’ varied according the variables voltinism [log (n + 1) transformed to normalize the data] and ‘time since release’. The latter was included because Travis et al. (2009) demonstrated how accelerating invasion rates should result from the evolution of density-dependent dispersal.

We subsequently used the 56 species data set for which fecundity data were available to perform a linear regression to test the hypothesis that agent dispersal rate varied according to the variable agent fecundity [log (n + 1) transformed to normalize the data] and we used the 42 species data set for which parasitoid diversity data were available to perform a linear regression that tested the hypothesis that dispersal rate varied according to the variable parasitoid diversity [log (n + 1) transformed].

We then performed analyses of covariance that included multiple predictors, using the complete data set of 66 species (using the predictors taxon, habitat, dispersal-type, life-style category, voltinism and time since release) and the smaller subsets that included fecundity (56 species, using the same predictors as the previous analysis, plus fecundity) and parasitoid diversity (42 species, using the same predictors as the 66 species analysis, plus parasitoid diversity). When conducting stepwise regression to produce a minimum adequate model, the order of parameter entry (or deletion), and the number of candidate parameters, can affect the selected model, particularly where predictors are correlated (Whittingham et al. 2006). Therefore, rather than relying on stepwise regression to produce a minimum adequate model for each data set we constructed alternative models, using all possible subsets of the predictors as candidate models and compared them using Akaike’s information criterion, corrected for small sample size (AICc), where the model with the lowest AICc value has the ‘best’ relative fit, given the number of parameters included. If the difference between the AICc values (Δi) of two models >2, this indicates that the model with the lowest AICc is preferable, whereas Δi values <2 indicate that models are similar in their ability to describe the data (see Burnham & Anderson 2000; Whittingham et al. 2006).

Finally, we reviewed the scientific literature to determine the success of the arthropod agents. Agents were classified as highly successful (i.e. having a major impact) if published quantitative information indicated an agent caused a statistically significant reduction in a weed’s population, percentage cover or biomass of 70% or more, and moderately successful or unsuccessful if published information indicated an agent had a lesser impact or no appreciable effect on a target weed (Table S1, Supporting information). We then performed a logistic regression (specifying binomial errors and a logit link; where a highly successful agent was coded = 1 and a moderately successful or unsuccessful agent = 0) to determine whether the probability of an agent being highly successful was correlated with dispersal rate [log (n + 1) transformed], excluding agents for which there was insufficient data to assess success.

Pathogen agents

For pathogens, we performed separate linear regressions to test the hypotheses that dispersal rate varied according to the variables voltinism and time since release. We could not test for differences due to habitat (all examples attack terrestrial weeds), guild (all but one were leaf pathogens) and taxon (all but two pathogens belong to the Uridenales; see Table S2, Supporting information). As above, we categorized pathogens according to their impact on the target weed (Table S2, Supporting information) and we tested whether faster dispersal was positively associated with agents’ success by performing a logistic regression.

Results

Dispersal rates of arthropod biocontrol agents ranged from just 22 m year−1 for the scale insect Dactylopius opuntiae to 400 km year−1 for the moth Arcola malloi Pastrana (Table S1, Supporting information). Approximately 30% of agents dispersed <1 km year−1, while c. 10% of agents dispersed more than 100 km year−1 (Fig. 1). Preliminary analysis, using the full data set of 66 species, indicated that the factors ‘habitat’ and ‘dispersal type’ significantly influenced dispersal: aquatic/wetland agents dispersed faster (back-transformed mean dispersal rate = 22·4 km year−1) than terrestrial agents (back-transformed mean dispersal rate = 4·85 km year−1; anova, F1,64 = 10·96, < 0·01) and agents capable of flight dispersed faster (back-transformed mean dispersal rate = 8·09 km year−1) than wind dispersed or crawling species (back-transformed mean dispersal rate = 1·75 km year−1; anova, F1,64 = 4·56, < 0·05). Taxon also influenced dispersal rate (anova, F6,59 = 2·75, < 0·05; Fig. 2a); of the main arthropod groups used in biocontrol programmes, Diptera and Lepidoptera were the fastest dispersers and Coleoptera the slowest. In addition, Acarina and Hemiptera dispersed significantly more slowly than Diptera (Fig. 2a). Dispersal rate varied according to life-style category, with miners, borers and leaf-rollers dispersing faster than external feeders, root and rosette-feeders and seed-feeders, but not gall-inducers, which also dispersed significantly faster than root and rosette-feeders and seed-feeders (anova, F4,61 = 4·51, < 0·01; Fig. 2b). Dispersal rate was positively correlated with voltinism (Fig. 3a), but not with time since release (anova, F1,62 = 2·35, n.s.).

Figure 1.

 Frequency distribution of the dispersal rates of 66 arthropod weed biocontrol agents.

Figure 2.

 Mean dispersal rate (±SE; back-transformed from log) of arthropod biocontrol agents according to (a) taxon and (b) life-style category, calculated from the complete data set of 66 arthropod species. Columns with the same letter are not significantly different (LSD). Numbers in parentheses refer to the sample size (number of species) in each category.

Figure 3.

 Mean dispersal rate of (a) arthropod and (b) pathogen biocontrol agents vs. agent voltinism. (a) anova, F1,63 = 33·46, < 0·001; log dispersal rate (km year−1) = 1·57 × log (no. generations year−1) − 0·163; (b) anova, F1,9 = 6·48, < 0·05; log dispersal rate (km year−1) = 0·099 × (no. generations year−1) − 0·024

Using the data set of 66 arthropod species, the model with the best relative fit included the factors ‘habitat’, ‘life-style category’ and ‘taxon’, and the variable ‘voltinism’ (Table S3, Supporting information and Table 1a). Analysis of the subset of 56 agents for which there was fecundity data indicated that dispersal rate was not correlated with agent fecundity alone (anova, F1,54 = 0·42, n.s.). The model with the best relative fit, however, included fecundity as well as ‘life-style category’ and ‘taxon’ and ‘voltinism’ (Table 1b). An analysis of the subset of 42 agents for which there was parasitoid diversity data indicated that dispersal rate was strongly correlated with the number of parasitoid species attacking the agent in the native range (anova, F1,40 = 14·4, < 0·001; Fig. 4). Alternative models were ranked according to AICc value and the model with the best relative fit included the predictors ‘parasitoid diversity’, ‘voltinism’, ‘taxon’ and ‘habitat’ (Table 1c). Parasitoid diversity varied significantly according to life-style category (anova, F4,37 = 7·09, < 0·001; Fig. S1a, Supporting information) and taxon (anova, F5,36 = 4·64, < 0·01; Fig. S1b, Supporting information).

Table 1.   The two best models for: (a) the full data set of 66 arthropod agents and for the subsets of 56 and 42 arthropod agents for which there were data regarding (b) fecundity and (c) parasitoid diversity in the native range
ModelModel parametersR2AICcΔi
(a)
1Log (voltinism); Taxon; Habitat; Life-style category0·584−84·9810
2Log (voltinism); Taxon; Habitat0·558−83·3981·58
(b)
1Log (voltinism); Log (fecundity); Life-style category; Taxon0·621−69·4670
2Log (voltinism); Log (fecundity); Taxon; Habitat; Life-style category0·636−69·8660·40
(c)
1Log (voltinism); Log (no. parasitoids); Taxon; Habitat0·726−57·1160
2Log (voltinism); Taxon; Habitat; Life-style category0·725−57·0310·09
Figure 4.

 Mean dispersal rate of arthropod biocontrol agents in their introduced range vs. parasitoid diversity in their native ranges: Log dispersal rate (km year−1) = 1·380 × log (No. parasitoid spp.) + 0·127.

Pathogen dispersal rates varied from c. 70 m year−1 for Puccinia myrsiphylli (Thuem) Wint to 280 km year−1 for Puccinia chondrillina Bubak & Syd. (Table S2, Supporting Information) and averaged 59·9 km year−1. As for arthropod biocontrol agents, dispersal rate was positively correlated with agent voltinism (anova, F1,9 = 5·65, < 0·05; Fig. 3b), but not with the time since release (anova, F1,9 = 0·21, n.s.).

Dispersal rate was in no way associated with agent’s success (i.e. both successful and unsuccessful agents were equally likely to be fast or slow dispersers). This held true for both arthropod (χ2 = 1·08, d.f. = 1, n.s.) and pathogen agents (χ2 = 0·40, d.f. = 1, n.s.).

Model validation

We used the best models from the three (64, 56 and 42 species) data sets of arthropod biocontrol agents to predict the dispersal rates of a suite of seven accidentally introduced insect herbivores for which native range parasitism, voltinism and fecundity data were available, that we could compare with published dispersal data (Table 2). Our predicted dispersal rates were generally close (same order of magnitude) to the actual dispersal rates with the exceptions of the emerald ash-borer Agrilus planipennis Fairmaire which dispersed two orders of magnitude more rapidly than predicted and Diaprepes abbreviates (L.), which dispersed at c. 20% of the predicted rate (see Table S4, Supporting information for a list of parameters used and worked examples showing how dispersal rates were estimated for models a1, b1 and c1).

Table 2.   Predicted and actual dispersal rates of accidentally introduced insect herbivores. Models used to derive parameter estimates: a1 & a2 = ‘best’ and ‘second best’ models, using the data set of 66 agents; b1 & b2 = ‘best’ and ‘second best’ models, using the data set of 56 agents including fecundity data; c1 & c2 = ‘best’ and ‘second best’ models, using the data set of 42 agents including parasitoid diversity data (see Table 1). V, voltinism (number of generations per year); F, fecundity; P, parasitoid diversity in the native range. All species attack terrestrial plants
AgentGuildVFPPredicted dispersal rate(km year−1) and model usedMeasured dispersal rate (km year−1)
a1a2b1b2c1c2
  1. 1Cappaert et al. (2005); 2Anon (2010); 3Liu et al. (2003); 4Muirhead et al. (2006);5Beavers (1982); 6Wolcott (1937); 7Hall et al. (2001); 8Miller & Tenhumberg (2010); 9Syrett & Harman (1995); 10Agwu (1974); 11Colbert, Xu & Jiang (1995); 12Forbush & Fernald (1896); 13Hoch et al. (2001); 14Liebhold, Halverson & Elmes (1992); 15Richards (1940); 16Andow et al. (1990); 17Clausen & King (1924);18Fleming (1972); 19Shigesada & Kawasaki (1997); 20Suckling et al. (2005); 21Farr (2002); 22Allen (1990).

Agrilus planipennis (Coleoptera)Wood-borer0·5175·02331·21·01·10·91·21·2up to 2504
Diaprepes abbreviatus (Coleoptera)Root-feeder0·9155793·55,6271·41·53·12·51·51·60·2968
Leucoptera spartifoliella (Lepidoptera)Stem-miner1930·67106103·23·03·02·56·55·76·439
Lymantria dispar (Lepidoptera)External feeder111450·01230132·13·02·02·013·43·7c. 3–2114
Pieris rapae (Lepidoptera)External feeder3·515356·0158156·59·09·48·419·810·715–15016
Popillia japonica (Coleoptera)Root-feeder11750·01810171·41·51·01·02·91·45–619
Uraba lugens (Lepidoptera)External feeder220403·72111225·75·24·44·113·96·41220

Discussion

The identification of factors closely correlated with insect dispersal, especially those that are relatively easy to measure/extract from known biological information, can provide attractive management tools. Our preliminary data exploration indicated that arthropod biocontrol agent dispersal was influenced by six of the eight explanatory factors and variables that we hypothesized were important. The model with the best relative fit for the 66 species data set, however, indicated that the most important predictors were ‘taxon’, ‘voltinism’, ‘habitat’ and ‘life-style category’. The best models for the 56 and 42 species data sets also included ‘taxon’ and ‘voltinism’, plus ‘fecundity’ and ‘life-style category’ and ‘parasitoid diversity’ and ‘habitat’, respectively (Table 1). Although it may seem intuitive that dispersal should vary according to taxon, this factor does not explain the underlying causal mechanisms associated with dispersal ability. We believe predictions according to taxon should be made cautiously, given the low sample sizes. For example, although Hemipteran biocontrol agents have generally dispersed only slowly, only one aphid has been used as a weed biocontrol agent (Aphis chloris; Table S1, Supporting information). Aphids can disperse rapidly (e.g. on air currents; Teulon & Stufkens 2002) and if another aphid was to be used in a biocontrol programme, it may disperse more rapidly than our current models would predict.

The strong correlation of other factors to dispersal is perhaps easier to interpret. As predicted, both arthropod and pathogen biocontrol agent dispersal rates were strongly correlated to voltinism which, in turn, is correlated to an organism’s intrinsic rate of increase (Shapiro 1975). Fecundity alone was not significant, but this predictor became important in models, perhaps because the interactions between fecundity and other factors such as voltinism are additive or multiplicative. The importance of the factor ‘habitat’ supports the hypothesis that arthropods adapted to aquatic habitats have undergone selection for good dispersal ability because they exploit ephemeral habitats that can dry out during droughts or become inundated or washed away by floods. The correlation between dispersal rate and parasitoid diversity in the area of origin may confirm our hypothesis that parasitoids exert a strong selection pressure for hosts to become good dispersers. An alternative hypothesis is that agents with a high parasitoid diversity in the native range are more likely to experience enemy release, and therefore, proliferate and disperse faster in their introduced ranges (Paynter et al. 2010). This could be tested by quantifying parasitism in both the native and introduced ranges, to determine if agents which escape parasitism disperse faster than those which do not. There are currently insufficient data to rigorously test this, although some of the most rapid dispersers are parasitized in their introduced range (e.g. Aphis chloris, Hydrellia pakistanae, Mesoclanis polana, Phytomyza vitalbae; McFadyen & Jacob 2004).

It is unclear why the factor life-style category is important; it may prove to be an indirect measure of parasitism: as previously noted (Hill & Hulley 1995), parasitoid diversity varied significantly according to life-style (Fig. S1, Supporting information) and the most heavily parasitized guilds (e.g. gall-inducers; miners, borers, leaf rollers and tiers) were among the fastest dispersers. Life-style may also be correlated to voltinism. For example, voltinism of seed or flower-feeders is constrained by the phenology of host plant reproduction. Food quality may also vary according to guild and, in turn, affect developmental time and therefore voltinism. There may also be a connection to the movement restriction of larval stages in some feeding guilds. For example, perhaps leaf-miners need stronger dispersal abilities to locate optimal hosts on which their offspring are confined, compared to species with larvae that can disperse to other host plants, should host-plant quality decline.

Model validation, predicting dispersal rate and biosecurity

Our predicted dispersal rates were generally close to measured dispersal rates (Table 2) with the exception of the emerald ash-borer Agrilus planipennis, for which human-mediated dispersal (for example in firewood) is known to transport significant numbers of individuals considerably further than they can naturally disperse (Muirhead et al. 2006). This reinforces the notion that the potential for human-mediated dispersal should be taken into account when predicting the dispersal of an invasive pest as well as when trying to establish a biocontrol agent.

Improving the models

Improving the predictive power of our models would allow optimization of release strategies: the geographic scale at which to release agents could be selected to best suit agents’ ability to close the gaps by natural dispersal. Our models could be improved in a number of ways. First, they often relied on records of voltinism in the native range, which may be an unreliable predictor of voltinism in the introduced range, or on simplistic estimates of voltinism, based on the published development times for each life cycle stage and the stated activity period of the organism. Temperature-based phenology models could be used to predict voltinism more accurately and, thereby, potentially improve on our current correlation between dispersal rate and voltinism. Such models, however, require knowledge of an insect’s base threshold of development and degree-day requirements for each life stage (Nietschke et al. 2007) and, while not impossible, it may prove impractical to gather this information during the host-range testing and mass-rearing stage of a candidate biocontrol agent.

Secondly although the variable ‘time since release’ did not explain variation in dispersal rate in our data sets, dispersal can certainly change over time. For example, Holloway & Huffaker (1951) noted that Chrysolina beetle dispersal changed from slow (crawling) to rapid (flight), once food became exhausted. It may be that many observations of dispersal were too short-term to detect changes in agent behaviour. It is also likely that the way much of the data were collected or reported would not have detected or indicated changes over time. For example, it was reported that Apion ulicis (Forster) dispersed 10 km in 6 years (Table S1, Supporting information), but it is not clear if dispersal was constant (as we have assumed, when calculating a dispersal rate of 1·67 km year−1) or accelerated over this period.

The importance of natural enemies is worthy of further research. In weed biocontrol programmes, native parasitoids of the candidate biocontrol agents are often studied on an ad hoc basis, if at all, and rearing is often confined to the biocontrol agent’s larval stage (Paynter et al. 2010). Differences in sampling effort may be a major cause of variation in our estimates of dispersal rate, because sampling effort should correlate with parasitoid richness on host species (Cornell & Hawkins 1993). Standardized estimates of parasitism in the native range may enhance our understanding of this factor. Moreover our argument that parasitoid diversity exerts a strong selection pressure for hosts to become good dispersers, assumes that parasitoid diversity equates to the impact that parasitism has on agent populations, which may not be the case. Percentage parasitism may be a better predictor of dispersal ability than parasitoid diversity, but it is difficult to measure accurately (Paynter et al. 2010). Moreover, we have not quantified the potential impact of other specialist natural enemies on dispersal as data regarding disease and predation is scant in the literature.

One potential criticism of our analysis of dispersal rate and agent success is that we failed to take into account the dispersal ability of the host weed. This is because an agent’s dispersal rate relative to that of its host plant may be critical for success. We were unable to test this rigorously due to a lack of data: We used CAB Abstracts® and Web of Science (ISI) (with the weed species and ‘spread’ or weed species and ‘dispersal’ as search terms) to search for quantitative spread data for the weed hosts of the biocontrol agents for which we had quantitative dispersal and biocontrol impact data. We found weed dispersal data that corresponded with dispersal data for only eight arthropod agents (Table S5, Supporting information). All but one of these dispersed more rapidly than its host weed, with the exception being Dactylopius opuntiae. Despite being a poor disperser D. opuntiae is considered to be a highly successful biocontrol agent in both Australia (Hosking, Sullivan & Welsby 1994) and in South Africa (Zimmermann & Moran 1991). However, it is likely that by dramatically reducing flowering, fruit and vegetative cladode production (Hosking, Sullivan & Welsby 1994), D. opuntiae will have significantly impaired the ability of its Opuntia host plants to disperse. Measuring the dispersal rate of a weed prior to the introduction of a biocontrol agent may not, therefore, be a good predictor of whether biocontrol will succeed or fail.

In conclusion, the results of this study are based on a limited data set and we hope that our findings will encourage further research to enhance our predictive ability. The fact that the slowest dispersing insects Dactylopius opuntiae, Apthona spp. and Cyphocleonus achates are all highly successful at suppressing their target weeds (Table S1, Supporting information) illustrates how effort spent optimizing the release and redistribution of biocontrol agents that are slow to disperse (e.g. Foxcroft & Hoffmann 2000; Batchelor et al. 2004) is likely to be worthwhile.

Acknowledgements

We thank Dr Ronny Groenteman for reviewing an earlier draft of this manuscript. We thank AgResearch Limited for supporting this study as part of the programme Undermining Weeds (contract C10X0811) funded by the Foundation for Research Science and Technology, New Zealand.

Ancillary