Response to Allen ‘An alternative hypothesis to the conclusion of Colman et al. (2014)’

We respond to the criticisms of Allen [1] regarding Colman et al. [2], examining the differences in forest mammal assemblages between areas where dingoes were controlled and not controlled in southeastern Australia. Allen [1] contends that: (i) Colman et al. [2] did not show any evidence of dingo control in ‘Treatment’ sites; (ii) the positive relationship between dingo activity and small mammal abundance observed by Colman et al. [2] was likely a product of bottom-up responses, and not due to indirect effects of dingo removal; and (iii) demonstrating trophic cascades resulting from top-predator control is only possible through manipulative experiments. In our response, we explain why we disagree with each of Allen's criticisms.

We reject Allen's [1] claim that the effect of poison baiting on dingoes was not shown. Allen [1] focuses on the activity of dingoes being influenced by the timing of poison baiting campaigns. Allen [1] contends that these only reduce dingo abundance in the short term (less than 1 year) and that rapid invasion of dingoes can occur soon after [3], (but see [4]). Allen [1] inappropriately used paired t-tests to analyse data from the electronic supplementary material of Colman et al. [2] to argue that dingo activity and fox activity did not differ between baited and unbaited pairs. Allen's paired t-tests tested whether the mean of the differences in the mean values of predator activity at the site-pairs differed from zero. His analyses did not take into account the variation in predator activity within the sites making up each pair. In Colman et al. [2], we treated each site-pair as an independent comparison and used the grand effect size to test the hypotheses that poison baiting influenced dingo and fox activity. We then calculated an effect size for each site-pair using the mean and standard deviation for the predator activity index within each site making up the pair, and then pooled the results of the seven comparisons using a meta-analytic approach to calculate a grand effect size. We used the log response ratio, a widely employed standardized metric of effect size as our response variable [5] (figure 1a,b), and in this response have also presented the same data using an alternative metric of effect size, Hedges’ d [6] (figure 1c,d). Tests for homogeneity of the effect sizes between site-pairs were conducted using the Q-statistic. Non-significant (p > 0.05) Q-values indicated that the effect sizes were homogeneous (figure 1).

Figure 1.

The effect of dingo control at the seven baited–unbaited site-pairs examined by Colman et al. [2] measured using the log response ratio (±95% CI) for (a) dingo activity (Q = 5.02, d.f. = 6, p = 0.541) and (b) fox activity (Q = 7.43, d.f. = 5, p = 0.190) and measured using Hedges’ d for (c) dingo activity (Q = 6.45, d.f. = 6, p = 0.375) and (d) fox activity (Q = 5.32, d.f. = 6, p = 0.379). The grand effect size shown in each graph represents the average effect size generated when the data from the seven sites were pooled. Negative values indicate variables that decreased in response to dingo control; positive values indicate variables that increased in response to control. The effect sizes were considered statistically significant if the 95% CIs excluded zero.

Our results showed that dingo and fox activity responded in different ways to the poison baiting treatment. Both effect size metrics showed that dingo activity was greater at unbaited sites than nbaited locations at three of the site-pairs, approached significance (p < 0.05) at Wollemi NP and displayed neutral responses at three sites. The grand effect size for both metrics showed that on average dingo activity was greater at unbaited than baited sites (figure 1a,c). Conversely, both effect size metrics showed that fox activity was greater at baited than unbaited locations at three of the site-pairs and showed neutral responses at three other site-pairs. Both effect size metrics showed that on average fox activity was greater at baited than unbaited sites (figure 1b,d).

Allen [1] also rejects the null hypothesis that the mean difference of both dingo and fox activity differ from zero at p < 0.05, but makes no effort to examine the effect sizes [6]. Plots of the effect size metric used by Allen [1] with 95% confidence intervals (CIs) suggest that dingo and fox activity respond in opposite ways to dingo control and indicate that the 95% confidence intervals of both variables only narrowly include zero (figure 2). The CIs for dingo and fox activity indicate that the 89% CI for dingo activity (dingo mean difference = −0.1557, 89% CI = −0.30627 to −0.00515) and 91% CI for fox activity (fox mean difference = −0.062866, 91% CI = 0.00106 to 0.12466) exclude zero. Hence, the CIs indicate that if the sampling regimen of Colman et al. [2] was repeated a large number of times, dingo activity would be lower at unbaited than baited sites on approximately 89% of occasions and that fox activity would be greater at baited than unbaited sites on approximately 91% of occasions.

Figure 2.

A plot of the effect size metric used by Allen [1] in his paired t-test on the data of Colman et al. [2] comparing dingo and fox activity at baited and unbaited sites. The dots represent the mean difference in the mean of dingo and fox activity at baited and unbaited sites examined by Colman et al. [2]. The error bars indicate 95% CIs. Negative values indicate variables that decreased in response to dingo control; positive values indicate variables that increased in response to control. The effect sizes were considered statistically significant if the 95% CIs excluded zero.

Allen [1] criticizes Colman et al.'s [2] structural equation model (SEM) for not investigating whether small mammal abundance was a potential driver of dingo abundance. In arid regions of Australia, synchronous increases in the abundances of dingoes and red foxes have been documented to occur following irruptions of rodent populations that accompany extreme rainfall events [7]. However, there is little evidence that population increases of small mammals are a driver of dingo abundance in the forested ecosystems of southeastern Australia, where dingoes prey primarily on larger species such as kangaroos and wallabies [8,9]. Indeed, this is why we did not include a pathway in the SEM predicting a positive effect of small mammals on dingoes. In addition, if bottom-up effects arising from higher small mammal abundances were the driver of predator abundances at the sites of Colman et al. [2], it would be expected that both dingo and fox numbers would be positively correlated with small mammal abundance in the final SEM of Colman et al. [2]. However, it was found that fox activity correlated negatively with small mammal abundance, coinciding with their predilection for smaller prey.

We conducted an SEM to test Allen's [1] hypothesis that small mammals may have influenced the abundance of dingoes (figure 3a) (see [2] for SEM protocol). We predicted in the a priori SEM that baiting should have a negative effect on dingo activity and applying Allen's hypothesis predicted that small mammal abundance should have a positive effect on dingo activity. After stepwise elimination of non-significant terms, the most parsimonious SEM showed that baiting was the best predictor of dingo activity (figure 3b).

Figure 3.

(a) A priori piecewise structural equation model (SEM) describing dingo responses to small mammal abundance. (b) Most parsimonious SEM. Dashed lines represent negative interaction pathways, and solid lines represent positive interaction pathways. Path co-efficient estimates are shown alongside arrows and deviance explained (d.e.) is shown for all variables.

We agree with Allen's comment that a replicated manipulation of dingo abundance with before and after observations would provide stronger inference on the direct and indirect effects of dingo control than the ‘natural experiment’ evaluated by Colman et al. [2]. However, we disagree with his view that no inference regarding the effects of dingo control can be made from non-manipulative experiments. Observational studies in which there is no experimental manipulation are widely used to provide inference in ecology and other disciplines. Observational studies are particularly useful in situations where the scale of the manipulation required to conduct an experiment is difficult or impossible to achieve or in situations where legal and ethical considerations may preclude experimentation [10]. In Colman et al. [2], we tested predictions generated from trophic cascade theory and the mesopredator release hypothesis. The results showed strong concordance with our a priori predictions, and hence support for the hypothesis that the direct and indirect effects of lethal control of dingoes structure forest ecosystems. Moreover, support for the hypotheses tested in Colman et al. [2] was bolstered by concordance between these results and those observed in previous studies [11–15]. We argue that it is unlikely that some source of variation other than the presence/absence of dingo control could have caused the consistent effects on dingo activity, fox activity, macropod activity, small mammal abundance and understorey vegetation that have been reported in comparisons of areas where dingoes are controlled and not controlled in forest and desert ecosystems [11,14,15].

In addition to the logistic challenges of manipulating dingo abundance, legal and ethical issues make the manipulation of dingo abundance problematic. Because dingoes are perceived to be a threat to livestock producers [16], laws in New South Wales (where Colman et al. was undertaken) mandate that dingoes must be controlled in all areas except those designated for dingo conservation. Similar laws exist in other jurisdictions in Australia. The primary method for dingo control is the distribution of poison baits. Thus, there are few areas where dingoes are now baited where it would be lawful to cease baiting. Similarly, if dingoes were experimentally reintroduced, it would be illegal not to control them under existing legislation. The alternative, experimental approach then is to commence dingo control in areas where they are currently not controlled. Given the evidence derived from observational studies suggesting that dingo control has strong, potentially adverse effects on ecosystems [11,13,15], studies reducing dingo populations in areas where they are currently not controlled could have unacceptable consequences. The logistic, legal and ethical constraints on the conduct of manipulative experiments on dingoes highlight the practicality of using well-designed observational approaches and opportunities made possible by existing contrasts or variations in dingo activity/control regimes to evaluate competing hypotheses generated from theory and previous studies.