Correction

STATISTICS, NEUROSCIENCE
Correction for “A study of problems encountered in Granger
causality analysis from a neuroscience perspective,” by Patrick A.
Stokes and Patrick L. Purdon, which was first published August
4, 2017; 10.1073/pnas.1704663114 (Proc Natl Acad Sci USA 114:
E7063–E7072).
The authors wish to note the following: “In the Discussion
section of our paper, we misunderstood the method described in
Barnett and Seth (25) to compute the conditional Granger cau-
sality. We realize now that the spectral factorization method they
describe can be used to obtain the conditional Granger causality
with a single model fit, which would avoid the computational
problems associated with separate model fits. We apologize for
this error.”

25. Barnett L, Seth AK (2014) The MVGC multivariate Granger causality toolbox: A new
approach to Granger-causal inference. J Neurosci Methods 223:50–68.

Published under the PNAS license.

Published online July 9, 2018.

www.pnas.org/cgi/doi/10.1073/pnas.1809324115

E6964 | PNAS | July 17, 2018 | vol. 115 | no. 29 www.pnas.org


A study of problems encountered in Granger causality
analysis from a neuroscience perspective
Patrick A. Stokesa,b,1 and Patrick L. Purdonb,1
a
Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA 02139; and b Department of Anesthesia, Critical Care,
and Pain Medicine, Massachusetts General Hospital, Boston, MA 02144
Edited by Apostolos P. Georgopoulos, University of Minnesota Medical School, Minneapolis, MN, and accepted by Editorial Board Member Thomas D.
Albright June 26, 2017 (received for review March 29, 2017)
Granger causality methods were developed to analyze the flow
of information between time series. These methods have become
more widely applied in neuroscience. Frequency-domain causal-
ity measures, such as those of Geweke, as well as multivariate
methods, have particular appeal in neuroscience due to the preva-
lence of oscillatory phenomena and highly multivariate experi-
mental recordings. Despite its widespread application in many
fields, there are ongoing concerns regarding the applicability of
Granger causality methods in neuroscience. When are these meth-
ods appropriate? How reliably do they recover the system struc-
ture underlying the observed data? What do frequency-domain
causality measures tell us about the functional properties of oscil-
latory neural systems? In this paper, we analyze fundamental
properties of Granger–Geweke (GG) causality, both computational
and conceptual. Specifically, we show that (i) GG causality esti-
mates can be either severely biased or of high variance, both lead-
ing to spurious results; (ii) even if estimated correctly, GG causal-
ity estimates alone are not interpretable without examining the
component behaviors of the system model; and (iii) GG causal-
ity ignores critical components of a system’s dynamics. Based on
this analysis, we find that the notion of causality quantified is
incompatible with the objectives of many neuroscience investiga-
tions, leading to highly counterintuitive and potentially mislead-
ing results. Through the analysis of these problems, we provide
important conceptual clarification of GG causality, with implica-
tions for other related causality approaches and for the role of
causality analyses in neuroscience as a whole.
Granger causality | time series analysis | neural oscillations |
connectivity | system identification
G ranger causality is a statistical tool developed to analyze
the flow of information between time series. Neuroscien-
tists have applied Granger causality methods to diverse sources
of data, including electroencephalography (EEG), magnetoen-
cephalography (MEG), functional magnetic resonance imaging
(fMRI), and local field potentials (LFP). These studies have
investigated functional neural systems at scales of organization
from the cellular level (1–3) to whole-brain network activity (4),
under a range of conditions, including sensory stimuli (5–7),
varying levels of consciousness (8–10), cognitive tasks (11), and
pathological states (12, 13). In such analyses, the time series data
are interpreted to reflect neural activity from a particular source,
and Granger causality is used to characterize the directionality,
directness, and dynamics of influence between sources.
Oscillations are a ubiquitous feature of neurophysiological sys-
tems and neuroscience data. They are thought to constrain and
organize neural activity within and between functional networks
across a wide range of temporal and spatial scales (14–17). Oscil-
lations at specific frequencies have been associated with different
states of arousal (18), as well as different sensory (5) and cog-
nitive processes (19, 20). The prevalence of these neural oscil-
lations, as well as their frequency-specific functional associa-
tions, spurred initial neuroscientific interest in frequency-domain
formulations of Granger causality, such as those developed by
Geweke (21, 22). In addition, neuroscience data are often highly
multivariate, making it difficult to distinguish direct from indi-
rect influences between system components. Geweke’s condi-
www.pnas.org/cgi/doi/10.1073/pnas.1704663114
PNAS PLUS
tional causality measure (22) suggested a means to assess direct
influences within a larger network. Hence, the Granger–Geweke
approach seemed to offer neuroscientists precisely what they
wanted—an assessment of direct, frequency-dependent, func-
tional influence between time series—in a straightforward exten-
sion of standard time series techniques. Current, widely applied
tools for analyzing neuroscience data are based on this Granger–
Geweke (GG) paradigm (23–25).
Many limitations of the GG approach are well known, includ-
ing the requirements that the generative system be approxi-
mately linear, stationary, and time invariant. Analyses of the
GG approach applied to more general processes, such as neural
spiking data (2, 3), continuous-time processes (26), and systems
with exogenous inputs and latent variables (4, 27), have been
shown to produce results inconsistent with the known functional
structure of the underlying system. These examples illustrate the
perils of applying GG causality in situations where the genera-
tive system is poorly approximated by the vector autoregressive
(VAR) model class. Other problems, such as negative causality
estimates (28), have been observed even when the generative
system belongs to the finite-order VAR model class. Together,
these problems raise several important questions for neuroscien-
tists interested in using GG methods: Under what circumstances
are these methods appropriate? How reliably do these methods
recover the functional structure underlying the observed data?
And what do frequency-domain causality measures tell us about
the functional properties of oscillatory neural systems?
In this paper, we perform an analysis of GG causality meth-
ods to help address these questions. We show how, due to the
Significance
Granger causality analysis is a statistical method for investi-
gating the flow of information between time series. Granger
causality has become more widely applied in neuroscience, due
to its ability to characterize oscillatory and multivariate data.
STATISTICS
However, there are ongoing concerns regarding its applicabil-
ity in neuroscience. When are these methods appropriate? How
reliably do they recover the functional structure of the sys-
tem? Also, what do they tell us about oscillations in neural
systems? In this paper, we analyze fundamental properties of
Granger causality and illustrate statistical and conceptual prob-
lems that make Granger causality difficult to apply and inter-
pret in neuroscience studies. This work provides important con-
NEUROSCIENCE
ceptual clarification of Granger causality methods and suggests
ways to improve analyses of neuroscience data in the future.
Author contributions: P.A.S. and P.L.P. designed research, performed research, analyzed
data, and wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission. A.P.G. is a guest editor invited by the Editorial
Board.
Freely available online through the PNAS open access option.
1
To whom correspondence may be addressed. Email: patrickp@nmr.mgh.harvard.edu or
pstokes2@mgh.harvard.edu.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1704663114/-/DCSupplemental.
PNAS | Published online August 4, 2017 | E7063–E7072

practical modeling choices required to compute GG causality,
the structure of the VAR model can lead to bias in conditional
causality estimates. We also show how, for similar reasons, the
GG causality measures can be highly sensitive to variations in
the estimated model parameters, particularly in the frequency
domain, leading to spurious peaks and valleys and even nega-
tive values. Finally, we show how the functional properties of the
underlying system are not clearly represented in Granger causal-
ity measures. As a result, the GG causality measures are prone to
misinterpretation. We conclude by discussing how these findings
apply to other related measures of causality and how many neu-
roscience investigations might be more meaningfully analyzed by
characterizing system dynamics rather than causality.
Background
Granger Causality. Granger causality developed in the field of
econometric time series analysis. Granger (29) formulated a sta-
tistical definition of causality based on the premises that (i) a
cause occurs before its effect and (ii) knowledge of a cause
improves prediction of its effect. Under this framework, a time
series {xi,t } is Granger causal of another time series {xj ,t } if
inclusion of the history of xi improves prediction of xj over
knowledge of the history of xj alone. Specifically, this is quanti-
fied by comparing the prediction error variances of the one-step
linear predictor, x̂j ,t , under two different conditions: one where
the full histories of all time series are used for prediction and
another where the putatively causal time series is omitted from
the set of predictive time series. Thus, {xi,t } Granger causes
{xj ,t } if
var (xj ,t − x̂j ,t | xj ,0:t−1 ) > var (xj ,t − x̂j ,t |xj ,0:t−1 , xi,0:t−1 ).
Granger causality can be generalized to multistep predictions,
higher-order moments of the prediction distributions, and alter-
native predictors (30, 31).
In practice, the above linear predictors are limited to finite-
order VAR models, and Granger causality is assessed by com-
paring the prediction error variances from separate VAR mod-
els, with one model including all components of the vector time
series data, which we term the full model, and a second one includ-
ing only a subset of the components, which we term the reduced
model. To investigate the causality from {xi,t } to {xj ,t }, let
 f
Aj ,j (p) Afj ,i (p) xj ,t−p
" #  f 
xj ,t wj ,t
" # P
X causal components. The frequency-domain definition requires
=   +  [1]
xi,t f f x f
p=1 Ai,j (p) Ai,i (p) i,t−p wi,t
be the full VAR(P) model of all time series components, where
the superscript f is used to denote the full model. This model
may be written more compactly as xt = P f f
P
p=1 A (p) xt−p + wt .
f f
The noise processes {wj ,t } and {wi,t } are zero mean and tempo-
   
T
rally uncorrelated with covariance E wtf1 wtf2 = Σf δt1 −t2 .
Thus, the full one-step predictor of xj ,t in the above causality def-
inition is x̂jf,t = P f
P P
p=1 m∈{i,j } Aj ,m (p) xm,t−p . Similarly, let
P
X r (i)
xj ,t = Ar (i) (p) xj ,t−p + wj ,t
p=1
be the reduced VAR(P) model of the xj (putative effect) compo-
nents of the time series, omitting the xi (putative cause) compo-
nents. We use the superscript r (i) to denote this reduced model
r (i)
formed by omitting xi . The noise process {wj ,t } is zero mean
and temporally uncorrelated with covariance Σr (i) . The reduced
r (i)
one-step predictor of xj ,t is thus x̂j ,t = P
P r (i)
(p) xj ,t−p .
p=1 A
Geweke Time-Domain Causality Measures. Building on Granger’s
definition, Geweke (21) defined a measure of directed causality
(what he referred to as linear feedback) from {xi,t } to {xj ,t } to be
E7064 | www.pnas.org/cgi/doi/10.1073/pnas.1704663114
 
r (i) r (i)
var xj ,t − x̂j ,t Σj ,j
Fxi −→xj = ln   = ln ,
var xj ,t − x̂jf,t Σfj ,j
√
where |Σ| = det Σ0 Σ. Geweke (22) expanded the previous
definition of unconditional (bivariate) causality to include condi-
tional time series components {xk ,t }, purportedly making it pos-
sible to distinguish between direct and indirect influences. For
example, consider three time series {xi,t }, {xj ,t }, and {xk ,t }.
By conditioning on xk , it is to some extent possible to distin-
guish between direct influences between xi and xj , as opposed
to indirect influences that are mediated by xk . The conditional
measure, Fxi −→xj |xk , has a form analogous to the unconditional
r (i)
case, except that the predictors, x̂jf,t and x̂j ,t , and associated
r (i)
prediction-error variances, Σfj ,j and Σj ,j , all incorporate the
conditional time series {xk ,t }.
These time-domain causality measures have a number of
important properties. First, they are theoretically nonnegative,
equaling zero in the case of noncausality. Second, the total lin-
ear dependence between two time series can be represented as
the sum of the directed causalities and an instantaneous causality
(details in S2. Instantaneous Causality and Total Linear Depen-
dence). Finally, these time-domain causality measures can fur-
ther be decomposed by frequency, providing a frequency-domain
measure of causality.
Geweke Frequency-Domain Causality Measures. The above time-
domain measure of causality affords a spectral decom-
position, which allowed Geweke (21) to also define an
unconditional frequency-domain measure of causality. Let
X (λ) = H f (λ) W f (λ) be the frequency-domain representation
of the moving-average (MA) form of the full model in Eq. 1.
X (λ) and W f (λ) are the Fourier representations of the vec-
(f )
tor time series {xt } and noise process {wt }, respectively, and
f
H (λ) is the transfer function given by
P
!−1
f
X f −ipλ
H (λ) = I − A (p) e .
p=1
As alluded to above, the model can contain instantaneously
removal of the instantaneous causality components by trans-
forming the system with a rotation matrix, as described in S3.
Removal of Instantaneous Causality for Frequency-Domain Com-
putation and ref. 21. For clarity, we omit this rotation from the
present overview of frequency-domain causality, but fully imple-
ment the transformation in our computational studies that fol-
low. The spectrum of {xj ,t } is then
Sxj ,xj (λ) = Hjf,j (λ) Σfj ,j Hjf,j∗ (λ) + Hjf,i (λ) Σfi,i Hjf,i∗ (λ),
where ∗ denotes conjugate transpose. The first term is the com-
ponent of the spectrum of xj due to its own input noise pro-
cess, whereas the second term represents the components intro-
duced by the xi time series. The unconditional frequency-domain
causality from {xi,t } to {xj ,t } at frequency λ is defined as
|Sxj ,xj (λ)|
fxi −→xj (λ) = ln
Hjf,j (λ)Σfj ,j Hjf,j∗ (λ)
Hjf,j (λ)Σfj ,j Hjf,j∗ (λ) + Hjf,i (λ)Σfi,i Hjf,i∗ (λ)
= ln .
Hjf,j (λ)Σfj ,j Hjf,j∗ (λ)
[2]
If {xi,t } does not contribute to the spectrum of {xj ,t } at a given
frequency, the second term in the numerator of Eq. 2 is zero
at that frequency, resulting in zero causality. Thus, the uncon-
ditional frequency-domain causality reflects the components of
Stokes and Purdon
 PNAS PLUS
the spectrum originating from the input noise of the putatively ditional GG causality measures. We demonstrate them with sim-
causal time series. ulated examples and discuss the practical implications of using
As with the time-domain measure, Geweke (22) expanded the such methods for analyzing real data, especially neuroscience
frequency-domain measure to include conditional time series. data. The simulations use finite-order VAR systems to ensure
And like the conditional time-domain measure, the conditional that the problems demonstrated are not due to inaccurate mod-
frequency-domain definition requires separate full and reduced eling of the generative process, but rather are inherent to GG
models. Let H f (λ) and Σf be the system function and noise causality. We place some emphasis on the Geweke conditional
covariance of the full model, and let H r (i) (λ) and Σr (i) be the approach (22) for several reasons. Although the unconditional
system function and noise covariance of the reduced model. As approach (21) is more commonly used, the conditional approach
in the unconditional case, instantaneous causality components enables some degree of distinction between direct and indirect
must be removed, this time from each model, as described in S3. influence and is therefore more relevant to the highly multivari-
Removal of Instantaneous Causality for Frequency-Domain Com- ate datasets common in neuroscience. Furthermore, the issues
putation and ref. 22. Again, for clarity, we omit this rotation in being presented are partly what have limited the use of the con-
the equations below, but fully implement this transformation in ditional approach and have motivated the proposal of related
our computational studies that follow. alternative causality measures. We also emphasize the Geweke
The derivation for the conditional form of frequency-domain frequency-domain measures, because they have motivated the
causality relies on the time-domain equivalence demonstrated adoption of Granger causality analysis and related methods in
by Geweke (22), Fxi −→xj |xk = F x w r (i) −→w r (i) . The conditional the neurosciences.
i k j
frequency-domain causality is then defined by fxi −→xj |xk (λ) = Results
fx w r (i) −→w r (i) (λ). Thus, we must rewrite our original VAR Impact of VAR Model Properties on Conditional Causality Estimates.
i k j
r (i) r (i)
In this section, we illustrate how the properties of the VAR
model in terms of the time series {wj ,t },
{xi,t }, and {wk ,t }. model impact estimation of GG causality. We show that the
Geweke (22) did this by cascading the full model with the inverse structure of the VAR model class, in combination with the choice
of an augmented form of the reduced model, to use separate full and reduced models to estimate causality,
 r (i)   r (i) r (i) −1 introduces a peculiar bias–variance trade-off to the estimation
Wj (λ) Hj ,j (λ) 0 Hj ,k (λ) of conditional frequency-domain GG causality. In particular, use
 Xi (λ)  =  0 I 0  [3] of the true system model order results in bias, whereas increas-
r (i)
Wk (λ) H
r (i)
(λ) 0 H
r (i)
(λ) ing the model order to reduce bias results in increased variance
k ,j k ,k
 f  that introduces spurious frequency-domain peaks and valleys.
Hj ,j (λ) Hjf,i (λ) Hjf,k (λ) Wjf (λ) The separate model fits also introduce an additional sensitivity to

 f f
×  Hi,j (λ) Hi,i (λ) Hi,k f
(λ)  Wif (λ)
 parameter estimates, exacerbating the problem of spurious peaks
and valleys and even producing negative values.
f f f
Hk ,j (λ) Hk ,i (λ) Hk ,k (λ) Wkf (λ) Example 1: VAR(3) three-node series system. In this simulation
study, we analyze time series generated by the VAR(3) system
# W f (λ)
Gj ,j (λ) Gj ,i (λ) Gj ,k (λ) shown in Fig. 1. In this system, node 1, {x1,t }, resonates at fre-
"
j
= Gi,j (λ) Gi,i (λ) Gi,k (λ) Wif (λ). [4] quency f1 = 40 Hz and is transmitted to node 2 with an approx-
Gk ,j (λ) Gk ,i (λ) Gk ,k (λ) W f (λ) imate high-pass filter. Node 2, {x2,t }, is driven by node 1, res-
k onates at frequency f2 = 10 Hz, and is transmitted to node 3 with
This combined system relates the reduced-model noise processes an approximate high-pass filter. Node 3, {x3,t }, is driven by node
and the putatively causal time series to the full-model noise pro- 2 and resonates at frequency f3 = 50 Hz. All nodes also receive
cess. It can be viewed as the frequency-domain representation of input from independent, white Gaussian noises, {wi,t }, with zero
r (i) r (i)
a VARMA model of the time series {wj ,t }, {xi,t }, and {wk ,t }. mean and unit variance. The VAR equation of the system is thus
r (i)
The spectrum of {wj ,t } is then
x1,t 2r1 cos (θ1 ) 0 0 x1,t−1
" # " #" #
Sw r (i) ,w r (i) (λ) = Gj ,j (λ) Σfj ,j Gj∗,j (λ) [5] x2,t = −0.356 2r2 cos (θ2 ) 0 x2,t−1
j j x3,t 0 −0.3098 2r3 cos (θ3 ) x3,t−1
+Gj ,i (λ) Σfi,i Gj∗,i (λ)
| {z }
A1

+Gj ,k (λ) Σfk ,k Gj∗,k (λ).

 "

STATISTICS
−r120 0 x1,t−2
#
+ 0.7136 −r22 0  x2,t−2
Hence, the conditional frequency-domain causality is 0 0.5 −r32 x3,t−2
| {z }
Sw r (i) ,w r (i) (λ) A2
j j
fxi −→xj |xk (λ) = ln 0
"
0 0
#"
x1,t−3
# "
w1,t
#
Gj ,j (λ) Σfj ,j Gj∗,j (λ) + −0.356 0 0 x2,t−3 + w2,t ,
0 −0.3098 0 x3,t−3 w3,t
NEUROSCIENCE

r (i)
Σj ,j
= ln
Gj ,j (λ) Σfj ,j Gj∗,j (λ)
The last equality follows from the assumed whiteness of
r (i)
with covariance Σj ,j .
This equation is similar in form to the expression for uncon-
ditional frequency-domain causality in Eq. 2, suggesting, at least
superficially, an analogous interpretation. However, the transfor-
mations detailed in Eq. 3 suggest a more nuanced interpretation.
Overview of Analysis. In the next section, we explore several
issues that arise in the application of the conditional and uncon-
| {z }
. A3
where r1 = 0.9, r2 = 0.7, and r3 = 0.8; θi = fi ∆t2π for i = 1, 2, 3;
1
r (i) and the assumed sampling rate is ∆t = 120 Hz. See S5. Construc-
{wj ,t }
tion of Simulated Examples for the selection of parameter values
for this example.
We generated 1,000 realizations of this system, each of length
N = 500 time points (4.17 s). These simulated data have oscil-
lations at the intended frequencies that are readily visible in the
time domain, as shown in Fig. 1. The conditional frequency-
domain causality was computed for each realization, using sepa-
rate full- and reduced-model estimates using the true model order
P = 3, as well as model orders 6 and 20. The VAR parameters for

Stokes and Purdon PNAS | Published online August 4, 2017 | E7065

 Internal Dynamics Internal Dynamics Internal Dynamics the mean causality curve for high model orders (6 and 20) is con-
of x1 (dB) of x2 (dB) of x3 (dB) sistent with this idea (Fig. 2, Column 4). A pronounced bias is
20 20 20 evident when using the true model order 3, reflected in the large
10 10 10 qualitative differences in the shape of the distribution compared
0 0 0 with the model order 20 case (Fig. 2). This is somewhat discon-
−10 −10 −10 certing as one might expect the causality analysis to perform best
0 50 0 50 0 50
when the true model order is used.
Frequency (Hz) Frequency (Hz) Frequency (Hz)
Whereas the bias decreases as the model order is increased,
the number of parameters increases, each with its own variance.
x1 x2 x3 This leads to increased total uncertainty in the full and reduced
Channel Channel models and in the GG causality estimate. This impacts both the
x1 to x2 (dB) x2 to x3 (dB) detection of causal connections and the estimation of frequency-
0 0 domain GG causality. For causality detection, the analysis cor-
−20 −20 rectly identifies the significant connections under model orders 3
−40 −40
−60 and 6 for all realizations. At model order 20, the variance in the
−60 causality estimates has increased, reflected in the width of the
0 50 0 50
Frequency (Hz) Frequency (Hz) causality distribution and its null distribution, raising the 95%
significance threshold obtained by permutation. As a result, 7%
Spectrum (dB) Realization of Time Series of the realizations fail to reject the null hypothesis that no causal
20 10
connection is present from node 2 to node 3.
We also observe that this increased variance takes a partic-
x1

0 ular form, where the estimates for the individual realizations

0 show numerous causality peaks and valleys as a function of fre-
−10 −10 quency. This occurs because, as the model order increases, so
20 10 too does the number of modes in the dynamical system. In the
frequency domain, this is reflected in the number of poles in the
x2

0 system, each of which represents a peak in the spectral represen-

0
tation (33). Thus, with increasing model order, we see not only an
−10 −10 increase in the variance of the causality distribution, but also an
20 10 increase in the number of oscillatory components, resulting in
an increase in the number of peaks and valleys in the causality
x3

0
0 estimates.
−10 −10 The bias in the conditional GG causality distribution under the
0 50 0 1 2 true model order occurs because, whereas GG causality is the-
Frequency (Hz) Time (sec) oretically defined using infinite histories, practical computation
requires using a finite order. Unfortunately, the reduced models
Fig. 1. VAR(3) three-node series system of example 1. Top shows the net- are being used to represent a subset of a VAR process, and sub-
work diagram of the system, with plots showing the frequency response of sets of VAR processes are not generally VAR, but instead belong
the internal dynamics (univariate AR behavior) of each node and the fre-
to the broader class of vector autoregressive MA (VARMA) pro-
quency responses of the channels connecting the nodes. Bottom shows the
cesses. VAR representations of VARMA processes are generally
spectrum of each node and a corresponding time series of a single realiza-
tion from the system. The high signal-to-noise ratio of this system and these
infinite order. Thus, when a finite-order VAR model is used for
data is evident in the highly visible oscillations in the time series and in the
the reduced process, as required in conditional Geweke causality
spectral peaks. computation, some terms will be omitted.
This can be seen more clearly by rewriting the full VAR model,
xj ,t Aj ,j (p) Aj ,i (p) Aj ,k (p) xj ,t−p wj ,t
" # P
" #" # " #
X
each model order were well estimated from single realizations of xi,t = Ai,j (p) Ai,i (p) Ai,k (p) xi,t−p + wi,t .
the process (Table S1). We also performed model order selec- xk ,t p=1 Ak ,j (p) Ak ,i (p) Ak ,k (p) xk ,t−p wk ,t
tion using a variety of methods—including the Akaike informa-
tion, Hannan–Quinn, and Schwarz criteria—all of which correctly Focusing on the reduced components, we see the VARMA form,
identified, on average, the true model order to be P = 3 (Fig. S1).   XP   
Fig. 2 shows the distributions of causality estimates for differ- xj ,t Aj ,j (p) Aj ,k (p) xj ,t−p
ent model orders for the truly causal connections from node 1 x = A (p) A (p) x +
k ,t k ,j k ,k k ,t−p
p=1
to node 2 and from node 2 to node 3, as well as the truly non- | {z }
causal connection from node 3 to node 1. Results for the true Autoregressive Term
model order 3 are shown in Fig. 2, Column 1, whereas those for   X P  
the order 6 and 20 models are shown in Fig. 2, Columns 2 and 3, wj ,t Aj ,i (p)
respectively. Each plot shows the median (blue line) and 90% wk ,t + Ak ,i (p) xi,t−p .
central interval (purple shading) of the distribution of causal- p=1
ity estimates. In addition, we show estimates for a single real-
| {z }
Moving-Average Term
ization (red line) and the 95% significance level (black dashes)
estimated by permutation (32). The details for the significance For the reduced model, the MA terms, specifically those contain-
level computation are described in S9. Permutation Significance ing xi , need to be appropriately projected onto the typically infi-
for Noncausality. Fig. 2, Column 4 shows the mean causality val- nite histories of the remaining xj and xk components. Truncation
ues for each order. The causality estimates for all cause–effect of the reduced VAR to finite order eliminates nonzero, although
pairs for model orders 3, 6, and 20 are shown in Figs. S2, S3, and diminishing, terms. Thus, when using a finite order, especially the
S4, respectively. true model order for the full system, the reduced model com-
The differences in these causality distributions reflect the ponents can be poorly represented, leading to biased causality
trade-off in bias and variance that would be expected with estimates.
increasing model order. Because the GG causality measure is The stochastic nature of the separate fits can further con-
defined for infinite-order VAR models, high-order VAR mod- tribute to artifactual peaks and valleys and even negative causal-
els are approximately unbiased. The similarity in the shape of ity (28). This is because when the full and reduced models

E7066 | www.pnas.org/cgi/doi/10.1073/pnas.1704663114 Stokes and Purdon

 PNAS PLUS
Fig. 2. Comparisonofconditionalfrequency-domain Model Order P = 3 P=6 P = 20 Means
causality estimates for the VAR(3) three-node series Causality from x1 to x2 (truly causal)
system of example 1 using model orders 3, 6, and 6 7 7 7
20. Rows 1, 2, and 3 show the results for the truly 6 6 6
causal connections from node 1 to node 2 and from 4
4 4 4
node 2 to node 3 and a truly noncausal connection
2 2
from node 3 to node 1, respectively. Columns 1, 2, 2 2
and 3 show the results from using the true model 0 0 0 0
order 3 and the increased model orders 6 and 20, 0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
respectively. Each subplot shows the median (blue Causality from x2 to x3 (truly causal)
line) and 90% central interval (purple shading) of 0.8 0.7 0.8 0.8
the distribution of causality estimates, the estimate 0.6
0.6 0.6 0.6
for a single realization (red line), and the 95% sig- 0.4 0.4
0.4 0.4
nificance level (black dashes) estimated by permuta-
0.2 0.2
tion. For further comparison, Column 4 overlays the 0.2 0.2
0
mean causality estimates for three model orders: 3 0 0 -0.1 0
(red), 6 (blue), and 20 (black). The bias from using 0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
the true model order is indicated by the qualitative Causality from x3 to x1 (truly non-causal)
differences between the mean estimates of Column -0.3 -0.3 0.5
-0.2 -0.2 0.4 0.04
4 and between the distributions of Column 1 com- 0.02
pared with those of Columns 2 and 3, where larger 0 0 0.2
0
orders were used. Whereas increasing the model order 0
-0.2 -0.2 -0.02
diminishes the bias, the variance is increased, indi-
-0.2 -0.04
cated by the increase in the width of the distribu- -0.4 -0.4 -0.3
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
tions, the increased number of peaks and valleys of
Frequency (Hz)
the individual realizations, and the increase in the
null significance level. The additional sensitivity to variability in the model parameter estimates from the separate full and reduced model fits is most
evident from the occurrence of negative causality estimates. This occurs predominantly for the noncausal connection (Row 3), but when the model
order is increased to 20, instances of negative causality also occur for the truly causal connection from node 2 to node 3 (Row 2, Column 3).

are estimated separately, their frequency structure can be mis-
aligned, producing spurious fluctuations in the resulting causal-
ity estimate. As seen in Fig. 2, negative values are extensive for
the truly noncausal connection from node 3 to node 1 and even
appear at model order 20 for the truly causal connection from
node 2 to node 3. This sensitivity to variability in the model
parameter estimates can be quite dramatic, depending on the
specific system, and compounds the increased variability due to
high model order discussed above.
Causality Interpretation: How Are the Functional Properties of the
System Reflected in GG Causality? Given a frequency-domain
causality spectrum over a range of frequencies, a face value inter-
pretation would be that peaks represent frequency bands that
are more causal or effective, whereas valleys represent frequency
bands of lesser effect. At the same time, we know that causal-
ity measures are functions of the VAR system parameters and
reflect a combination of dynamics from the different components
of the system. How then do causality values relate to the under-
examine a bivariate case, such that the unconditional causality
can be computed correctly from a single model using the true
VAR parameters. By using the true VAR parameters, we also
eliminate the uncertainties associated with VAR parameter esti-
mation. Thus, in this example, we focus solely on the functional
relationship between system structure and the causality measure.
We compare the causality of a set of similar unidirectional
two-node systems, all with the structure shown in Fig. 3, Left.
To isolate the influence of transmitter and receiver behaviors,
we fix the channel as all pass (i.e., all frequencies pass through
the channel without influencing the signal amplitude). In these
systems, node 1, {x1,t }, is driven by a white noise input, {w1,t };
resonates at frequency f1 ; and is transmitted to node 2. Node 2,
{x2,t }, is driven by both node 1 and a white noise input, {w2,t },
and resonates at frequency f2 . The channel is fixed as all-pass
unit delay, but the transmitter and receiver resonance frequen-
cies are varied. The set of transmitter resonance frequencies is
f1 ∈ {10, 20, 30, 40, 50} Hz. The set of receiver resonance fre-
quencies is f2 ∈ {10, 30, 50} Hz. The VAR equation of these

STATISTICS
lying structure or dynamics of the generative system? systems is thus
In this subsection, we examine the relationship between sys-     
tem behavior and causality values by systematically varying the x1,t 2r1 cos (θ1 ) 0 x1,t−1
x2,t = 1 2r2 cos (θ2 ) x2,t−1 +
parameters of different parts of the system and assessing the
dependence of the causality on the component behaviors. We
 2    
−r1 0 x1,t−2 w1,t
use the terms “transmitter” to denote putatively causal nodes, + w2,t ,
0 −r22 x2,t−2
“receiver” to denote putative effect nodes, and “channel” to
denote the linear relationship linking the two in a particular where ri = 0.65 and θi = fi ∆t2π for i = 1, 2, and the assumed
NEUROSCIENCE

direction of connection. We will show that GG causality reflects
only the dynamics of the transmitter node and channel, with no
dependence on the dynamics of the receiver node. This suggests
that, even in simple bivariate AR systems, GG causality may
be prone to misinterpretation and may not reflect the intuitive
notions of causality most often associated with these methods.
Example 2: Transmitter–receiver pair with varying resonance
alignments. In this simulation study, we investigate how GG
causality values change as the system’s frequency structure varies,
to determine whether the changes in the causality values agree
with the face value interpretation of causality described above.
We structure the analysis to avoid the model order and model
subset issues described in the previous section. Specifically, we
1
sampling rate is ∆t = 120 Hz. The driving inputs are inde-
pendent white Gaussian noises with zero mean and unit vari-
ance. The frequency-domain causality is computed for each sys-
tem using the single-model unconditional approach with the
true parameter values. Fig. 3, Right shows the true frequency-
domain causality values for the different systems. Each plot
in Fig. 3, Right represents a single receiver resonance fre-
quency, with the different transmitter frequencies. For instance,
Fig. 3, Top Right shows the values for all five transmit-
ter settings with the receiver resonance set at 10 Hz. Simi-
larly, Fig. 3, Middle Right and Bottom Right shows the values
with the receiver resonance held fixed at 30 Hz and 50 Hz,
respectively.

Stokes and Purdon PNAS | Published online August 4, 2017 | E7067

Receiver Independence: GG Causality Does Not Depend on Receiver Note how the A2,2 (λ) system component and hence the
Dynamics. Perhaps the most notable aspect of Fig. 3 is that the {A2,2 (p)} system parameters, which characterize the receiver
causality plots in Fig. 3, Right are identical, indicating that the dynamics, are absent. Thus, the unconditional frequency-domain
frequency-domain causality values are independent of the reso- causality does not reflect receiver dynamics.
nance frequency of the receiver. This suggests that the uncondi- Instead, the causality is a function of the transmitter and chan-
tional causality is actually independent of the receiver dynamics. nel dynamics and the input variances. In Fig. 3, we see that
We can see this by explicitly writing the causality function in terms the causality parallels the transmitter spectrum, with the causal-
of the VAR system components. The transfer function of the sys- ity peak located at the transmitter resonance frequency. This
tem is the inverse of the Fourier transform of the VAR equation, makes sense, as the channel remains fixed as an all-pass unit
!−1 delay, so the causality is determined by the transmitter spectrum,
P
−1
X −ipλ A−1 −1∗
1,1 (λ)Σ1,1 A1,1 (λ), in the second term of the numerator in
H (λ) = A (λ) = I − A(p) exp .
p=1
Eq. 6.
Even in this simple bivariate example, we see that the causal-
As shown in S6. Receiver Independence of Unconditional ity analysis is potentially misleading. For example, consider the
Frequency-Domain GG Causality, the unconditional causality (Eq. cases where the 50-Hz transmitter drives either a 10-Hz receiver
2) from node 1 to node 2 can be expressed in terms of the AR matri- (Fig. 3, Top row, purple lines) or a 50-Hz receiver (Fig. 3, Bottom
ces associated with each (possibly vector-valued) component row, purple lines). The spectra we observe at the effect nodes in
∗ ∗ these scenarios are different (“Spectrum 2” in Fig. 3, Middle col-
H2,2 (λ) Σ2,2 H2,2 (λ) + H2,1 (λ) Σ1,1 H2,1 (λ) umn); in one case, the spectrum at the effect node has a peak
fx1 −→x2 (λ) = ln ∗
H2,2 (λ) Σ2,2 H2,2 (λ) at 10 Hz (Fig. 3, Top Center), whereas in the other one, it has
a peak at 50 Hz (Fig. 3, Bottom Center). However, the causal-
Σ2,2 +A2,1 (λ) A−1 −1∗ ∗
1,1 (λ) Σ1,1 A1,1 (λ) A2,1 (λ) ity functions are identical. Thus, GG causality will fail to make
= ln . a distinction between these scenarios. If the analyst’s notion of
|Σ2,2 | what should be causal is in any way related to the properties of
[6] the observed effect—e.g., 10-Hz alpha waves vs. 50-Hz gamma

System Causality from 1 to 2

3
10, 20, 30, 40, 10 10 50 10
and 50 Hz 10 Hz
1 time-step
1 2 2
20 10 40 10
Spectrum 1 (dB) Spectrum 2 (dB) 30 10
15 25
10 20
1
10
0
0
−10 −10 0
0 20 40 60 0 20 40 60 0 20 40 60

3
10, 20, 30, 40, 10 30 50 30
and 50 Hz 30 Hz
1 time-step
1 2 2
20 30 40 30
Spectrum 1 (dB) Spectrum 2 (dB) 30 30
15 12
10 8 1
4
0
0
−10 −4 0
0 20 40 60 0 20 40 60 0 20 40 60
3 Fig. 3. System and unconditional frequency-
10, 20, 30, 40, 10 50 50 50 domain causality values for different resonance
and 50 Hz 50 Hz alignments of the VAR(2) unidirectional trans-
1 time-step mitter–receiver pair of example 2. The channel is
1 2 2 all-pass unit delay. Three receiver resonances are
20 50 40 50 tested: 10 Hz, 30 Hz, and 50 Hz, shown in the
Spectrum 1 (dB) Spectrum 2 (dB) 30 50 Top, Middle, and Bottom rows, respectively. For
15 25 each setting of the receiver resonance, five trans-
10 20 mitter resonances are tested: 10 Hz, 20 Hz, 30 Hz,
1
10 40 Hz, and 50 Hz. Left column shows the spec-
0 tra. Right column shows the causality. As discussed
0 in the text, the causality reflects the spectrum of
−10 −10 the transmitter and channel, independent of the
0
0 20 40 60 0 20 40 60 0 20 40 60 receiver dynamics. Hence, all three plots of the
Right column are identical and similar to the node
Frequency (Hz) Frequency (Hz) Frequency (Hz) 1 spectrum plots of the Left column.

E7068 | www.pnas.org/cgi/doi/10.1073/pnas.1704663114 Stokes and Purdon


waves—the face value interpretation of GG causality would be
grossly misleading.
Said another way, the receiver independence property of GG
causality implies that two systems can have identical causality
functions, but completely different receiver dynamics. The fre-
quency response and causality function in these examples reflect
merely simple parametric differences in the transmitter dynam-
ics. In a more general scenario, the form of the transmitter and
channel VAR parameters can be more complicated, making it
even more difficult to interpret causality values without first con-
sidering the dynamics of the underlying system components.
Receiver Independence: Unconditional and Conditional Time Domain
and Conditional Frequency Domain. We now show how the remain-
ing forms of GG causality—unconditional time domain, con-
ditional time domain, and conditional frequency domain—also
appear to be independent of the receiver dynamics. For time-
domain GG causality, we have derived closed-form expressions
that illustrate receiver independence for both the unconditional
and conditional cases. The details for these derivations are
described in S7. Receiver Independence of Time-Domain GG
Causality. Briefly, the time-domain GG causality compares the
prediction error variances from full and reduced models. The
prediction error variance for the full model is simply the input
noise variance, so the dependence of the GG causality on the
VAR parameters arises through the prediction error variance
of the reduced model. In general, the reduced model can be
obtained only numerically (25), which obscures the form of its
dependence on the full-model VAR parameters. Instead, we use
a state-space representation of the VAR to derive an implicit
expression for the reduced-model prediction error variance in
terms of the VAR parameters (S7. Receiver Independence of
Time-Domain GG Causality). We then show that the reduced-
model prediction error variance is independent of the receiver
node VAR parameters. Hence, the time-domain GG causality
is independent of receiver dynamics. This is true for both the
unconditional and conditional cases.
At present, we do not have a closed-form expression for the
conditional frequency-domain reduced model in terms of the
full model parameters. As a consequence, we cannot determine
analytically whether the conditional frequency-domain causality
depends on the receiver node parameters. However, we can use
the relationship between the conditional time- and frequency-
domain causality measures to help analyze receiver dependence.
We note that the frequency-domain causality is a decomposition
of the time-domain causality. In the unconditional case (21),
1
Z π
fx −→xj (λ) d λ≤ F xi −→xj ,
2π −π i
withequality if the transmitter
 node is stable—i.e., iff the roots of
PP p
det I − p=1 Ai,i (p) z are outside the unit circle—which is
the case for the systems in example 2. Similarly, in the conditional
case (22),
1
Z π
fx −→xj |xk (λ)d λ ≤ Fxi −→xj |xk , [7]
2π −π i
P 
∞ p
with equality iff the roots of det p=1 G (p) z are inside
the unit circle, where G is the system function of Eq. 4. Hence,
the frequency-domain causality is a decomposition of the time-
domain causality.
Because the time-domain conditional causality is indepen-
dent of the receiver node, the relationship expressed in Eq. 7
would, at a minimum, impose strict constraints on the form
of any receiver dependence in the frequency-domain condi-
tional causality. Example 1 from the previous subsection qual-
itatively agrees with this (Fig. 2, Column 3). We see that
receiver dynamics do not appear to be reflected in the con-
ditional frequency-domain causality. In the causality estimates
from node 1 to node 2 (Fig. 2, Top), the 10-Hz receiver reso-
Stokes and Purdon
PNAS PLUS
nance of node 2 appears to be absent. Similarly, in the causal-
ity estimates from node 2 to node 3 (Fig. 2, Middle), the 50-Hz
receiver resonance of node 3 appears to be absent. Thus, the
conditional frequency-domain causality appears to be indepen-
dent of the receiver dynamics, similar to the other cases con-
sidered above.
Example 1 also helps to illustrate the challenge in interpret-
ing GG causality when both the transmitter and the channel have
dynamic properties, i.e., a frequency response. From node 1 to
node 2, we see that the causality (Fig. 2, Column 3, Top) primar-
ily reflects the spectrum of the transmitter, showing a prominent
peak at 40 Hz (node 1, Fig. 1). In contrast, the causality from
node 2 to node 3 shows a nadir at ∼15 Hz (Fig. 2, Column 3, Mid-
dle), which resembles the channel dynamics more than the node 2
transmitter dynamics (Fig. 1). This example reinforces the earlier
observation that causality values cannot be appropriately inter-
preted without examining the estimated system components.
Discussion
Summary of Results. In this paper we analyzed several problems
encountered during Granger causality analysis. We focused pri-
marily on the Geweke frequency-domain measure and the case
where the generative processes were finite VARs of known
order. We found the following:
i) There is a troublesome bias–variance trade-off as a function
of the VAR model order that can produce erroneous condi-
tional GG causality estimates, including spurious peaks and
valleys in the frequency domain, at both low and high model
orders. At low model orders, the GG causality can be biased,
due to the fact that subsets of VAR processes—e.g., the puta-
tive effect and conditional nodes represented by the reduced
model—are VARMA processes, which in general can be rep-
resented only by infinite-order VAR models. Crucially, sig-
nificant biases occur even when the data-generating process
is VAR and the true underlying model order is used. As the
model order increases, the bias decreases, but each model
order introduces new oscillatory modes, whose variance can
then introduce spurious peaks in the frequency-domain GG
causality.
ii) GG causality estimates are independent of the receiver
dynamics, which can run counter to intuitive notions of
causality intended to explain observed effects. Instead, GG
causality estimates reflect a combination of the transmitter
and channel dynamics, whose relative contributions can be
understood only by examining the component dynamics of the
estimated model. As a result, causality analyses, even for the
simplest of examples, can be difficult to interpret.
Our results illustrate fundamental problems associated with
STATISTICS
GG causality, in terms of both its statistical properties and its
conceptual interpretation. These problems have important prac-
tical consequences for neuroscience investigations using GG
causality and related methods. Moreover, as we discuss below,
these problems can be avoided in large part by paying more care-
ful attention to the modeling and system identification steps that
are fundamental to data analysis (Fig. 4).
NEUROSCIENCE
Alternative Approaches for Estimating GG Causality. The potential
for negative causality estimates and spurious peaks in frequency-
domain GG causality has been recognized previously (28). The
problem has been framed primarily as one in which conditional
causality estimates are sensitive to mismatches between the sepa-
rate full and reduced models; i.e., spurious values occur when the
spectral features of the full and reduced models do not “line up”
properly. Our work identifies the bias–variance trade-off in the
VAR model order as a more fundamental difficulty. Model order
selection is problematic even in the context of single-model para-
metric spectral estimation. AR models have the property that
as model order is increased, the number of potential oscillatory
modes increases, which can alter estimated peak locations. This
PNAS | Published online August 4, 2017 | E7069
problem is compounded when the full model must be compared
with an approximate reduced model.
Alternative approaches have been proposed for estimating
conditional frequency-domain GG causality to eliminate the
need for separate model fits. Chen et al. (28) suggested using
a reduced model featuring a colored noise term formed from the
components of the estimated full model. Whereas this method
successfully eliminates negative causality values, it does not cor-
rectly partition the variance of the putative causal nodes within
the reduced model (34). Thus, it is unclear whether the method
described in Chen et al. (28) can accurately estimate conditional
GG causality. Barnett and Seth (25) have proposed fitting the
reduced model and using it to directly compute the spectral
components of Eq. 5. However, for the same reasons described
above, these reduced-model estimates will generally require high
model order representations that would be susceptible to spu-
rious peaks. Moreover, this method would remain sensitive to
mismatches between the spectral estimates of the components in
Eqs. 3–5.
Potential Problems in Interpreting GG Causality Analyses in Neuro-
science Studies. The statistical problems illustrated in this paper
raise obvious concerns for the reliability of Granger causality
estimates. However, the problems of interpreting causality val-
ues, even if they are estimated accurately, are more fundamen-
tal. Receiver independence for a scalar unconditional frequency-
domain example was previously noted in ref. 35. In this work, we
have shown that this holds more generally, for the vector uncon-
ditional frequency-domain, the unconditional time-domain, and
the conditional time-domain cases. We have also shown evidence
for receiver independence in the conditional frequency-domain
case as well. Here we also discuss the conceptual difficulty this
poses in neuroscience.
That GG causality is unaffected by and not reflective of the
receiver dynamics is entirely understandable from the original
principle underlying the definition of Granger causality, that
of improved prediction. The use of improvement of prediction
makes sense in the context of econometrics, from which GG
causality originated, where the objective is often to construct par-
simonious models to make predictions. The functional interpre-
tation of the models is considered less important in this context
than the accuracy of the predictions. However, in neuroscience
investigations, the opposite is generally true: Fundamentally, a
mechanistic understanding of the neurophysiological system is
sought, and as a result the structure and functional interpre-
tation of the models are most important. In particular, neuro-
science investigations seek to determine the mechanisms that
produce “effects” observed at a particular site within a neural
system or circuit as a function of inputs or “causes” observed
at other locations. The effect node dynamics are central to this
question, as they determine how the effect node responds to dif-
ferent inputs. In contrast, the predictive notion of causality that
underlies Granger causality methods explicitly ignores the effect
node dynamics and effect size.
There are numerous examples in neuroscience that illustrate
the importance of the effect or receiver node dynamics and
effect size. For example, the dynamics of high-threshold thala-
mocortical neurons are governed by the degree of membrane
hyperpolarization, producing tonic firing at the lowest levels of
hyperpolarization, burst firing at alpha (∼10 Hz) frequencies at
intermediate levels of hyperpolarization, and burst firing at low
frequencies (<1 Hz) at even greater levels of hyperpolarization
(36). If we viewed the firing behavior or postsynaptic fields gener-
ated by these neurons as the effect to be explained, the Granger
methods would be blind to their frequency response, which
is fundamental to their neurophysiology. As another example,
studies of cognition and memory using fMRI have demonstrated
that the amplitude of blood–oxygen-level–dependent responses
can be modulated by stimulus novelty and cognitive complexity
(37). Characterizations of these effects with Granger causality
methods would be similarly indeterminate in explaining these
amplitude-dependent effects that are central to the cognitive
E7070 | www.pnas.org/cgi/doi/10.1073/pnas.1704663114
processes being studied. As a final example, we note a recent
study by Epstein et al. (38) in which Granger causality was used
to analyze electrocorticogram recordings of seizure activity in
epilepsy patients, with the goal of informing epilepsy surgery
planning. Epileptic seizures are observed electrophysiologically
by the localized onset and propagation of characteristic oscil-
lations, often of large amplitude. The seizure effect is there-
fore defined by both the system dynamics and the amplitude of
the electrophysiological response. Unfortunately, because GG
causality is indifferent to the receiver dynamics and the ampli-
tude of the measured effects, it cannot specify which influences,
for instance at specific foci and frequencies, contribute to the
seizure effect. In particular, an epileptogenic focus would not
necessarily have large GC values, because the seizure onset and
propagation may be due to the effect node’s internal dynamics.
At the same time, large GC values do not necessarily suggest a
focus, because large GC values relate to prediction error vari-
ance and not to the effect at the recording site.
These examples illustrate how Granger causality methods, due
to the receiver-independence property, can fail to character-
ize essential neurophysiological effects of interest and lead to
misinterpretation of the causes for those effects. These exam-
ples are representative of typical neuroscience problems seeking
the “cause” for an “effect,” implying that the misapplication of
Granger methods is likely widespread within neuroscience.
Implications for Other Causality Methods. The conceptual dis-
agreement between GG causality and the mechanistic interpreta-
tion of data from neuroscience studies such as Epstein et al. (38)
stems fundamentally from a disconnect between the system prop-
erties of interest, given the scientific question, and those repre-
sented by GG causality. Beyond GG causality, numerous alter-
native causality measures have also been proposed, including the
directed transfer function (DTF) (39, 40), the partial directed
coherence (PDC) (41), and the direct directed transfer function
(dDTF) (42, 43). Like GG causality, these causality measures are
based on a VAR model, but each of these methods represents a
different functional combination of the individual system com-
ponents and thus a conceptually distinct notion of causality. The
DTF represents the transfer function, not from the transmitter to
the receiver, but instead from an exogenous white noise input at
the transmitter to the receiver. Depending on the normalization
used, it may actually include the dynamics of all systems compo-
nents. The DTF essentially captures all information passing from
the transmitter directed to the receiver, but this information flow
is not direct (42, 43) (Fig. S5). The dDTF claims to characterize
direct influences by normalizing the DTF by the partial coher-
ence. The PDC, on the other hand, represents the actual channel
between two nodes (44) but, depending on how it is normalized,
may contain the frequency content of other components (Fig.
S5). These measures all differ from GG causality, which repre-
sents a combination of the transmitter and direct channel dynam-
ics, as well as the conditional node variances and their indirect
channel dynamics.
Each of the above measures, with the exception of the dDTF,
is based on a single-model estimate and thus does not suffer
from the model subset or sensitivity issues we have described
for GG causality. However, similar to GG causality, all of these
methods can be problematic to interpret, because each measure
represents a combination of different system components whose
contributions cannot be disentangled by examining the causal-
ity values alone. Thus, studies using these methods may be sub-
ject to a mismatch between the system properties represented
by the measures and the properties of interest to the investiga-
tor. Despite the clear fundamental differences in how these and
other causality measures are defined, these methods are often
discussed as interchangeable alternatives or referred to generi-
cally as “Granger causality” (35, 44, 45), making it even more
difficult to interpret studies using these methods.
In many cases, investigators may wish to detect connections
between nodes in a system and may not be concerned with
Stokes and Purdon

(6) Misinterpreting relationship between
Scientific Question scientific question and causality measures;
(7) Conflation of causality measures
Design Postulate
(1) Overlooking modeling,
Experiment Models
jumping ahead to causality
Model Fit Properties of
Data
Set Criterion Interest
Model Model
Fit Model Analyze Model
Estimate Statistics
e.g. examine component
e.g. fit VAR(P)
dynamics, compute
causality statistics
Check Model
e.g. check residuals,
posterior predictive checks
(2) Model poorly represents (4) Component dynamics are entangled
underlying process; within the causality measure;
(3) Model estimated poorly (5) Causality measures have poor
statistical properties
interpreting the form or dynamics of such connections. In such
cases, under a VAR model, the direct connection between nodes
is represented by the corresponding cross-components of the
VAR parameter matrices, which indicate the absence of a con-
nection if all of these parameters are equal to zero. Hypothesis
tests for such connections have been described in the time series
analysis literature (30, 46). This approach is analogous to that pro-
posed by Kim et al. (1) for point process models of neural spik-
ing data. To our knowledge, the VAR versions (30, 46) of this
approach have not been used in neuroscience applications, but
may offer potential advantages in computational and statistical
efficiency compared with more widely applied permutation-based
tests on causality measures, such as those outlined in ref. 25.
A number of causality approaches have been developed that
purportedly handle a wider range of generative dynamic systems,
beyond those that might be well approximated by VAR models.
Transfer entropy methods (31, 47, 48) can be viewed as a gener-
alization of the Granger approach and are applicable to nonlin-
ear systems. However, such methods require significant amounts
of data, are computationally intensive, and present their own
model identification issues, such as selection of the embedding
order. These approaches are strictly time-domain measures, and
their similarity to GG causality (31) suggests that they would face
similar problems with interpretation such as independence of
receiver dynamics. Dynamic causal modeling (DCM) uses neuro-
physiologically plausible dynamic models to characterize causal
relationships (49). The models, estimation methods, and notions
of causality involved in DCM are different from the time series-
based causality approaches discussed here. The interpretation
and statistical properties of DCM are the subject of ongoing
analysis (50–53).
Causality Analysis in the Framework of System Identification. The
problems in causality methods that we have discussed arise when
these methods are applied without clear reference to the underly-
ing models, their estimation methods and statistical measures, or
the scientific question of interest—that is, without clear reference
to the process of system identification (54, 55). System identifi-
cation comprises the set of steps undertaken to study a scientific
Stokes and Purdon
PNAS PLUS
Fig. 4. System identification framework. The
processes of modeling and data analysis, some-
times referred to as system identification,
involve a number of steps shown here (54, 55).
Models of one form or another underlie all data
analyses, and model selection involves several
crucial choices: the class of models from which
a model structure is selected, the method(s) by
which candidate models are estimated, and the
criteria by which model fits are optimized and
a structure selected. Subsequent analysis steps
also involve important choices, including the
properties and statistics to be investigated and
the methods by which they are computed. These
choices are shaped by the scientific question
Make Inferences
of interest. Unfortunately, the use of causality
Based On Model
analysis methods often hides this process and
the associated choices from the investigator (1).
Consequently, the many possible points of fail-
ure for causality methods are also hidden, such
as inappropriate model structure (2) and poor
model estimation (3), poor representation of
the features of interest (4), and poor statistical
properties (5). More fundamentally, application
of causality methods can obscure the scientific
question of interest, resulting in misinterpreta-
tion (6). In addition, different causality measures
with different properties or interpretations are
often conflated (7).
question (Fig. 4), including the processes of model formulation
and model selection. Although these steps can be challenging and
complex, they are essentially a quantitative formalization of the
scientific method. Unfortunately, causality methods have been
used, in effect, to circumvent many of these painstaking steps (Fig.
4, annotation 1). As we have seen in our analysis, this can mask
many potential points of failure, including poor representation
of the system by the model, and computational and interpreta-
tional problems with statistics computed from the model. Perhaps
more importantly, the use of causality methods supplants a clear
statement of the scientific question of interest with an ambiguous
statistic and a loaded concept that carries philosophical connota-
tions, detracting from the question at hand.
In many cases, the scientific question of interest is simply to
determine the presence or absence of a direct connection from
one region to another. This question can often be answered
directly from the estimated model. In the case of VAR mod-
els, as discussed earlier, a hypothesis test could be performed
STATISTICS
on the off-diagonal VAR components (30). In the transfer func-
tion case, this is given by the unnormalized PDC. In DCM, this is
explicitly represented in the selected model’s connections (49).
In other cases, the scientific question centers on characterizing
the mechanistic interactions between components of a system.
As we have argued, this can be ascertained by examining the
component dynamics from the estimated model. For instance,
in the analysis of electrocorticographic oscillations in nonhu-
NEUROSCIENCE
man primate somatosensory cortex described in ref. 5, one could
examine the component node and channel dynamics to charac-
terize more specific interactions between system components.
Yet in other cases, the scientific question may relate to pre-
dicting the effects that structural changes might have on system
dynamics. For instance, Epstein et al. (38) wanted to evaluate
whether surgical resections could alleviate seizures. This question
could also be investigated from the estimated model, by removing
connections or nodes, simulating data using the model, and then
evaluating the influence of these interventions on the generation
of seizure oscillations. The validity of such statements would be
incumbent on the applicability of the model and would require
experimental validation. However, by structuring the problem in
PNAS | Published online August 4, 2017 | E7071
terms of the dynamic systems model, this validation is straight-
forward, more so than with causality measures.
In the end, causality measures, however elaborate in construc-
tion, are simply statistics estimated from the model (Fig. 4, model
statistics). If the model does not adequately represent the system
properties of interest, subsequent analyses based on the model
will fail to address the question of interest. The VAR model class
underlying GG causality and other causality measures represents
linear relationships within stationary time series data. Although
VAR models can, in many cases, adequately approximate prop-
erties of nonlinear systems, in general this structure limits the
system behaviors that can be modeled. For example, VAR mod-
els cannot explicitly represent cross-frequency coupling. Thus,
although a model structure may fit some features of the data, its
representation of other features and applicability to the system as
a whole may not be adequate and must be considered carefully
(Fig. 4, postulate models). The inability of the model to represent
1. Kim S, Putrino D, Ghosh S, Brown EN (2011) A Granger causality measure for point
process models of ensemble neural spiking activity. PLoS Comput Biol 7:e1001110.
2. Kispersky T, Gutierrez GJ, Marder E (2011) Functional connectivity in a rhythmic
inhibitory circuit using Granger causality. Neural Syst Circuits 1:9.
3. Gerhard F, et al. (2013) Successful reconstruction of a physiological circuit with known
connectivity from spiking activity alone. PLoS Comput Biol 9:e1003138.
4. Eichler M (2005) A graphical approach for evaluating effective connectivity in neural
systems. Philos Trans R Soc Lond B Biol Sci 360:953–967.
5. Brovelli A, et al. (2004) Beta oscillations in a large-scale sensorimotor cortical net-
work: Directional influences revealed by Granger causality. Proc Natl Acad Sci USA
101:9849–9854.
6. Roebroeck A, Formisano E, Goebel R (2005) Mapping directed influence over the
brain using Granger causality and fMRI. Neuroimage 25:230–242.
7. Wang X, Chen Y, Bressler SL, Ding M (2007) Granger causality between multiple inter-
dependent neurobiological time series: Blockwise versus pairwise methods. Int J Neu-
ral Syst 17:71–78.
8. Baccalá LA, Sameshima K, Ballester G, do Valle AC, Timo-Iaria C (1998) Studying the
interaction between brain structures via directed coherence and Granger causality.
Appl Signal Process 5:40–48.
9. Barrett AB, et al. (2012) Granger causality analysis of steady-state electroencephalo-
graphic signals during propofol-induced anaesthesia. PLoS One 7:e29072.
10. Seth AK, Barrett AB, Barnett L (2011) Causal density and integrated information as
measures of conscious level. Philos Trans R Math Phys Eng Sci 369:3748–3767.
11. Graham S, et al. (2009) Role of medial cortical, hippocampal and striatal interactions
during cognitive set-shifting. Neuroimage 45:1359–1367.
12. Cadotte AJ, et al. (2010) Granger causality relationships between local field poten-
tials in an animal model of temporal lobe epilepsy. J Neurosci Methods 189:121–
129.
13. Wang S, et al. (2007) Revealing the dynamic causal interdependence between neural
and muscular signals in Parkinsonian tremor. J Franklin Inst 344:180–195.
14. Kopell NJ, Gritton HJ, Whittington MA, Kramer MA (2014) Beyond the connectome:
The dynome. Neuron 83:1319–1328.
15. Cannon J, et al. (2014) Neurosystems: Brain rhythms and cognitive processing. Eur J
Neurosci 39:705–719.
16. Buzsáki G, Wang XJ (2012) Mechanisms of gamma oscillations. Annu Rev Neurosci
35:203–225.
17. Buzsáki G, Watson BO (2012) Brain rhythms and neural syntax: Implications for effi-
cient coding of cognitive content and neuropsychiatric disease. Dialogues Clin Neu-
rosci 14:345–367.
18. Olbrich E, Achermann P (2005) Analysis of oscillatory patterns in the human sleep EEG
using a novel detection algorithm. J Sleep Res 14:337–346.
19. Anderson KL, Rajagovindan R, Ghacibeh GA, Meador KJ, Ding M (2010) Theta oscil-
lations mediate interaction between prefrontal cortex and medial temporal lobe in
human memory. Cereb Cortex 20:1604–1612.
20. Bollimunta A, Mo J, Schroeder CE, Ding M (2011) Neuronal mechanisms and
attentional modulation of corticothalamic alpha oscillations. J Neurosci 31:4935–
4943.
21. Geweke J (1982) Measurement of linear dependence and feedback between multiple
time series. J Am Stat Assoc 77:304–313.
22. Geweke JF (1984) Measures of conditional linear dependence and feedback between
time series. J Am Stat Assoc 79:907–915.
23. Bressler SL, Seth AK (2011) Wiener–Granger causality: A well established methodol-
ogy. Neuroimage 58:323–329.
24. Seth AK (2010) A MATLAB toolbox for Granger causal connectivity analysis. J Neurosci
Methods 186:262–273.
25. Barnett L, Seth AK (2014) The MVGC multivariate Granger causality toolbox: A new
approach to Granger-causal inference. J Neurosci Methods 223:50–68.
26. McCrorie JR, Chambers MJ (2006) Granger causality and the sampling of economic
processes. J Econom 132:311–336.
27. Guo S, Seth AK, Kendrick KM, Zhou C, Feng J (2008) Partial Granger causality – elimi-
nating exogenous inputs and latent variables. J Neurosci Methods 172:79–93.
E7072 | www.pnas.org/cgi/doi/10.1073/pnas.1704663114
key features of interest can in turn lead to erroneous assessments
of causality (2, 3), independent of the computational and inter-
pretational problems that we describe in this paper. We note that
such failures of causality analysis are frequently discussed in the
literature, often without identifying modeling misspecification as
the key point of failure.
Overall, the problems of GG causality analyzed in this paper,
the related problems of other causality methods, and the ongo-
ing confusion surrounding these methods can, in most cases, be
resolved by a careful consideration of system identification prin-
ciples (Fig. 4). Moreover, careful modeling and analysis likely
obviate the need for causality measures and their many potential
pitfalls.
ACKNOWLEDGMENTS. P.A.S. was supported by Training Grant T32
HG002295 from the National Human Genome Research Institute. P.L.P. was
supported by the National Institutes of Health Director’s New Innovator
Award DP2-OD006454.
28. Chen Y, Bressler SL, Ding M (2006) Frequency decomposition of conditional Granger
causality and application to multivariate neural field potential data. J Neurosci Meth-
ods 150:228–237.
29. Granger CWJ (1969) Investigation causal relations by econometric models and cross-
spectral methods. Econometrica 37:424–438.
30. Lütkepohl H (2005) New Introduction to Multiple Time Series Analysis (Springer,
Berlin).
31. Barnett L, Barrett AB, Seth AK (2009) Granger causality and transfer entropy are
equivalent for Gaussian variables. Phys Rev Lett 103:238701.
32. Blair RC, Karniski W (1993) An alternative method for significance testing of wave-
form difference potentials. Psychophysiology 30:518–524.
33. Kay SM (1988) Modern spectral estimation: Theory and application (Prentice Hall,
Englewood Cliffs, NJ).
34. Stokes PA (2015) Fundamental problems in Granger causality analysis of neuroscience
data. PhD thesis (Massachusetts Institute of Technology, Cambridge, MA).
35. Hu S, Liang H (2012) Causality analysis of neural connectivity: New tool and limitations
of spectral Granger causality. Neurocomputing 76:44–47.
36. Hughes SW, Crunelli V (2005) Thalamic mechanisms of EEG alpha rhythms and their
pathological implications. Neuroscientist 11:357–372.
37. Wagner AD, Koutstaal W, Maril A, Schacter DL, Buckner RL (2000) Task-specific repe-
tition priming in left inferior prefrontal cortex. Cereb Cortex 10:1176–1184.
38. Epstein CM, Adhikari BM, Gross R, Willie J, Dhamala M (2014) Application of high-
frequency Granger causality to analysis of epileptic seizures and surgical decision
making. Epilepsia 55:2038–2047.
39. Kamiński MJ, Blinowska KJ (1991) A new method of the description of the informa-
tion flow in the brain structures. Biol Cybern 65:203–210.
40. Kamiński M, Ding M, Truccolo WA, Bressler SL (2001) Evaluating causal relations in
neural systems: Granger causality, directed transfer function and statistical assessment
of significance. Biol Cybern 85:145–157.
41. Baccalá LA, Sameshima K (2001) Partial directed coherence: A new concept in neural
structure determination. Biol Cybern 84:463–474.
42. Korzeniewska A, Manczak M, Kamiński M, Blinowska KJ, Kasicki S (2003) Determi-
nation of information flow direction among brain structures by a modified directed
transfer function (dDTF) method. J Neurosci Methods 125:195–207.
43. Korzeniewska A, Crainiceanu CM, Kus R, Franaszczuk PJ, Crone NE (2008) Dynamics
of event-related causality in brain electrical activity. Hum Brain Mapp 29:1170–1192.
44. Astolfi L, et al. (2007) Comparison of different cortical connectivity estimators for
high-resolution EEG recordings. Hum Brain Mapp 28:143–157.
45. Florin E, Gross J, Pfeifer J, Fink GR, Timmermann L (2011) Reliability of multivariate
causality measures for neural data. J Neurosci Methods 198:344–358.
46. Dufour JM, Renault E (1998) Short run and long run causality in time series: Theory.
Econometrica 66:1099–1125.
47. Schreiber T (2000) Measuring information transfer. Phys Rev Lett 85:461–464.
48. Quinn CJ, Coleman TP, Kiyavash N, Hatsopoulos NG (2011) Estimating the directed
information to infer causal relationships in ensemble neural spike train recordings.
J Comput Neurosci 30:17–44.
49. Friston KJ, Harrison L, Penny W (2003) Dynamic causal modelling. Neuroimage
19:1273–1302.
50. Daunizeau J, David O, Stephan KE (2011) Dynamic causal modelling: A critical review
of the biophysical and statistical foundations. Neuroimage 58:312–322.
51. Lohmann G, Erfurth K, Müller K, Turner R (2012) Critical comments on dynamic causal
modelling. Neuroimage 59:2322–2329.
52. Friston K, Daunizeau J, Stephan KE (2013) Model selection and gobbledygook:
Response to Lohmann et al. Neuroimage 75:275–278, discussion 279–281.
53. Lohmann G, Müller K, Turner R (2013) Response to commentaries on our paper: Criti-
cal comments on dynamic causal modelling. Neuroimage 75:279–281.
54. Box GEP, Jenkins GM, Reinsel GC (2008) Time Series Analysis: Forecasting and Control
(Wiley, Hoboken, NJ).
55. Ljung L (1999) System Identification: Theory for The User (Prentice Hall PTR, Upper
Saddle River, NJ), 2nd Ed.
56. Priestley MB (1994) Spectral Analysis and Time Series (Academic, London).
Stokes and Purdon