Multivariate Autoregressive State-space Models for Analyzing Time-series Data by. Holmes, E. E., Ward, E. J., & Wills, K. C. In 2012.
abstract   bibtex   
MARSS is a package for fitting multivariate autoregressive state-space models to time-series data. The MARSS package implements state-space models in a maximum likelihood framework. The core functionality of MARSS is based on likelihood maximization using the Kalman filter/smoother, combined with an EM algorithm. To make comparisons with other packages available, parameter estimation is also permitted via direct search routines available in ’optim’. The MARSS package allows data to contain missing values and allows a wide variety of model structures and constraints to be specified (such as fixed or shared parameters). In addition to model-fitting, the package provides bootstrap routines for simulating data and generating confidence intervals, and multiple options for calculating model selection criteria (such as AIC). The MARSS package (Holmes et al., 2012) is an R package for fitting linear multivariate autoregressive state-space (MARSS) models with Gaussian errors to time-series data. This class of model is extremely important in the study of linear stochastic dynamical systems, and these models are used in many different fields, including economics, engineering, genetics, physics and ecology. The model class has different names in different fields; some common names are dynamic linear models (DLMs) and vector autoregressive (VAR) state-space models. There are a number of existing R packages for fitting this class of models, including sspir (Dethlefsen et al., 2009) for univariate data and dlm (Petris, 2010), dse (Gilbert, 2009), KFAS (Helske, 2011) and FKF (Luethi et al., 2012) for multivariate data. Additional packages are available on other platforms, such as SsfPack (Durbin and Koopman, 2001), EViews (www.eviews.com) and Brodgar (www.brodgar.com). Except for Brodgar and sspir, these packages provide maximization of the likelihood surface (for maximum-likelihood parameter estimation) via quasi-Newton or Nelder-Mead type algorithms. The MARSS package was developed to provide an alternative maximization algorithm, based instead on an Expectation-Maximization (EM) algorithm and to provide a standardized modelspecification framework for fitting different model structures. The MARSS package was originally developed for researchers analyzing data in the natural and environmental sciences, because many of the problems often encountered in these fields are not commonly encountered in disciplines like engineering or finance. Two typical problems are high fractions of irregularly spaced missing observations and observation error variance that cannot be estimated or known a priori (Schnute, 1994). Packages developed for other fields did not always allow estimation of the parameters of interest to ecologists because these parameters are always fixed in the package authors’ field or application. The MARSS package was developed to address these issues and its three main differences are summarized as follows. First, maximum-likelihood optimization in most packages for fitting state-space models relies on quasi-Newton or Nelder-Mead direct search routines, such as provided in optim (for dlm) or nlm (for dse). Multidimensional state-space problems often have complex, non-linear likelihood surfaces. For certain types of multivariate state-space models, an alternative maximization algorithm exists; though generally slower for most models, the EMalgorithm (Metaxoglou and Smith, 2007; Shumway and Stoffer, 1982), is considerably more robust than direct search routines (this is particularly evident with large amounts of missing observations). To date, no R package for the analysis of multivariate state-space models has implemented the EM algorithm for maximum-likelihood parameter estimation (sspir implements it for univariate models). In addition, the MARSS package implements an EM algorithm for constrained parameter estimation (Holmes, 2010) to allow fixed and shared values within parameter matrices. To our knowledge, this constrained EM algorithm is not implemented in any package, although the Brodgar package implements a limited version for dynamic factor analysis. Second, model specification in the MARSS package has a one-to-one relationship to a MARSS model as written in matrix form on paper. Any model that can be written in MARSS form can be fitted without extra code by the user. In contrast, other packages require users to write unique functions in matrix form (a non-trivial task for many non-expert R users). For example, while dlm includes linear and polynomial univariate models, multivariate regression is not readily accessible without these custom functions; in MARSS, all models written in matrix form are fitted using the same model specification. The R Journal Vol. 4/1, June 2012 ISSN 2073-4859 12 CONTRIBUTED RESEARCH ARTICLES The MARSS package also allows degenerate multivariate models to be fitted, which means that some or all observation or process variances can be set to 0. This allows users to include deterministic features in their models and to rewrite models with longer lags or moving averaged errors as a MARSS model. Third, the MARSS package provides algorithms for computing model selection criteria that are specific for state-space models. Model selection criteria are used to quantify the data support for different model and parameter structures by balancing the ability of the model to fit the data against the flexibility of the model. The criteria computed in the MARSS package are based on Akaike’s Information Criterion (AIC). Models with the lowest AIC are interpreted as receiving more data support. While AIC and its small sample corrected version AICc are easily calculated for fixed effects models (these are standard output for lm and glm, for instance), these criteria are biased for hierarchical or state-space autoregressive models. The MARSS package provides an unbiased AIC criterion via innovations bootstrapping (Cavanaugh and Shumway, 1997; Stoffer and Wall, 1991) and parametric bootstrapping (Holmes, 2010). The package also provides functions for approximate and bootstrap confidence intervals, and bias correction for estimated parameters. The package comes with an extensive user guide that introduces users to the package and walks the user through a number of case studies involving ecological data. The selection of case studies includes estimating trends with univariate and multivariate data, making inferences concerning spatial structure with multi-location data, estimating inter-species interaction strengths using multi-species data, using dynamic factor analysis to reduce the dimension of a multivariate dataset, and detecting structural breaks in data sets. Though these examples have an ecological focus, the analysis of multivariate time series models is cross-disciplinary work and researchers in other fields will likely benefit from these examples.
@inproceedings{holmes_multivariate_2012,
	title = {Multivariate {Autoregressive} {State}-space {Models} for {Analyzing} {Time}-series {Data} by},
	abstract = {MARSS is a package for fitting multivariate autoregressive state-space models to time-series data. The MARSS package implements state-space models in a maximum likelihood framework. The core functionality of MARSS is based on likelihood maximization using the Kalman filter/smoother, combined with an EM algorithm. To make comparisons with other packages available, parameter estimation is also permitted via direct search routines available in ’optim’. The MARSS package allows data to contain missing values and allows a wide variety of model structures and constraints to be specified (such as fixed or shared parameters). In addition to model-fitting, the package provides bootstrap routines for simulating data and generating confidence intervals, and multiple options for calculating model selection criteria (such as AIC). The MARSS package (Holmes et al., 2012) is an R package for fitting linear multivariate autoregressive state-space (MARSS) models with Gaussian errors to time-series data. This class of model is extremely important in the study of linear stochastic dynamical systems, and these models are used in many different fields, including economics, engineering, genetics, physics and ecology. The model class has different names in different fields; some common names are dynamic linear models (DLMs) and vector autoregressive (VAR) state-space models. There are a number of existing R packages for fitting this class of models, including sspir (Dethlefsen et al., 2009) for univariate data and dlm (Petris, 2010), dse (Gilbert, 2009), KFAS (Helske, 2011) and FKF (Luethi et al., 2012) for multivariate data. Additional packages are available on other platforms, such as SsfPack (Durbin and Koopman, 2001), EViews (www.eviews.com) and Brodgar (www.brodgar.com). Except for Brodgar and sspir, these packages provide maximization of the likelihood surface (for maximum-likelihood parameter estimation) via quasi-Newton or Nelder-Mead type algorithms. The MARSS package was developed to provide an alternative maximization algorithm, based instead on an Expectation-Maximization (EM) algorithm and to provide a standardized modelspecification framework for fitting different model structures. The MARSS package was originally developed for researchers analyzing data in the natural and environmental sciences, because many of the problems often encountered in these fields are not commonly encountered in disciplines like engineering or finance. Two typical problems are high fractions of irregularly spaced missing observations and observation error variance that cannot be estimated or known a priori (Schnute, 1994). Packages developed for other fields did not always allow estimation of the parameters of interest to ecologists because these parameters are always fixed in the package authors’ field or application. The MARSS package was developed to address these issues and its three main differences are summarized as follows. First, maximum-likelihood optimization in most packages for fitting state-space models relies on quasi-Newton or Nelder-Mead direct search routines, such as provided in optim (for dlm) or nlm (for dse). Multidimensional state-space problems often have complex, non-linear likelihood surfaces. For certain types of multivariate state-space models, an alternative maximization algorithm exists; though generally slower for most models, the EMalgorithm (Metaxoglou and Smith, 2007; Shumway and Stoffer, 1982), is considerably more robust than direct search routines (this is particularly evident with large amounts of missing observations). To date, no R package for the analysis of multivariate state-space models has implemented the EM algorithm for maximum-likelihood parameter estimation (sspir implements it for univariate models). In addition, the MARSS package implements an EM algorithm for constrained parameter estimation (Holmes, 2010) to allow fixed and shared values within parameter matrices. To our knowledge, this constrained EM algorithm is not implemented in any package, although the Brodgar package implements a limited version for dynamic factor analysis. Second, model specification in the MARSS package has a one-to-one relationship to a MARSS model as written in matrix form on paper. Any model that can be written in MARSS form can be fitted without extra code by the user. In contrast, other packages require users to write unique functions in matrix form (a non-trivial task for many non-expert R users). For example, while dlm includes linear and polynomial univariate models, multivariate regression is not readily accessible without these custom functions; in MARSS, all models written in matrix form are fitted using the same model specification. The R Journal Vol. 4/1, June 2012 ISSN 2073-4859 12 CONTRIBUTED RESEARCH ARTICLES The MARSS package also allows degenerate multivariate models to be fitted, which means that some or all observation or process variances can be set to 0. This allows users to include deterministic features in their models and to rewrite models with longer lags or moving averaged errors as a MARSS model. Third, the MARSS package provides algorithms for computing model selection criteria that are specific for state-space models. Model selection criteria are used to quantify the data support for different model and parameter structures by balancing the ability of the model to fit the data against the flexibility of the model. The criteria computed in the MARSS package are based on Akaike’s Information Criterion (AIC). Models with the lowest AIC are interpreted as receiving more data support. While AIC and its small sample corrected version AICc are easily calculated for fixed effects models (these are standard output for lm and glm, for instance), these criteria are biased for hierarchical or state-space autoregressive models. The MARSS package provides an unbiased AIC criterion via innovations bootstrapping (Cavanaugh and Shumway, 1997; Stoffer and Wall, 1991) and parametric bootstrapping (Holmes, 2010). The package also provides functions for approximate and bootstrap confidence intervals, and bias correction for estimated parameters. The package comes with an extensive user guide that introduces users to the package and walks the user through a number of case studies involving ecological data. The selection of case studies includes estimating trends with univariate and multivariate data, making inferences concerning spatial structure with multi-location data, estimating inter-species interaction strengths using multi-species data, using dynamic factor analysis to reduce the dimension of a multivariate dataset, and detecting structural breaks in data sets. Though these examples have an ecological focus, the analysis of multivariate time series models is cross-disciplinary work and researchers in other fields will likely benefit from these examples.},
	author = {Holmes, Elizabeth E. and Ward, Eric John and Wills, Kellie C.},
	year = {2012}
}

Downloads: 0