Harvard Forest Research

Testing the Effects of Priors on the Prediction Errors of Bayesian Demographic Models

Principal Investigator: Aaron Ellison
Harvard Forest: Jun 01 2009 - May 01 2009:

Abstract:
Ecologists and conservation planners have a long-standing interest in modeling population dynamics and forecasting population sizes. When applied to the management of rare species, it is critical that the estimates from population viability analyses have high accuracy and low uncertainty. Long-term data sets help to reduce uncertainty in estimates of demographic models, but are often difficult to collect because of funding and logistical constraints. One possible solution to reducing uncertainty in these models is to include information from similar studies as prior data in a Bayesian demographic model. Prior data may be as difficult to come by as long-term data for rare species, but there may be available data on closely related species with similar natural or evolutionary histories. While this method is appealing, given the challenges of collecting long-term data, its validity relies upon choosing appropriate prior data. There are few – if any – studies that empirically test the effects of different priors on the prediction error of Bayesian demographic models. The overall objective of the proposed research is to evaluate the ability of prior probability distributions to provide more accurate and less uncertain population viability analyses of rare species.
This goal will be accomplished using a three-step approach. (1) Construction of Bayesian demographic models for two rare plant species, Furbish’s Lousewort (Pedicularis furbishiae) and Tiburon Mariposa Lily (Calochortus tiburonensis), that were the focus of demographic studies in the 1980s. For P. furbishiae, prior data will be derived from a demographic study on P. lanceolata collected as part of the S. Record’s dissertation research. For C. tiburonensis, prior data will be derived from studies of C. albus, C. obispoensis and C. pulchellus. (2) The originally studied populations of C. tiburonensis and P. furbishiae will be re-censused using adaptive cluster sampling methods to account for observation error in our ability to detect individuals. (3) Estimates of the quasi-extinction risks and 2009 population sizes of the different models will be compared with observed information on the populations from contemporary censuses using goodness-of-fit tests.
For ecologists and conservation planners to confidently use Bayesian inference to model population dynamics and forecast extinction risk and population sizes, knowledge about the effects of priors on the reliability of model projections is sorely needed. For population biologists the ultimate test of a Bayesian model is its ability to predict real future data. The proposed research provides such a test.
: