A MATLAB toolbox for Granger causal connectivity analysis
Introduction
The accelerating availability of neuroscience data is placing increasing demands on analysis methods. These data are generated at multiple levels of description, from spike trains to local field potentials to functional MRI (fMRI) BOLD signals. A key challenge when analyzing such data is to determine the functional connectivity of the underlying mechanisms. Many methods for functional connectivity analysis identify undirected connectivity; examples include synchrony (Engel and Singer, 2001) and phase coherence (Nunez et al., 2001, Doesburg et al., 2009). However, a satisfactory understanding of neural mechanisms is likely to require identification of directed functional connectivity. A powerful technique for extracting such connectivity from data is Granger causality (G-causality) (Granger, 1969, Ding et al., 2006, Seth, 2007). According to G-causality, a variable ‘Granger causes’ a variable if information in the past of helps predict the future of with better accuracy than is possible when considering only information in the past of itself (Granger, 1969, Seth, 2007). This paper describes a freely available software toolbox, programmed in the MATLAB (Natick, MA) environment, which allows application of a range of G-causality analyses to neuroscience data broadly construed. Taken together, the analysis methods incorporated in the toolbox form ‘Granger causal connectivity analysis’ (GCCA). A first version of the toolbox was released in 2005 (Seth, 2005); the present paper describes the first significant revision and extension of the software, as well as the theoretical infrastructure underpinning its functionality.
It should be noted that identifying directed functional connectivity is not equivalent to identifying physically instantiated causal interactions in systems. Although the two descriptions are intimately related (Seth and Edelman, 2007, Cadotte et al., 2008), physically instantiated causal structure can only be unambiguously identified by perturbing a system and observing the consequences (Pearl, 1999). Directed functional connectivity in general, and G-causality analysis in particular, is therefore best understood as a statistical relationship among observed variables that reflects but may not be identical to the underlying physical mechanism.
The importance of identifying causal structure within data, especially during exploratory analysis phases, indicates a need for easy-to-apply, transparent, and extensible software methods. Such methods are provided by the GCCA toolbox described here. The toolbox includes several different types of function. The core functions implement G-causality analysis given multivariate time series data. Other functions test whether the provided data satisfies necessary assumptions, assess the statistical significance and validity of inferred interactions, generate network-level descriptions of patterns of causal interactions, and graphically display analysis results. Functions are also included to apply various preprocessing techniques and to demonstrate the toolbox capabilities. The toolbox is intentionally small when compared to several other brain signal analysis toolboxes (e.g., Delorme and Makeig, 2004, Cui et al., 2008), in order to facilitate assimilation of the methods into the repertoire of researchers. Also, the toolbox is not targeted specifically to any particular experimental technique. G-causality analysis has been successfully applied to a wide range of data types including spike trains (Cadotte et al., 2008, Nedungadi et al., 2009), local field potentials (Kaminski et al., 2001, Brovelli et al., 2004), EEG signals (Babiloni et al., 2005, Keil et al., 2009), fMRI BOLD signals (Roebroeck et al., 2005, Sato et al., 2006), among others. Several such studies have utilized an earlier version of the present toolbox (Sridharan et al., 2008, Shanahan, 2008, Gow et al., 2008, Gow and Segawa, 2009, Gaillard et al., 2009, Chang et al., 2008, Stevens et al., 2009) or have followed the procedures it advocated (Keil et al., 2009). By November 2009, over 600 researchers from many different countries had downloaded the toolbox. The present toolbox extends the earlier version in several important ways, including (i) frequency decomposition of causal interactions, (ii) partial Granger causality, (iii) ‘Granger autonomy’, (iv) analysis of event-related data, (v) additional preprocessing and validation techniques, for example for removal of single frequency line-noise from data and for checking model consistency, and (vi) built-in functions for bootstrap and permutation statistical testing. The toolbox is also enhanced with respect to the efficiency of its core algorithms.
A screenshot of the toolbox demonstration function is provided in Fig. 1. In this figure, data are generated according to the model described in Baccalá and Sameshima (2001) (Eq. (1.1), see top-left panel).
The model generates five time-series which are shown in the bottom-right panel of the figure. are independent normally-distributed processes with zero mean and unit standard deviation. The remaining panels show the results of GCCA of the time-series. The top panels show G-causalities in both matrix and network form. In this case the inferred causalities are identical to the underlying physical network structure (as noted this need not be true in general). The bottom panels show network-level summary descriptions of the causal patterns, described further in Section 4. Briefly, causal flow reflects the extent to which a variable is influenced by or influences the remainder of the system; causal density expresses the overall degree of causal interactivity, either averaged across an entire network or assessed on a per-variable basis.
The remainder of this paper is organized as follows. Section 2 describes the underlying theory of G-causality and of the various extensions represented in the toolbox. Section 3 discusses the statistical assumptions involved in G-causality analysis and describes various preprocessing and validation techniques. Section 4 introduces some network-level descriptors of G-causality interaction patterns which can be useful for inferring macroscopic dynamical properties of neural systems. Section 5 discusses the important issue of filtering. Section 6 then outlines modality-specific issues involved in application of G-causality to data acquired from functional MRI (fMRI), neuroelectric and neuromagnetic signals, and spike train data. Limitations of the present instantiation of the toolbox are noted in Section 7, and Section 8 sets G-causality in the context of other frameworks for identifying directed functional connectivity. For detailed descriptions of the individual functions in the toolbox the reader is referred to the accompanying manual, available online at www.anilseth.com.
Section snippets
Bivariate and conditional G-causality
In 1969 Granger introduced the idea of G-causality as a formalization, in terms of linear regression modelling, of Wiener’s (and Akaike’s) intuition that ‘causes’ if knowing helps predict the future of (Granger, 1969, Seth, 2007). According to G-causality, causes if the inclusion of past observations of reduces the prediction error of in a linear regression model of and , as compared to a model which includes only previous observations of . To illustrate
Preprocessing and validation
Meaningful application of the GCCA toolbox requires, minimally, that (i) the data satisfy certain preconditions and (ii) the MVAR models describe the data in a statistically satisfactory manner.
Causal networks and their visualization
Methods for analyzing causal connectivity are especially powerful when applied in combination with graph-theoretic and network-theoretic techniques which allow their quantitative characterization (Seth, 2005). In causal networks, nodes represent variables or system elements and directed edges represent causal interactions. There is wide range of graph theory concepts that may find useful application to causal networks (Eichler, 2005, Bullmore and Sporns, 2009). Here, we describe those that are
Filtering
It is critical to ensure that preprocessing steps do not introduce spurious correlation structure into the data which can result in artifactual causal connectivity. In general, procedures that preserve the fine-grained timing relations among variable are safe, whereas those that do not, are not. One common preprocessing step that can cause problems is that of bandpass and/or notch filtering. According to Geweke (1982), the estimation of MVAR models in asymptotic conditions should not be
Functional MRI
Application of GCCA to fMRI data is particularly attractive given the high spatial specificity of the fMRI BOLD signal. However, G-causality as applied to fMRI faces distinctive challenges and constraints. These challenges arise from the facts that the fMRI BOLD signal has relatively poor temporal resolution (on the order of seconds) as compared to other noninvasive neuroimaging techniques, and that it is an indirect measure of neural dynamics usually modelled by a convolution of underlying
Limitations
The current GCCA toolbox does not support several potentially useful analyses. One limitation is that multivariate spectral G-causality analysis cannot be performed. The best method for performing this analysis remains an active area of investigation. The standard approach is to analyze spectral features of autoregressive models following (Geweke, 1982), as implemented here for pairwise analysis by the function cca_pwcausal. However, in the multivariate case Geweke spectral causality can
Discussion
The GCCA toolbox provides a range of MATLAB functions enabling the application of G-causality analysis to a broad range of neuroscience data. Together with the theoretical and practical context provided in this paper, it is well placed to facilitate progress in a wide cross-section of the neuroscience community. However, it is critical to ensure that the toolbox is not treated as ‘a black box’ whereby its output is assumed to be a valid and informative transformation of its input. As with any
Acknowledgements
I am grateful to A. Barrett, L. Barnett, and R. Giacomini for guidance and comments, and to many others for useful feedback on the first version of the GCCA toolbox (see the manual for individual acknowledgements). Many thanks to H. Liang and colleagues for providing links to the functions pwcausal.m and armorf.m which are part of the BSMART toolbox (Cui et al., 2008). The author’s research is funded by EPSRC Leadership Fellowship EP/G007543/1 and by a donation from the Mortimer and Theresa
References (76)
- et al.
The variability of human, BOLD hemodynamic responses
Neuroimage
(1998) - et al.
Estimation of the cortical functional connectivity with the multimodal integration of high-resolution EEG and fMRI data by directed transfer function
Neuroimage
(2005) - et al.
Autonomy: an information theoretic perspective
Biosystems
(2008) - et al.
Frequency decomposition of conditional Granger causality and application to multivariate neural field potential data
J Neurosci Methods
(2006) - et al.
BSMART: a Matlab/C toolbox for analysis of multichannel neural time series
Neural Netw
(2008) - et al.
EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis
J Neurosci Methods
(2004) - et al.
Analyzing information flow in brain networks with nonparametric Granger causality
Neuroimage
(2008) - et al.
Temporal binding and the neural correlates of sensory awareness
Trends Cogn Sci
(2001) - et al.
Testing non-linearity and directedness of interactions between neural groups in the macaque inferotemporal cortex
J Neurosci Methods
(1999) - et al.
Dynamic causal modelling
Neuroimage
(2003)
DEM: a variational treatment of dynamic systems
Neuroimage
Articulatory mediation of speech perception: a causal analysis of multi-modal imaging data.
Cognition
Lexical influences on speech perception: a Granger causality analysis of MEG and EEG source estimates
Neuroimage
Economic processes involving feedback
Inform Control
Spurious regressions in econometrics
J Econom
Partial granger causality: eliminating exogenous inputs and latent variables
J Neurosci Methods
The use of time-variant EEG Granger causality for inspecting directed interdependencies of neural assemblies
J Neurosci Methods
Comparing dynamic causal models
Neuroimage
Nonlinear multivariate analysis of neurophysiological signals
Prog Neurobiol
Mapping directed influence over the brain using granger causality and fmri
Neuroimage
A method to produce evolving functional connectivity maps during the course of an fMRI experiment using wavelet-based time-varying Granger causality
Neuroimage
Inferring neural activity from BOLD signals through nonlinear optimization
Neuroimage
A new look at the statistical model identification
IEEE Trans Autom Control
Radial basis function approaches to nonlinear granger causality of time series
Phys Rev E
Partial directed coherence: a new concept in neural structure determination
Biol Cybern
Granger causality and transfer entropy are equivalent for Gaussian variables
Phys Rev Lett
Controlling the false discovery rate: a practical and powerful approach to multiple testing
J R Stat Soc: Ser B (Methodol)
An alternative method for significance testing of waveform difference potentials
Psychophysiology
Empirical model building and response surfaces
Beta oscillations in a large-scale sensorimotor cortical network: directional influences revealed by Granger causality
Proc Natl Acad Sci USA
Complex brain networks: graph theoretical analysis of structural and functional systems
Nat Rev Neurosci
Causal measures of structure and plasticity in simulated and living neural networks
PLoS One
Mapping and correction of vascular hemodynamic latency in the BOLD signal
Neuroimage
Unconscious errors enhance prefrontal-occipital oscillatory synchrony
Frontiers in Human Neuroscience
Identifying neural drivers with functional MRI: an electrophysiological validation
PLoS Biol
Multivariate Granger causality analysis of fMRI data
Hum Brain Mapp
Short-window spectral analysis of cortical event-related potentials by adaptive multivariate autoregressive modeling: data preprocessing, model validation, and variability assessment
Biol Cybern
Granger causality: basic theory and application to neuroscience
Cited by (633)
A survey on intelligent management of alerts and incidents in IT services
2024, Journal of Network and Computer ApplicationsGranger Causality for prediction in Dynamic Mode Decomposition: Application to power systems
2023, Electric Power Systems ResearchEnhanced Myocardial Adenylyl Cyclase Activity Alters Heart-Brain Communication
2023, JACC: Clinical Electrophysiology