Using expert elicitation to predict feral cat, Felis catus, responses to management

Abstract

BACKGROUND

Domestic cats that have become wild are primary drivers of species' decline globally. Multiple tools and strategies exist for managing cat populations; however, in large, unconstrained rural and remote areas effectiveness is often highly variable. Previous work showed cat populations should be reduced by at least 57% annually, on average, to achieve a sustained population reduction. We assess whether available lethal management scenarios in Australia were likely to meet this reduction threshold. We held a 2‐day expert elicitation with 24 experts. Experts gave informed estimates of cat population reduction for 648 management scenarios combining lethal management techniques (e.g., aerial poison baiting), management decisions (e.g., poison bait type) and environmental conditions. We extrapolated the outcomes to an additional 216 scenarios (total 864) and combined the results to estimate when a scenario was likely to meet the population reduction threshold.

RESULTS

Expert estimates indicated 71 of 864 management scenarios met the population removal threshold. All 71 scenarios involved more than one management technique using combinations of cat‐targeted baits, Eradicat (n = 51) and Curiosity (n = 17), or fox baits (n = 3). Scenarios in ‘Deserts and xeric scrublands’ and ‘Grasslands, Savannas and Shrublands’ were the most likely to be successful, particularly in below average rainfall years.

CONCLUSION

The expert elicitation and analytical techniques used here provide managers with practical guidance to assess the likely success of planned programs for effective lethal feral cat control. This study presents a flexible framework using expert knowledge to explore management options for more effective conservation strategies. © 2026 The Author(s). Pest Management Science published by John Wiley & Sons Ltd on behalf of Society of Chemical Industry.

We generated estimates of lethal feral cat management success for 864 scenarios, only 71 lethal management scenarios met a population reduction target of 57%. Cat‐targeted poison baits, in dry conditions and arid areas were most successful. Feral cat management requires continued adaptive strategies to reduce cat impacts.

1. INTRODUCTION

Globally, uncontrolled wild populations of domestic cats (Felis catus) are implicated in the decline and extinction of bird, mammal and reptile species, ¹ , ² through predation, ³ , ⁴ competition, ⁵ and to a lesser extent, disease transmission. ⁶ , ⁷ Consequently, reducing cat impacts is recognized as a global conservation priority. ⁸ However, there is no universally effective method for reducing cat impacts. In remote areas, and where people occur at low‐densities broad‐scale reduction of cat impacts is mostly achieved by lethal control. The most successful lethal cat control programs have occurred on islands ⁹ or within fenced areas. ¹⁰ Broadscale management of cats across large, open continental areas is more difficult and remains a global challenge.

In Australia, broadscale lethal management of unowned or feral cat populations in rural or remote areas usually involves some combination of poison baiting, trapping and shooting. ¹¹ , ¹² Poison baiting is often used as it can cost‐effectively reduce cat populations at large scales. ¹³ , ¹⁴ However, poison baiting practice and success is also limited by a range of factors including: lack of access by managers to target‐specific baits, ¹⁵ non‐target species' impacts, ¹⁶ and reduced uptake or longevity of baits under certain conditions. ¹⁷ , ¹⁸ Alternative methods, such as trapping or shooting, can also remove cats from a population. ¹⁹ However, the effectiveness of baiting, trapping and shooting varies based on timing, ²⁰ , ²¹ local resources ¹⁵ and the skill of the staff implementing the program. ¹¹ , ²² Consequently, while land managers have management options available to them, there is uncertainty about management program success arising from these options due to variability in local populations and conditions.

Feral cat population dynamics are driven by changes in reproductive rates, prey availability and local environmental conditions, and these affect the outcome of management programs. ²³ , ²⁴ For example, high rainfall years can lead to increased prey densities; this can make feral cats less likely to consume baits, thus reducing the efficacy of poison baiting. ¹⁷ Conversely, in lower rainfall years, lower prey availability likely increases feral cat consumption of meat‐based baits or other food lures in poison baiting or trapping programs. ¹⁹ When feral cat population suppression is insufficient due to low management technique success or high reproductive rates, populations may not decline, or may quickly rebound after management. ²² Additional factors such as: reinvasion times ¹³ , ²⁵ ; dispersal ability ²⁶ ; and management area characteristics (e.g., land size ²⁷ ) can further influence the speed of feral cat population recovery. Understanding how local conditions and population dynamics impact management success is important for choosing effective strategies and consistently suppressing feral cat populations below target threshold levels.

Evaluating effective population suppression and consequently management program success is dependent on the program objective. Common objectives include the control of feral cats for eradication ¹⁰ or the targeted protection and conservation of species. ¹³ , ²⁸ A frequently used measure of success in these programs is feral cat population reduction, or the number of individual feral cats removed during management. ²⁹ However, assessing whether population reduction is successful can be difficult. One method to assess the success of management programs uses estimates of the annual population removal rates required to suppress population growth. ²⁴ Although the exact population reduction required will vary based on the local management objective and breeding rates of local cats, it has been calculated that an average of 57% (95% credible interval: 24–93%) of the population needs to be removed annually to reduce feral cat populations. ²⁴

Empirical data from replicated field experiments and real‐world management programs is highly valuable, but differences in how lethal cat control is conducted and reported make it challenging to synthesize and draw out consistent patterns. Recent work has attempted to summarize the efficacy of different baiting strategies for feral cats. ³⁰ However, research into the use of poison baiting in combination with other lethal techniques is limited. Further, reported measures of population change within the literature following management are highly variable, making robust inferences about population reductions difficult. ³⁰ For example, population change in many reports is measured using activity indices based on feral cat tracks ¹³ , ²⁵ while more recent studies tend to use a change in occupancy or density, based on camera trap imagery. ¹⁴ , ²⁸ Many researchers and practitioners hold valuable management insights that are not formally documented. Expert elicitation methods provide a means to document this knowledge in a structured workshop setting and combine local information in a standardized manner. ³¹

In this study, we aimed to estimate the probability that aerial and ground baiting programs – either alone or in combination with the supplementary methods of cage trapping, padded leghold trapping or shooting – would meet the threshold removal rate for feral cats proposed by Hone et al. ²⁴ In doing so, we aimed to assess how many management scenarios were likely to meet the threshold population reduction required to achieve a feral cat population decline. We worked with 24 experts to elicit reductions in feral cat populations under different Australian environmental and management conditions.

2. METHODS

We convened an 8‐h workshop in May 2023, with 24 experts and two observers with experience in feral cat management, research or policy. Experts were asked to quantify their knowledge on the efficacy of aerial and ground baiting programs alone or in combination with cage trapping, padded leghold trapping or shooting. Observers, while knowledgeable on issues of feral cat management, were not considered practitioners or researchers. They participated in discussions but did not complete the expert surveys. Five experienced facilitators and one note‐taker led the structured elicitation overseeing workshop design, discussion phases and collation of results. Facilitators did not contribute opinions or data but instead guided discussions using pre‐written key questions. The research was approved by the Human Research Ethics Committee at the University of New England (HREC Project Number: HE23‐012).

2.1. Expert elicitation

Participants were drawn from every Australian state and territory, with additional expertise sought from New Zealand. Experts were identified from an Australian National advisory body (the Feral Cat Taskforce) and literature searches for the five key management techniques examined in this study. To meet the criteria for inclusion in the workshop, experts were required to have research, management or policy experience relating to at least two of the following management techniques: aerial baiting, ground baiting, leghold trapping, cage trapping and shooting. In total, 68 potential experts were identified to meet these criteria and were invited to the workshop, with 24 experts agreeing to attend. The experts' experience ranged from <5 to >30 years for the listed management techniques (Table S1). Participants had experience from several different ecological regions around the country. Experts came from a range of sectors including Government, Universities and other conservation‐focused organizations (Table S1).

We used three key ecoregion groupings to allocate the participants to groups relating to their experience of cat management during the workshop. The three groups included: (1) Deserts and Xeric Shrublands (2) Forests, woodlands and scrublands (3) Grasslands, Savannas and Shrublands. These groups were identified and defined based on the work of a previous workshop, ¹¹ and were formed from the ecoregion descriptions previously used by the Australian Government. ³²

2.1.1. Management technique definition

Experts agreed on definitions for five key management techniques: aerial baiting, ground baiting, leghold trapping, cage trapping and shooting (Table 1). The techniques were selected based on the outcomes of a previous workshop identifying these techniques as the most broadly applied methods of lethal feral cat management around Australia. ¹¹ Definitions were developed from four key sources including: the standard operating procedures for each technique ³³ , ³⁴ , ³⁵ , ³⁶ ; methods reported in peer‐reviewed and grey literature; State and Territory legislation; and Australian Pesticides and Veterinary Medicines Authority permits. The definitions were circulated to all participants one week before the workshop to allow them to provide feedback on the suitability of the definitions.

Table 1.

Management technique definitions provided to experts during the workshop. Definitions were developed based on the standard operating procedures for each technique; the methods reported in peer‐review and grey literature; State and Territory legislation; and Australian Pesticides and Veterinary Medicines Authority

Technique	Definition
Aerial Baiting	Applied in late‐autumn or early winter at a rate of 50 baits/km2. Baits dropped from a fixed wing or helicopter. Flight lines are spaced at 500 m or 1 km intervals to meet density target for baiting region. Bait type varies based on State/Territory approvals. All baiting activity must avoid waterways and residential areas.
Ground Baiting	Applied along vehicle accessible trails and other linear features in the landscape. Depending on State/Territory can be 50 baits/km2 or 25 baits/km2. Bait type varies based on State/Territory approvals. All baiting activity must avoid waterways and residential areas.
Padded Leghold Trapping	Padded or soft jaw leghold traps installed as either single or paired units in the ‘cubby’ or ‘walk‐through’ formation placed along or adjacent to tracks or linear features. Traps are distributed at a density of 50 traps within 10 000 ha separated by a minimum of 200 m. Traps remain open for 5–10 consecutive days. Traps should be lured with a scent lure (most commonly cat urine/ feces mix). For the purposes of this workshop, we consider leghold trapping to be conducted within 4 weeks after baiting.
Cage Trapping	Wire cage traps located within 50 m of tracks or linear features (e.g., creek lines, fire breaks, fence lines). Traps are distributed at a density of 100 traps within 10 000 ha separated by a minimum of 200 m. Traps remain open for 5–10 consecutive days. Traps are baited with a food lure (e.g., chicken or fish). For the purposes of this workshop, we consider cage trapping to be conducted within 4 weeks after baiting.
Shooting	Nocturnal with aid of spotlight or thermal scope, can occur from a vehicle or on foot. Involves a team made up of 1 hunter and 1 spotter. Hunting should occur for 1 week and result in 4‐h hunting per night. For the purposes of this workshop, we consider shooting to be conducted within the 4 weeks after baiting. This does not include opportunistic shooting – only the targeted shooting of cats.

2.1.2. IDEA protocol

We used the IDEA protocol (‘Investigate’, ‘Discuss’, ‘Estimate’ and ‘Aggregate’) ³⁷ to quantify the impact of 25 scenarios using aerial or ground baiting programs alone or combined with cage trapping, leghold trapping or shooting. Experts were asked to independently estimate the expected change in the feral cat population (using number of cats removed) for a 10 000‐ha area. Experts were asked to parameterize their responses by specifying: (1) location (i.e., State, Territory, or Other); (2) ecoregion group; (3) bait type and (4) feral cat density (derived from maps ³⁸ ; literature; or prior knowledge). For the area described, experts generated estimates of the number of feral cats removed in each management scenario after 12 months under lower‐ and higher‐than‐average rainfall conditions (Round 1). Experts provided their estimates as the (1) ‘best’ guess, (2) minimum and, (3) maximum number of cats removed, along with their confidence (scored from 10% to 100%) that the range captured the true number removed. Experts were not required to provide estimates for all scenarios as management techniques could be outside their subject knowledge. The survey was hosted on the online platform Qualtrics ³⁹ (see Supporting Information for survey questions).

A discussion of the results was conducted within the expert's ecoregion groups after they had completed the investigation phase. During the discussion experts were shown the anonymized results of the round 1 estimates, and the potential reasons for variation in the survey responses were explored. Following this, a larger group discussion summarizing the outcomes from within each ecoregion group was conducted. During this discussion, experts emphasized that the time after which the impact of a program was measured should be adjusted from 12‐months after management to the end of the management program (as defined by the techniques in Table 1). Experts were asked to consider only a single, intensive round of management as described in the technique definitions, rather than ongoing or repeated management programs. Following this phase, experts were provided with access to their round 1 survey responses to review their estimates for the new end‐of‐program definition and any additional insights they gained from the discussion (Round 2).

2.2. Data analysis

Data processing was performed in R (v4.4.2) ⁴⁰ using the ‘tidyverse’ package (v2.0.0) ⁴¹ to evaluate the impact of each management scenario based on estimates from the second round of expert elicitation. The data preparation process is outlined in Supporting Information, and the full procedure is provided in the code repository online (https://github.com/adorph/A11_now-that-we-have-talked). We used each expert's confidence values to calculate the upper and lower bounds of feral cat removals within an 80% credible interval (recommended by Hemming et al. ³⁷ ). These bounds were derived by adjusting the experts’ upper and lower estimates using linear extrapolation to account for over‐ or under‐confidence of the experts (Hemming et al. ³⁷ ). Final values were truncated at 0 and 1 since removing more or fewer feral cats than the total population is not possible.

We then calculated feral cat management outcomes under average rainfall conditions by calculating the mean of experts' best, minimum and maximum estimates for wet and dry conditions and dividing this by the mean density estimate averaged across the wet and dry periods. The adjusted estimates were loaded into Microsoft Excel, ⁴² where experts' comments were reviewed. Some responses were excluded from further analysis based on experts' comments. Specifically, responses from the ecoregion ‘Tropical and Subtropical Moist Broadleaf Forests’ ³² were removed due to significant differences in the available management strategies, vegetation structure, and rainfall compared to other regions within this ecoregion grouping. One response for the ‘Other’ bait type, when a bait intended for rabbits was used, was also excluded due to its differences to the other meat‐based baits. Finally, responses for ‘QLD Curiosity’, ‘Fox 1080,’ and ‘Other dried meat baits’ were combined into the single category, ‘Fox 1080,’ due to high overlap identified from the literature in bait sizes (60–100 g), matrix (fresh meat), toxin and dosage (3–6 mg of 1080). This resulted in four bait categories for the final analysis, two baits manufactured for feral cat control – Eradicat (15 g manufactured chipolata style bait containing 4.5 mg of 1080) and Curiosity (15 g manufactured chipolata style bait containing 4.5 mg of para‐aminopropiophenone within a hard‐shelled delivery vehicle), and two meat baits for canid control – Fox 1080 (as described above) and Dog 1080 (meat bait of 200–500 g in size containing 6 mg of 1080).

Expert responses for each management scenario combination were summarized using equal‐weight aggregation. ⁴³ This method is considered suitable for summarizing expert responses when calibration questions (i.e. a set of questions closely related to the elicitation questions than can be used to evaluate expert judgement) are not used. ⁴⁴ Best estimates were averaged for each combination. Upper and lower confidence limits were derived by averaging the adjusted upper and lower bounds, respectively. Finally, we transformed the estimates to ‘proportions of the population removed’ by dividing the estimates by the number of feral cats in the area, calculated from the average of the densities provided by experts. We used the resulting values in a copula‐based resampling procedure (see below) to extract the relative influence of the baiting strategies from the supplementary techniques.

2.2.1. Copula‐based resampling procedure

We calculated the relative contribution of baiting and supplementary methods from the expert estimates to improve analysis within a Bayesian Network model (see below). To estimate total proportion removed and the contributions of each method, we used a copula‐based resampling approach. For each ecoregion, rainfall and bait type group, we defined a baseline proportion removed from baiting only scenarios and compared expert estimates of the total proportion removed to this baseline. To preserve non‐linear correlations in shape between the total removed relative to the baseline, we used a Gaussian copula to generate 1000 dependent samples from a uniform transformation of the triangular distribution. ⁴⁵ , ⁴⁶ These samples were back transformed to produce a joint realization of the total and baseline proportion removed.

Since supplementary techniques could only be applied in the weeks following a baiting event (Table 1 ), their removal contributions must be additional to the baseline proportion removed. To reflect this, we sorted each sampled pair, so the total proportion removed was greater than or equal to the proportion removed from baiting (this occurred for ~11% of resampled estimates). In baiting‐only scenarios, the total proportion removed was constrained to equal the baiting‐only proportion. Finally, the proportion removed by supplementary methods was calculated by subtracting the total from the baiting‐only proportion removed. This method preserved expert uncertainty while maintaining the dependency structure between management scenarios.

Separating the data into the relative contributions of baiting and supplementary methods allowed for more effective examination of expert responses in the Bayesian Network model. For supplementary management techniques expert responses were applied across broader ecological and climatic contexts only (three ecoregion groups × three rainfall groups), resulting in larger sample sizes. This allowed for more stable and representative estimates of their effectiveness. In contrast, baiting techniques were examined across a finer set of groupings (three ecoregion groups × three rainfall groups × four bait types), resulting in smaller sample sizes per group. This finer resolution was necessary to capture the greater variability in baiting outcomes, but also meant estimates were based on fewer expert responses within each subgroup.

2.2.2. Imputing missing values

About 25% of scenarios had no data provided by experts. From discussions during the workshop, we assumed the relationships between management approach, environmental variables and the scale of feral cat reduction were related. Therefore, we imputed the data for these scenarios using Multivariate Imputation by Chained Equations [MICE] using the ‘mice’ package in R. ⁴⁷ This approach uses the correlations between variables to predict the outcome in missing scenarios. For example, if expert knowledge indicated that aerial baiting effectiveness was lower in Forests and Woodlands relative to Deserts and Xeric Shrublands across multiple bait types, MICE would use this relationship to estimate outcomes for combinations where specific ecoregion‐bait type scenarios were not directly assessed by experts.

Missing values for the total proportion removed, proportion removed (baiting methods), and proportion removed (supplementary methods) were imputed, employing the random forest (RF) method. Random forests are a non‐parametric modelling technique with reduced risk of overfitting compared to alternative parametric methods. ⁴⁸ , ⁴⁹ Model tuning with the ‘randomForest’ package in R (v4.7–1.2) ⁵⁰ indicated the default RF parameters were optimal for use in the imputation. For each scenario with missing data, multiple imputation was performed, generating 1000 complete chains, each with 10 iterations to check convergence. All variables were used to impute the missing values. Convergence was assessed across iterations, and the imputed distributions were compared to initial expert estimates to ensure plausibility (Fig. S1). The resulting imputed dataset was combined with the expert estimates for downstream analyses.

2.2.3. Bayesian network modelling

A Bayesian Network is a directed acyclic graph representing variables and their conditional dependencies. ⁵¹ A discrete Bayesian Network was constructed using the resampled expert estimates and the imputed missing data. The combined expert and imputed proportion removed estimates were discretized to 10% increments. Discretization was preferred over a continuous Bayesian Network approach as it (1) reduced uncertainty in the estimates, (2) simplified interpretation of the results, and (3) increased computational efficiency. ⁵²

The Bayesian Network was built in R using the ‘bnlearn’ package (v5.0.2). ⁵³ Input (parent) nodes represented the management scenario variables, including the management techniques, ecoregion group, annual rainfall condition, and bait type. The response (child) nodes contained the resampled estimates of feral cat population reduction from experts and the imputed data. Arcs, representing the potential direction of causal relationships between variables (i.e., nodes), were added to the model based on the structure of the online questionnaire, feedback from experts during the elicitation procedure and following a structure learning procedure in ‘bnlearn’.

To learn the Bayesian Network structure, we applied both whitelist and blacklist model constraints. Whitelist conditions specify mandatory arcs (i.e., dependencies) within the network and were applied exclusively to management technique nodes to reflect the expert elicitation survey structure. Blacklist conditions specify prohibited arcs (i.e., independence) and was used to prevent implausible arcs in the network structure. For the model structure learning procedure, we specified the ‘whitelist’ condition that Leghold Trapping, Cage Trapping, and Shooting were dependent on Aerial or Ground Baiting as experts were unable to answer questions without the use of Aerial or Ground baiting. This meant the final network structure had to contain arcs from Aerial or Ground baiting to each of the supplementary management techniques. Additionally, we specified the following ‘blacklist’ conditions:

1. Ecoregion group, Rainfall group and Bait type were independent of management techniques.

2. Measures of the proportion of the population could not be parent nodes to environmental or management nodes.

3. Proportion removed (baiting methods) was independent of supplementary technique decisions.

4. Proportion removed (supplementary methods) was independent of, and therefore not directly affected by, baiting decisions.

5. Total proportion removed was treated as a derived variable (the sum of baiting and supplementary effects) and was therefore independent of all environmental and management decisions.

We implemented these rules in a structure learning procedure following the method outlined by Leonelli et al. ⁵⁴ using an 80% training and 20% test dataset randomly sampled from each scenario. In learning Bayesian network structures from data, two approaches are typically used. The first is a score‐based method which searches all possible network structures and evaluates each using a statistical scoring function (e.g., AIC or BIC). The second method uses constraint‐based algorithms which infer network structures by testing conditional independence relationships within the data and constructing structures based on these. ⁵⁵ We tested seven structure building procedures: five used different score‐based evaluations (Bayesian Information Criterion [BIC], Akaike Information Criterion [AIC], factorized normalized maximum likelihood [FNML], Cooper and Herskovits's K2, locally averaged Bayesian Dirichlet [BDLA]) within a tabu search‐based algorithm, and two used commonly employed constraint‐based algorithms (PC‐stable and Grow‐Shrink). ⁵⁵ , ⁵⁶

Model structures were compared using a global monitor from the ‘bnmonitor’ package. ⁵⁴ The model structures and the global monitor outputs are shown in the Figs S2 and S3. The structure learning procedure showed all ‘tabu’ search‐based learning algorithms returned the same global monitor. The stability of these models and the implausible structures for exploring population reduction proposed by the constraint‐based algorithms meant the ‘tabu’ structures were chosen. The arc between Aerial Baiting and Ground Baiting was the only unstable relationship in all ‘tabu’ search‐based learning algorithms. As the strength of relationship between the two variables was close to 0.5 in all models, we considered the arc undirected and removed it from the structure. Using the final structure, we learnt the conditional probabilities for all nodes using the complete dataset.

We evaluated the Bayesian Networks performance using both classification accuracy and the area under the receiver operating curve (AUC). ⁵⁷ AUC was used in addition to classification accuracy as it provided a more comprehensive assessment of how well the model distinguished between population reduction intervals across probability thresholds. This was important to account for uncertainty in the original expert estimates, where strict classification to a single category did not always reflect the range of possible outcomes covered by the uncertainty in an expert's original estimated interval. To generate these metrics, 500 simulations using average likelihood weighting were performed with all the available nodes as evidence to predict the feral cat population reduction interval. The simulations returned two outputs for each scenario (1) the reduction interval with this highest conditional probability and (2) the conditional probability distribution for all reduction intervals. The first output was compared with the original value provided to the model to generate a confusion matrix (comparing the model predicted values against the actual values) and classification accuracy using the ‘caret’ package (v7.0–1). ⁵⁸ The second output was used to calculate the AUC for multiple classes following Hand and Till ⁵⁹ using the ‘pROC’ package (v1.18.5). ⁶⁰

Finally, we used the conditional probability tables (CPTs) for population reductions across all scenarios to evaluate whether they could meet or exceed the maximum removal rate required to suppress feral cat populations, P = 0.57 as estimated by Hone et al.. ²⁴ We applied a reverse cumulative sum to all CPTs to assess the probability of achieving each 10% population reduction interval. A threshold of 50% was used to determine whether a given outcome was more likely to occur. To assess whether a scenario's predicted population reduction was likely to meet a threshold of >57%, we focused on scenarios with a probability greater than 50% of achieving a reduction of 60–70% or higher (i.e., P ≥ 0.6–0.7). Finally, the CPTs of the response node were visualized for all scenarios using ‘ggplot2’ (v3.5.1). ⁶¹ Summaries of whether a scenario was likely to meet the extremes of the 95% credible interval for the population reduction threshold (i.e. 24–93%) were also generated and are provided in Figs S4 and S5. All code used for data preparation, resampling, imputation and analysis, and the anonymous resampled dataset is provided online https://github.com/adorph/A11_now-that-we-have-talked.

3. RESULTS

Twenty expert responses to the online survey were used to evaluate the impact of each management scenario. Responses provided information for 75% (n = 648) of the possible 864 combinations for the management techniques used, ecoregion group, annual rainfall conditions group and bait type. Estimates for the remaining 25% of combinations were obtained from the MICE imputation procedure (Fig. S1). The combined expert estimates and imputed data were used in a Bayesian Network to predict which management scenarios had a chance of meeting a population reduction threshold of >57% for feral cats. The final Bayesian Network structure is presented in Fig. 1. The model returned an AUC of 91.1% for the response variable ‘Proportion Removed’ (classification accuracy = 52.97%, Table S2). The high AUC indicates the model reliably ranked scenarios according to their likelihood of success. However, the overall classification accuracy of the model was moderate. In practice, this means it was able to predict the relative outcomes of management scenarios based on the input variables, but it could not perfectly assign each case to its correct category.

Figure 1.

Structure of the Bayesian Network built from expert provided estimates that was used to predict the outcome of management scenarios of feral cat populations.

Seventy‐one management scenarios met the annual population removal rates required to suppress cat population growth (Fig. 2). Forty‐five of these scenarios used aerial baiting in combination with ground baiting, 26 used aerial baiting only. No scenarios using ground baiting only. Of the 108 baiting only scenarios, only two were able to reduce cat populations below the threshold rate without the addition of supplementary methods. The supplementary technique most used in successful scenarios was Leghold Trapping, or Leghold Trapping in combination with Cage Trapping or Shooting. Most successful scenarios used Eradicat baits (n = 51), followed by Curiosity baits (n = 17) then Fox baits (n = 3). None of the successful scenarios used Dog 1080 baits. A much greater number of scenarios were able to meet the reduction threshold when a lower reduction target was used, with 653 scenarios meeting the lower credible interval of the population reduction threshold (i.e. 24%; Fig. S4). No scenarios were more likely than not to exceed the population reduction threshold when the upper credible interval of 93% was used (Fig. S5).

Figure 2.

Summary of management scenarios with orange squares indicating when experts consider managers are more likely than not to reduce the feral cat population by at least 57%. Management scenarios are divided into groups for Ground Baiting, Aerial Baiting or combined Aerial and Ground baiting for each bait type, ecoregion group and annual rainfall condition for below average, average and above average rainfall. Grey cells indicate scenarios where no supplementary management techniques were used. AB, Aerial Baiting; CT, Cage Trapping; GB, Ground Baiting; LT, Leghold Trapping; SH, Shooting. Imputed results are shown with red text and a red dashed border.

Scenarios within ‘Deserts and xeric scrublands’ and ‘Grasslands, Savannas and Shrublands’ were the most likely to be successful, particularly in below average rainfall years (Fig. 2). Only four scenarios were identified as effective in ‘Forests, Woodlands and Scrub’, all using Eradicat baits in below average rainfall years.

Overall, the greatest number of successful scenarios occurred when annual rainfall conditions were below average (n = 52; Fig. 2). When annual rainfall conditions were above average, successful scenarios (n = 5) used Eradicat baits in ‘Deserts and xeric scrublands’ or ‘Grasslands, Savannas and Shrublands’ with at least two additional management techniques. For the full set of probability distributions predicted by the model see Figs S6–S14.

4. DISCUSSION

Our study explored expert judgement on the efficacy of various lethal management scenarios in Australia. Of the 864 potential scenarios, experts considered that only 71 surpassed a 57% population reduction threshold. These successful scenarios primarily utilized feral cat specific baits (‘Eradicat’ and ‘Curiosity’) and were more likely during below‐average rainfall years or in ‘Deserts and xeric shrublands’ or ‘Grasslands and savannas’. Programs using Fox 1080 baits, Dog 1080 baits or ground baiting alone were mostly considered insufficient for achieving necessary population reductions. While combining all available techniques increased the proportion of the cat population removed, these integrated management scenarios still often failed to meet the threshold for population reduction. This reflects the challenges that managers face in designing effective lethal feral cat population control programs. Future research should focus on improving the effectiveness of control techniques, exploring innovative methods, and assessing the long‐term impacts of suppression strategies to develop more effective management frameworks for feral cat populations.

4.1. Baiting program success

Experts judged that poison baits manufactured specifically for feral cat control, namely Eradicat and Curiosity (in Australia), contributed most to the success of management scenarios. Although experts recognize the potential for Fox or Dog 1080 baits to reduce feral cat populations, their perceived impact was insufficient in all but three scenarios to meet the average threshold for population reduction. Moreover, these canid‐specific baits are not approved for feral cat management programs in Australia, ⁶² limiting their practical applicability. Consequently, regions without registered cat‐targeted baits may face challenges in controlling feral cats. Expert assessments of canid‐ versus cat‐specific baits align with research showing greater efficacy of cat specific baits for feral cat population reduction. ⁶³ However, the relative success of these baits may vary under different conditions, with other studies showing low uptake of both feral cat baits and alternative bait types (i.e., Fox 1080 baits). ⁶⁴ Further refinement of feral cat baits and deployment methods is needed to maximize their impact across different conditions.

Environmental conditions contributed to experts' opinions on overall management scenario efficacy, with poison baiting programs being sensitive to changes in annual rainfall conditions and ecoregion group. Experts considered programs more effective when the annual rainfall condition was below average, presumably because poison baiting is usually more effective in lower rainfall conditions when prey availability is low and hunger‐driven bait consumption increases. ¹⁷ , ²⁰ Some experts noted the effects of rainfall on lethal control may be exacerbated for different demographics of the population, with younger or naïve cats more likely to take a bait, compared to older, more experienced hunters. These patterns highlight the importance of integrating environmental conditions and population dynamics into baiting strategies to optimize outcomes.

Expert judgements also highlighted how ecoregion group in combination with different annual rainfall conditions could further influence the success of management programs. For example, scenarios in below average annual rainfall conditions and arid vegetation, such as ‘Deserts and xeric shrublands’, were more effective at reducing cat populations than their counterparts. Conversely, experts considered baiting programs in high rainfall ecosystems, like ‘Tropical and subtropical rainforests’, ineffective due to dense canopy cover, numerous waterways, and rapid bait degradation. Similar considerations of vegetation and site‐specific constraints also apply to the other management techniques experts evaluated.

According to expert opinions, management scenarios using aerial versus ground baiting strategies had very different outcomes. Aerial baiting was considered to cause greater population reductions in all scenarios due to its more extensive coverage and independence from site accessibility constraints (e.g., road or track access). Ground baiting was deemed by some as redundant or less cost‐effective compared to aerial baiting programs. However, others noted linear features, such as roads, can facilitate feral cat movements ²⁶ , ⁶⁴ and that targeting these areas with ground baiting can contribute to increased feral cat population knockdown. Consequently, while experts indicated ground baiting alone could not achieve population reduction thresholds, it was considered valuable as part of a combined approach depending on context, with additional benefits apparent under specific conditions.

4.2. Supplementary management techniques

Cage trapping programs contributed minimally to population reduction in most scenarios. This result likely stems from neophobic behavior in cats, ⁶⁵ rainfall‐driven prey availability, population structure, ¹⁹ and the resource‐intensive nature of most cage trapping programs. These programs demand considerable time and resources yet often yield minimal returns, especially under certain environmental conditions. ¹⁵ Further, like poison baiting programs, cage traps usually rely on a food lure to attract feral cats, ³³ making them susceptible to similar efficacy limitations. Experts noted high variability in success, with cage trapping most effective in management programs when targeted to areas adjacent to farmland, where feral cat capture rates improve, and non‐target bycatch can be reduced. Even with potential improved outcomes in agricultural settings (not explicitly tested in this study), experts generally considered cage trapping too resource‐intensive and impractical in feral cat management when alternative methods were available. However, if implemented at large scales, potentially assisted with technology to alert managers when traps are activated, cage trapping can be effective in some contexts.

Leghold traps were considered by experts as complementary to management scenarios using baiting or cage trapping, offering a management option unaffected by prey availability (legholds used only scent lures). However, their use was restricted to accessible areas and demands significant skill, time and labor, resulting in a low overall relative contribution to feral cat removal in management scenarios. Leghold trap use is further limited, under certain circumstances, by environmental, non‐target, or legislative constraints. For instance, non‐target species are sometimes reported by‐catch in leghold traps, and may be susceptible to serious injury, ⁶⁶ while conditions, such as heavy rainfall, can impair their functionality and effectiveness e.g. McCarthy. ⁶⁷ as cited by McCarthy et al.. ⁶⁸ Nonetheless, they remain a valuable tool to complement other methods in some scenarios, enhancing the overall adaptability and effectiveness of feral cat management programs and increasing cat removal.

Shooting was another tool that experts noted was not limited by prey availability or population dynamics and was less impacted by low road accessibility, making it valuable in certain scenarios. In combination with leghold traps and baiting, scenarios incorporating shooting increased cat removal above the threshold required to reduce cat populations. Experts emphasized visibility was a critical factor influencing the success of shooting programs, making them more viable along agricultural perimeters, or within grasslands and open forests rather than dense vegetation. One expert indicated shooting was ineffective in the dense vegetation and closed canopies of some ‘Tropical and subtropical moist broadleaf forests’, where they had recorded only 25 cat detections from spotlighting efforts over 30 years. Thermal imaging technology could improve outcomes in some environments, with increased ability to detect individuals. ⁶⁹ However, further work is needed to determine the relative benefits of shooting programs in these denser ecosystems.

4.3. The use of population reduction thresholds

Throughout this paper, we have used the average population reduction threshold of >57% proposed by Hone et al. ²⁴ to evaluate the effectiveness of management strategies. However, we acknowledge this threshold was applied uniformly in our elicitation study, without explicit consideration of how environmental variability might influence its suitability. Factors such as cat reproductive rate, rainfall variability, prey abundance, or predator–prey dynamics may affect the level of population reduction needed. ⁷⁰ , ⁷¹ For example, higher rainfall years with greater prey availability and feral cat density may require reductions closer to 90% (the upper credible interval proposed by Hone et al. ²⁴ ), while lower rainfall years may only need reductions of 20% (lower credible interval). Adjusting thresholds based on the ecological conditions could lead to better evaluations of management success. However, there is still a critical knowledge gap in understanding density‐damage functions—i.e., the relationship between predator density and the magnitude of ecological impact. More empirical data are required to understand and estimate how much the predator density must be reduced to meaningfully alleviate pressure on different prey species and communities under varying environmental conditions. ²²

Operational and logistic constraints may also influence our ability to identify useful population reduction thresholds. This study asked experts to evaluate the impact of a single application of various management techniques—baiting, trapping, and shooting—on feral cat populations. However, ‘one‐off’ events, per year, are rarely sufficient for meaningful knockdown and sustained or repeated management is required to meet program objectives. ²² Additionally, if control programs continually suppress population densities below the reproductive rate, the effort needed to locate and remove the remaining individuals can increase substantially. ¹⁰ This makes achieving a proportional population reduction target more difficult as operational efficiency declines over time.

4.4. Performance of imputed scenarios

We used imputed values for 25% of management scenarios where data were missing because experts either lacked knowledge of these specific combinations or had already provided feedback on alternative scenarios and could not respond twice. The missing combinations included scenarios for Fox 1080 baits in Forest and Woodlands, Dog 1080 baits in Deserts and Xeric Shrublands, and Curiosity baits within Forests and Woodlands. Of these, relevant information could only be found in published and grey literature on the performance of Curiosity baits for cat population reduction in Forests and Woodlands. Reports from the south‐east coast of Australia indicate that bait uptake of Curiosity baits during aerial or ground baiting programs is low (0–2% ⁷² , ⁷³ ). However, evidence also suggests that in wetter than average conditions a combined aerial and ground baiting program could reduce the feral cat population by 50%. ⁷⁴ This is supported by an aerial and ground baiting trial run on Kangaroo Island which, depending on the metric used to measure reduction, indicated a population reduction between 50% (density change with potential immigration of new individuals) and 75% (radio‐tracked cat mortality). ⁷⁵ These estimates mostly fall within the range of values predicted by the model produced here but indicate that the imputed data may be underestimating the impact of these management scenarios. However, the limited number of reported results and different responses measured between the studies makes robust conclusions about the most likely population reduction outcome from these scenarios difficult. Consequently, the results from the imputation process should be treated with a degree of caution, and further examination of potential population reductions under different management scenarios should be conducted.

5. CONCLUSIONS

Feral cat management in large, open landscapes remains a complex challenge with ecological, sociological, ethical and legal considerations. ⁷⁶ Consequently, management responses are affected by a number of different factors, and they vary by region. ⁷⁷ Feral cats have impacts globally and it is often necessary to manage feral populations with lethal control for biodiversity conservation outcomes. ²⁸ Some reviews have examined the success of lethal feral cat management across multiple management areas – particularly for islands and poison baiting programs. ⁹ , ³⁰ However, there is a lack of research on the efficacy of combined lethal control methods for suppressing feral cat populations.

Empirical data from replicated field experiments and real‐world management programs remain highly desirable but differences in how lethal cat control is conducted and reported make it challenging to synthesize and identify consistent patterns. To address this, we used expert knowledge to model likely outcomes for five of the commonly used lethal feral cat management techniques, under different rainfall and ecoregion groupings. Using a previously calculated average population reduction threshold for cats, most management scenarios were unlikely to result in feral cat population decline. In the management scenarios where cat population decline was likely, the use of cat‐targeted baits was most often the key to success. Conducting baiting in drier than average conditions may also help managers reduce cat populations. Integrating other management methods with cat‐targeted baits did not always help to achieve required population reductions. Our findings can aid managers, and funding bodies, to refine their expectations and strategies for planning and implementing lethal feral cat control.

The absence of a universal approach to cat management, inconsistent reporting and limited resources make it difficult to assess the effectiveness of management programs. This inhibits the ability of land managers and researchers' to learn from past efforts and refine their strategies. ³⁰ To overcome these challenges, we require flexible, evidence‐based frameworks to guide decision making. While the results presented in this study do not have global applications, the approach offers a promising way to overcome data gaps and funding limitations. Importantly, these estimates are based on expert elicitation and carry inherent uncertainty, continued field experiments will be essential to test and refine the model with empirical data – particularly for the scenarios relying on imputed data. The framework supports the development of management strategies that can be refined over time through an adaptive management approach.

FUNDING INFORMATION

This project was supported with funding from the Australian Government under the National Environmental Science Program's Resilient Landscapes Hub and the NSW Government through the Environmental Trust.

CONFLICT OF INTEREST

The authors declare no conflicts of interest.

ETHICS STATEMENT

This research was approved by the Ethics, Grants and Research Integrity team at the University of New England (HREC Project Number: HE23‐012, Valid to: 31/12/2023).

Supporting information

Data S1. Supporting Information.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.