Effects of model complexity and priors on prediction using sequential importance sampling/resampling for conservation of species
State-space models can be used to estimate complex, unobservable population processes and make predictions about future populations based on limited data. These models account for process and measurement error through the system model and observation model which links observation level data to the state process. To better understand the utility of state space models, we used them in a Bayesian framework and compared the accuracy of models of differing complexity with and without informative priors for estimating the state and parameter values. Count data was simulated for 25 years using known parameters and the observation process for both models. The first model was a linear birth and death model, and the more complex model was a two-sex, two-stage structured model with adults of both sexes able to be observed. We initialized the model using the informative and non-informative priors, and estimated state variables and parameters using sequential importance sampling-resampling (SISR) conditional on the observed data. We also used kernel smoothing to reduce the effect of particle depletion that is common when estimating both states and parameters with SISR.
The posterior parameter and state estimates of both models were more accurate and precise when informative priors were used in the models. Posterior distributions of parameters and states for all models included truth; however, uncertainty was greater in the estimates of parameters and states for the models with non-informative priors. While these models all produced posterior estimates that included truth, the posterior estimates from the structured model with informative priors performed best. For many species, demographic rates are known with some degree of certainty according to general life history strategies or could be elicited from similar species or experts. When this information is available it should be used to obtain more precise estimates of population state and demographic rates; however, this model framework still provides reasonable estimates when little to no information is available. Further, understanding population structure can be critical to the conservation and management of species, but incorporating model complexity can affect precision and accuracy when estimating latent population variables and predicting population dynamics. These results are important to consider when using this approach to model populations when census data is available and generally little is known of population dynamics and demographic rates.