First you get the money, then you get the power: Comparing the cost and power of monitoring programs to detect changes in occupancy of a threatened marsupial predator

Ecological monitoring is crucial for tracking changes in the status of species over time. However, ensuring that monitoring programs possess adequate statistical power—capacity to detect changes in populations with a high level of confidence—remains a challenge for many wildlife managers globally. While new monitoring technologies potentially offer cost effective solutions to this problem, transitioning to these methods requires careful calibration with existing techniques, such that differences in power and cost can be measured and assessed accurately. Here, we compare new (camera traps) and conventional (live trapping) survey methods in terms of cost and statistical power in tracking occupancy declines in an endangered marsupial predator, the northern quoll (Dasyurus hallucatus). We show that camera trap monitoring designs can detect northern quoll occupancy declines of 30%, 50%, and 80% at reduced cost when compared to live trap designs, without compromising statistical power. Overall, we find support for the cost‐effectiveness of camera traps for species monitoring and its potential to replace existing live trap sampling of species when measuring changes in occupancy. Additionally, we offer a robust framework to compare new monitoring techniques against pre‐existing methods on the basis of statistical power.


| INTRODUCTION
Ecological monitoring-the conscientious and consistent collection of ecological information over time-is a vital component of effective biodiversity conservation (Likens & Lindenmayer, 2018). Ecological monitoring is crucial for understanding a species conservation status, detecting changes in species' populations, assessing species' responses to management interventions, and for tracking changes in the status of species over time (Lindenmayer & Likens, 2010;Reynolds et al., 2011). However, ensuring that monitoring programs have the ability to achieve these objectives, often with limited resources, remains a challenge for many wildlife managers globally (Dobson, 2005;Robinson et al., 2018;Singh & Milner-Gulland, 2011).
Inadequate design of monitoring programs can lead to low statistical power to detect trends of interest. Power refers to the probability that a given monitoring program will detect a statistically significant change of a prespecified magnitude (Guillera-Arroita & Lahoz-Monfort, 2012), and it may differ in response to species prevalence and detectability, the monitoring method, and the total survey effort applied-the number of sites multiplied by the number of repeat visits (Guillera-Arroita & Lahoz-Monfort, 2012). Disregarding the statistical power of a given monitoring program can lead to inaccurate conclusions about species' population trends. This may lead to inaction or implementation of inappropriate management, and consequently, misguided resource allocation (Einoder et al., 2018). It is thus crucial to understand if a given monitoring program has the ability to detect changes in populations through time.
New monitoring technologies can offer cost effective alternatives for long term monitoring programs that aim to detect population trends (Southwell et al., 2019), but transitioning to new technology presents several challenges. First, emerging and existing survey methods often differ in detection probabilities (e.g., bird surveys cf. acoustic recorders; Celis-Murillo et al., 2012), which makes comparisons between detection rates before and after transitions erroneous. Second, advances in remote surveillance mean that emerging and existing survey methods may differ in their deployment lengths (i.e., the number of sample occasions), leading to differences in cumulative detection probabilities over the course of a sampling session . Third, emerging technologies-particularly those that allow for remote surveillance-often have high costs for outlay (i.e., the purchasing of equipment), which can impact the number of sites that can be sampled under a limited budget (Welbourne et al., 2020). Each of these three factorsdetection probability, number of sample occasions, and the number of sites targeted-strongly influence the power of a monitoring program to detect changes in wildlife populations (Guillera-Arroita & Lahoz-Monfort, 2012). It is therefore crucial to understand how the detectability, power, and cost of emerging methods compare to more conventional approaches, as this will assist in determining whether we can successfully transition to using new technology, and whether this transition can be accommodated within the existing budget (and preferably at a reduced cost).
Passive infrared triggered cameras (hereafter referred to as camera traps) are a prime example of technological progress driving changes in survey methods. Camera traps hold considerable cost saving potential when compared to traditional survey techniques, such as live trapping, because they are able to operate remotely for long periods of time with few additional resources, leading to the collection of data over larger temporal scales (Rowcliffe & Carbone, 2008;Wearn & Glover-Kapfer, 2019). There are also distinct differences between the two techniques when considering welfare impacts; live trapping may require an individual to be constrained from its normal activities for many hours, whereas camera traps offer a much lower impact survey method. Many studies have directly compared the effectiveness of live and camera trapping surveys for monitoring wildlife populations (De Bondi et al. 2010, Welbourne et al., 2015, although cost comparisons between the two methods have primarily focused on resource requirements for optimizing detection probabilities. For example, Hohnen et al. (2019) found that camera traps are more cost effective than live traps for detecting Kangaroo Island dunnarts (Sminthopsis fuliginosus aitkeni). However, costing designs based solely on optimizing detectability can be problematic (Quinn and Keough 2002). For example, while increased detectability may lead to improved statistical power, there are limits by which further improvements can be achieved only by increasing the survey effort (i.e., the number of trap occasions or the number of sites), and these trade-offs are important considerations in all monitoring programs (Guillera-Arroita & Lahoz-Monfort, 2012).
In this study, we used data collected from an existing live and camera trap monitoring program to compare occupancy and detectability estimates of an endangered marsupial predator, the northern quoll (Dasyurus hallucatus). We focused on species occupancy-the proportion of an area occupied-as it is more cost effective and less technically demanding, particularly when compared to other approaches that aim to measure changes in population abundance or density through time (Alexander et al., 2016;Karanth et al., 2011;Southgate et al., 2019). Furthermore, occupancy is commonly used in large-scale monitoring programs (MacKenzie et al., 2003;McGrath et al., 2015;Royle & Kéry, 2007;Wibisono et al., 2011), and is recognized as an indicator of species extinction risk (IUCN, 2012), often mirroring trends in abundance (MacKenzie & Nichols, 2004).
Using the data collected from two different survey approaches, we estimated the survey effort requirements for detecting declines in northern quoll occupancy of differing magnitudes with high statistical power. Finally, we compared detectability, power, and resource requirements of the live and camera trap designs to assess our capacity to transition from a monitoring program based on live-trapping to a monitoring program based on camera trapping.

| Study species and system
The northern quoll, or ngawungawau (Nyamal), is a small-to medium-sized (300-1200 g) marsupial carnivore, endemic to northern Australia. Northern quolls are also one of the largest semelparous species on earth (Oakwood et al., 2001)-males exhibit partial-complete die-off on an annual basis. The species is currently listed as Endangered against International (IUCN, 2016) and National (EPBC, 2020) criteria, primarily due to having undergone a recent and severe decline in distribution. Populations now persist only in patches of northern Queensland, the top end of the Northern Territory, the north-west region of the Kimberley, and in the Pilbara bioregion .
Here, we focus on the Pilbara population of northern quolls, a genetically distinct population (Spencer et al., 2017) comprising the last remaining stronghold for the species . This is largely because it is the only population not yet impacted by cane toads (Rhinella marina); a toxic invasive species that northern quolls are prone to consuming, often resulting in death (Cramer et al., 2016). The Pilbara bioregion covers 178,060 km 2 , and is comprised mostly of hummock grasslands, intersected by rocky ranges and granite outcrops (McKenzie et al., 2009). Annual rainfall varies between 250 and 450 mm, with the majority falling between December and March (Australian Bureau of Meteorology, 2020). Primary land uses include rangeland cattle pastoralism, mining, Indigenous reserves, and nature conservation areas.

| Survey data
We collected live trap data as part of the Pilbara Northern Quoll Monitoring (PNQM) program; a long-term monitoring program initiated in 2011 (Cramer et al., 2016). A key objective of the PNQM is to track natural changes in northern quoll occupancy across this Pilbara as a result of seasonal/annual fluctuations in resource availability, as well as to collect data relevant to the biology and life history of the Pilbara populations. This data can then be used to identify unnatural changes in northern quoll occupancy in areas impacted by disturbances, such as mining. Mining activity is recognized to pose a significant threat to northern quolls in the Pilbara because rocky landforms, which provide important denning habitat for this species, are often the focus of ironore extraction, or are used for road and rail bedding materials (Cramer et al., 2016). The overlap of potential mining activity and northern quoll habitat is large-79.7% of the Pilbara bioregion is currently occupied by mining tenement (DMIRS, 2020).
In 2012, we undertook a large baseline survey across 100 monitoring sites to identify suitable locations for establishment of the PNQM. The location of these sites was guided by expert knowledge and northern quoll species distribution models (Molloy et al., 2017). From the 100 sites sampled as part of this widespread survey, only sites that had one or more quoll detections (12 sites) were selected for inclusion in the PNQM. Live trapping sites were separated from each other by at least 50 km. Either the entire set of 12 sites, or a subset thereof, were sampled annually using a trapping protocol developed specifically for northern quolls (Dunlop et al., 2014). The protocol comprised of two transects of 25 traps, with each trap separated by 50 m, and each transect separated by 200 m (Appendix 1). Total area covered by live traps at each site was approximately 25 ha. Traps were opened for four consecutive nights, and baited with a mixture of peanut butter, oats, and sardines. All live trap surveys were conducted between May and October to avoid times after male die-off, trapping females carrying young, and extreme summer temperatures.
We collected camera trap data as part of a concurrent project that aimed to reveal factors determining northern quoll occupancy at the patch and landscape scale in the Pilbara. A total of 12 sites (three of which were also live trapping sites), and 11 sites were sampled between 2017 and 2018 ( Figure 1). Each site consisted of five Reconyx PC900 Hyperfire (Reconyx, WI) cameras, spread across a 75-ha circular area (Appendix 1). Sites were separated from each other by at least 1 km, and cameras were deployed for a period of up to 8 months. For the purposes of this study, we used only the first 10 nights of data collected by each camera as 10 nights of repeat sampling represents a more realistic amount of effort that may be applied more widely by land managers conducting presence absence surveys. This also ensured that the occupancy status of northern quolls at each site was closed (i.e., no immigration, emigration, births, or deaths) over the duration of sampling, an assumption underlying occupancy models (MacKenzie et al., 2002). Camera data were taken from sampling conducted in August 2017, such that it would be as temporally comparable as possible to live trap data (collected May-October 2016).
Cameras were attached to a wooden stake 1.5 m above the ground, orientated downward, with at least 200 m between each camera, as per Moore et al. (2020). Orientating cameras vertically facilitates individual identification of quolls by capturing unique spot patterning located on the dorsal surface of animals, without compromising detectability (Moore et al., 2020;Moore et al., 2021). All cameras were placed within rocky habitat (the preferred habitat of the northern quoll) and were baited with a sealed polyvinyl chloride canister containing pilchards.
In response to the growing recognition of the versatility and cost-effectiveness of camera traps as a monitoring tool for northern quolls (Austin et al., 2017;Diete et al., 2016;Hernandez-Santin et al., 2016;Hohnen et al., 2013;Moore et al., 2020), the PNQM program had considered transitioning from a live trapping approach to using camera traps as the primary survey method for monitoring northern quolls in the Pilbara. However, before this transition could occur, it was deemed important to calibrate the two methods based on their ability to detect northern quolls, their power to detect changes in occupancy, and their resource requirements. Rather than undertaking a series of potentially costly trials in the field, we make use of the existing datasets, given the close temporal proximity of the two programs.

| Existing designs
We included live trap data from the 2016 PNQM surveys in our analyses (collected May-October), as all 12 sites were sampled during this year. We included camera trap data collected from the 12 sites sampled in August 2017. Average annual rainfall across the study area was $38% higher in 2017 (421 mm) when compared to 2016 (261 mm) (SILO, 2022). While increased rainfall may have positively impacted quoll populations within the study area by increasing resource availability (Moore et al., 2021), it did not have a substantial impact on quoll detectability in this study. This is demonstrated by the fact northern quoll camera trap detectability from the same study area in 2018 was very similar to camera trap detectability in 2017, despite annual rainfall being $30% lower (290 mm) (SILO, 2022). Similarly, live trap F I G U R E 1 Distribution of live and camera trap sites used to survey northern quolls within the Pilbara Bioregion in Western Australia. detectability stayed very consistent from 2016 to 2018, despite varying rainfall (Appendix 2).
We used single-species, single-season occupancy models (MacKenzie et al., 2002) that account for imperfect detection to estimate detectability and occupancy of northern quolls for the live and camera trap designs, independently. Detection histories for each site were assembled by pooling detections across all traps within a site (i.e., detections pooled across 50 trap units deployed at each live trap site, and five trap units deployed at each camera trap site) into a single measure of detection/nondetection for each trap night. Occupancy models were formulated in terms of parameters ψ i (occupancy) and p ij (detectability), where ψ i is the probability that site i is occupied by the species, and p ij is the probability of the species being detected at site i on trap night j, conditional upon its presence. The models assume independence of detections between sites, that there are no false detections, and that occupancy remains constant over the sampling period. The objective of this study was not to measure the impact of environmental variables of northern quoll occupancy or detectability, and thus no covariates were included in the models. Occupancy models were fit using the "unmarked" package (Fiske & Chandler, 2011) in R version 3.6.2 (R Core Team, 2019).
To estimate the statistical power of live and camera trap designs, we used Equation (1) outlined in Guillera-Arroita and Lahoz-Monfort (2012). This equation can be solved to calculate the statistical power (G) of a given monitoring design, in detecting occupancy declines of varying magnitudes (R), based on the number of sites (S) and sampling nights (K) (i.e., overall sampling effort), and underlying sample values of occupancy (Ψ) and detectability (p). This equation can be written as: where Ψ 1 and Ψ 2 are equal to the underlying occupancy probabilities of the two occasions sampled, z α=2 is the upper 100α/2-percentage point for the standard normal distribution, ϕ (x) is the cumulative distribution function for the standard normal distribution, and Geyle et al. (2019), we set alpha α to 0.2, and β to 0.8 for all analyses, as it better reflects the costs of committing a Type II error (failing to detect a decline), which, for threatened species, could lead to extinctions (di Stefano, 2003). We set R values to reflect decline thresholds as defined by IUCN criterion A2c for classifying a species as threatened; an observed, estimated, inferred, projected or suspected reduction in the area of occupancy of ≥80% (Critically Endangered), ≥50% (Endangered), or ≥30% (Vulnerable) over any 10 years or three generations (whichever is longer). However, we note that the temporal period over which we intended to measure occupancy declines (12 months) was much shorter than timing specified in the IUCN criterion A2c for classifying species as threatened, and therefore we do not suggest the method described in this paper could inform IUCN assessments directly.
We calculated the cost ($AUD) associated with undertaking10-years of live and camera trap monitoring based on total trap effort (S Â K). We chose 10 years because this is considered the shortest period of time over which effective ecological monitoring can be conducted (Lindenmayer & Likens, 2010) and because this is the timeframe for the original northern quoll monitoring program (Cramer et al., 2016). Cost calculations accounted for all expenses related to equipment, consumables, travel, staff salaries, and so forth. See Appendix 3 for a detailed list of costs. Equipment costs were only included in the calculations for the first year of monitoring. We capped the number of traps to 120 camera units and 50 live trap units based on real-world limitations associated with equipment costs and time spent in the field. We chose a maximum annual budget of $100,000 because most monitoring programs (including the PNQM) are unlikely to have budgets greater than this. A single Reconyx PC900 Hyperfire camera unit (including batteries, SD cards, and mounting equipment) costs approximately $830. Therefore the maximum number of camera traps that can be set at any given time (based on this budget) is 120. We capped the number of live traps to 50 because, based on field experience, this is approximately the maximum number of traps that can be deployed at sites with high densities of quolls using two staff, while remaining adherent to animal ethics guidelines (Dunlop per obs).

| Optimized designs
We explored optimized live and camera trap survey designs to identify the amount of survey effort required to detect occupancy declines of 30%, 50%, and 80% in northern quoll population with a power of ≥0.8. To do this, we first assumed the live/camera trap set-up used at individual sites would be identical to that used as part of existing sampling-50 live traps or 5 camera traps per site (see "Current design" section of methods). Our second assumption was that the probability of detecting quolls, per site, per trap night (calculated as per the methods described in the previous section) would also be identical to that of the existing designs. Next, we solved Equation (1) for all combinations of S (sites = 1:100), K (nights = 1:21), and R (proportionate decline in occupancy between occasions = 0.3, 0.5, or 0.8), giving a total of 6300 monitoring design simulations for each trap type (100 Â 21 Â 3). Initial occupancy was set to 0.75 for both live and camera trap designs to allow the two methods to be compared. Although slightly higher than occupancy rates measured using existing designs, an occupancy rate of 75% is a realistic expectation for future targeted monitoring, based on increased knowledge of quoll occurrence in the Pilbara. The total number of sites (S) was limited to 100 because surveying over 100 sites per year was not a feasible option for the PNQM. The total number of nights (K) was limited to 21 to avoid breaching assumptions of northern quoll population closure across sampling periods for occupancy models. We identified the live and camera trap designs with the lowest cost that achieved a power of ≥0.8 for occupancy declines of 30%, 50%, and 80%.
The existing live and camera trap designs had limited statistical power in detecting a 30% decline in occupancy (power <0.5), or a 50% decline in occupancy (power <0.7), but high power in detecting an 80% decline in occupancy (power >0.9) ( Table 1). The camera trap design had consistently greater power for all levels of decline, however differences were not significant (overlapping CI). The estimated 10-year cost of the existing camera trapping design was almost three times lower ($343,720) than the existing live trap design ($936,000).

| Optimized designs
The optimal number of sites required for live and camera trap designs to achieve a power of ≥0.8 in detecting varying magnitudes of occupancy decline were very similar. For example, to detect occupancy declines of 30% the live trap design required 32 sites, while the camera trap design required 33 sites. To detect occupancy declines of 50% and 80%, both methods required 12 and 4 sites respectively ( Table 2).
The optimal live trap designs consistently required a lower number of nights than the optimal camera trap designs to achieve a power of ≥0.8 in detecting occupancy changes of each magnitude (Figures 3 and 4). To detect a decline in northern quoll occupancy of 30% the camera trap design required over three times more trap nights (7, CI = 5-9) than the live trap design (2, CI = 2-2) ( Table 2). To detect an occupancy decline of 80%, the camera trap design required over twice as many trap nights (5, CI = 4-6) than the cheapest live trap design (2, CI = 2-2) ( Table 2). Optimal live trap designs were always more expensive than optimal camera trap designs, F I G U R E 2 Occupancy and detectability estimates (circles) and 95% confidence intervals (lines) for northern quolls (Dasyurus hallucatus) in the Pilbara bioregion of Western Australia. Data derived from 12 live trap sites (two transects of 25 cage traps open for 4 consecutive nights) and 12 cameras trap sites (5 Reconyx PC900 Hyperfire cameras spread across a 75 ha area for 10 consecutive nights).
T A B L E 1 Statistical power of existing live and camera trap designs to detect occupancy declines of 30%, 50%, and 80% in Pilbara northern quolls (Dasyurus hallucatus). and the relative cost difference decreased with increasing decline in occupancy (Figures 3 and 4).

| DISCUSSION
New monitoring technologies offer cost-effective alternatives to achieve long-term monitoring goals. However, before transitioning to these new approaches, it is vital to compare them to existing trapping designs to ensure that they are able to achieve identical or higher detection rates and power, ideally at lower costs (Lindenmayer & Likens, 2009). Here, we show that an existing camera trap design was almost three times cheaper over 10 years for detecting northern quoll occupancy declines, with similar statistical power. When optimized, camera trap designs were consistently cheaper than live trap designs for detecting occupancy declines of each magnitude with a power of ≥0.8. Overall, we find support for the cost-effectiveness of camera traps for species monitoring (Steenweg et al., 2017;Wearn & Glover-Kapfer, 2019) and its potential to replace existing live trap sampling of northern quolls as part of the PNQM. Additionally, we offer a robust framework to compare new monitoring techniques against pre-existing methods on the basis of statistical power.

| Existing designs
The nightly probability of detecting northern quolls using the existing live trapping design was almost double the nightly probability of detecting northern quolls using the T A B L E 2 Optimized live and camera trap designs with the lowest 10-year costs ($AUD) to detect occupancy declines of 30%, 50%, and 80% with a power of ≥0.8 in Pilbara northern quolls (Dasyurus hallucatus). F I G U R E 3 The statistical power (as a function of nights and sites) of optimized live trap (top) and camera trap (bottom) designs to detect occupancy declines of 30%, 50%, and 80% in Pilbara northern quolls (Dasyurus hallucatus). Points and labels indicate monitoring designs with the lowest cost ($AUD) that achieve a power of ≥0.8. camera trap design. Detectability is an important consideration in survey design, as species with low detection probability typically require more sampling nights to detect trends than species with higher detection probability (Field et al., 2005;MacKenzie & Royle, 2005). The observed differences in detectability between the two trap methods are likely a result of the differences in the density of live traps ($1 trap per 0.5 ha), which was 30 times greater than the density of camera traps ($1 trap per 15 ha). Thus, the likelihood of a quoll encountering a live trap (assuming quolls were present at the site), was also greater. Whilst a number of previous studies have directly compared the capacity of live and camera traps to detect species (Hohnen et al., 2019;Richardson et al., 2018;Welbourne et al., 2015), the results of these studies are mixed, likely due to a combination of varying species habits, camera models, live trap methods, intensities in trapping effort, and bait use. Comparisons with our own results are therefore unlikely to be useful. Despite having a lower nightly detection probability, the camera trap design was cheaper for detecting northern quoll occupancy declines when compared to the live trap design. This is likely a result of the camera trap designs capacity to accrue greater nightly replicates than the live trap design (live trap = 48, camera trap = 120) with negligible increases in cost. For example, if you were to double the number of trap nights surveyed using the existing camera trap design (i.e., 20 trap nights instead of 10), the difference in costs over a 10-year period would be $720, which equates to less than 0.25% of total cost. By contrast, doubling the number of trap nights surveyed using the existing live trap design (i.e., 8 trap nights instead of 4) increases cost by $931,000, which equates to almost double the original cost. These results further highlight the cost efficiencies of camera traps as survey tools, and support previous findings which suggest ecological survey effort (in terms of sites multiplied by nights) can be a poor surrogate for financial cost (G alvez et al., 2016).

| Optimized designs
The optimized camera trap designs were consistently cheaper than optimized live trap designs, with smaller differences associated with increasing magnitudes of decline in occupancy. This is probably because more nights can be surveyed without high additional costs using camera traps, compared to live traps. Our results are in contrast to a recent meta-analysis that compared 62 studies, finding that camera traps were less cost effective than other survey methods, including live traps (Wearn & Glover-Kapfer, 2019). However, the cost estimates in that study were not standardized across the trapping methods using a universal metric, such as statistical power (Wearn & Glover-Kapfer, 2019). Consequently, the reported differences in cost may not have accurately reflected actual differences in cost associated with each of the trapping methods (Wearn & Glover-Kapfer, 2019), nor do those costs necessarily reflect the ability of a given survey method to meet the objectives of a monitoring program. Calculating the power of a design allows us to determine whether our methods have the potential to produce a statistically significant result, when the effect size (in this case a change in occupancy) is biologically important (Guillera-Arroita & Lahoz-Monfort, 2012). Therefore, power analysis provides an important and useful metric to compare alternate trap designs using data from existing monitoring programs (G alvez et al., 2016;Hohnen et al., 2019;Ward et al., 2017).

| Further considerations
We demonstrate that camera traps are substantially cheaper than live traps at detecting occupancy declines of differing magnitudes in northern quolls. Thus, transitioning from live traps to camera traps offers a cost-effective way for monitoring programs to meet their long-term F I G U R E 4 (a) Most cost efficient live and camera trap optimized designs that are capable of detecting varying magnitudes of occupancy decline in Pilbara northern quoll (Dasyurus hallucatus) populations with power ≥0.8. (b) Total effort required for monitoring designs that are depicted in panel a. K = nights, S = sites objectives. However, it is important to consider that monitoring programs are often required to achieve multiple objectives in addition to detecting changes in occupancy (Likens & Lindenmayer, 2018), such as the collection of genetic (Schwartz et al., 2007;Chan 2020), morphological (Dunlop & Morris, 2018), dietary (Bannister et al., 2020), mutations , and breeding information (Robinson et al., 2020), which may not be possible through the exclusive use of camera traps. In such cases it may be important for managers to take a broader approach when assessing the applicably of cameras traps to their monitoring design, and potentially consider the use of multiple monitoring techniques in combination.