Biometrics & Population Modeling III

Contributed Paper
ROOM: CC, Room 20
SESSION NUMBER: 43
 

1:10PM A New Tool for Measuring Autocorrelation in Binary Data: Using Lorelograms to Assess Dependence in Camera Trap Data
Fabiola Iannarilli; Todd W. Arnold; John R. Fieberg
Most ecological data are correlated in space or time. The strength of these dependencies (i.e., correlation) as a function of distance or time can provide insights into important ecological processes. In addition, the distance (in time or space) at which observations no longer exhibit correlation can inform sampling and experimental designs aimed at generating independent data. Although a variety of statistical methods are commonly used to study spatial and temporal dependence of continuous variables, few ecologists seem to be aware of similar tools available for binary data. We demonstrate how the lorelogram (Heagerty & Zeger 1998) can be used to quantify and visually explore dependencies in binary data generated from camera traps. Using real and simulated data, we show how the lorelogram can describe temporal dependence at different time scales, quantify time-to-independence between sequential detection events of the same individual or species, and provide insights into behavioral responses to extrinsic factors, such as the presence of sympatric species (which might induce either positive or negative correlations), use of lures to enhance detection probabilities, or visitations by investigators to a trapping or field site. Ecologists need to think critically about how spatio-temporal components in binary data can affect data collection and analyses. Previous studies have shown that naively assuming data are independent can lead to underestimation of variance and increased type I errors. The lorelogram could be used to help select appropriate time intervals that would allow data to be treated as independent. We argue, however, that spatial and temporal correlations are strong components of the data-generating process, and they should not be considered as nuisance parameters, but rather as interesting properties of the data that can yield novel insights into behavioral and spatio-temporal patterns.
1:30PM Getting a Handle on Telemetry Error: Speed, Distance and Flight-Height Estimation with Tracking Data
Christen H. Fleming; Michael J. Noonan; Justin M. Calabrese
Animal tracking data are being collected at an increasing rate and with increasing detail. However, for finely sampled data, it is under appreciated that most of the information conveyed by simple movement metrics, such as step lengths and turn angles can be dominated by error. Here we describe methods for outlier detection, path reconstruction, distance/speed estimation, and home-range estimation based on continuous-time movement models that decouple the sampling schedule from the movement process and tease apart movement from error. Critical for the success of these methods is the task of error calibration, which needs to become standard practice when collecting animal tracking data. We highlight with empirical examples the dangers of non-existent and unreliable error models, demonstrating the inflated outlier detection rate, overestimation of distance/speed, inaccuracy of path reconstruction, overestimation of home-range, and the general inability to distinguish between genuine animal behavior and error, with the potential for generating spurious/artifactual results. Example species include wood turtles (Glyptemys insculpta), lowland tapir (Tapirus terrestris), night heron (Nycticorax nycticorax), barn owl (Tyto alba), and Bornean orangutan (Pongo pygmaeus). In summary, finely sampled tracking data needs to be paired with equally sophisticated analytic methods, and what constitutes “finely sampled” data is both species and question dependent.
1:50PM An Unsupervised Machine-Learning Algorithm for Behavioral Classification From Animal-Borne Accelerometers
Jane Dentinger; Roy Jafari; Luca Borger; Brian K. Smith; Garrett M. Street; Bronson K. Strickland
Resource availability is one of the most important predictors of reproductive success, distribution, and survival of organisms from individuals to populations. Studies of animal distributions typically use our understanding of animal habitat requirements and behavioral ecology to deduce the most likely explanations of observed habitat use. Early studies linking behavior to landscapes frequently focused on identifying where animals are in space. However, a critical limitation of these data is that they only provide information about where an animal was observed but not why the animal was in that location. An animal’s behavior can be distinguished by a series of distinct signals in an accelerometry data set. Past research has focused on using supervised machine-learning techniques such as decision trees, Random Forests, and artificial neural networks to identify behavioral groups. These approaches are limited to predefined behaviors with sufficient training data to identify and characterize the accelerometry signals. Unsupervised approaches benefit from not requiring known behaviors prior to classification, allowing the model to identify as many unique signal types as permitted by the data. Here we present a framework that couples data sets of known behaviors collected through direct observation of animals to unique behavioral signal bins obtained via unsupervised machine learning on a large accelerometry data set (>1 billion observations). Specifically, we use orthogonal transformation to derive the principle components of the data set, which are subsequently aggregated into groups of similar signal types using an unsupervised artificial neural network. Unique groups are identified using spatial clustering analysis performed on a Self-Organizing Map (SOM). Because observed behaviors may occur across multiple clusters, we construct a “confusion matrix” describing the proportion of observed behaviors occurring within a signal cluster to explicitly characterize the error in classification, and we examine the predictive accuracy of classifications both within and between individual animals.
2:10PM Location Uncertainty in Azimuthal Telemetry Data and the Spatial Ecology of the Threatened Gunnison Sage Grouse.
Brian D. Gerber; Mevin Hooten; Christopher Peck; Mindy Rice; Mindy Rice; James Gammonley; Anthony Apa; Amy Davis
Telemetry technologies are fundamental tools used by ecologists to observe spatial locations of animals. While the use of satellite-based telemetry technology is becoming more ubiquitous, researchers have historically relied on, and still often use, radio telemetry to collect spatial location information. Common practice is to use available software to estimate animal locations, which are then used for subsequent analyses, such as resource selection, home range, density, or site fidelity. This process ignores animal location uncertainty, which can be highly complex, and has far-reaching implications on ecological inference and conservation decision making. We develop hierarchical Bayesian models for radio telemetry data that accommodate multiple sources of uncertainty. We evaluate our model using simulation, comparing common estimators under a variety of study designs. We also demonstrate the importance of accounting for telemetry-based uncertainty in home range analyses, resource-selection functions, and site fidelity using data on the Gunnison sage-grouse (Centrocercus minimus), a federally threatened species. We find good performance of our telemetry model in estimating animal locations across study designs and the only modeling approach that has appropriate measures of coverage. Ignoring animal location uncertainty when estimating resource selection or home range can have pernicious effects on ecological inference. We show that home range estimates can be overly confident and conservative when ignoring location uncertainty and resource selection coefficients can lead to incorrect inference and over confidence in the magnitude of selection. Our findings and model development have important implications to ecological inference from historical analyses using this type of data and the future design of radio-telemetry studies. Lastly, we demonstrate that the Gunnison sage-grouse uses an ‘always-stay’ site fidelity strategy, which has important implications for spatial population structure and dynamics.
2:30PM Refreshment Break
3:20PM Using Semi-Supervised Machine Learning Anomaly Detection for Prediction of Parturition of Ungulates
Alison C. Ketz; Daniel J. Storm; Daniel P. Walsh
Animal tracking technology is used in wildlife ecology for inferring animal movement, linking behavior to measures of population dynamics, resource use, and fitness. We apply a machine learning algorithm to predict parturition of white-tailed deer (Odocoileus virginianus) using features derived from movement data collected in nearly real time. Machine learning methods are powerful prediction tools that have rarely been used in ecology because the underlying focus is to predict an outcome, rather than to obtain inference about mechanisms driving ecological patterns. We use data recorded from GPS devices deployed on adult females, including latitude, longitude, altitude, and temperature, and calculate quantities to use in a predictive machine learning algorithm. We used vaginal implant transmitter data from the first year of the study to train a semi-supervised anomaly detection algorithm using leave one out cross validation. We automated the algorithm to aid daily sampling efforts for the capture of neonates immediately following predicted births for the subsequent year of the study. Prediction of parturition to facilitate surveys for inferring neonate survival of ungulates has previously been developed for ungulate species that have a drastic movement prior to birth, but the behavioral patterns of white-tailed deer provide a difficult challenge because the pattern of movement is the contraction of an already highly localized home-range. The method we describe is a highly flexible and effective approach for detecting changes in behavior with limited data and could be applied to any wildlife species where location data and changes in patterns of behavior reflect different ecological states. We present the novel application of a machine learning algorithm for detection of changes in movements and space use to predict a common ecological process, birth, and to facilitate surveys, thereby improving our understanding of processes important in population ecology.
3:40PM Carnivore Territoriality: Emergence of Population-Level Patterns From Individual Behaviors
Sarah N. Sells; Michael S. Mitchell; Kevin M. Podruzny
We developed individual-based models to understand how wolves (Canis lupus) select territories. Territorial animals are expected to be adapted to select territories efficiently based on spatially distributed resources. This means animals should select economic territories that maximize benefits acquired from resources (e.g., food) against costs of acquiring them (e.g., competition for resources). Accordingly, our models capture the hypotheses that wolves select territories based on the benefit of prey and costs of travel, humans, and intraspecific competition. The models reveal how benefits and costs interact to influence territorial behavior and generate population-level patterns in territoriality. The models show that on average, territories are expected to be smaller and of higher quality where food resources are more clumped. Territory sizes are also more variable in these scenarios, and there is more overlap among territories. Additionally, territories are expected to be smaller where food resources are more abundant. Wolf territories are therefore expected to vary based on ungulate ecology, behavior, and abundance. The models also show that as human use increases in a given area, territory size generally increases as well. Furthermore, if wolves associate humans with higher costs, they select territories to avoid humans at greater rates. Wolf territories are therefore expected to vary based on human use and management decisions, e.g., degree of harvest pressure. We also present application of these predictions to empirical observations of patterns in wolf territoriality. Ultimately, we will use our models and a patch occupancy modeling framework to predict territorial behavior and abundance of wolves in Montana and Idaho.
4:00PM An Interesting Nuisance: A Bayesian Hierarchical Model for Distance Sampling When a Detection Covariates Is Also a Demographic Property of Interest
Heather E. Gaya; Bryan L. Nuse; Clinton T. Moore
Line-transect distance sampling (LTDS) is a popular method for estimating abundance and density of animal populations. Covariates occurring in statistical models, including models for LTDS data, typically serve as nuisance adjustments to improve inference about the response by accounting for sources of variability potentially affecting it. However, inference on the covariate itself, such as its distributional properties, is sometimes of interest if the attribute relates to population status. To exemplify this sampling situation, we focus on the gopher tortoise (Gopherus polyphemus), a terrestrial reptile endemic to the southeastern United States that digs its own burrows. We present a Bayesian model for line-transect distance sampling in which a covariate associated with the animal, burrow width, helps to account for detection bias of all burrows present and improves precision of estimated density. By modeling the number of burrows missed, by size interval, as a latent quantity, we also produce a detection-adjusted size frequency distribution of occupied burrows, an indicator of population age structure. We compare the raw distribution of burrow diameter found during LTDS and the estimated posterior distribution produced when detection probabilities are accounted for. We further compare the density point estimates and precision produced from DISTANCE 6.2 and our own model for both a simulated and real dataset. Approximated burrow size distributions for the smallest covariate values (≤10 cm) from our model were consistently closer to the true distribution than estimates produced from the raw LTDS data regardless of population density. Except at lowest population density (1 individuals/hectare), the point estimates from our covariate model were more accurate than those from DISTANCE 6.2. Our covariate model provides a powerful tool for understanding the distribution of an individual-level attribute that dually affects the detection of organisms and is informative about population status.
4:20PM Negative Bias in Survival Estimates Due to Individual Misidentification in Long-Term Mark-Resight Studies
Anna M. Tucker; Conor P. McGowan; Robert A. Robinson; Jacquie A. Clark; James E. Lyons; Audrey DeRose-Wilson; Richard du Feu; Graham E. Austin; Philip W. Atkinson; Nigel A. Clark
Long term capture-recapture datasets can provide valuable ecological insights and monitoring for conservation and management. For mark-resight datasets relying on noninvasive observations (e.g. field readable color bands, natural marks) there is the potential to accumulate false positive detections of individuals as the length of the study and number of tags deployed increases, particularly when data are collected by observers with varying levels of experience. Here we estimate the range of possible apparent misread rates for shorebird leg flags using multiple estimation methods. We assessed apparent error rates in mark-resight observations of red knot (Calidris canutus rufa) marked with plastic leg flags inscribed with field-readable alphanumeric codes in Delaware, USA. Flag resights for this monitoring program are collected by volunteer observers with a wide range of expertise and experience. We assess flag- and observer- specific attributes associated with higher apparent misread probabilities in our dataset, and quantify the effect of misreads on estimates of apparent annual survival using a simulation study. The apparent error rate in our data set ranged from 0.45%-7.1% depending on method used to estimate apparent misreads. Apparent error rates and among-observer variation in error rates were negatively associated with the total number of resightings an observer had recorded. Our simulation study indicated that misidentification may cause negative bias in estimated apparent annual survival of up to 20-30% and that spurious support for models with time-varying survival and detection (when data were simulated with time-constant survival and detection probabilities) could arise when misread rates were ≥ 5%. We recommend methods for quantifying error rate in a long-term mark-recapture dataset, suggest data cleaning methods that may reduce the occurrence of false positive detections, and discuss model-based options to reduce bias from false detections.
4:40PM The Fast and the Spurious: Scale-Free Estimation of Speed and Distance Travelled From Animal Tracking Data
Michael J. Noonan; Chris H. Fleming; Justin M. Calabrese
Speed and distance travelled provide quantifiable links between behavior and energetics, and are among the metrics most routinely estimated from animal tracking data. These are typically quantified by summing the straight-line distance (SLD) between locations, where distance is divided by time to estimate speed. Problematically however, this approach suffers from severe, scale-dependent bias, such that estimates are not only influenced by how far/fast the animal actually travelled, but also by the sampling frequency, the tortuosity of the animal’s movement, and the amount of telemetry error. Moreover, SLD estimates do not come equipped with confidence intervals to quantify their uncertainty. Therefore, it is currently not possible to determine if a set of SLD-based estimates are statistically different from one another. These issues present a serious problem for any comparative analyses of SLD estimates. To alleviate the scale-dependency of SLD estimation, we have developed a new scale-free method for estimating speed and distance travelled that builds upon the existing continuous-time movement modeling framework. Using a combination of simulations, and empirical data from wood turtles (Glyptemys insculpta), and white nosed coatis (Nasua narica), we demonstrate how our novel estimator provides accurate, scale-free estimates with reliable confidence intervals. The net result are rigorous estimates of speed and distance travelled that can validly be compared across studies, sites, species, sampling schedules, and tracking devices. In addition to being statistically rigorous, this method also benefits from being computationally efficient, a property that is well suited to the growing volume of data used in these analyses.
12:50PM Which Home Range Estimator Should I Use? an Analysis of Autocorrelation and Bias in Home Range Estimation
Justin M. Calabrese; Michael J. Noonan; Christen H. Fleming
Home range estimation is a routine practice in ecological research. While advances in animal tracking technology have increased our capacity to collect data to support home range analysis, the question of which method should be applied to these data has yet to be resolved. A primary reason for this is that tracking data are often strongly autocorrelated, while most home range estimators assume independently sampled data. Here, we provide a comprehensive evaluation of the effects of autocorrelation on home range estimation. We base our study on recently developed tools for working with and simulating autocorrelated tracking data, which allow both the amount and nature of autocorrelation to be precisely manipulated, and a dataset of GPS relocations from 369 individuals spanning 27 species. We define objective measures of bias induced by autocorrelation, including half-sample cross-validation, and compare autocorrelated KDE (AKDE), multiple conventional KDE bandwidth optimizers, Minimum Convex Polygon (MCP), and Local Convex Hull (LoCoH) methods. We found that AKDE 95% area estimates were significantly larger than conventional KDE and MCP estimates by, on average, a factor of 2, and LoCoH estimates by a factor of 13. Importantly, the median number of cross-validated relocations included in AKDE 95% (or 50%) contour estimates was 95.3% (or 50.1%), confirming the larger AKDE range estimates were appropriately selective at the specified quantile. Conversely, conventional estimates exhibited significant negative bias, and cross-validation regularly failed to capture an animal’s future space use. Simulations varying sampling duration, frequency, and movement processes also demonstrated that AKDE was generally more accurate than conventional methods. Crucially, in the frequently encountered scenario of small effective sample sizes and highly autocorrelated data, AKDE stood alone as the only method capable of producing consistently accurate home range estimates.

 

Contributed Paper
Location: Cleveland CC Date: October 9, 2018 Time: 12:50 pm - 5:00 pm