Delaying conservation actions matters for species vulnerable to climate change

Handling Editor: Céline Bellard Abstract 1. Climate change vulnerability assessments are commonly used to identify species or populations at risk from global climate change, but few translate impact assessments to climate change adaptation actions. Furthermore, most climate change adaptation efforts emphasize where to implement management actions, whereas timing remains largely overlooked. The rate of modern climate change introduces urgency in evaluating whether delaying conservation actions compromises their efficacy for reaching important conservation targets. 2. We evaluated the importance of multiple climate change adaptation strategies including timing of actions on preventing extinctions for a threatened climate-sensitive species, the Eastern Massasauga rattlesnake (Sistrurus catenatus). We parameterized a range-wide population viability analysis model that related demographic sensitivities to drought events and human-modified land cover to assess vulnerability to future climate change. Using simulations, we assessed the efficacy and trade-offs associated with alternative climate adaptation strategies aimed at maximizing the number of future populations including when to initiate conservation actions, duration of management, number of managed populations, and local management effectiveness. 3. Population-level projections under future climate change scenarios revealed a broad-scale pattern of range contraction in the southwestern portion of the current range. Along the extinction gradient, we identified demographic strongholds and refugia critical for population persistence under climate change as well as populations at high risk of extinction and candidates for climate change adaptation actions. 4. In the context of future climate change, the timing of conservation actions was crucial; acting earlier maximized chances of achieving conservation targets. Even considering uncertainty in climate change projections, delaying actions was less efficient and introduced undesirable trade-offs including the need to implement conservation actions for longer or targeting more populations to achieve a similar conservation target. 5. Synthesis and applications. Our findings highlight how acting quickly reduces risk and improves outcomes for a highly vulnerable species under future climate change. Climate change vulnerability assessments require translation of


| INTRODUC TI ON
During periods of rapid environmental change, conservation actions that are not implemented in a timely manner may miss windows of opportunity resulting in inefficient use of resources, or at worst, failure to reach conservation goals. Delays can occur at various points in decision-making processes for threatened species management, including listing a species (Martin et al., 2012) or delayed implementation of on-the-ground conservation measures (Martin, Camaclang, Possingham, Maguire, & Chadès, 2017). Decisions on how quickly or how long to act, or when actions should change, such as shifting resources from monitoring to alternative actions, impact strategic interventions aimed at threatened species management and ultimately biodiversity conservation (Lindenmayer, Piggott, & Wintle, 2013;Ng, McCarthy, Martin, & Possingham, 2014). Managers faced with limited resources must make these decisions in the context of other considerations such as how many sites or populations to manage, and how aggressively to manage. Improved understanding of the long-term consequences of timing of conservation actions, and of the trade-offs involved in a broader management context, can aid decision making during periods of rapid environmental change.
Anthropogenic climate change introduces both uncertainty and urgency to the timing of management actions. Climate change varies not only regionally, but across multiple ecologically relevant temporal scales (e.g., past and future changes in extremes and variability) (Garcia, Cabeza, Rahbek, & Araújo, 2014). There is increasing evidence of "tipping points" when small changes in the climate system result in strongly nonlinear responses and rapid shifts in novel climate space (Lenton, 2011). Consequently, both the rate and magnitude of exposure to climate change characterize how stressors or opportunities for adaptation vary across a species' range over time. Even independent of the additional complexities added by synergies with land-use change (Brook, Sodhi, & Bradshaw, 2008), spatial heterogeneity and temporal nonlinearities in climate change necessitate that assessments and actions be optimized both regionally and for specific time periods.
Climate change vulnerability assessments (CCVAs) evaluate the propensity and susceptibility of multiple species to be adversely impacted by modern climate change (Pacifici et al., 2015). CCVAs incorporate a combination of intrinsic and extrinsic factors to capture sensitivity, exposure, and adaptive capacity, and link the spatial and temporal heterogeneity of future climate change to species traits or population trends (Williams, Shoo, Isaac, Hoffmann, & Langham, 2008). Population viability analysis (PVAs) is one quantitative approach used to model extinction risk under climate change and associated range shifts through the interaction between population dynamics and changes in habitat suitability over space and time (Keith et al., 2008). Although PVA models are one of the more data intensive approaches used in a CCVA context, they are amenable to integrating the important components of vulnerability: sensitivity, exposure, and adaptive capacity (e.g., McCauley, Ribic, Pomara, & Zuckerberg, 2017;Naujokaitis-Lewis et al., 2013). From a practical standpoint, these models present a powerful approach to compare alternative climate change adaptation strategies using a common (probabilistic) currency of extinction risk (Pe'er et al., 2013).
CCVAs can support decisions for managing climate-sensitive species and serve as a platform to assist with prioritization of adaptation actions. Given the resource constraints in conservation planning, quick management decisions need to be made regarding the most effective and efficient actions for reducing threats associated with future climate change (Pacifici et al., 2015). For individual species that are deemed to be especially vulnerable, translating model-based vulnerability to practical climate change adaptation strategies requires quantifying climate-demographic relationships and simulating the potential benefits of specific conservation actions (Game, Kareiva, & Possingham, 2013); however, moving from impact assessment to climate change actions is a step that still few CCVAs consider (Akçakaya, Butchart, Watson, & Pearson, 2014). Previous studies applying PVA models to evaluate alternative actions to reduce climate change impacts tend to emphasize the spatial dimensions of management such as how many populations to manage, which populations to translocate, and where and how much habitat is required to offset climate change impacts (Fordham et al., 2013;Naujokaitis-Lewis et al., 2013;Regan et al., 2012). By comparison, the consideration of the timing of conservation actions, has been less frequently assessed in PVA-based management scenarios (but see McDonald-Madden, Runge, Possingham, & Martin, 2011). PVA model outputs can include estimates of expected time to extinction and can be used to inform early warning signals of climate risk, which are relevant for listing and categorizing species at risk of extinction (Stanton, Shoemaker, Pearson, & Akçakaya, 2015). However, such metrics do not directly inform decisions related to when to start or how long to implement an action. In the context of climate change and other rapidly changing threats it is imperative to consider the timing of management actions. model-based outputs into tractable information for climate change adaptation planning. Quantifying trade-offs associated with the multidimensional decision space related to species conservation and recovery planning is a critical step in climate change adaptation.
The goal of our study was to develop a novel species-specific CCVA for comparing trade-offs associated with alternative adaptation strategies used for promoting persistence of a climate-sensitive species. In doing so, we evaluate a set of decision points that managers commonly face including (a) when to implement actions, (b) how long to manage, (c) how many populations to target for management actions, and (d) how aggressively to manage. We first built a range-wide PVA model that incorporates relationships between demographics and climate change for the Eastern Massasauga Rattlesnake (Sistrurus catenatus-hereafter EMR), a threatened species that has demonstrated rapid range contraction and vulnerability to past climate change (Pomara, LeDee, Martin, & Zuckerberg, 2014). Using our range-wide predictions of extinction risk under climate change, we explored the trade-offs associated with realistic management scenarios, while accounting for uncertainties in climate change projections. Our work addresses the heretofore overlooked issue of optimal timing of conservation actions in the context of climate change by better linking population-specific outcomes to tangible, concrete adaptation strategies that inform threatened species conservation and management into the future.

| Study area and species
EMR is found throughout the Great Lakes Region and is federally threatened in Canada (COSEWIC, 2002) and the U.S.A. (US Fish and Wildlife Service, 2016). Habitat loss is a primary driver of recent declines, and a dependency on semi-wetland habitats confers sensitivity to long-term changes in climate (Szymanski et al., 2016). Specifically, drought and flooding events pose a risk as EMR is dependent on stable water levels, especially during winter hibernation. Alongside habitat restoration and vegetation management, direct water-table manipulation is a candidate management action aimed at improving EMR persistence (Szymanski et al., 2016). We performed a CCVA for EMR by modelling range-wide population dynamics using demographic relationships linked to climate conditions and land use to assess extinction probability under different scenarios of future climate change (Figure 1). Our approach builds on the demographic models of Pomara et al. (2014), who found that historic range-wide declines in EMR were associated with demographic sensitivities to both winter drought and summer flooding.

| Environmental variables
Previous research highlights the importance of environmental drivers of EMR adult active season survival rates including winter minimum temperature, summer cumulative precipitation, anthropogenic land cover, and winter drought (Pomara et al., 2014). Winter

| Climate change projections
Using the delta method, we produced finer-resolution and bias corrected annual climate projections for each climate variable, including winter SPEI, winter minimum temperature, and summer cumulative precipitation. This ensured a continuous dataset from the observationbased data (recent historical climate data) and model-based climate projections (future data). Methodology followed (Harris, Grose, et al., 2014) and details are included in the Supporting Information. Gridded projections of winter minimum temperature and summer cumulative precipitation were downloaded at a resolution of 12 km (Reclamation, 2013). Gridded projections of the original SPEI data varied in spatial resolution, but were downscaled to a common resolution of 0.5° (Table S1; Cook, Smerdon, Seager, and Coats (2014)).
Global climate projections were based on the World Climate

Research Programme's (WCRP's) Coupled Model Intercomparison
Project phase 5 (CMIP5) multimodel dataset, for the highest Representative Concentration Pathway (RCP) 8.5. RCP8.5 corresponds to a radiative forcing of approximately 8.5 W/m 2 and represents the largest increases in greenhouse gases across all RCPs. Current emissions continue to track this high end emission scenario (Peters et al., 2013). Given the need to develop robust adaptation strategies we selected this single yet currently realistic scenario and applied a larger number of Global Circulation Models (GCMs) (n = 11) to capture higher inter-model climate model uncertainties (Table S1, Lutz et al., 2016).

| Survival modelling
We modelled the relationship between adult active season survival estimates and environmental variables using binomial generalized linear models with a logit link function. Survival rates from across the species range from Jones et al. (2012) were expressed as proportions, and each observation (n = 17) was weighted by the number of cases (i.e., telemetered snakes, which ranged from 12 to 48).
Specification of main effects was constrained to a maximum of three variables to facilitate interpretation, and all variable combinations were considered using an information-theoretic approach. We used Akaike's information criterion for small sample sizes (AIC c ) to identify highest ranked models, given the data. To account for model-based uncertainty with our models of EMR active season survival, and given our primary goal of prediction, we model-averaged parameter estimates of the top 95% confidence model set (i.e., cumulative AIC c weight of models ≤0.95) (Symonds & Moussalli, 2011). We projected active season survival rates based on model-averaged parameter estimates annually through 2100 using the adjusted future annual time series based on climate anomalies (Banner & Higgs, 2017

| Climate-driven PVA model
We parameterized a population-level range-wide demographic model for EMR where fecundity and survival estimates were combined to parameterize a females-only, age-based with 11 classes, stochastic population model across the geographical distribution of EMR (Table S2,  We synthesized results over multiple future time points (current: 2010, mid-century: 2050, late-century: 2100) by averaging outcomes 10 years prior to minimize variation associated with annual variability in projections.

| Simulating management decision points and actions
Managers often require making decisions around number of populations to manage, the timing of management actions, and which type of management action to implement. We translated these decisions into population-level consequences using a simulation approach where investments in conservation actions were varied to explore trade-offs in the decision landscape ( Figure 1). We represented these decision points by simulating improvements to adult survival given the evidence for demographic sensitivities in relation to various climatic factors (Pomara et al., 2014). For the EMR, restoring wetland habitat and direct water-table management are two different actions that might improve survival by minimizing local drought effects.
We randomly sampled the number of populations to target for management actions in a given simulation. We applied a random uniform distribution where the minimum number of populations was set to 5 with a maximum representing the number of populations with a corresponding predicted quasi-extinction probability of ≥0.1. This upper limit threshold reflects the criteria used to identify self-sustaining populations based on the recent status assessment of EMR in the USA whereby populations with a probability of persistence >0.9 were qualitatively considered robust (Szymanski et al., 2016). EMR consists of three genetically distinct subunits, where each subunit is considered to represent an area of unique adaptive diversity (Western, Central, and Eastern; Ray et al., 2013). To ensure representation across this gradient of genetic diversity, we randomly sampled populations by genetic subunit.
To address the importance of timing of conservation actions, we varied two parameters: the year that the conservation action began and the duration of management. We applied a uniform distribution to randomly sample the start year of management, which ranged from 2011 to 2090. The number of years an action was implemented varied between 1 and 50 years and was sampled from a uniform distribution. Our conservation actions were initiated during the sampled start year and implemented in successive years until the number of sampled years was reached. Our timing variables reflected when to start and how long to manage, but did not consider when to switch between alternative conservation actions.
We introduced two levels of conservation effectiveness to capture the variable effect of local (i.e., population-level) management to improve survival. Two different actions that may improve survival by minimizing local drought effects are restoring wetland habitat and direct water-table management. However, one might be more effective than the other, they might be implemented in tandem or separately, and either may be implemented with varying intensity or success. We distinguished between the two levels by modifying adult survival rates to increase to 0.78 (mean of survival estimates; moderate level) or to 0.90 (this represents the top 10th percentile of predicted estimates based on modelled outcomes; high level), when and where they fell below these rates; higher rates were not altered.
This range of values represents a realistic range of survival estimates for EMR across its range (Jones et al., 2012).

| Demographic climate change refugia
The 6 best supported models of active season survival based on the top 95% confidence model set (i.e., cumulative AIC weight of models  (Table 2). Model projections to both mid-(2050) and late-century (2100) depicted a strong spatial gradient in quasi-extinction risk across the EMR's range with increasing risk over time (Figure 2). Quasi-extinction risk was lowest in the northeast with a distinct extinction risk gradient increasing to the southwest, highlighting a broad-scale pattern of range contraction towards the northern periphery of the range. The general spatiotemporal pattern of quasi-extinction risk was consistent across GCMs, but the largest sources of model uncertainty associated with choice of GCM occurred among populations within the south-central portion of the range, and were highest across late-century projections ( Figure S1). Validation outcomes of our population dynamics model included an AUC of 0.78, indicating acceptable discrimination (Hosmer & Lemeshow, 2000). Classification metrics included a sensitivity value of 0.93, and specificity of 0.62. These outcomes indicate that the model was better at classifying true presences (extant) than absences (extirpation). We also found trade-offs between start year and the duration of management whereby delaying conservation actions would   (Table S4, Figure 3c, Figure S4). To achieve a target of approximately 160 extant populations would require implementing actions over 70 to 90 populations for a 30 to 50 year period.

| Trade-offs between climate change adaptation decision points
Although the choice of GCM was included in two interactions (with start year and number of populations respectively), the interaction size was negligible (Table S4). Overall, our results emphasize diminishing returns and loss of conservation opportunities as actions are delayed into the future.

| D ISCUSS I ON
There is a critical need to develop species-specific models of climate change vulnerability and translate those model outputs into tractable information for conservation decision making. Here we demonstrate that testing alternative decision points around climate change adaptation actions through simulations can provide this linkage. Using a range-wide PVA model built on climate-demography relationships, we identified geographic regions of EMR vulnerability to future climate change and potential refugium critical for species persistence (Keppel et al., 2015). While accounting for uncertainties in future climate change projections, we illustrate the relative importance of timing of management actions in comparison to other more commonly assessed management decision points. For EMR, delaying implementation of management actions meant increased effort was needed to achieve a similar conservation target, with more lost opportunities and fewer options as delays grew longer. Our findings suggest that timing of conservation is crucial and targeted actions can buffer the effects of future climate change on range-wide persistence, but their effectiveness is mediated by interactions among different decision points and future climate uncertainty.
We documented a range-wide extirpation front that was consistent across GCMs; however, GCM selection introduced substantial variation in extinction risk ( Figure S1). This variation was evident towards the contracting range edge and was most pronounced in latecentury projections, an expected finding given divergence among GCM projections over time (Beaumont, Hughes, & Pitman, 2008).

The resulting continuum of outcomes based on multiple individual
GCMs present an envelope of possible climate futures, which would not be evident using a GCM ensemble approach . While variation across GCMs is thus evident, the disproportionate sensitivity of populations located at the contracting range can include actions that are spatial (e.g., how many and which populations to target) and temporal (e.g., when and how long to implement an action). Moving beyond impact assessment to the selection of climate change adaptation strategies that will maximize conservation outcomes is a complex process, but one that would benefit from a comparison of anticipated actions using scenario-/simulationbased approaches. In the context of CCVAs, this component remains largely over-looked (but see Fordham et al., 2013;Regan et al., 2012), especially with respect to the timing of conservation actions.
Knowing the critical management decision points, such as when it is too late to start acting, is a pervasive question for decision makers and conservation scientists. Adequate warning times for preventing extinctions will depend on a combination of factors, including political will, socio-economic considerations, species' expected responses to management actions, management objectives set for a species, and anticipated magnitude of climate change in a focal area (Akçakaya et al., 2014). We presented our results by highlighting the trade-offs associated with the multidimensional decision space. As an example, we used a conservation target of 160 extant Managers have several different options to conserve species threatened by climate change including in situ approaches that have the potential to offset both current and impending threats (Greenwood, Mossman, Suggitt, Curtis, & Maclean, 2016). In the case of EMR, direct water-table manipulation and vegetation management are proposed in situ strategies aimed at minimizing drought and flooding effects on existing EMR populations (Faust et al., 2011).
These types of habitat modifications could provide effective changes to local climatic conditions experienced by EMR and help to minimize negative outcomes associated with climate change. Managing species or populations in situ can present challenges as actions may not translate immediately into improved recovery outcomes and should also be robust to future climate change and associated uncertainties. Furthermore, knowledge associated with species' responses to a specific action is typically sparse, and different actions could result in being more or less effective for species recovery and adaptation to climate change (Bonebrake et al., 2018 (Shoo et al., 2013). More risk-averse management may be appropriate for populations showing higher disagreement among GCMs, as management outcomes are less certain.
Identifying trade-offs associated with alternative conservation actions requires several simplifying assumptions in our simulations.
We simulated actions that improved active season survival rates, assuming that implementation translated to an immediate increase in survival rates. Additionally, once the duration of an action was complete, survival rates returned to the original projections rather than staying elevated. Depending on the specific management action, population-level responses could lag behind initial implementation. For example, direct water-table manipulation may result in immediate effects on survival while attempting to improve watertable levels indirectly through restoration of vegetation is more likely to produce a more lagged (i.e., slower) response at the population level. This assumption may have resulted in overly optimistic results, but reinforces the need to act quickly. While our approach assists in prioritizing and evaluating temporal relative to spatial dimensions of conservation actions, we did not explicitly consider costs and subsequent trade-offs in a cost-efficiency framework (Sebastián-González et al., 2011), nor when to shift between management actions, which are important next steps. Despite these limitations, our approach to simulating both climate-driven threats and the effectiveness of adaptation actions in a single framework can be readily extended to other species and systems, which include species currently known to be climate-sensitive and those anticipated to be most vulnerable to future climate change.
Real-world situations where decisions have been delayed have clearly contributed to species extinctions (Martin et al., 2012). These delays in conservation action have even greater implications during a time of rapid climate change that is unprecedented over decades and millennia (IPCC, 2014). Conservation prioritizations for climatethreatened species have largely not addressed timing of conservation, yet we show here that timing is critical for improving persistence of a climate-threatened species, even while accounting for uncertainties of future climate change. Delays in decisions and actions onthe-ground are likely to have significant negative impacts on both currently declining climate-sensitive species and those vulnerable to unprecedented changes in climate and land-use practices. There is an urgent need to make decisions related to the management of climatesensitive species while there is still an opportunity to act.

ACK N OWLED G EM ENTS
We gratefully acknowledge support for this work from the Upper F I G U R E 4 Trade-offs between decision points of number of populations to manage and start year of management actions on range-wide number of populations predicted extant for the moderate scenario GCM (CSIRO-MK3). Isolines represent increments of five populations. As an example trade-off using a conservation target of 160 populations (z-axis), implementing management in 2020 (x-axis) at 40 populations (y-axis) depicts the advantages of acting quickly (black-filled dot) and is contrasted by consequences of delaying actions to 2040 which would require management of 75 populations to achieve the same target (greyfilled dot). The black-outlined dot represents the lowered expected conservation target (154 predicted populations extant) if the number of populations managed remained constant (40) but the start year of management was delayed. The difference between the grey-filled and grey-outlined dot is the (greater) lost opportunity associated with delaying actions at higher level of number of populations managed
Note: Eastern Massasauga Rattlesnake population status and location data are sensitive and have not been archived.