Chapter 11 EM algorithm

The MARSS package fits models via maximum likelihood. The MARSS package is unusual among packages for fitting MARSS models in that fitting is performed via a constrained EM algorithm based on a vectorized form of Equation (See Chapter for the vectorized form used in the algorithm). Although fitting via the BFGS algorithm is also provided using method="BFGS" and the optim function in ,, the examples in this guide use the EM algorithm primarily because it gives robust estimation for datasets replete with missing values and for high-dimensional models with various constraints. However, there are many models/datasets where BFGS is faster and we typically try both for problems. The EM algorithm is also often used to provide initial conditions for the BFGS algorithm (or an MCMC routine) in order to improve the performance of those algorithms. In addition to the main model fitting function, the MARSS package supplies functions for bootstrap and approximate confidence intervals, parametric and non-parametric bootstrapping, model selection (AIC and bootstrap AIC), simulation, and bootstrap bias correction.