Presentation Title
A candle in the dark: Accounting for undetected species using Bayesian hierarchical models
Abstract
Species occupancy, or the presence or absence of a species at a location, is important for informing conservation decisions. However, observations of ecological processes, such as occupancy, species richness, or species’ responses to environmental factors, are often biased due to observer error. Multi-species occupancy models (MSOMs) account for observer error by jointly modeling an ecological process— in this case, occupancy— and a survey process that is dependent on an ecological process. MSOMs can also model rare species by assuming all species come from a community-wide distribution: essentially, the rare species “borrow” information from those that are common. Using the same assumption, MSOMs can model species that were never detected during sampling using a method known as data augmentation. However, the lack of data for augmented species poses additional challenges: 1) augmentation provides no information about species identity, which must be assigned post analysis, and 2) the model tends to “pull” data-deficient species towards the center of the community distribution. Using a Bayesian approach may resolve these challenges by using informed priors for augmented species, but the efficacy of this method has yet to be tested. I will use simulated data to test whether the use of informed priors improves the accuracy of occupancy estimates. Specifically, I will compare the results of MSOMs using 1) uninformed priors, 2) informed priors, and 3) misinformed priors. Finally, I will apply data augmentation and informed priors to a real dataset of small mammal communities in Vermont to model undetected species that are known to occur in the study region.
Primary Faculty Mentor Name
Nicholas Gotelli
Status
Graduate
Student College
College of Agriculture and Life Sciences
Program/Major
Biology
Primary Research Category
Biological Sciences
A candle in the dark: Accounting for undetected species using Bayesian hierarchical models
Species occupancy, or the presence or absence of a species at a location, is important for informing conservation decisions. However, observations of ecological processes, such as occupancy, species richness, or species’ responses to environmental factors, are often biased due to observer error. Multi-species occupancy models (MSOMs) account for observer error by jointly modeling an ecological process— in this case, occupancy— and a survey process that is dependent on an ecological process. MSOMs can also model rare species by assuming all species come from a community-wide distribution: essentially, the rare species “borrow” information from those that are common. Using the same assumption, MSOMs can model species that were never detected during sampling using a method known as data augmentation. However, the lack of data for augmented species poses additional challenges: 1) augmentation provides no information about species identity, which must be assigned post analysis, and 2) the model tends to “pull” data-deficient species towards the center of the community distribution. Using a Bayesian approach may resolve these challenges by using informed priors for augmented species, but the efficacy of this method has yet to be tested. I will use simulated data to test whether the use of informed priors improves the accuracy of occupancy estimates. Specifically, I will compare the results of MSOMs using 1) uninformed priors, 2) informed priors, and 3) misinformed priors. Finally, I will apply data augmentation and informed priors to a real dataset of small mammal communities in Vermont to model undetected species that are known to occur in the study region.