Académique Documents
Professionnel Documents
Culture Documents
Jin-Lung Lin
Institute of Economics, Academia Sinica
Department of Economics, National Chengchi University
May ,
Abstract
Tis note reviews the defnition, distribution theory and modeling strategy of test-
ing causality. Starting with the defnition of Granger Causality, we discuss various
issues on testing causality within stationary and nonstationary systems. In addi-
tion, we cover the graphical modeling and spectral domain approaches which are
relatively unfamiliar to economists. We compile a list of Do and Dont Do on causal-
ity testing and review several empirical examples.
Introduction
Testing causality among variables is one of the most important and, yet, one of the
most difcult issues in economics. Te difculty arises fromthe non-experimental
nature of social science. For natural science, researchers can perform experiments
where all other possible causes are kept fxed except for the sole factor under in-
vestigation. By repeating the process for each possible cause, one can identify the
causal structures among factors or variables. Tere are no such luck for social sci-
ence, and economics is no exception. All diferent variables afect the same variable
simultaneously and repeated experiments under control are infeasible (experimen-
tal economics is no solution, at least, not yet).
Two most difcult challenges are :
. Correlation does not imply causality. Distinguishing between these two is by
no means an easy task.
. Tere always exist the possibility of ignored common factors. Te causal
relationship among variables might disappear when the previously ignored
common causes are considered.
While there are no satisfactory answer to these two questions and there might
never be one, philosophers and social scientists have attempted to use graphical
models to address the second issue. As for the frst issue, time series analysts look
for rescue from the unique unidirectional property of time arrow: cause precedes
efect. Based upon this concept, Clive W.J. Granger has proposed a working def-
nition of causality, using the foreseeability as a yardstick which is called Granger
causality. Tis note examines and reviews the key issues in testing causality in eco-
nomics.
In additional to this introduction, Section discusses the defnition of Granger
causality. Testing causality for stationary processes are reviewed in Section and
Section focuses on nonstationary processes. We turn to graphical models in sec-
tion . A to-do and not-to-do list is put together in Section .
Defning Granger causality
. Two assumptions
. Te future cannot cause the past. Te past causes the present or future. (How
about expectation?)
i=
i
u
ti
,
= I
l
i=
A
i
Z
ti
+ u
t
If there are only two variables, or two-group of variables, j and k, then a necessary
and sufcient condition for variable k not to Granger-cause variable j is that A
jk,i
=
, for i = , , . Te condition is good for all forecast horizon, h.
Note that for a VAR() process with dimension equal or greater than , A
jk,i
=
, for i = , , is sufcient for non-causality at h = but insufcient for h > .
Variable k might afect variable j in two or more period in the future via the efect
through other variables. For example,
y
t
y
t
y
t
.
. . .
. .
y
t
y
t
y
t
u
t
u
t
u
t
Ten,
y
; y
= A
.
.
; y
= A
.
.
.
To summarize,
. For bivariate or two groups of variables, IR analysis is equivalent to applying
Granger-causality test to VAR model;
. For testing the impact of one variable on the other within a high dimensional
( ) system, IR analysis can not be substituted by the Granger-causality test.
For example, for anVAR() process withdimensiongreater than, it does not
sufce to check the upper right-hand corner element of the coefcient matrix
in order to determine if the last variable is noncausal for the frst variable.
Test has to be based upon IR.
See Lutkepohl() and Dufor and Renault () for detailed discussion.
(B} A
(B}
A
(B} A
(B}
_ _
y
t
x
t
_ + _
u
yt
u
xt
_
= _
(B}
(B}
(B}
(B}
_ _
u
yt
u
xt
_ + _
u
yt
u
xt
_
x
t
does not Granger-cause y
t
if
(B} = or
,i
= , for i = , , . Tis
condition is equivalent to A
,i
= , for i = , , ,p. In other words, this corre-
sponds to the restrictions that all cross-lags coefcients are all zeros which can be
tested by Wald statistics.
We nowturn to determining the causal direction for bivariate VAR system. For
ease of illustration, we shall focus upon bivariate AR() process so that A
i j
(B} =
A
i j
, i, j = , as defned above. Te results can be easily generalized to AR(p) case.
Four possible causal directions between x and y are:
. Feedback, H
, x y
H
= _
A
_
. Independent, H
x y
H
= _
A
_
. x causes y but y does not cause x, H
, y ] x
H
= _
A
, x ] y
H
= _
A
_
Caines, Keng and Sethi() proposed a two-stage testing procedure for deter-
mining causal directions. In frst stage, test H
(null) against H
, H
(null) against
H
, and H
(null) against H
. If necessary, test H
(null) against H
, and H
(null)
against H
, . . . , y
n
, rank these variables
according to the decreasing order of their specifc gravity which is the inverse
of MFPE(X, y
i
};
. For eachcaused variable process, X, frst construct the optimal univariate AR
model using FPE to determine the lag order. Te, add the causal variable,
one at a time according to their causal rank and use FPE to determine the
optimal orders at each step. Finally, we get the optimal ordered univariate
multivariate AR model of X against its causal variables;
. Pool all the optimal univariate AR models above and apply the Full Infor-
mation Maximum Likelihood (FIML) method to estimate the system. Fi-
nally perform the diagnostic checking with the whole system as maintained
model.
. Causal analysis for Vector ARMA model (h = )
Let X be n vector generated by
(B}X
t
= (B}a
t
X
i
does not cause X
j
if and only if
det(
i
(z},
(j)
(z}} =
where
i
(B} is the ith column of the matrix (z} and
(j)
(z} is the matrix (z}
without its jth column.
For bivariate (two-group) case,
_
(B}
(B}
(B}
(B}
__
X
it
X
t
_ = _
(B}
(B}
(B}
(B}
__
a
t
a
t
_
Ten, X
i
does not cause X
j
if and only if
(z}
(z}
(z}
(z} =
If n
= n
= , Ten, X
i
does not cause X
j
if and only if
(z}
(z}
(z}
(z} =
General testing procedures is:
. Build a multivariate ARMA model for X
t
,
. Derive the noncausality conditions in term of AR and MA parameters, say
R
j
(
l
} = , j = , . . . , K
. Choose a test criterion, Wald, LM or LR test.
Let
T(
l
} = (
R
j
(B}
l
|
l
}
kk
Let V(
l
} be the asymptotic covariance matrix of
N(
l
=
l
}. Ten the Wald
and LR test statistics are:
W
= NR(
l
}
|T(
l
}
V(
l
}T(
l
}|
R(
l
},
LR
= (L(
, X} L(
, X}}
where
is the MLE of under the constraint of noncausality.
To illustrate, let X
t
be a invertible -dimensional ARMA(,) model.
_
B
__
X
t
X
t
_ = _
B
__
a
t
a
t
_
X
if and only if
(z}
(z}
(z}
(z} =
(
}z + (
}z
= ,
=
For the vector,
l
= (
, the matrix
T(
l
} =
.
Remarks:
Te conditions are weaker than
= is sufcient condi-
tion and
= , &
.
Let H
= ;
H
=
H
= , and
=
Ten, H
=
H
, H
, H
.
Testing procedures:
. Test H
at level
. If H
is rejected, then H
is rejected. Stop.
. If H
at level
. If H
is not rejected, H
cannot be
rejected. Stop
. If H
is rejected, test
H
= at level
. If
H
is rejected, then H
is
also rejected. If
H
i=
i
Y
ti
+ D
t
+ U
t
Y
t
= C
t
i=
(U
t
+ D
i
} + C
(B}(U
t
+ D
t
} + P
A
p
(} = =
C =
with
n
, n
, n
i=
J
i
X
ti
+ u
t
Te null hypothesis of X
J
,
= J
,
= = J
k,
=
Let F
LS
be the Wald statistics for testing H
.
. If X
t
has unit root and is not cointegrated, F
LS
converges to a limiting distri-
bution which is the sum of
then F
LS
k
. , More specif-
cally, let A = (A
, A
, A
} = n
, ie. all X
(B}X
t
+ A
X
t
+ u
t
where , A are respectively the loading matrix and cointegration vector. Partition
, A conforming to X
, X
, X
. Ten, if rank(A
} = n
or rank(
} = n
, F
ML
k
. In other words, testing with ECM the usual asymptotic distribution hold
when there are sufcient cointegrations or sufcient loading vector.
Remark: Te Johansen test seems to assume sufcient cointegration or suf-
cient loading vector.
Toda and Yamamoto () proposed a test of causality without pretesting coin-
tegration. For an VAR(p) process and each series is at most I(k), then estimate the
augmented VAR(p+k) process even the last k coefcient matrix is zero.
X
t
= A
X
t
+ + A
p
X
tk
+ + A
p+k
X
tpk
+ U
t
and perform the usual Wald test A
kj,i
= , i = , , p. Te test statistics is
asymptotical
, A
and A
are:
A
; A
; A
Several search algorithms are available and the PC algorithm seems to be the
most popular one (see Pearl (), and Spirtes, Glymour and Scheines () for
the details). In this paper, we adopt the PC algorithm and outline the main algo-
rithm as shown below. First, we start with a graph in which each variable is con-
nected by an edge with every other variable. We then compute the unconditional
correlation between each pair of variables and remove the edge for the insignifcant
pairs. We then compute the -th order conditional correlation between each pair
of variables and eliminate the edge between the insignifcant ones. We repeat the
procedure to compute the i-th order conditional correlation until i = N-, where
N is the number of variables under investigation. Fishers z statistic is used in the
signifcance test:
z(i, j|K} = ](n |K| }
()
l n(
| + r|i, j|K||
| r|i, j|K||
}
where r(|i, j|K|} denotes conditional correlation between variables, which i and j
being conditional upon the K variables, and |K| the number of series for K.
Under some regularity conditions, z approximates the standard normal distri-
bution. Next, for each pair of variables (Y, Z} that are unconnected by a direct
edge but are connected through an undirected edge through a third variable X, we
assign Y X Z if and only if the conditional correlations of Y and Z condi-
tional upon all possible variable combinations with the presence of the X variable
are nonzero. We thenrepeat the above process until all possible cases are exhausted.
If X Z, Z Y and X and Y are not directly connected, we assign Z Y. If there
(B}
(B}
(B}
(B}
_ _
e
xt
e
yt
_
Rewrite the above as
_
x
t
y
t
_ = _
(B}
(B}
(B}
(B}
_ _
e
xt
e
yt
_
where
_
(B}
(B}
(B}
(B}
_ = _
(B}
(B}
(B}
(B}
_ _
_
and
_
e
xt
e
yt
_ = _
_ _
e
xt
e
yt
_
f
x
(w} =
(z}|
+ |
(z}|
}}
where z = e
iw
.
Hosoyas measure of one-way causality is defned as:
M
yx
(w} = l og|
f
x
(w}
]|
(z}|
|
= l og| +
|
(z}|
}
|
(z} +
(z}
|
|
u
t
+ a
(B}x
t
+ b
(B}y
t
+ e
xt
y
t
=
u
t
+ a
(B}x
t
+ b
(B}y
t
+ e
yt
_
D(B}x
t
D(B}y
t
_ = _
( B}( b
B}
B + b
B( B}
( B}a
AB
AB + ( a
B}( B}
_ _
e
xt
e
yt
_
where D(B} arises from matrix inversion. Ten,
M
yx
(w} = l og| +
|
+ b
( z}|
}
|( z}( z b
} +
+ b
( z}}||
|
where z = e
iw
.
Sofwares
Again, the usual disclaim applies. Tey are subjective. Your choices might be as
good as mine. See Lin() for a detailed account.
. Impulse responses: Reduced form and structural form
VAR.SRC/RATS by Norman Morin
SVAR.SRC/RATS by Antonio Lanzarotti and Mario Seghelini
VAR/View/Impulse/Eviews
FinMetrics/Splus
. Cointegration:
CATS/RATS
COINT/GAUSS
VAR/Eviews
urca/R
FinMetrics/Splus
. Ofen graph the residuals and check for abnormality and outliers.
. Be aware of seasonality for data not seasonally adjusted.
. Apply the Wald test within the Johansen framework where one can test for
hypothesis on long- and short- run causality.
. When you employ several time series methods or analyze several similar
models, be careful about the consistency among them.
. Always watch for balance between explained and explanatory variables in
regression analysis. For example, if the dependent variable has a time-trend
but explanatory variables are limited between and , then the regression
coefcient can never be a fxed constant. Be careful about mixing I() and
I() variables in one equation.
. For VAR, number of parameters grows rapidly with number of variables and
lags. Removing the insignifcant parameters to achieve estimation efciency
is strongly recommended. Te resulting IR will be more accurate.
Empirical Examples
. Evaluating the efectiveness of interest rate policy in Taiwan: an impulse re-
sponses analysis
Lin(a).
. Modelling information fow among four stock markets in China
Lin and Wu ().
. Causality between export expansion and manufacturing growth (if time per-
mits)
Liang, Chou and Lin ().
Reference Books:
. Banerjee, Anindya and David F. Hendry eds. (), Te Econometrics of
Economic Policy, Oxford: Blackwell Publishers
. Hamilton, James D. Time Series Analysis, NewJersey: Princeton University
Press,