PARENT SESSION
Symposium #9: Uncertainty and information in ecological forecasting.

Organized by: JS Clark and C Brewer
Tuesday, August 6. 8:00 AM to 11:30 AM. Crystal Ballroom, TCC.

Uncertainty in population forecasting: the hierarchical approach.

Clark, James^*,1, ¹ Duke University, Durham

ABSTRACT- Estimates of uncertainty are the basis for inference concerning populations at risk. Uncertainty is estimated from models fitted to data that typically include a deterministic model (e.g., population growth) and stochastic elements, which should accommodate errors in sampling and any sources of variability that affect observations. Error propagation from fitted models (of, say, demography) to new variables (say, population growth) require propagation of these stochastic elements. In fact, ecological models ignore most forms of variability, because they make statistical models complex, and they pose considerable computational challenges. Variability caused by space, time, and among individuals is not accommodated by current demographic models, making growth rate predictions unrealistic. I demonstrate that Bayesian hierarchical models can be readily adapted to the common problem of estimating population growth rates and their uncertainties when individuals vary and the causes for that variability cannot be assigned to specific variables. In contrast to an overfitted model that would assign a different parameter value to each individual, hierarchical models accommodate individual differences, but assume that those differences derive from an underlying distribution; they belong to a "population". The hierarchical model can be implemented in classical (frequentist) and Bayesian frameworks and computed using Markov Chain Monte Carlo (MCMC) simulation. The MCMC results provide a straightforward basis for uncertainty estimates on population growth rates. My demonstration of the method uses as examples individual growth rates based on resource availability, populations growth rates estimated from demographic data, and rates of population spread based on life history and dispersal data. It shows that the common population growth models that rely on standard propagation of estimation error, but ignore process variability among individuals, provide uncertainties that will mislead managers and decision makers and (eventually) erode credibility of ecological analyses and forecasting.

KEY WORDS: Bayesian, demograpy, population growth