Browsing by Subject "Bayesian Models"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Evaluation of Dead-Recovery Models for Estimating the Survival of Canada Geese
    (2024-07) Struble, Megan
    Population modeling of birds with deferred maturity, such as Canada geese (Branta canadensis), requires consideration of multiple age classes due to age-related variation in survival and encounter probabilities. The lack of recapture and release data from subadult age classes complicates the estimation of these parameters, particularly in dead-recovery models where individuals are only reencountered once. Individuals are categorized into two age classes when initially marked: “hatch-year” (HY) and “after-hatch-year” (AHY), wherein AHY individuals may include subadults and adults, leading to heterogeneous survival and encounter probabilities within cohorts if subadult survival or encounter probabilities differ from adults. Previous studies have highlighted the importance of estimating age-specific parameters accurately and the limitations of traditional modeling approaches. Dooley et al. (2019) modeled survival and encounter probabilities for three age classes– HY, “second-year” (SY), and “after-second-year” (ASY)– using data from only two marked age classes; however, unrealistic model assumptions led to significant bias, particularly regarding subadult survival rates and encounter probabilities. We ran simulations for each of three scenarios: (1) where the true value for SY survival is greater than HY survival, but SY recovery is less than HY recovery; (2) where the true values for SY survival and recovery are less than HY survival and recovery; and (3) where the true value for SY survival is less than HY survival, but SY recovery is greater than HY recovery. For each of these scenarios, we ran 100 20-year simulations of twelve random-effects models, including two-age-class models that presumed SY and ASY individuals had similar survival and encounter probabilities; three-age-class models that assumed no correlation among age classes on the logit scale; three three-age-class offset models that assumed perfect correlation among age classes on the logit scale (as in Dooley et al. 2019); and three-age-class models that allowed correlations among parameters to vary from -1 to +1. We compared these models against a “truth” model in which all three age classes are banded and correctly identified at time of marking. For all three scenarios, the three-age-class models yielded negative bias for HY survival, while SY survival and recovery suffered from positive bias, with the direction of bias determined by the true differences in parameter values among age classes. In particular, the model proposed by Dooley et al. displayed the least accurate parameter estimates, while the three-age-class model assuming no correlation between age classes exhibited the least precision. Our results demonstrate that there is no way to reliably estimate these parameters using a dead-recovery model. For future models, we recommend a joint-encounter approach utilizing both dead encounters and live recapture data.