MARSS

Download the current release of MARSS 1.1 from CRAN.  This project page is where we are posting code for MARSS 1.2 (bug fixes) before uploading to CRAN. 

MARSS is an R package which fits mulitvariate state-space models to multivariate time-series data. 

x(t) = B x(t-1) + u + v(t), v(t)~MVN(0,Q)

y(t) = Z x(t) + a + w(t), w(t)~MVN(0,R)

where all elements of these equations are matrices as this is a multivariate auto-regressive model.

What you need to use this code: MARSS is an R package, thus you need to install R from CRAN in order to use the package  Once you have R installed, then install the MARSS package using the standard R package instructions (if you are using an R GUI, then you use the "Install Packages" menu.)  If you have never done this, see the instructions on CRAN.

About this code: This code is used to estimate maximum likelihood parameters for multivariate state space time series models via an EM algorithm using the Kalman filter+smoother. We have a number of online workshops on multivariate state-space models available with case studies and examples for estimating trends, evaluating population structure, estimating interactions, and analyzing movement data: WORKSHOPS  Our EM algorithm is similar to that in Shumway and Stoffer (1982) but actually was inspired by Ghahramani and Hinton (1996). Most other software uses the BFGS algorithm (a quasi-Newton method) for maximization, which often works great but for some models needs a bit of fiddling to get it to work (not throw numerical errors).  Some researchers will use the EM algorithm to "get close" and then polish off with the BFGS algorithm.  MARSS includes functions for bootstrapping (parametric and innovations), model selection (AIC, AICc, and bootstrap bias-corrected AICb), confidence intervals (approximate via Hessian, parametric bootstrap, and innovations bootstrap), parameter bias estimation (via bootstrapping), simulation, and initial condition searching.  MARSS was developed by Eli Holmes, Eric Ward, and Kellie Wills.

Learning how to use the code: The user manual is given below. Scripts and data for all the case studies are included in the package (see the manual for instructions). The online workshop includes pdfs of our lectures.

Project news:  March 10, 2011, A version of MARSS 2.0 is ready for public testing.  Download MARSS 2.0.  Mar 1, 2011, Testing of MARSS 2.0 is going well.  I'm finishing up a dynamic factor analysis example to show Z estimation.   Feb 16, 2011, Uploaded new Derivations.pdf that deals with 0s on R and Q matrices and uses a more elegant approach to missing values.  Oct 2010, Coding of MARSS 2.0 is now 95% done (all algorithms coded and testing in progress). 

See our personal websites for group news on papers coming out of this work: EE Holmes website, Eric Ward website, Brice Semmens website, and Mark Scheuerell website.  (download stats from CRAN)

Bugs and issues fixed in 1.1:  See Changes.pdf in MARSS/inst/doc folder.  Issues being fixed in 2.0:  adding algorithm for fully constrained estimation for all parameters (per algorithm in Derivations.pdf).