eMVAR – Extended Multivariate Autoregressive
Modelling Toolbox |
Introduction Multivariate (MV) time series analysis is nowadays extensively
used to investigate the concept of connectivity in dynamical systems through approaches which are ubiquitous,
for instance, in the study of physiological time series. The analysis
of connectivity is not only important to detect coupling, i.e., the presence or absence of interactions, between the
considered processes, but also to identify causality, i.e., the presence of driver-response relationships. Quantitative
assessment of connectivity is commonly performed representing the
considered MV time series as a realization of a linear MV autoregressive
(MVAR) process, and deriving measures of coupling and causality from
the frequency domain representation of the process. The most popular
of these measures are the Coherence (Coh)1, the Partial Coherence (PCoh)2,3, the Directed Coherence (DC)4, the Partial Directed Coherence (PDC)5 and related measures (directed transfer function (DTF)6,7, generalized PDC (gPDC)8). Coh and PCoh are symmetric measures, which can be
decomposed into factors eliciting directionality, these factors being
exactly the DC and the PDC. More specifically, these four measures
describe in the frequency domain the time domain concepts of coupling
(Coh), direct coupling (PCoh), causality (DC), and direct causality
(PDC). Therefore, PCoh and PDC measure direct connectivity between
two processes, while Coh and DC account for both direct and indirect
connections. The parametric model traditionally used to compute Coh,
PCoh, DC and PDC is a strictly causal MVAR model, whereby only lagged effects are modeled, while instantaneous
(i.e., not lagged) effects among the time series are not described
by any model coefficients. Nevertheless, neglecting instantaneous
effects may lead to detection of misleading connectivity patterns.
We have recently shown that causality measures may be adversely affected
by the exclusion of instantaneous effects in the model representation9. To overcome this limitation, we have proposed the utilization
of an extended MVAR model which combines both instantaneous and lagged
effects in order to achieve a full description of the correlation
structure of the observed set of time series9-11. The extended model allows computation of the same connectivity
measures than the traditional strictly causal one: while Coh and PCoh
are identical, DC and PDC may be evaluated either including or excluding
instantaneous effects in the computation of the causality measure;
in the first case the resulting measures, lDC and lPDC, consider only
lagged effects, in a similar way to DC and PDC but resolving the problems
related to model mis-specification; in the second case the resulting
measures, eDC and ePDC, are novel extended measures of causality and
direct causality in which instantaneous causality is accounted for
in combination with the traditionally studied lagged causality12. Note that, in the absence of instantaneous effects,
both lagged and extended measures of causality and causality are equivalent
to the to traditional measures,
because the extended model reduces to the classic strictly causal
model. |
The
eMVAR Toolbox The eMVAR Matlab Toolbox performs both traditional MVAR
analysis and extended MVAR analysis, deriving the corresponding frequency
domain measures of connectivity from the time domain model coefficients.
The toolbox provides also several algorithms for the identification
of the two models from time series data, and is completed with algorithms
for model validation and for the estimation of frequency-domain significance
thresholds. It contains a set of functions realizing model identification
and validation and frequency domain analysis, as well as a set of
scripts illustrating the utilization of the various functions. The
code conforms to methods and notation as described in Faes and Nollo12, which is available from this
link. DOWNLOAD: Zip file with all scripts and functions: eMVAR.zip Note: the “functions” and “external” directories need
to be added to the MATLAB path for proper working Description of the Toolbox Functions:
Scripts:
External
functions: The toolbox makes use of a set of external functions
taken or modified from existing MATLAB toolboxes:
|
References
(1) Kay SM. Modern
spectral estimation. Theory & application. New Jersey: Prentice
Hall, Englewood Cliffs, 1988.
(2) Yacoub.K. Relationship
between multiple and partial coherence functions. IEEE Trans Inf Theory
1970; 16(6):668-672.
(3) Eichler M, Dahlhaus
R, Sandkuhler J. Partial correlation analysis for the identification
of synaptic connections. Biol Cybern 2003; 89(4):289-302.
(4) Baccala LA, Sameshima
K, Ballester G, Valle AC, Timo-Iaria C. Studying the interaction between
brain structures via directed coherence and Granger causality. Appl
Signal Process 1998; 5:40-48.
(5) Baccala LA, Sameshima
K. Partial directed coherence: a new concept in neural structure determination.
Biol Cybern 2001; 84(6):463-474.
(6) Kaminski MJ, Blinowska
KJ. A new method of the description of the information flow in the
brain structures. Biol Cybern 1991; 65(3):203-210.
(7) Kaminski M, Ding
M, Truccolo WA, Bressler SL. Evaluating causal relations in neural
systems: granger causality, directed transfer function and statistical
assessment of significance. Biol Cybern 2001; 85(2):145-157.
(8) Baccala L, Sameshima
K, Takahashi DY. Generalized partial directed coherence.Proceedings
of the 2007 15th International Conference on Digital Signal Processing.
Cardiff, UK: 2007.
(9) Faes L, Nollo G.
Extended causal modelling to assess Partial Directed Coherence in
multiple time series with significant instantaneous interactions.
Biol Cybern 2010; 103(5):387-400.
(10) Erla S, Faes L,
Tranquillini E, Orrico D, Nollo G. Multivariate autoregressive model
with instantaneous effects to improve brain connectivity estimation.
Int J Bioelectromag 2009; 11(2):74-79.
(11) Faes L, Nollo
G. Assessing frequency domain causality in cardiovascular time series
with instantaneous interactions. Methods Inf Med 2010; 49(5):453-457.
(12) Faes L, Nollo
G. Multivariate frequency domain analysis of causal interactions in
physiological time series. In: Laskovski AN, editor. Biomedical Engineering, Trends
in Electronics, Communications and Software. Rijeka, Croatia: InTech, 2011:
403-428.
(13) Schlogl A. A comparison
of multivariate autoregressive estimators. Sign Proc 2006; 86:2426-2429.
(14) Akaike H. A new
look at the statistical model identification. IEEE Trans Autom Contr
1974; 19:716-723.
(15) Rissanen J. An
universal prior for the integers and estimation by minimum description
length. Ann Stat 1983; 11:417-431.
(16) Lutkepohl H. Introduction
to multiple time series analysis. Berlin Heidelberg New York: Springer,
1993.
(17) Faes L, Pinna
GD, Porta A, Maestri R, Nollo G. Surrogate data analysis for assessing
the significance of the coherence function. IEEE Trans Biomed Eng
2004; 51(7):1156-1166.
(18) Faes L, Porta
A, Nollo G. Testing Frequency Domain Causality in Multivariate Time
Series. IEEE Trans Biomed Eng 2010; 57(8):1897-1906.
(19) Faes L, Erla S,
Tranquillini E, Orrico D, Nollo G. An identifiable model to assess
frequency-domain granger causality in the presence of significant
instantaneous interactions. Conf Proc IEEE Eng Med Biol Soc 2010;
1:1699-1702.
|