Wednesday, March 9, 2011

MARSS 2.1 ready for testing

Download from the MARSS Dev Site
The big update of MARSS to allow B and Z estimation is finally ready for testing. It'll be awhile before its ready for CRAN since I need to work on a couple new case studies and fix bugs and stuff that comes up in testing. I added a mock up of a new case study on dynamic factor analysis (sensu Alain Zuur) since that is of special interest to fisheries folks.  See the editlog on the dev site to see the recent changes

MARSS is an R package which fits multivariate state-space models to time series data. It is different than other state-space packages in that it does maximization via a constrained Expectation-Maximization (EM) algorithm (an extension of the EM algorithm developed by Shumway and Stoffer 2000). The EM algorithm is slower (much) than the BFGS algorithm used in other packages, but can be more stable (less likely to get hung up on numerical problems) for some types of state-space models also the EM algorithm used to produce good initial conditions for Bayesian or Nelder-Mead algorithms.

New in version 2.1
* B and Z estimation
* Dynamic factor analysis (sensu Harvey 1989 and Zuur et al 2003)
* Allows degenerate Q and R (variances=0)
* Fewer constraints on what type of Q and R matrices you can fit

MARSS 2.1 still does not allow covariates. Including that is next on my to-do list, but it will be awhile. MARSS 2.1 also does not allow temporally varying parameters. Again, it's easy enough to include but is not high on my to-do list.