Documente Academic
Documente Profesional
Documente Cultură
Modelling
Approach to Cointegration
Analysis
M. Hashem Pesaran
Trinity College, Cambridge, ngland
!ongcheol "hin
Department o# Applied conomics, $niversity o# Cambridge, ngland
%irst &ersion' %ebruary, ())*, +evised' ,anuary,
())-
Abstract
This paper e.amines the use o# autoregressive distributed lag /A+DL0 mod1
els #or the analysis o# long1run relations 2hen the underlying variables are 3/(0.
3t sho2s that a#ter appropriate augmentation o# the order o# the A+DL model,
the 4L" estimators o# the short1run parameters are
p
T 1consistent 2ith the as1
ymptotically singular covariance matri., and the A+DL1based estimators o# the
long1run coe5cients are super1consistent, and valid in#erences on the long1run pa1
rameters can be made using standard normal asymptotic theory. The paper also
e.amines the relationship bet2een the A+DL procedure and the #ully modi6ed
4L" approach o# Phillips and Hansen to estimation o# cointegrating relations, and
compares the small sample per#ormance o# these t2o approaches via Monte Carlo
e.periments. These results provide strong evidence in #avour o# a rehabilitation
o# the traditional A+DL approach to time series econometric modelling. The
A+DL approach has the additional advantage o# yielding consistent estimates o#
the long1run coe5cients that are asymptotically normal irrespective o# 2hether
the underlying regressors are 3/(0 or 3/70.
,L Classi6cations' C(8, C(9, C(*, C88.
:ey ;ords' Autoregressive distributed lag model, Cointegration, 3/(0 and 3/70
regressors, Model selection, Monte Carlo simulation.
This is a revised version o# a paper presented at the "ymposium at the Centennial o# +agnar
%risch, The <or2egian Academy o# "cience and Letters, 4slo, March 91*, ())*. ;e are grate#ul
to Peter =os2i>?, Clive @ranger, Alberto Holly, :yung "o 3m, =rendan McCabe, "teve "atchell,
+ichard "mith, +on "mith and an anonymous re#eree #or help#ul comments. Partial 6nancial
support #rom the "+C /@rant <o. +777899A7B0 and the 3saac <e2ton Trust o# Trinity
College, Cambridge is grate#ully ac?no2ledged.
( C
(. 3<T+4D$CT34<
conometric analysis o# long1run relations has been the #ocus o# much
theoreti1 cal and empirical research in economics. 3n the case 2here the
variables in the long1run relation o# interest are trend stationary, the general
practice has been to de1trend the series and to model the de1trended series as
stationary distributed lag or autoregressive distributed lag /A+DL0 models.
stimation and in#erence concerning the long1run properties o# the model are
then carried out using stan1 dard asymptotic normal theory. /%or a
comprehensive revie2 o# this literature see Hendry, Pagan and "argan /()BD0
and ;ic?ens and =reusch /()BB00. The analysis becomes more complicated
2hen the variables are dierence1stationary, or integrated o# order ( /3/(0 #or
short0. The recent literature on cointegration is concerned 2ith the analysis o#
the long run relations bet2een 3/(0 variables, and its basic premise is, at least
implicitly, that in the presence o# 3/(0 variables the traditional A+DL approach
is no longer applicable. ConseEuently, a large number o# alternative estimation
and hypothesis testing procedures have been speci6cally developed #or the
analysis o# 3/(0 variables. /"ee the pioneering 2or? o# ngle and @ranger /()B-0,
,ohansen /())(0, Phillips /())(0, Phillips and Hansen /())70 and Phillips and
Loretan /())(0.0
3n this paper 2e re1e.amine the use o# the traditional A+DL approach #or the
analysis o# long run relations 2hen the underlying variables are 3/(0. ;e consider
the #ollo2ing general A+DL/pF E0 model'
p
y
t
G H
7
I H
(
t I
J
K
i
y
t
i
I L
7
.
EC(
I
J
L
7
5. I u F /(.(0
C
iG(
t
i
iG7
tCi t
5.
t
G P
(
5.
tC(
I P
8
5.
tC8
I 5 5 5 I P
s
5.
tCs
I M
t
F /(.80
2here .
t
is the ?1dimensional 3/(0 variables that are not cointegrated among
themselves, u
t
and M
t
are serially uncorrelated disturbances 2ith Nero means and
constant variance1covariances, and P
i
are ? O ? coe5cient matrices such that
the vector autoregressive process in 5.
t
is stable. ;e also assume that the
roots o#
P
p
iG(
K
i
N
i
G 7 all #all outside the unit circle and there e.ists a stable uniEue
long1run relationship bet2een y
t
and .
t
.
;e consider the problem o# consistent estimation o# the parameters o# the
A+DL model both 2hen u
t
and M
t
are uncorrelated, and 2hen they are corre1
lated. 3n the #ormer case 2e 2ill sho2 that the 4L" estimators o# the short1
run
parameters, H
7
, H
(
, L, L
F '''F
L
i
G 7 #or i W E, and P
i
G 7 #or i W s. 3n
this augmented speci6cation V
t
and M
t
are uncorrelated and the results stated
above 2ill be directly applicable to the 4L" estimators o# the short1run and
long1run parameters o# /(.90. 4nce again traditional methods o# estimation and
in#erence, originally developed #or trend1stationary variables, are applicable to
6rst1dierence stationary variables. The estimation o# the short1run eects still
reEuires an e.plicit modelling o# the contemporaneous dependence bet2een u
t
and M
t
. 3n practice, an appropriate choice o# the order o# the A+DL model is
crucial #or valid in#erence. =ut once this is done, estimation o# the long1run
parameters and computation o# valid standard errors #or the resultant
estimators can be carried out either by the 4L" method, using the so1called
XdeltaY method /51method0 to compute the standard errors, or by the
=e2elyZs /()-)0 regression approach. These t2o procedures yield identical
results and a choice bet2een them is only a matter o# computational
convenience.
The use o# the A+DL estimation procedure is directly comparable to the
semi1 parametric, #ully1modi6ed 4L" approach o# Phillips and Hansen /())70
to esti1 mation o# cointegrating relations. 3n the static #ormulation o# the
cointegrating regression,
y
t
G [ I Rt I S
7
.
t
I v
t
F /(.D0
P8Q
t
2here 5.
t
G e
t
, and \
t
G /v
t
F e
7
0
7
#ollo2s a general linear stationary process, the
9
4L" estimators o# R and S are T
8
1 and T 1consistent, but in general the asymp1
totic distribution o# the 4L" estimator o# S involves the unit1root
distribution as 2ell as the second1order bias in the presence o# the
contemporaneous correla1 tions that may e.ist bet2een v
t
and e
t
. There#ore, the
6nite sample per#ormance o# the 4L" estimator is poor and in addition, due
to the nuisance parameter dependencies, in#erence on S using the usual t1
tests in the 4L" regression o# /(.D0 is invalid. To overcome these problems
Phillips and Hansen /())70 have suggested the #ully1modi6ed 4L" estimation
procedure that asymptotically ta?es account o# these correlations in a semi1
parametric manner, in the sense that the #ully1modi6ed estimators have the
@aussian mi.ture normal distribution asymp1 totically, and in#erences on the long
run parameters using the t1test based on the limiting distribution o# the #ully1
modi6ed estimator is valid.
The A+DL1based approach to estimation and in#erence, and the #ully1modi6ed
4L" procedure are both asymptotically valid 2hen the regressors are 3/(0, and a
choice bet2een them has to be made on the basis o# their small sample properties
and computational convenience. To e.amine the small sample per#ormance o# the
t2o estimators 2e have carried out a number o# Monte Carlo e.periments.
"ince in practice the XtrueY orders o# the A+DL/pF m0 model are rarely ?no2n a
priori, in the Monte Carlo e.periments 2e also consider a t2o1step strategy
2hereby p and m are 6rst selected /estimated0 using either the A?ai?e
3n#ormation Criterion /A3C0, or the "ch2arN =ayesian Criterion /"C0, and then
the long1run coe5cients and their standard errors are estimated using the
A+DL model selected in the
6rst step. ;e re#er to these estimators as A+DL1A3C and A+DL1"C. The main
6ndings #rom these e.periments are as
#ollo2s'
/i0 The A+DL1A3C and the A+DL1"C estimators have very similar small1sample
per#ormances, 2ith the A+DL1"C per#orming slightly better in the ma>ority
o# the e.periments. This may re]ect the #act that the "ch2artN criterion
is a consistent model selection criterion 2hile A?ai?e is not.
/ii0 The A+DL test statistics that are computed using the 51method /or eEuiv1
alently by means o# the so1called =e2leyZs regression0, generally per#orm
much better in small samples than the test statistics computed using
the asymptotic #ormula that e.plicitly ta?es account o# the #act that the
regres1 sors are 3/(0.
/iii0 The A+DL1"C procedure 2hen combined 2ith the 51method o# comput1
ing the standard errors o# the long1run parameters generally dominates the
Phillips1Hansen estimator in small samples. This is in particular true o# the
siNe1po2er per#ormance o# the tests on the long1run parameter.
P9Q
u
/iv0 The Monte Carlo results point strongly in #avor o# the t2o1step estimation
procedure, and this strategy seems to 2or? even 2hen the model under con1
sideration has endogenous regressors, irrespective o# 2hether the regressors
are 3/(0 or 3/70.
(
The plan o# the paper is as #ollo2s' "ection 8 e.amines the asymptotic prop1
erties o# the 4L" estimators in the conte.t o# a simple autoregressive model 2ith
a linear deterministic trend and the ?1dimensional strictly e.ogenous 3/(0 regres1
sors. "ection 9 considers a more general A+DL model, allo2ing #or residual serial
correlations and possible endogeneity o# the 3/(0 regressors, and develops the re1
sultant asymptotic theory. 3n "ection D the A+DL1based approach is compared
to the cointegration1based approach o# Phillips and Hansen /())70. "ection *
reports and discusses the results o# Monte Carlo e.periments. "ome concluding
remar?s are presented in "ection A. Mathematical proo#s are provided in an
Appendi..
8. The Lagged Dependent &ariable Model 2ith the
Deter1 ministic Trend and .ogenous 3/(0 +egressors
3nitially 2e consider the simple A+DL/(,70 model containing 3/(0 regressors and
a linear deterministic trend,
K/L0y
t
G H
7
I H
(
t I L
7
.
t
I u
t
F t G (F '''F T F /8.(0
2here y
t
is a scalar, K/L0 G ( C KL, 2ith L being the one period lag operator, .
t
is a ? O ( vector o# regressors assumed to be integrated o# order ('
8
.
t
G .
tC(
I e
t
F /8.80
and L is a ? O ( vector o# un?no2n parameters. "uppose that the #ollo2ing
assumptions hold'
/A(0 The scalar disturbance term, u
t
, in /8.(0 is iid/7F ^
8
0,
(
The case 2here the regressors are 3/(0 and cointegrated among themselves presents ad1
ditional identi6cation problems and is best analyNed in the conte.t o# a system o# long1run
structural eEuations. 4n this see Pesaran and "hin /())*0.
8
"peci6cations /8.(0 and /8.80 can easily be adapted to allo2 #or inclusion o# a dri#t term in
the .
t
process. Consider, #or e.ample, the process 5.
t
G [
.
I e
t
F and note that it can also be
2ritten as .
t
G [
.
t I ._
t
, 2here 5._
t
G e
t
' There#ore, substituting .
t
in /8.(0 2e have
K/L0y
t
G H
7
I /H
(
I L
7
[
.
0t I L
7
._
t
I
u
t
F
2here ._
t
#ollo2s an 3/(0 process 2ithout a dri#t.
PDQ
C t t
`
C
/A80 The ?1dimensional vector, e
t
, in /8.80 has a general linear multivariate
stationary process,
/A90 u
t
and e
t
are uncorrelated #or all leads and lags such that .
t
is strictly
e.ogenous 2ith respect to u
t
,
/AD0 The 3/(0 regressors, .
t
, are not cointegrated among themselves, and
/A*0 >K> a (, so that the model is dynamically stable, and a long1run relationship
bet2een y
t
and .
t
e.ists.
9
;e shall distinguish bet2een t2o types o# parameters, the parameters
capturing the short1run dynamics /H
7
F H
(
F L and K0, and the long run
parameters on the trended regressors, t and .
t
, de6ned by
R G
H
(
( C K
F S G
L
( C
K
' /8.90
Applying the decomposition ( C KL G /( C K0 I K/( C L0 to /8.(0, y
t
can
be e.pressed as
2here
y
t
G [ I Rt I S
7
.
t
I v
t
F /8.D0
and
[ G
H
7
( C
K
(
S
K
b
RF
( C
K
(
v
t
G
J
K
i
u
t i
C K
J
K
i
S
7
e ' C
iG7 iG7
tCi
%rom /8.(0 and /8.D0 it is clear that y
t
and .
t
are individually 3/(0, but must be
cointegrated #or /8.(0 to be meaning#ul.
D
"imilarly, 2e obtain
y
tC(
G [
(
I Rt I S
7
.
t
I `
t
F /8.*0
2here [
(
G [ C R, `
t
G v
t (
C S
7
e , and ` is an 3/70 process 2ith variance ^
8
.
4ur main aim is to derive the asymptotic properties o# the 4L" estimators o#
the short1run as 2ell as the long1run parameters in the conte.t o# the A+DL/(,70
9
Tests o# the e.istence o# long1run relationships bet2een y
t
and .
t
, 2hen it is not ?no2n a
priori 2hether .
t
are 3/70 or 3/(0, are discussed in Pesaran, "hin and "mith /())A0.
D
A relationship bet2een 3/(0 variables is said to be Xstochastically cointegratedY i# it is trend
stationary, 2hile Xdeterministic cointegrationY re#ers to the case 2here the cointegrating relation
is level stationary. %or a discussion o# these t2o types o# cointegrating relations see Par? /())80.
P*Q
^
^
8
uT
/P
uT T
8
c
T
model, /8.(0. %or e.positional convenience, 2e trans#orm /8.(0 to the partitioned
regression model in the matri. #orm as,
y
T
G c
T
b I y
T C(
K I u
T
F /8.A0
2here y
T
G /y
(
F '''F y
T
0
7
, y
T C(
G /y
7
F '''F y
T C(
0
7
, d
T
G /(F '''F (0
7
, t
T
G /(F '''F T
0
7
, J
T
G /.
(
F '''F .
T
0
7
, c
T
G /d
T
F t
T
F J
T
0, u
T
G /u
(
F '''F u
T
0
7
, and b G /H
7
F H
(
F
L
7
0
7
. "ince our main interest is in the long1run coe5cients on trended regressors, t
and .
t
, 2e also partition
c
T
G /d
T
F "
T
0F "
T
G /t
T
F J
T
0F b G
S
H
7
c
b
F c G
S
H
(
b
F
L
2here the dimensions o# c
T
, "
T
, b and c are T O /? I 80, T O /? I (0, /? I 80 O
(
and /? I (0 O (, respectively.
Theorem 8.(. $nder the assumptions /A(0 1 A/*0, the 4L" estimators o# K and
c G /H
(
F L
7
0
7
in /8.A0, denoted by K
e
T
and ec
T
, respectively, are
p
T 1consistent,
and
have the #ollo2ing asymptotic distributions'
p
a
f
^
8
^
T /K
e
T
C K0 \
<
7F
u
`
F /8.-0
p
a
f
^
8
^
T /ec
T
C c0 \
<
7F
u
WW
7
`
F /8.B0
2here W G /RF S
7
0
7
is a /? I (0 O ( vector o# the long run parameters on trended
regressors, t and .
t
, and ran?/WW
7
0 G (. 3n addition, the 4L" estimator o#
H
7
in /8.A0, denoted by He
7T
, is also
p
T 1consistent, but has the mi.ture
normal
distribution. De6ning h G /b
7
F K0
7
and P
c
T
G /c
T
F y
T C(
0, and denoting the
4L"
estimator o# h by h
e
T
, the covariance matri. o# h
e
T
can be consistently
estimated
by
&
e
/h
e
T
0 G ^e
8 7
P
c
T
0
C(
F
2here
^e
8
G T
C(
/y
T
C
P
c
h
e
T
0
7
/y
T
C
P
c
T
h
e
T
0, and &
e
/h
e
T
0 is asymptotically
sin1
gular 2ith ran? eEual to 8.
Theorem 8.( sho2s that despite the presence o# stochastic and deterministic trends
in the A+DL model, the 4L" estimators o# the short1run parameters are
p
T 1 consistent.
*
The second and more important 6nding is that the 4L" estimators
*
"imilar results can also be obtained in the case o# regressors 2ith higher order trend terms
such as t
8
F t
9
F '''F or 3/80, 3/90, ..., variables.
PAQ
T
C W G
/( K 0
9
e
u
T
9
R
T
L
L
L
(
9
o# the coe5cients on the trended regressors, H
(
and L, in /8.(0 are asymptoti1
cally per#ectly collinear 2ith the 4L" estimator o# the coe5cient on the lagged
dependent variable, KF namely,
T
n o p
/ec
T
C c0 I
W/K
e
C K
0
G o
p
/(0' /8.)0
4ne interesting implication o# this result is that the t1statistics #or testing the
sig1 ni6cance o# individual impact coe5cients on the 3/(0 regressors are
asymptotically
A
eEuivalent, namely t
e
i
C t
L
e
>
G o
p
/(0 #or i G >, and
t
e
i
C t
He
(
G o
p
/(0.
%urthermore,
t
e
i
C t
/(CK
e
0
G o
p
/(0. +elation /8.)0 in con>unction
2ith
W
e
/ec
T
C c0 I W/K
e
T
C
K0
C
e
T
F /8.(70
also yields an important result #amiliar #rom the cointegration literature, 2hich
2e set out in the #ollo2ing theorem'
Theorem 8.8. $nder assumptions /A(0 1 /A*0, the A+DL1based estimators o#
the long1run parameters, given by
e
R
T
G He
(T
G/( C K
e
T
0, and S
e
T
G L
e
T
G/( C K
e
T
0, converge to their true values R and S, respectively, at the rates, T
8
and T . Also
9
asymptotically, T
8
/R
T
CR0 and T /S
e
T
CS0 have the /mi.ture0 normal
distributions,
and there#ore,
g
8 C(
e
a
f
^
8
^
"
_
T
D
"
T
/W
T
C W0 \
<
7F
/( C
3
?I(
K0
8
F /8.((0
2here W
e
T
G /
e
R
T
F S
e
7
0
7
,
g
_
G D
"
"
7
H
T
"
T
D
"
F "
T
G /t
T
F J
T
0, H
T
G 3
T
C
T "
T
T T T
( 7 C C(
d
T
/d
7
d
T
0
C
d
T
F and D
"
T
G Diag/T
8
F T 3
?
0'
The 6nding that the estimator o# S is T 1consistent is ?no2n as the Xsuper1
consistencyY property in the cointegration literature. "ince the limiting
distri1
butions o# T
8
/
e
R
T
C R0 and T /S
e
T
C S0 are /mi.ture0 normal, optimal t2o1
sided
in#erences concerning and S are possible. <otice also that the covariance matri.
o# the estimator o# W simply depends on the inverse o# the /scaled0 demeaned
data matri. and the spectral density at Nero #reEuency o# /( C KL0
C(
u
t
,
namely
^
8 8
u
G/( C K0 . 4nce again, this 6nding is in line 2ith the results already #amiliar
L
#rom the cointegration literature. /"ee "ection D #or #urther discussions.0
A
%or large enough T 2e have t
e
i
U /( C K0 /^
`
G^
u
0 ' This e.plains the relatively lo2 t1ratios
o#ten obtained #or short1run coe5cients in A+DL regressions 2ith 3/(0 variables, especially 2hen
K is close to unity.
P-Q
^
e
T
i
T ii
Hypothesis testing on the general linear restrictions involving the ? I ( di1
mensional long1run parameter vector, W, can be carried out in the usual manner.
Consider the g linear restrictions on W,
+W G rF
2here + is a g O /? I (0 matri. and r is a g O ( vector o# ?no2n constants. These
restrictions can be tested using the ;ald statistic,
C(
; G /+W
e
T
C r0
7
n
+Cov/W
e
T
0
+
7
o
/+W
e
T
C r0 /8.(80
G /+W
e
T
C
r0
7
/
/( C
K
e
T
8
uT
0
8
0
/"
7
H
T
"
T
0 /+W
e
T
C r0'
4# special interest is the t1statistic on the individual coe5cients given by
W
e
iT
C
W
i
t
i
G
se
i
F i G (F '''F ? I (F /8.(90
2here the standard error o# the i1th coe5cient is consistently estimated by
se
i
G
s
^e
8
uT
/"
7
H
T
"
T
0
C(
F
/( C K
e
T
0
8
and /"
7
H
T
"
T
0
C(
denotes the i1th diagonal element o# /"
7
H
T
"
T
0
C(
. =y Theorem
T ii T
8.8, the ;ald statistic in /8.(80 #ollo2s the asymptotic h
8
distribution 2ith g
degrees o# #reedom, and t
8
in /8.(90 is distributed asymptotically as a h
8
variate
2ith one degree o# #reedom.
3t is 2orth noting that the results in Theorem 8.8 eEually apply to the purely
autoregressive model 2ith deterministic trend,
y
t
G H
7
I H
(
t I Ky
tC(
I u
t
F t G (F '''F T F /8.(D0
and to the A+DL/(,70 model 2ithout a deterministic trend,
y
t
G H
7
I L
7
.
t
I Ky
tC(
I u
t
F t G (F '''F T ' /8.(*0
%or completeness the asymptotic results #or these models are summariNed in
The1
orems 8.9 and 8.D.
Theorem 8.9. $nder the assumptions /A(0 and /A*0, the 4L" estimators o#
H
7
F H
(
and K in /8.(D0, denoted by He
7T
, He
(T
, and K
e
T
,are all
p
T 1consistent,
and asymptotically normally distributed. 3n addition,
p
T /He
(T
C H
(
0 and
p
T
/K
e
T
C K0
PBQ
u
u
/( C K
e
T
0
8
P
T
u
(
(
(
are per#ectly collinear asymptotically and the covariance matri. o# /He
7T
, He
(T
,
K
e
T
0 is asymptotically singular 2ith ran? eEual to 8. %urthermore, the estimator
o# the long run parameter R, computed by He
(T
G/( C K
e
T
0, has the #ollo2ing
asymptotic distribution'
9
a
f
(8^
8
^
T
8
/
e
R
T
C R0 \
<
7F
u
/( C
K0
8
' /8.(A0
Theorem 8.D. $nder assumptions /A(0 1 /A*0, the 4L" estimators o# H
7
F L and
K in /8.(*0, denoted by He
7T
, L
e
T
, and K
e
T
are
p
T 1consistent, and have the
asymp1 totic /mi.ture0 normal distributions. 3n addition,
p
T /He
(T
C H
(
0 and
p
T /K
e
T
C K0
are per#ectly collinear asymptotically and so the covariance matri. o# /He
7T
,
L
e
T
, K
e
T
0 is asymptotically singular 2ith ran? eEual to 8. %urthermore, the
estimator o# the long run parameter S, given by S
e
T
G L
e
T
G/( C K
e
T
0F has the
mi.ture normal
distribution asymptotically, and
g
8
e
a
f
^
8
^
J
_
T
T /S
T
C S0 \
<
8 7
7F
/( C
3
?
K0
8
F /8.(-0
2here g
J
_
T
G T
C
J
T
H
T
J
T
'
=e#ore considering a more general speci6cation o# the A+DL model, 2e e.amine
the relation bet2een the standard errors o# the estimator o# the long1run para1
meter, S, obtained #rom our asymptotic results and the standard errors obtained
#rom the so called XdeltaY method /51method #or short0. %or ease o# e.position
2e consider the simple model /8.(*0, and 2ithout loss o# generality #ocus on the
case 2here .
t
is a scalar /i.e., ? G (0. %rom Theorem 8.D 2e have
g
8
e
( M
T
i
8
J
8
e
a
f
^
8
^
J
_
T
T /S
T
C S0 G
8
P
T
tG(
/.
t
C
.[0
8
/S
T
C S0 \
<
C(
P
T
7F
/( C
-
K0
8
F /8.(B0
2here g
J
_
T
G T
C
tG(
/.
t
C
.[0
and .[ G
T
tG(
.
t
. Hence a consistent
estimator o# the variance o#
e
S
T
is given by
^e
8
(
&
e
/
e
S
T
0 G
uT
' /8.()0
T
tG(
/.
t
C .[0
8
-
3n the case 2here .
t
is 3/70 2e have the same asymptotic result given by /8.(B0F that is,
since T
C(
.
7
H
T
.
T
G 4
p
/(0 and
p
T /
e
S
T
C S0 G 4
p
/(0, hence
/T
C(
.
7
8
p
( M
T
i
8
J
8
a
f
^
8
^
T
H
T
.
T
0 T /
e
S
T
C S0
G
tG(
/.
t
C
.[0
/
e
S
T
C S0 \
<
7F
/( C
K0
8
'
P)Q
7
^
e
T
C
^
e
t
C S
0
t
C S
0
._ 0
/
T
uT
8
P P
T
/
8
The computation o# the variance o#
e
S
T
by the 51method involves appro.imating
L
e
T
T
G g/j
e
T
0 G F
( C K
e
T
by a linear #unction o# j
e
T
G /L
e
T
F K
e
T
0
7
, and then appro.imating the variance o#
e
S
T
by the variance o# the resulting linear #unction. Denoting the estimator o# the
variance o#
e
S
T
by &
e
5
/
e
S
T
0, 2e have
&
e
5
/
e
S
T
0
G
k
l g /j
e
T
0
m
&
e
/j
e
T
0
k
l g /j
e
T
0
m
lj
e
T
M
( L
e
T
lj
e
T
i
8
(
9
(CK
e
T
G F
( C K
e
T
/( C K
e
T
0
8
8
/+
7
H
T
+
T
0
C(
A
D
-
F
L
e
T
*
/(CK
e
T
0
8
2here +
T
G /.
T
F y
T C(
0. A#ter some algebra &
e
5
/
e
S
T
0 can be e.pressed as
^e
8
h i
(
` P
/y
t (
C y[0
8
C
P
/y C y[0
/.
C .
[0
W `
(
W
&
e
5
/
e
S
T
0 G
uT
(F
e
S
T
C tC( t
F
/( C K
e
T
0
8
D
T
C
P
/y
tC(
C y[0/.
t
C .[0
P
/.
t
C
.[0
T
/8.870
2here the bar over the variable denotes the sample mean, and
D
T
G
M
T
J
/.
t
C
.[0
8
tG(
i M
T
J
/y
t
tG(
i
(
C
y[0
8
M
T
C
J
/y
tG(
tC(
C y[0
/.
t
i
8
C .[0 '
$sing /8.*0, recalling that R G 7 and de6ning y_
tC(
G y
tC(
C y[F ._
t
G .
t
C .[ and
`_
t
G `
t
C `[, 2e also have
y_
tC(
G S._
t
I `_
t
F /8.8(0
2here `_
t
#ollo2s a general linear stationary process. "ubstituting this result in
/8.870, 2e obtain
8
&
e
5
/ 0 G
uT
P
T
tG(
`_
8
I
/
8
T
T
tG(
._
8
C
8/
T
T
tG(
._
t
`_
t
' /8.880
/( C K
e
T
0
8
P
T
8
P
tG(
t
tG(
`_
t
0 C
/
P
T
tG(
._
t
`_
t
0
8
"ince `_
t
is 3/70 and ._
t
is 3/(0, using the results #amiliar in the literature /see,
#or e.ample, Phillips and Durlau# /()BA00, 2e have
T T T
T
C(
J
`_
8
G 4
p
/(0F T
C8
J
._
8
G 4
p
/(0F T
C(
J
._
t
`_
t
G 4
p
/(0'
t
tG (
t
tG( tG(
P(7Q
/( C K0
8
T
C8
P
T
tG(
u
Also #rom the result o# Theorem 8.D 2e ?no2 that T /
e
S
T
C S0 G 4
p
/(0.
Hence, ta?ing probability limits o# the right hand side o# /8.880 as T m (, 2e
have
^
8
(
&
e
5
/
T
0 G
u
I o
p
/(0'
tG(
/.
t
C .[0
8
There#ore, the standard error #or the estimator o# the long run parameter, S,
obtained using the 51method is asymptotically the same as that given by /8.()0,
2hich 2as derived assuming that .
t
is 3/(0. 4ne important advantage o# the
variance estimator obtained by the 51method over the asymptotic #ormula /8.()0
lies in the #act that it is asymptotically valid irrespective o# 2hether .
t
is 3/(0 or
3/70, 2hile the latter estimator is valid only i# .
t
is 3/(0.
The t2o variance estimators clearly dier in 6nite samples. <otice that /
P
T
._
t
`_
t
0
8
is asymptotically negligible compared to other terms in /8.880, but it may not be
negligible in 6nite samples, especially 2hen ._
t
and `_
t
are correlated. %or a
com1 parison o# the small sample properties o# the t2o variance estimators
see the Monte Carlo results reported in "ection *.
9. @eneral Autoregressive Distributed Lag Models 2ith
a
Deterministic Trend and 3/(0 +egressors
"o #ar 2e have derived the estimation and asymptotic results #or the simple
A+DL/(,70 model under the t2o strong assumptions /A(0 and /A90. These as1
sumptions, ho2ever, are too restrictive in the time series analysis, and so the
estimation procedures developed in "ection 8 are not e.pected to be robust to
the violation o# these assumptions, because the limiting distributions o# the 4L"
estimators 2ould then be inconsistent andnor depend on nuisance parameters.
;e 6rst rela. the assumption /A(0 and allo2 #or the possibility o# the error
process in /8.(0 to be serially correlated. To deal 2ith this serial correlation 2e
consider the A+DL/pF E0 model,
B
K/L0y
t
G H
7
I H
(
t I L
7
/L0.
t
I u
t
F /9.(0
2here K/L0 G ( C P
p
K L
>
, and L/L0 G
P
E
L L
>
, and assume
> G( > >G7 >
/A(0
7
The scalar disturbance, u
t
F in the A+DL/pF E0 model /9.(0 is iid/7F ^
8
0.
B
%or convenience 2e use the same notation u
t
#or the disturbance terms in /8.(0 and /9.(0. 3n
practice the order o# the lag polynomials operating on dierent elements o# .
t
could be dierent.
=ut this does not aect the asymptotic theory presented belo2.
P((Q
>G7
K
>
G
C C C
>
G
$sing the decomposition L/L0 G L/(0 I /( C L0L
/L0, 2here L/(0 G
P
E
L
>
F
L
/L0 G
P
EC(
L
L
>
and L
G C
P
E
L F /9.(0 can be re2ritten as
>G7 > > iG> I( i
EC(
K/L0y
t
G H
7
I H
(
t I L
7
.
t
I
J
L
7
5.
t
>
I u F /9.80
>
C t
>G7
2here 2e have used L G L/(0. "imilarly, applying the decomposition K/L0 G
K/(0 I /( C L0K
/L0 to /9.80, 2here K/(0 G ( C
P
p
K F K
/L0 G
P
pC(
K
L
>
and
iG( i
p
>G7 >
P
iG> I(
K
i
F 2e have
EC(
K/(0y
t
G H
7
I H
(
t I L
7
.
t
I
J
L
7
5.
t
>
C K
/L0 PK/L0Q
C(
L/L0
j
7
5.
t
I
K/(0
o
( /( L0K
/L0 PK/L0Q
C(
j
u
t
F
K/(0
/9.D0
H
7
C K
/(0R H
(
L
[
7
G
<o2 it is easily seen that
F R G F S G S/(0 G '
K/(0 K/(0
K/(0
(
and
/( C L0L
/L0 C /( C L0K
/L0
PK/L0Q
C
K/(0
L/L0
G S/L0 C SF
( C /( C L0K
/L0 PK/L0Q
C(
K/(0
( C #K/L0 C K/(0g PK/L0Q
C(
K/(0
G PK/L0Q
C(
F
2here S/L0 G L/L0GK/L0. $sing these results and the decomposition S/L0 G
S/(0 I /( C L0S
/L0, 2here S
/L0 G
P
(
S
L
>
and S
G C
P
(
S
i
in /9.D0 2e
obtain
> G7 > > iG>I(
(
y
t
G [
7
I Rt I S
7
.
t
I S
7
/L05. I
PK/L0Q
C
u
t
' /9.*0
Matching the regressors on the right1hand1side o# /9.80 2ith those in /9.*0 2e
6nally obtain
EC(
y
t
G [
7
I Rt I S
7
.
t
I
J
S
7
5.
t
C>
I `
7t
F /9.A0
> G7
P(8Q
>GE >
C
i>
S
T
uT
/P
8
@
T
(
2here `
7t
G
P
(
S
7
e
t >
I PK/L0Q
C(
u
t
. "imilarly,
EC(
J
y
tCi
G [
i
I Rt I S
7
.
t
I
2here [
i
G [
7
C iR, i G (F '''F pF
> G7
g
7
5.
tC>
I `
it
F i G (F '''F pF /9.-0
and
g
i>
G
f
CS i# i T >
^
F 7 >
> C(
i# i ` >
`
` E C(F i G (F '''F pF
/
P
(
S
7
e
t i >
I PK/L0Q
C(
u #or i ` E
0
`
it
G
>GECi >
C C
CS
7
P
iCEC( 7
tCi
C(
' /9.B0
>G7
e
tCEC>
I S /L0e
tCi
I PK/L0Q u
t
#or i T E
As in the previous section, 2e re2rite the A+DL/pF E0 model /9.80 in matri.
notations in the partitioned regression #orm,
y
T
G @
T
# I !
T
K I u
T
/9.)0
G H
7
d
T
I "
T
c I ;
T
L
I !
T
K I u
T
F
2here y
T
G /y
(
F '''F y
T
0
7
, y
TFCi
G /y
(Ci
F '''F y
T Ci
0
7
F #or i G (F '''F pF !
T
G /y
TFC(
F '''F y
TF
Cp
0F
5J
TFC>
G /5.
(C>
F '''F 5.
T C>
0 #or > G 7F '''F EC(F ;
T
G /5.
TF7
F 5.
TFC(
F '''F 5.
TFCEI(
0F
d
T
G /(F '''F (0
7
, t
T
G /(F '''F T 0
7
, J
T
G /.
(
F '''F .
T
0
7
, @
T
G /d
T
F t
T
F J
T
F ;
T
0 G
/d
T
F "
T
F ;
T
0, u
T
G /u
(
F '''F u
T
0
7
, # G /H
7
F c
7
F L
7
0
7
, c G /H
(
F L
7
0
7
, L
G /L
7
F '''F L
7
0
7
7 EC(
and K G /K
(
F '''F K
p
0
7
' <ote that the dimensions o# !
T
, @
T
, K and # are T O pF T O
/? I ?E I 80F p O ( and /? I ?E I 80 O (, respectively.
Theorem 9.(. $nder assumptions /A(0
7
and /A80 1 /A*0, the 4L" estimators o#
K and c G /H
(
F L
7
0
7
in the A+DL/pF E0 model /9.)0 are
p
T 1consistent and have
the #ollo2ing asymptotic distributions'
p
a
o j
T /K
e
T
C K0 \
<
7F ^
8
g
C(
F /9.(70
u :
2here g
:
is the pOp positive de6nite covariance matri. o# /`
(t
F `
8t
F '''F `
pt
0
7
de6ned
by /9.B0, and
p
a
o j
T /ec
T
C c0 \
<
7F ^
8
d
7
g
C(
d
p
WW
7
F /9.((0
u p :
2here W G /RF S
7
0
7
, d
p
is the p1dimensional unit vector, and ran?/WW
7
0 G (. The
4L" estimators o# H
7
and L
, denoted by He
7T
and L
e
F are also
p
T 1consistent, and
have the mi.ture normal distributions, asymptotically. The covariance matri. #or
all the short1run parameters, h G /#
7
F K0
7
, is asymptotically singular 2ith ran?
eEual to ?E I 8, and can be consistently estimated in the usual 2ay by
&
e
/h
e
T
0 G ^e
8 7
P
@
T
0
C(
F
2here P
@
T
G /@
T
F !
T
0F and ^e
uT
G
T
C
/y
T
C P
@
T
h
e
T
0
7
/y
T
C P
@
T
h
e
T
0.
P(9Q
2here K
e
T
/(0 G ( C
P
T
u
u
^
iG(
T
T
(
u
%rom Theorem 9.( 2e also 6nd that
p
T /He
(T
C H
(
0 and
p
T /L
e
T
C L0 are
asymp1 totically per#ectly collinear 2ith
p
T /K
e
T
C K0F that is,
T
n o p
/ec
T
C c0 I
WPK
e
/(0 C
K/(0Q
G o
p
/(0' /9.(80
p
iG(
K
e
i
T
. 3t is also straight#or2ard to sho2 that
/ec
T
C c0 I WPK
e
T
/(0 C K/(0Q
W
e
T
C W G
K
e
T
/(0
' /9.(90
$sing Theorem 9.(, and results /9.(80 and /9.(90, 2e have'
Theorem 9.8. $nder the assumptions /A(0
7
and /A80 1 /A*0, the 4L" estimators
o# the long1run parameters,
W
e
T
G /
e
R
TF
S
e
7
0
7
G
ec
T
GK
e
/(0 in /9.)0, converge to
their true values at #aster rates than the estimators o# the associated short1run
parameters, and #ollo2 the mi.ture normal distribution asymptotically. There#ore,
g
8
C(
e
a
f
^
8
^
"
_
T
D
"
T
/W
T
C W0 \
<
7F 3
?I(
PK/(0Q
8
F /9.(D0
2here g
"
_
T
and D
"
T
are as de6ned in Theorem
8.8.
Comparing Theorems 8.8 and 9.8, 2e 6nd that the presence o# the 3/70 stationary
regressors in /9.)0 /i.e., additional lagged changes in y
t
and the lagged changes
in .
t
2hich are introduced to deal 2ith the residual serial correlation problem0
does not aect the asymptotic properties o# the 4L" estimator o# the long run
coe5cients, R and S. There#ore, in#erences concerning the long1run parameters
can be based on the same standard tests as given by /8.(80 and /8.(90. 3n this
more general case, ho2ever, the e.pression #or the asymptotic variance o# W
e
T
is still given by /8.((0, but 2ith ^
8
G/( C K0
8
replaced by the more general
e.pression,
8
GPK/(0Q
8
.
;e no2 rela. assumption /A90 and allo2 #or the possibility o# endogenous
regressors, but con6ne our attention to the case 2here 5.
t
can be represented by
a 6nite order vector A+/s0 process,
)
2here P/L0 G 3
?
C
P
s
P/L05.
t
G M
t
F /9.(*0
P
i
, and P
i
, i G (F '''F s, are the ? O ? coe5cient matrices
such that the vector autoregressive process in 5.
t
is stable. Here M
t
are assumed
)
4ur analysis can also allo2 #or the inclusion o# lagged 5yZs and a dri#t term in /9.(*0 2ithout
aecting the results presented belo2. 4n this see =os2i>? /())*0 and Pesaran, "hin and "mith
/())A0.
P(DQ
t
^
>
V
to be serially uncorrelated, but possibly contemporaneously correlated 2ith u
t
F
namely, 2e assume that p
t
G /u
t
F M
7
0
7
#ollo2s the multivariate iid process
2ith mean Nero and the covariance matri.,
`
8
q
pp
G
u
q
uM
W
' /9.(A0
q
Mu
q
MM
;e 2ill, ho2ever, continue to assume that Cov/u
tC>
F M
tCi
0 G 7 #or i G >. <o1
tice that despite this assumption the model is still general enough to allo2
not only #or the contemporaneous but also #or cross1autocorrelations bet2een u
t
and
5.
t
. ;ith assumption /A90 rela.ed, the 4L" estimators in /9.(0 are no longer
consistent. To correct #or the endogeneity o# .
t
, 2e model the
contemporaneous
correlation bet2een u
t
and M
t
by the linear regression o# u
t
on M
t
u
t
G d
7
M
t
I V
t
F /9.(-0
2here using /9.(A0 2e have d G q
C(
q
7
, and M
t
is strictly e.ogenous 2ith respect
MM uM
to V
t
.
(7
"ubstituting /9.(*0 in /9.(-0 2e obtain'
u
t
G d
7
P/L05.
t
I V
t
F /9.(B0
2here 5.
tCi
Zs, i G 7F '''F sF are also strictly e.ogenous 2ith respect to V
t
. The
parametric correction #or the endogenous regressors is then eEuivalent to
e.tending the A+DL/pF E0 model /9.80 to the more general A+DL/pF m0 speci6
cation,
mC(
K/L0y
t
G H
7
I H
(
t I L
7
.
t
I
J
U
7
5.
t
C>
I V
t
F /9.()0
>G7
2here m G ma./EF s I (0, and U
i
G L
C P
7
d, i G 7F (F 8F '''F m C (, P
7
G 3
?
,
i i
L
i
G 7 #or i W E, and P
i
G 7 #or i W s.
;e no2 replace assumption /A90 by
/A90
7
The scalar disturbance V
t
in /9.()0 is iid/7F ^
8
0, and 5.
t
#ollo2s the general
stationary process given by /9.(*0. %urthermore, V
t
and M
t
are uncorrelated
such that .
t
and 5.
tC>
Zs > G 7F '''F m C (F are strictly e.ogenous 2ith
respect to V
t
in the A+DL/pF m0 model /9.()0.
There are t2o main dierences bet2een the A+DL models de6ned by /9.80 and
/9.()0. %irstly, the order o# lagged 5.
t
Zs in the t2o models can dier, and
secondly, the coe5cients on 5.
t
Zs and their lagged values have dierent
interpretations. Although this alters the dynamic structure o# the model, the basic
#rame2or? #or estimating the long1run parameters and carrying out statistical
in#erence on them is the same as be#ore.
(7
The relation /9.(-0 2ill be e.act 2hen the >oint distribution o# u
t
and M
t
is normal.
P(*Q
K
T
iG(
T
^
V
> >
(
f
V
C
Theorem 9.9. $nder the assumptions /A90
7
, /AD0 and /A*0, the 4L" estimators
p
o# the short1run parameters in /9.()0, H
7
, H
(
, L, K
(
F '''F K
p
, U
7
F '''F U
mC(
, are T 1
consistent, and asymptotically have the /mi.ture0 normal distributions. %urther1
i
more,
p
T /ec
T
C c0 is asymptotically per#ectly collinear 2ith
p
T h
e
/(0 C K/(0 ,
2here c G /H
(
F L
7
0
7
and K/(0 G ( C
P
p
K
i
, such that the covariance matri. #or
the estimators o# the short1run parameters is asymptotically singular 2ith ran?
eEual to ?m I 8. The asymptotic distribution o# the 4L" estimators o# the long1
run parameters,
there#ore,
W
e
T
G
/
TF
S
e
7
0
7
G
ec
T
GK
e
T
/(0 in /9.()0, are mi.ture normal and
g
8 C(
e a
8
^
"
_
T
D
"
T
/W
T
C W0 \
<
7F 3
?I(
PK/(0Q
8
F /9.870
2here ^
8
is the variance o# V
t
in /9.()0, and g
"
_
T
and D
"
T
are as de6ned in
Theorem 8.8.
There are no #undamental dierences bet2een the results o# Theorems 8.8,
9.8 and 9.9, as #ar as the estimators o# the log1run parameters are concerned. A
com1 parison o# /8.((0, /9.(D0 and /9.870 sho2s that the asymptotic distributions o#
the
estimators o# the long1run parameters, W
e
T
, under various assumptions
discussed
above dier only by a scalar coe5cient.
3n sum, in the conte.t o# the A+DL model in#erence on the long run para1
meters, R and S, is Euite simple and reEuires a priori ?no2ledge or estimation o#
the orders o# the e.tended A+DL/pF m0 model. Appropriate modi6cation o# the
orders o# the A+DL model is su5cient to simultaneously correct #or the resid1
ual serial correlation and the problem o# endogenous regressors. &ariances o# the
4L" estimators o# the long1run coe5cients can then be consistently estimated
using either /9.870, or by means o# the 51method applied directly to the long1
run estimators. Alternatively, one could compute the estimates o# the long1run
coe5cients and their associated standard errors using =e2leyZs /()-)0 regression
procedure. =e2leyZs method involves re2riting /9.()0 as
H
7
(
mC(
J
p (
(
J
V
t
K/L0y
t
G
K/(0
I Rt I S
7
.
t
I
K/(0
>G7
U
7
5.
tC>
C
K/(0
>G7
K
5y
tC>
I
K/(0
F /9.8(0
and then estimating it by the instrumental variable method using /(, t, .
t
, 5.
t
,
5.
tC(
F '''F 5.
tCmI(
, y
tC(
, y
tC8
F '''F y
tCp
0 as instruments. 3t is easy to sho2 that
the
3& estimators o# R and S obtained using /9.8(0 are numerically identical to the
4L" estimators o# R and S based on the A+DL model /9.()0, and that the standard
errors o# the 3& estimators #rom the =e2leyZs regression are numerically identical
to the standard errors o# the 4L" estimators o# R and S obtained using the 51
method. /"ee, #or e.ample, =ardsen /()B)0.0 The main attraction o# the =e2leyZs
P(AQ
regression procedure lies in its possible computational convenience as compared to
the direct 4L" estimation o# /9.()0 and computation o# the associated standard
errors by the 51method.
((
%inally, 2e note in passing that the results developed in this section also apply
to the case 2here the underlying regressors, .
t
, given by /9.(*0, are 3/70. /"ee
#ootnote - and the Monte Carlo simulation results in "ection *.0
D. A Comparison o# A+DL and Phillips1Hansen Procedures
Here 2e #ocus on the case 2here there e.ists a uniEue cointegrating relation be1
t2een 3/(0 variables, y
t
and .
t
, possibly 2ith a deterministic trend. The case
2here there are multiple cointegrating relations among 3/(0 variables presents
ad1 ditional di5culties and 2ill not be discussed in this paper. /"ee Pesaran and
"hin /())*0, and the re#erences cited therein0.
Consider the #ollo2ing cointegrating relation
y
t
G [ I Rt I S
7
.
t
I v
t
F /D.(0
5.
t
G e
t
' /D.80
Although the 4L" estimator o# S is sho2n to be T 1consistent, /see "toc? /()B-00,
it has also been #ound that the 6nite sample behavior o# the 4L" estimator is
generally very poor /see, #or e.ample, =aner>ee et. al. /()BA00. specially, in the
presence o# non1Nero correlation bet2een v
t
and e
t
, 4L" estimators o# S in /D.(0
are o#ten heavily biased in 6nite samples, and in#erences based on them are invalid
because o# the dependence o# the limiting distribution o# the 4L" estimators on
nuisance parameters. %or details see Phillips and Loretan /())(0.
=roadly spea?ing, there are t2o basic approaches to cointegration analysis' ,o1
hansenZs /())(0 ma.imum li?elihood approach, and Phillips1HansenZs /())7, PH0
#ully modi6ed 4L" procedure.
(8
The A+DL approach to cointegration analysis
advanced in this paper is directly comparable to the PH procedure, and 2e shall,
there#ore concentrate on this method. PH assume that v
t
and e
t
in /D.(0 and /D.80
#ollo2 the general correlated linear stationary processes'
(9
v
t
G A
(
/L0u
t
F e
t
G A
8
/L0M
t
F /D.90
((
%or a computer implementation o# the A+DL approach using the 51method see
Pesaran and Pesaran /())-0.
(8
There are also other related procedures such as the original t2o1step method o# ngle and
@ranger /()B-0, the leads and lags estimation procedure suggested by "ai??onnen /())(0 and
"toc? and ;atson /())90, and the canonical method by Par? /())80.
(9
%or more details see Phillips and "olo /())80.
P(-Q
t
iG(
t
MM
t
(*
MM
2here p
t
G /u
t
F M
7
0
7
are serially uncorrelated random variables 2ith Nero means
and a constant variance matri. given by /9.(A0. Assuming A
(
/L0 and A
8
/L0 are
invertible, /D.(0 can be appro.imated as an A+DL speci6cation by truncating
the order o# the in6nite order lag polynomials PA
(
/L0Q
C(
and PA
8
/L0Q
C(
such that
K/L0 U PA
(
/L0Q
C(
and P/L0 U PA
8
/L0Q
C(
, 2here the orders o# the lag polyno1
mials K/L0 and P/L0 are denoted by p and s, respectively. Then 2e obtain the
appro.imate 6nite1dimensional A+DL/pF m0 speci6cation,
K/L0y
t
G #K/(0[ I RK
7
/(0g I RK/(0t I K/L0S
7
.
t
I q
uM
q
C(
P/L05.
t
I V F
/D.D0
MM t
2here K
7
/(0 G C
P
p
iK
i
, m G ma./pF s I (0, and by construction .
t
/and 5.
t
Zs0
are uncorrelated 2ith V
t
.
(D
<otice that /D.D0 is o# the same #orm as /9.()0, 2ith
the #ollo2ing relations among their parameters' H
7
G K/(0[ I RK
7
/(0, H
(
G RK/(0,
L G K/(0S, U
7
/L0 G K
/L0S
7
I q
uM
q
C(
P/L0, 2here K
c
uT
/P
T
T
T
e
T
@
T
K
(8
2here g
((
is the /(,(0 element o# g
C(
. The above result can be re2ritten sepa1
c c
rately #or He
7T
F ec
T
and L
e
as
p
C 5
p
p
e
V
T /He
7T
C
H
7
0 I
p
[
(
F [
8
F '''F [
p
T K
T
C
K
p
G d
cuF(
I o
p
/(0F /A.(D0
V
T /ec
T
C c0 I
Wd
7
T
p
K
e
C K G o
p
/(0F /A.(*0
p
T
p
e
V
p
p
e
V
T
C
L
I /g
(
F g
8
F '''F
g
p
0
T K
T
C
K
G g
C(
E
;u
I o
p
/(0F /A.(A0
;
2here d
p
is a p O ( vector o# unity and d
cuF(
is the 6rst element o# g
C(
E
cu
. $sing
/9.(70 in /A.(*0 2e obtain /9.((0. 3t is also clear #rom above results that the
4L" estimators o# H
7
and L
/standardiNed by
p
T 0 have the /mi.ture0
normal
distributions asymptotically.
%inally, using /9.(70, /9.((0, and /A.(901/A.(A0, it is easily seen that a consis1
tent estimator o# the variance o# h
e
T
is given by &
e
/h
e
T
0 G ^e
8
7
P
@
T
0
C(
2ith
the ran? o# &
e
/h
e
T
0 being eEual to ?E I 8.
Proo# o# Theorem 9.8.
Partition d
T
G /a
T
F s
7
F 2
7
0
7
con#ormably to @
T
G /d
T
F "
T
F ;
T
0F then s
T
is
given by
T
s
T
G
p
T
/ec
T
C c0 I Wd
7
p
p
K
e
C
V
' /A.(-0
T
p
T
T
$sing /A.(70 and /A.((0, /s
7
F 2
7
0
7
can be e.pressed as
T T
`
s
T
W `
"
7
G
7
W
(
`
T
W
/A.(B0
T
H
T
"
T
"
T
H
T
;
T
C
"
7
H
T
u
T
T
H
T
"
T
;
7
H
T
;
T
;
7
H
T
u
T
`
"
7 7
W
C(
` W
p V
T
H
T
"
T
"
T
H
T
;
T
"
7
H
T
:
T
K
e
K ' C
;
7 7 7 T
C
Let
T
H
T
"
T
;
T
H
T
;
T
;
T
H
T
:
T
E
"_
T
u
T
G D
"
T
"
T
H
T
u
T
F g
"_
T
G D
"
T
"
T
H
T
"
T
D
"
T
F
7 7
9
(
2here D
"
T
G Diag/T
C
8
F T
C
3
?
0. Then, it is also easily seen that as T m (,
M
+
(
(
i
E
"
_
T uT
0 E
"
_
u
G
7
/r C
8
0d=
u
/r0
+
(
=
_
7
F /A.()0
7
e
/r0d=
u
/r0
M
(
+
(
(
i
g
"
_T 0 g
"
_
G
+
( ( 7
7
/r C
8
0=
_
e
/r0dr
+
(
F /A.870
7
/r C
8
0=
_
e
/r0dr
7
=
_
7
/r0=
_
e
/r0dr
PA.DQ
D
"
7 C(
" "
_
"
t
C
t
C
t
C
(
(
T
2here =
_
e
/r0 G =
e
/r0 C
+
=
e
/r0dr is a ?1dimensional demeaned =ro2nian motion
7
on P7F (Q. "ince =
_
e
/r0 is also distributed independently o# =
u
/r0, 2e obtain as
in
/A.D0,
a
8
E
"
_
u
\ M<
C
7F ^
u
g
"
_
5
' /A.8(0
Multiplying /A.(B0 by the diagonal matri., Diag/D
C(
F T
(
0, using /A.()01/A.8(0
and noting that
"
T
8
T
" H
T
;
T
G 4
p
/(0F T ;
7
H
T
;
T
G 4
p
/(0F
T T
(
D
"T
"
7
H
T
:
T
G 4
p
/(0F T
C
8
;
7
H
T
:
T
G 4
p
/(0F
2e obtain
T
D
C( C(
a
T
p
8 C(
V
and there#ore,
"
T
s
T
0 g
_
E
"
_
u
\ M< 7F ^
u
g
_
F
g
8 C(
a
C
8
5
"
_
T
D
"
T
s
T
\ <
%inally, by /9.(90 and /A.(*0 2e have
7F ^
u
3
?I(
s
T
' /A.880
W
e
T
C W G
K
e
T
/(0
' /A.890
(
Multiplying /A.890 by g
8
"
T
D
C(
, using /A.880 and noting that
K
e
T
p
/(0 m K/(0F 2e
obtain /9.(D0.
Proo# o# Theorem 9.9 can be established in a similar manner and is omitted to
save space.
Proo# o# Theorem D.(.
Consider the dynamic A+DL/pF m0 model /9.()0 /or /D.D00, and its static coun1
terpart /D.(0. Applying the decomposition K/L0 G K/(0 I /( C L0K
/L0 to
/9.()0 2e have
H
7
U
7
/L0 V
t
K
/L0
y
t
G
K/(0
I Rt I S
7
.
t
I
5. I
K/(0 K/(0
5y
t
' /A.8D0
K/(0
"ubstituting #or 5y
t
G R I S
7
5.
t
I 5v
t
#rom /D.(0 in /A.8D0, 2e have
U
7
/L0 V
t
K
/L0
y
t
G [ I Rt I S
7
.
t
I
5. I
K/(0 K/(0
/S
7
5.
t
I 5v
t
0 ' /A.8*0
K/(0
$sing /A.8*0, v
t
in /D.(0 can be e.pressed as
U
7
/L0 C K
/L0S
7
V
t
K
/L0
v
t
G 5. I
K/(0 K/(0
5v
t
' /A.8A0
K/(0
PA.*Q
t
0 G
8
^
8
Vv
^
7
t
t
C
t
^
t
ee
^
v
V
De6ning ?
t
G /V F v
t
F 5.
7
0
7
G /V F v
t
F e
7
0
7
, and j/L
h
(
F
CK
/L0/(CL0
F
U
7
/L0CK
/L0S
7
i
,
t t t K/(0 K/(0 K/(0
then the spectral density o# v
t
G j/L0?
t
is given by
8U#
vv
/m0 G j/e
i2
0& ar/?
t
0j
7
/e
Ci2
0F
2here
V
^
Vv
7
9
& ar/?
t
0 G
D
^
7 8
q
ve
*
'
7 q
ve
q
ee
Hence, the spectral density o# v
t
at Nero #reEuency is given by
^
8 7
8U#
vv
/70 G
V
I PU /(0 C K
/(0S
7
Q q
ee
PU/(0 C K
/(0SQ
PK/(0Q
8
' /A.8-0
The Phillips1Hansen semi1parametric correction is eEuivalent to removing the
sec1 ond part o# /A.8-0, by subtracting the terms involving 5.
t
#rom v
t
. $sing
/A.8A0 2e have the #ollo2ing e.pression #or the modi6ed disturbance term,
v
I
, in the
Phillips1HansenZs procedure'
v
I
U
7
/L0 C K
/L0S
7
V
t
K
/L0
I I
t
G v
t
C 5. G
K/(0 K/(0
5v
t
G j
K/(0
/L0?
t
F
2here ?
I
G /V F v
t
0
7
F and j
I
/L0 G
h
(
F
CK /L0/(CL0
i
' There#ore, the spectral
t t
K/(0 K/(0
density o# v
I
at Nero #reEuency is given by
8
7
V
8U#
v
I
v
I /70 G j
I
/70& ar/?
I
0j
I
/70
G
PK/(0Q
8
'
$sing /D.-0 2e also have
#
v
I
v
I /70 G =#
\\
/70=
7
F
2here = G P(F C
ve
C(
Q. =y de6nition
\
G 8U#
\\
/70, and
8
8U#
v
I
v
I /70 G =
\
=
7
G m
vv
C
ve
C(
ev
G '
ee
PK/(0Q
8
8
8
Hence, by /D.B0 m
v5e
G ^
V
G PK/(0Q .
PA.AQ
+e#erences
P(Q =aner>ee, A., ,. Dolado, D. Hendry and @. "mith /()BA0, X.ploring Eui1
librium +elationships in conomics through "tatistical Models' "ome Monte
Carlo vidence,Y 4.#ord =ulletin o# conomics and "tatistics, DB' 8*918--.
P8Q =ardsen, @. /()B)0, XThe stimation o# Long1+un Coe5cients #rom rror
Correction Models,Y 4.#ord =ulletin o# conomics and "tatistics, *(' 9D*1
9*7.
P9Q =e2ley, +. /()-)0, XThe Direct stimation o# the Euilibrium +esponse in a
Linear Dynamic Model,Y conomics Letters, 9' 9*-19A(.
PDQ =os2i>?, H.P. /())*0, X5cient 3n#erence on Cointegration Parameters in
"tructural rror Correction Models,Y ,ournal o# conometrics, A)' (991(*B.
P*Q Cavanaugh, C.L., @. lliott and ,.H. "toc? /())*0, X3n#erence in Models 2ith
<early 3ntegrated +egressors,Y conometric Theory, ((' ((9(1((D-.
PAQ ngle, +.%. and C.;.,. @ranger /()B-0, XCointegration and rror Correction
+epresentation' stimation and Testing,Y conometrica, **' 8*(18-A.
P-Q Hendry, D., A. Pagan and ,. "argan /()BD0, XDynamic "peci6cationsuY Chap1
ter (B in Handboo? o# conometrics, &ol 33 /ed., c. @riliches and M.
3ntrili1 gator0, <orth Holland
PBQ Hsiao, C. /())*0, XCointegration and Dynamic "imultaneous Euations
Model,Y unpublished manuscript, $niversity o# "outhern Cali#ornia.
P)Q 3nder, =. /())90, Xstimating Long +un +elationships in conomics,Y ,our1
nal o# conometrics, *-' *91AB.
P(7Q ,ohansen, ". /())(0, Xstimation and Hypothesis Testing o# Cointegrating
&ectors in @aussian &ector Autoregressive Models,Y conometrica, *)' (**(1
B7.
P((Q Ltt?epohl H. /())(0, 3ntroduction to Multiple Time "eries Analysis, <e2
!or?, <.!. "pringer and &erlag.
P(8Q Par?, ,.!. /())80, XCanonical Cointegrating +egressions,Y conometrica, A7'
(()1(D9.
P(9Q Pesaran, M.H. /())-0, XThe +ole o# conomic Theory in Modelling the Long1
+un,Y The conomic ,ournal, (7-' (-B1()(.
P+.(Q
P(DQ Pesaran, M.H. and =. Pesaran /())-0, Micro6t D.7' 3nteractive conometric
Analysis, 4.#ord $niversity Press /#orthcoming0.
P(*Q Pesaran, M.H. and !. "hin /())*0, XLong1+un "tructural Modelling,Y un1
published manuscript, $niversity o# Cambridge.
P(AQ Pesaran, M.H., !. "hin and +.,. "mith /())A0, XTesting #or the .istence o# a
Long1+un +elationship,Y DA ;or?ing Papers Amalgamated "eries, <o.
)A88, $niversity o# Cambridge.
P(-Q Phillips, P.C.=. /())(0, X4ptimal 3n#erence in Cointegrated "ystems,Y cono1
metrica, *)' 8B9197A.
P(BQ Phillips, P.C.=. and ".<. Durlau# /()BA0, XMultiple Time "eries +egression
2ith 3ntegrated Processes,Y +evie2 o# conomic "tudies, *9' D-91D)A.
P()Q Phillips, P.C.=. and =. Hansen /())70, X"tatistical 3n#erence in 3nstrumental
&ariables +egression 2ith 3/(0 Processes,Y +evie2 o# conomic "tudies, *-'
))1(8*.
P87Q Phillips, P.C.=. and M. Loretan /())(0, Xstimating Long +un conomic
Euilibria,Y +evie2 o# conomic "tudies, *B' D7-1D9A.
P8(Q Phillips, P.C.=. and &. "olo /())80, XAsymptotic #or Linear Processes,Y An1
nals o# "tatistics ' )-(1(77(.
P88Q "ai??onnen, P /())(0, XAsymptotically 5cient stimation o# Cointegration
+egressions,Y conometric Theory, -' (18(.
P89Q "toc?, ,.H. /()B-0, XAsymptotic Properties o# Least "Euares stimates o#
Cointegrating &ectors,Y conometrica, **' (79*1(7*A.
P8DQ "toc?, ,.H. and M.;. ;atson /())90, XA "imple stimator o# Cointegrating
&ectors in Higher 4rder 3ntegrated "ystems,Y conometrica, A(' -B91B87.
P8*Q ;ic?ens, M.+. and T.". =reusch /()BB0, XDynamic "peci6cation, the Long
+un stimation o# the Trans#ormed +egression models,Y The conomic
,our1 nal, )B' (B)187*.
P+.8Q