Abstract:
|
Hierarchical models are widely used in the ecology. They have grown in popularity in connection with the advancement of MCMC and other developments in statistical computation. Unfortunately, our ability to specify and fit hierarchical models has not been matched by the capability to check the assumptions of the models, particularly as the hierarchical structure deepens. This is important because not all hierarchical models are created equal. On one hand, hierarchical modeling can allow for partial pooling and regularization. On the other, we can consider models with many more latent variables than we have data points and there is often a temptation to model ourselves out of any situation. In this talk, we will use several examples to consider which aspects of a hierarchical model are robust and well supported by the data, as well as which variables are likely to be dependent on assumptions and sensitive to model specification. This has implications for how we interpret model output and distinguish between different hierarchical specifications. We will discuss the interaction between model fitting, statistical computation and model specification.
|