Euring banner kiwi picture

EURING 2007

Software Session

14 January 2007

Dunedin, NZ




TimePresenterTitleslidesduration
9:00-9:15Jim HinesWelcome,Introduction.15
9:15-10:05Gary WhiteMARK MCMC & NewFeaturesna50
10:05-10:45Chris FonnesbeckPyMC -software for MCMC mark-recapture analysispdf(3.7Mb)40
10:45-11:00Break.15
11:00-11:40Olivier Gimenez Assessing population dynamics using WinBugspdf(2.5Mb)40
11:40-13:00Lunch.80
13:00-13:40Mark MaunderAD Model Builder -highlighting its use for random effects for time-specific parameters in mark-recapture modelspdf(1.1Mb)40
13:40-14:20Remi ChoquetE-SURGE :First steps for the practice of models with state uncertaintypdf(1.8Mb)40
14:20-14:35Break.15
14:35-15:15Len ThomasAnalysis of mark-recapture distance sampling surveys using DISTANCE5 pdf(1.6Mb)40
15:15-15:55Murray EffordDensity -spatially explicit capture–recapturena40

MCMC Estimation Procedures with Program MARK

Gary C. White

MCMC estimation routines have been installed in Program MARK. I will demonstrate how to specify inputs, including: prior distributions for the beta parameters; number of tuning, burn-in, and samples from posterior distributions; hyperdistribution parameters and their prior distributions (including correlations between sets of parameters and design matrices with covariates to model hyperdistribution parameters), and convergence diagnostics. Samples from posterior distributions for the beta, real, and hyperdistribution parameters are stored in a binary file (as well as multiple chains in a single analysis) that can be input to standard statistical packages. SAS and R codes are currently available. Because the FORTRAN code is optimized for the mark-encounter models coded in MARK, the computer time required to perform MCMC estimation is considerably less than more general Bayesian estimation programs. Users familiar with the MARK interface will be able to conduct Bayesian estimation with no modification of existing models. Other recent additions to MARK will be reviewed as time permits.

PyMC-software for MCMC mark-recapture analysis

Chris Fonnesbeck

We demonstrate the implementation of several CMR models, both open- and closed population, which include hierarchical covariate relationships among parameters, using PyMC. PyMC is an open-source module for the Python programming language that implements MCMC sampling via the Metropolis-Hastings algorithm. We highlight the relative ease with which even complex models are implemented in PyMC, as users are only required to specify the parameters and the joint log-likelihood of the model. Important features include relational database connectivity, convergence diagnostics, goodness-of-fit diagnostics, autocorrelation plots, and profiling tools.

Assessing population dynamics using WinBugs

Olivier Gimenez

The computer package WinBUGS is introduced. After reviewing the main concepts of Bayesian theory and MCMC methods that are needed to run this program, we illustrate the main steps of its utilisation by using simple linear and logistic regressions. We pay particular attention to assessing convergence, checking identifiability and testing goodness-of-fit as well as the possibility of running WinBUGS from program R or Matlab for more convenience. We then focus on 3 case studies to show how WinBUGS can be used when the classical theory is more difficult or impossible to implement. The first example deals with mark-recapture/recovery models to estimate survival and shows how to cope with unexplained variance through random effects. Recent progresses are also exposed including a flexible way of incorporating covariates using splines smoothing, variable selection among a large number of covariates using Reversible Jump MCMC and how to deal with missing values in covariates. The second example shows how to estimate population density while accounting for detectability by using distance sampling methods, with special emphasis on spatial modelling. Finally, the third case study involves modelling data using state-space models which provide an integrated framework for assessing population dynamics. WinBUGS code and real data are provided for all examples.

Comparison of estimators for mark-recapture models using AD Model Builder

Mark N. Maunder, Hans J. Skaug, and Simon D. Hoyle

AD Model Builder (ADMB) is a general modeling environment for fitting parameter rich nonlinear models to data. It uses automatic differentiation to provide a more efficient and stable parameter estimation framework. ADMB has been used to fit complex nonlinear models with thousands of parameters simultaneously to multiple types of data and to fit nonlinear models with fewer parameters to hundreds of thousands of data points. We describe the basic use of ADMB and its many features. Some of ADMB features include automatic profile likelihoods, random effects using Laplace approximation and importance sampling, and MCMC to implement Bayesian analysis. To demonstrate ADMB, we apply ADMB to simulated mark recapture data to compare maximum likelihood, random effect, and Bayesian estimators.  

Program E-SURGE:First steps for the practice of models with state uncertainty

Choquet Rémi, Rouan Lauriane

We introduce a new program called E-SURGE for fitting models to capture-recapture data. Built upon the concepts developed in program M-SURGE (Choquet et al. 2004), this software contains new features allowing to build a wider range of multistate models and the recently developed multievent models (Pradel, 2005). These last models generalize the multistate models by taking into account uncertainty in state assessment. The generalization goes through the replacement of the traditional encounter probabilities with “event” probabilities and the introduction of a new type of parameters, the initial state probabilities. Independently of the multievent generalization, E-SURGE adds more flexibility to model designing by allowing the decomposition of transitions into successive steps. To illustrate the potential of E-SURGE, we will detail the practical implementation of three examples with increasing degrees of complexity:
  1. Decomposition of dispersal as a two-step process (Grosbois & Tavecchia 2003): Does the animal disperse? If it does, where does it settle? This feature allows handling fidelity and direction of movement independently.
  2. Multievent formalization of a model with two classes of catchability (Pledger 2000, Pradel 2005): this example serves to introduce the initial state probabilities, which estimate here the proportions of individual belonging to each class of catchability.
  3. Implementation of the memory model (Brownie et al. 1993, Pradel 2005) as a multievent model: we focus here on the role of the event matrices (a generalization of the capture probabilities) which make the link between what is observed (the events) and the underlying states.

Analysis of mark-recapture distance sampling surveys using DISTANCE

Jeff Laake, David Borchers and Len Thomas

Distance is a windows-based software package for the design and analysis of distance sampling surveys of wildlife population size and/or density. In these surveys, observers move along a transect line or stand at a point, and record the distance to all objects of interest (usually animals) detected out to some truncation distance, w. These distances are used to estimate the average probability of detecting an object given it is within distance w. From there it is simple to estimate population size. A fundamental assumption of conventional methods is that all objects at zero distance are detected with certainty - but this is by no means assured for some wildlife taxa (e.g., marine mammals). Recently, methods have been developed that combine mark-recapture ideas with distance sampling, and allow estimation of probability of detection at zero distance. Two or more (semi-) independent observer teams "mark" objects by detecting and tracking them, and thereby set up a trial for the other team(s), who may either detect ("recapture") each object, or fail to detect it. We give an overview of these methods, and describe how they have been implemented in the Distance software, using as an example an analysis of aerial surveys of crabeater seals in Antarctica. The new analysis methods have been implemented as an R library, and we outline how to access them directly in R. Lastly, we give a brief overview of other new developments in the Distance software and plans for the future.

DENSITY-spatially explicit capture–recapture

Murray G Efford

Density provides capture–recapture methods for spatially distributed data such as those from grid trapping. A previous version implemented simulation-based methods for estimating population density (Efford 2004, Efford et al. 2004). The latest version (4.0) provides flexible likelihood-based model fitting and model selection (Borchers & Efford in review). Other closed population analyses have been added, and the simulation capability extended. This session will introduce participants to the Density interface through the analysis of a mist-netting dataset and the simulation of bird point counts.