Biometrics & Modeling I

Contributed Oral Presentations

Contributed paper sessions will be available on-demand for the duration of the conference, then again at the conclusion of the conference.


Should You Integrate Those Data Sources? a Decision Tree for Data Checking in Integrated Species Distribution Models
Brent Pease; Krishna Pacifici; Roland Kays
The expanse of biodiversity data available today has resulted in a diversity of animal occurrence records over large geographical areas that span many years, but each data source often has its own nuances and biases that must be accounted for. Model based data integration has emerged as a means to formally combine multiple data sources in estimates of species distributions while accounting for differences among each data source. Despite the promise of these methods, the more complex integrated models may not always provide improvements over simpler models. In some cases, it may be simply due to important but unaccounted for biases in the data. Other times there may be an apparent incompatibility or disagreement among the data sources leaving practitioners unsure how to proceed. Unfortunately, there are currently no established best practices to determine whether data integration will reduce parameter uncertainty or improve model predictive performance. Here we illustrate a decision tree for data checking in integrated species distribution models aimed to highlight important characteristics of data sources that tend to be suitable for data integration. This decision tree provides several simple and accessible checks that indicate likely outcomes of data integration prior to implementing models, including tests for the degree of spatial agreement, covariation among data sources, differential effects of estimated coefficients, spatial and/or temporal mismatch, and variable effort across data sources. To illustrate our approach, we use publicly available data sources including citizen science camera trapping observations and state agency monitoring programs such as hunter and trapper harvest records. Using four data sources and a range of species, we demonstrate cases when data integration is useful and scenarios where we caution against integration.
Spatial Capture-Recapture Density Estimation Using Acoustic Recording Units Without Individual Identification
Ben Augustine; J. Andrew Royle; Angella K. Fuller
Acoustic recording units (ARUs) have increasingly been used to monitor wildlife populations for species that are difficult or impossible to monitor using other noninvasive methods. Most studies using ARUs to date estimate occupancy rates, rather than population density, because the calls of most species are not identifiable to individual, a typical requirement for estimating population density. Recently, distance sampling and spatial capture-recapture (SCR) approaches have been developed for unidentified bioacoustic detections, but they require supplementary data on sound attenuation or call rates and movement parameters, which likely vary across space and time. Extending recent developments in SCR with unknown or partial individual identities, we developed a model that probabilistically assigns calls to individuals, allowing for the joint estimation of population density, individual call rate, and individual space use. This model consists of a Neyman-Scott type cluster point process, where individuals are “parents” and call locations are “children”. Unlike typical cluster point process models, we do not observe the children (calls) directly—their number, location, and individual membership are estimated using an observation model such as a sound attenuation function, whose parameters are also estimated from the data in hand. We demonstrate that density estimation using SCR-based methods for bioacoustic detections is possible without individual identities and supplementary information about sound attenuation or call rates and movement parameters, though supplementary information for these parameters can be included via informative priors. Further, call features that correlate with individual identity (e.g., spectrograph features) can be used as partial identity covariates to further improve estimation. Finally, this estimation framework may be applied to other capture-recapture problems such as area searches with imperfect detection and density estimation of group-living species whose group membership is not known when detected.
Modeling Community Occupancy from Line Transect Data: A Case Study with Large Mammals in Post-War Angola
Lisanne Petracca; Paul Funston; Phil Henschel; Jonathan Cohen; Seamus Maclennan; Jacqueline Frair
Human disturbance can have a profound effect on the occurrence and distribution of wildlife. Such disturbance often extends into protected areas, particularly in countries that have undergone civil strife and lack the institutional capacity to effectively mitigate anthropogenic threats. We demonstrate the first application of a multi-species hierarchical occupancy model to spatially-correlated detections from vehicle-based spoor transects, estimating species richness and species-specific drivers of occurrence of a large mammal community comprising five large carnivores (cheetah [Acinonyx jubatus], leopard [Panthera pardus], lion [Panthera leo], spotted hyena [Crocuta crocuta], and wild dog [Lycaon pictus]) and six large herbivores (buffalo [Syncerus caffer], eland [Taurotragus oryx], elephant [Loxodonta africana], giraffe [Giraffa giraffa], roan [Hippotragus equinus], and sable [Hippotragus niger]) in Luengue-Luiana and Mavinga National Parks, Angola. This area is the largest contiguous national park complex within a single African country and part of the largest transfrontier conservation area in the world, and is still recovering from the effects of civil war. In this post-war landscape, the most substantive drivers of community-level occupancy were anthropogenic, with occupancy associated with lower frequency of human sign, proximity to adjacent national park, and distance away from human settlement. In contrast, ecological variables (precipitation, vegetation cover, seasonal water availability) had less explanatory value. Our results highlight the deleterious effects of human incursion into protected areas on the richness and distribution of large mammal species, underscoring the need for intensive mitigation of anthropogenic threats (e.g. poaching, bushmeat hunting) to maintain species of high conservation value in areas impacted by war.
Predicting Migration Corridors: Using Maximum Likelihood to Fit Corridor Models to Movement Data
Tristan Nunez; Mark Hurley; Julien Fattebert; Jerod Merkle; Anna Ortega; Hall Sawyer; Tabitha Graves; Matthew Kauffman
Conserving migratory populations requires knowing the location of migratory corridors. GPS collars have greatly advanced knowledge of ungulate migration corridors by allowing direct observation of migratory movements. However, due to logistics and limited funding, many ungulate populations cannot be collared, making it difficult to identify their migration corridors. We developed a novel approach to predicting migration corridors that uses maximum likelihood to fit cost distance movement models to GPS telemetry data. Model selection can identify the best models, which can then be used to predict the corridors used by populations with limited collar data. We demonstrated using simulations that maximum likelihood estimation can recover the resistance parameters used to simulate movements. We then used data from multiple mule deer migrations in Idaho and Wyoming to fit cost distance models with the following covariates: date of peak green-up, elevation, percent shrub, snow-off date, and human footprint. We found that migration routes minimize stepwise change in date of peak green-up, follow areas of higher stepwise change in elevation and higher percentage shrub habitats, but that the strength of these relationships varied depending on the migration. In addition to the practical benefit of mapping corridors, this approach can address conceptual ideas in migration ecology that center around navigation, seasonal range decisions, fidelity, and movement constraints. Models of predictive corridors can inform management and planning in supporting healthy ungulate migrations.
Basic Machine Learning Models Using Spatially Optimized Landscape Variables Predict the Abundance of a Declining Sagebrush-Obligate Bird
Kimberly M. Serno; Kevin J. Gutzwiller
Machine learning statistical methods can identify complex patterns in data and derive predictive models from these patterns. Planners and managers can use machine learning models that predict species’ distributions and habitat associations to advance conservation. We compared the ability of basic random forest models and support vector machine models to predict the total abundance (abundance) and coefficient of variation in total abundance (CV) of Brewer’s Sparrow (Spizella breweri), a sagebrush-obligate species in decline. Using data from North American Breeding Bird Survey routes within the Brewer’s Sparrow’s breeding range for 2009-2013 (n = 356 routes), we assessed whether habitat, non-habitat, and drought variables at 6 biologically relevant spatial extents (0.4km, 1km, 4km, 5km, 8km, 20km) predicted Brewer’s Sparrow abundance and CV. For each explanatory variable, we included in subsequent modeling the measure at the spatial extent that was most correlated with each response variable. Based on 10-fold cross-validation, the best random forest model had a mean absolute error of prediction for test data (MAE) of 31 (6% error relative to the range) for abundance and 26% (11% relative error) for CV. Proportion of desert/semi-desert (0.4km) land cover was the most important predictor for both abundance and CV. The best support vector machine model had a MAE of 31 (6% relative error) for abundance and 18% (8% relative error) for CV. Proportion of dwarf sagebrush shrubland/steppe (1km) was the most important predictor for abundance, whereas proportion of montane sagebrush steppe (20km) was the most important predictor for CV. These basic models made reasonable predictions of Brewer’s Sparrow abundance and CV, and increasing their complexity has the potential to improve predictive performance. Good models can predict Brewer’s Sparrow abundance and CV under changing land-cover and climatic conditions. Such models would inform management strategies for Brewer’s Sparrow conservation.
Relating Spatial Predictors to Wildlife Abundance Using Sightability Models
Althea ArchMiller; John Fieberg
Sightability models account for imperfect detection in wildlife surveys by applying logistic regression models to observations of marked individuals. Abundance can be estimated using either a modified Horvitz Thompson (mHT) estimator or a model-based approach. The model-based approach offers several advantages, including: 1) the potential to model abundance as a function of habitat composition within the sample plots; 2) the ability to “borrow strength” across years when estimating abundance; and 3) the ability to estimate trends directly as part of the model. Our focus here is on the development of models with spatial covariates and comparisons of model-based and mHT approaches in terms of estimator performance. We compared mHT and model-based estimators across a range of simulated sampling efforts and data generating scenarios. We also reanalyzed data from 2006 and 2007 winter aerial surveys of moose (Alces alces) in Minnesota, including shrub cover as a predictor of plot-level abundance. Although the precision of model-based and mHT estimators were similar, the model-based approach was able to correctly identify an abundance-covariate relationship when present in simulated data. We found that Minnesota moose abundance in 2007 was positively related to plot-level shrub cover, but no such relationship was identified in 2006. We conclude that the model-based approach shows promise for testing hypotheses regarding spatial features that affect population abundance and distribution.
Accounting for Asymmetry in Ecology
Perry Williams
Asymmetric regression is an alternative to conventional linear regression that allows us to model the relationship between predictor variables and the response variable (e.g., abundance, survival, occupancy) while accommodating skewness. Advantages of asymmetric regression include robustness to model misspecification and less sensitivity to outliers. Bayesian asymmetric regression relies on asymmetric distributions such as the asymmetric Laplace (ALD) or asymmetric normal (AND) in place of the normal distribution used in classic linear regression models. Asymmetric regression concepts can be used for process and parameter components of hierarchical Bayesian models and have a wide range of applications in wildlife data analyses. In particular, asymmetric regression allows us to fit more realistic statistical models to skewed data and pairs nicely with Bayesian inference. Our objective is to describe asymmetric regression using the ALD and AND. Second, we show how the ALD and AND can be used for Bayesian quantile and expectile regression for continuous response data. Third, we consider an extension to generalize Bayesian asymmetric regression to count data. Fourth, we describe a regression model using the ALD, and show that it can be used to add needed flexibility, resulting in better predictive models compared to Poisson or negative binomial regression. We demonstrate concepts by analyzing a data set consisting of counts of Henslow’s sparrows following prescribed fire. Our results suggest Bayesian asymmetric regression is an essential component of an wildlife ecologist’s quantitative toolbox.
Estimating Demographic Parameters in Animal Communities Using Detection-Nondetection Data with a ‘Multi-Species N-Occupancy Model’
Matthew T. Farr; Timothy O’Brien; Charles B. Yackulic; Elise F. Zipkin
Unmarked analyses have expanded the capabilities of data to estimate both abundance (N-occupancy, Royle-Nichols model) and dynamics (Dail-Madsen model) of a species population. Recent progress in unmarked analyses has synthesized a dynamic N-occupancy model from the combination of the Royle-Nichols and Dail-Madsen models to jointly estimate population abundance and dynamics from only detection/nondetection data. The scope of community ecology could benefit by leveraging this advancement as many species are cryptic or elusive, and there may only be detection/nondetection data available. We extended a single-species dynamic N-occupancy framework to estimate population abundance and demography at a community-level. This multi-species model hierarchically joins species by specifying random effects that assume species-specific abundance and demography are distributed from a community-level. Additionally, environmental drivers of community and species-specific demography can be identified within the multi-species dynamic N-occupancy model. We demonstrated the capabilities of our model on a community of duikers (Cephalophus spp.) across Central and East Africa. It’s hypothesized that declines in duikers across this region is caused from both habitat loss and poaching and rising anthropogenic pressure along the borders of protected areas is concerning for the status of duikers. Unfortunately, the cryptic nature of duikers complicates the assessment of their abundance and demography where only detection/nondetection data collected from The Tropical Ecology Assessment and Monitoring (TEAM) Network exists. Using the multi-species dynamic N-occupancy model, we estimated community and species-specific effects of anthropogenic pressure on the abundances and dynamics of duikers from 2009 – 2017 across six national parks in Central and East Africa. Our results provide valuable demographic information on an understudied community of duikers, and our novel multi-species dynamic N occupancy framework provides an advancement to community ecology where frequently limited data and understanding can benefit from robust analyses.
Multistage Nest Survival: A Hidden Markov Model Approach When Nest Age Or Stage Is Uncertain
William L. Kendall; Reesa Y. Conrey; James H. Gammonley
Nest survival is an important component of assessing the reproductive performance, and hence population dynamics of many species. In avian studies, individual nests are often discovered at staggered times within the nesting season, and at various days since nest initiation. For many species the age of the nest can be determined by assessing the extent of development of the eggs, in which case standard nest survival models (e.g., Dinsmore et al. 2002) are effective. For other species, such as many raptors, frequently the nest cannot be directly accessed, and the age of the nest cannot be determined, and therefore the day of completion, if successful, of nesting cannot be predicted. Partial information on development, such as nesting stage (incubation or nestling) might be available for these cases. We present a hidden Markov model (HMM) approach to estimating stage-specific nest survival, where the specific state (nest age) is mostly latent, but partial information on the state (nest stage) could be known. We apply the model to a 5-year data set on bald eagles (Haliaeetus leucocephalus) in Colorado, collected by a mixture of citizen scientists and professional biologists, where nest stage (egg laying, incubation, chick rearing) was sometimes known. We also use simulation to assess the effectiveness of different patterns of frequency and timing of nest visits on bias and precision of survival estimates.
A Spatial Capture-Recapture Model for Group-Living Species
Robert L. Emmet; Ben Augustine; Briana Abrahms; Lindsey N. Rich; J.W. McNutt; Alan M. Wilson; Brett T. McClintock; Beth Gardner
Spatial capture-recapture (SCR) models have been applied to a large number of species, including those that are considered group-living species. Standard SCR models applied to group-living species assume that group members move independently, though this is not often the case. Recently, SCR models that include a group component have been proposed, thus allowing researchers to estimate density and home range sizes of individuals and groups. However, these models do not include explicit group movement dynamics. Group dynamics are an important part of the detection process, as individuals within groups likely have correlated movements, which may bias density estimates in group-living species. We developed a new SCR model for group-living species with an unknown number of groups. The model estimates the number of groups and group sizes using a clustering algorithm, assigning individuals to groups based on proximity to group activity centers. The model allows for the estimation of the density and home range sizes of both individuals and groups of animals. We examined the performance of this new model through a simulation study and applied the model to an empirical case study of African wild dogs in the Okavango Delta of Botswana. In this well-studied population of wild dogs, there is a known number of groups and individuals within each group. Standard SCR models were previously applied to this dataset and shown to perform well for estimating abundance. However, in our simulation study, we found that group movement dynamics can bias density estimates from standard SCR models. Thus, the new group SCR model promises to advance understanding of group movement dynamics and better estimate density of group-living species.


Contributed Oral Presentations
Location: Virtual Date: Time: -