1:55-2:20 PM Computing issues concerning hierarchical models - Ken Burnham The subject of capture-recapture (CR) data analysis is incomplete without well-developed, user capable software (such as MARK) that can fit models incorporating random effects as well as fixed effects. The MCMC (i.e., Bayesian) model fitting approach provides a sound theory for fitting all CR models. Why not just use it? Answer: lack of software suitable for all potential users, and the relative slowness of MCMC analysis (there are reasons why speed still matters). Pragmatically I maintain that likelihood analysis, including profile likelihood intervals, as needed, is eminently adequate for CR data analysis when models have all fixed effects: the inferential results are virtually the same as those from MCMC with vague priors; but the analyses are much faster. This talk considers what we might be able to do to obtain CR random effects inferences, say incorporated into MARK, that are suitable approximations to MCMC inferences at a fraction of the computing time. The technical issue is simple. If all effects are considered fixed, numerical likelihood inference only requires function evaluations over a K dimensional space. Exact inference for a set of K random parameters requires integration of an even more complex function (than the likelihood) over a K dimensional space. Such integration is computationally more demanding than function maximization. MCMC analysis has complexity at the same level as the integration approach. The random effects method in MARK can be reformulated and extended. So doing has value, even though I accede to the MCMC approach as the gold standard for rigor and generality. The talk will note several possible approximate analyses approaches for some types of random effects (along with fixed effects, so technically mixed models). Under any inference computations we will augment the fixed effects likelihood with a model for the "randomly" varying parameters, a model which includes fixed hyperparameters. Focus is usually on survival probabilities. In some sense the resultant analysis is either a type of penalized likelihood for inference about random effects, or a smoothed likelihood for inference about fixed effects, including hyperparameters. Thus, the issue can be thought of more as a computing issue than as a modeling issue: the same model underlies all inference methods for random effects. 2:20-2:45 PM Estimation of nest success rates and comparing methods avaialbe in SAS GLM, SAS NLMIXED, and MARK - Jay Rotella, Steve Dinsmore & Terry Shaffer Estimating nesting success and evaluating factors potentially related to nest survival are key aspects of many studies of avian populations. A strong interest in nest survival has led to a rich literature detailing a variety of estimation methods for this vital rate. In recent years, modeling approaches have undergone especially rapid development. Despite these advances, most recent field studies still employ Mayfield's ad-hoc method (1961) or, in some cases, the maximum-likelihood estimator of Johnson (1979) and Bart and Robson (1982). Such methods allow for analyses of stratified data but do not allow for more complex and realistic models of survival data that include covariates that vary by individual, nest age, time, etc. and that may be continuous rather than categorical. Methods that allow researchers to rigorously assess the importance of a variety of biological factors that might affect nest survival can now be readily implemented in Program MARK and in SAS's Proc GENMOD and Proc NLMIXED. In this paper, we first describe the likelihood used for these models and then consider the question of what the effective sample size is for computation of AICc. Next, we consider the advantages and disadvantages of these different programs in terms of ease of data input and model construction; utility/flexibility of generated estimates and predictions; ease of model selection; and ability to estimate variance components. Finally, we discuss improvements that would, if they became available, promote a better general understanding of nest survival. 2:45-3:10 PM M-SURGE: an integrated software for multistate recapture models - Remi Choquet, Anne-Marie Reboulet, Roger Pradel, Olivier Gimenez & Jean-Dominique Lebreton M-SURGE (along with its companion program U_CARE) has been written specifically to handle multistate capture-recapture models (Lebreton and Pradel 2002) with the ultimate concern to alleviate their inherent difficulties (model specification, quality of convergence, flexibility of parameterization, assessment of fit…). In its domain, MSURGE covers a broader range of models than a general program like MARK (White and Burnham 1999), while being more user-friendly than a research program like MS-SURVIV (Hines 1994). Among the main features of MSURGE is a wide class of models and a variety of parameterizations: M-SURGE integrates conditional models with probability of recapture depending on the current state (Arnason-Schwarz type models) but also on the current and previous state (Jolly-MoVement type models; (Brownie, Hines et al. 1993)). In both cases, age and/or time-dependence and multiple groups can be considered. Combined Survival-Transition probabilities can be represented as such or decomposed into transition and survival probabilities (Hestbeck, Nichols et al. 1991). Among the transition probabilities with the same state of departure, the one to be computed as 1 minus the others can be freely picked by the user. User-friendliness is enhanced by the easiness with which constrained models are built, using a language, interpreted by a generator of design matrices called GEMACO. This language is alike those in general statistical software such as SAS or GLIM, i.e., a formula such as t+g generates a model with additive effects of time and group, thus avoiding tedious and error-prone matrix manipulations using an editor or a spread-sheet. Examples of various types of multistate models are developed and presented. You can download M-Surge freely from ftp.cefe.cnrs-mop.fr/biom/Soft-CR. 3:10-3:45 PM DENSITY: software for fitting spatial detection functions to data from passive sampling - Murray Efford & Deanna Dawson Rigorous sampling of bird populations to estimate density raises the problem of incomplete detection (e.g., Pollock et al. 2002). Varying detectability is widely acknowledged, but variation in its spatial component (how detection declines with distance) is considered less often. Active methods (double sampling and distance sampling) require an observer to determine the instantaneous distance between the sampling point and each animal. Passive methods (mistnets or traps) rely on animals moving to the detector: instantaneous locations are unknown, and movement is an important unmeasured component of detectability. New methods have been developed to fit spatial detection functions to capture-recapture data from passive detectors. The methods are computer-intensive and depend on specialised software ('DENSITY'), available for download at www.landcareresearch.co.nz. DENSITY provides a graphical interface for the analysis of closed-population capture-recapture data from arrays of passive detectors. Its simulation capability enables users to perform power analysis of different sampling designs before going into the field. We demonstrate the use of DENSITY to estimate bird population density from mist netting data. Netting was conducted over 1992 to 1996 on a forest-pasture ecotone in Mexico. In each of 16 netting sessions, six local arrays of 20 nets were run for 2-3 consecutive days. Despite the large number of captures in total, within-session recaptures were rare for most species. This restricted application of the method to a few common species (e.g. Sporophila torqueola) and to species aggregates (e.g. 'winter residents'). Although the use of the method for mist netting data was experimental, it may lead through simulation in DENSITY to improved design of mist net arrays where density estimation is a study goal. The spatially explicit framework of DENSITY opens up new possibilities and we expect the software to evolve. An exciting prospect is the direct fitting of simple density surfaces during estimation, given suitable spatial covariates.