7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 1/14
Cambridge University Press is collaborating with JSTOR to digitize, preserve and extend access to Econometric Theory.
http://www.jstor.org
Estimating Error Component Models with General MA(q) DisturbancesAuthor(s): Badi H. Baltagi and Qi Li
Source: Econometric Theory, Vol. 10, No. 2 (Jun., 1994), pp. 396-408Published by: Cambridge University PressStable URL: http://www.jstor.org/stable/3532874Accessed: 23-01-2016 16:12 UTC
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/ info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of contentin a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.For more information about JSTOR, please contact [email protected].
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 2/14
Econometricheory,0, 1994, 96-408. rintedntheUnited tates fAmerica.
ESTIMATING ERROR COMPONENT
MODELS WITH GENERAL
MA(q) DISTURBANCES
BADI
H.
BALTAGI
Texas
A
&M University
Qi
Li
IndianaUniversity
This aper rovides simplestimationethodor n erroromponentegres-
sion
modelwith eneral A(q) remainderisturbances.he stimationethod
utilizes
he ransformationerived yBaltagi nd Li [3] for n error ompo-
nentmodelwith utoregressiveemainderisturbances,nd standardrthog-
onalizinglgorithm
or he
eneral A(q) model.
his stimation ethods
computationallyimple tilizingnlyeast-squaresegressions.his s mportant
for aneldataregressionshere rute orceGLS is n many asesnotfeasi-
ble.This stimationethod erforms
ell
elative
o
true LS
in
Monte-Carlo
experiments.
1.
INTRODUCTION
MacDonald and MacKinnon [14] argued that the AR(p) specifications
more
popular
han
he
MA(q) specification
n
empirical pplications,
ot
be-
cause t s more
plausible, ut
rather
ecause
t
s easier o compute.This pa-
per focuses
n
the
moving verageprocess,
which
s
a
viable
alternative
hat
naturally
rises
n
several conomic
xamples,
ee
Nicholls,Pagan,
and
Ter-
rell 17]. For the panel data error omponent egressionmodelwith emain-
der
disturbances ollowing
n
MA(q) process,
the
estimationmethod
s
further
omplicated y
the
presence
f
random
ndividual ffects.
ecently
Baltagi
nd
Li
[3]
obtained
simple ransformation
hat ransforms
he
uto-
regressiverror omponent
isturbancesnto
spherical
isturbances.
his
s
important
or
panel data studies
where
rute
orceGLS
is in
many
ases
not
feasible, specially
or
argepanels.Baltagi
nd
Li
[3]
showedhow this rans-
formation
an be
applied
when
he
remainder
isturbances
vit's)
ollow
n
AR(1),
an
AR(2),
or a
special AR(4) process
for
quarterly
ata.1
n
fact,
BaltagiandLi [3] argued hat,for heerror omponentmodel,their rans-
formation eneralizes
o
any
error
rocessgenerating
he
remainder istur-
We
thankMax King, Peter chmidt, eterPhillips, nd
two anonymous eferees
orproviding elpful om-
ments nd
suggestions hat mproved
he
paper.
396
( 1994Cambridge
University ress
0266-4666/94 5.00 +
.00
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 3/14
ERROR COMPONENTS WITH MA(q) DISTURBANCES 397
bances s long s there xists simple nownmatrix,ayC, that ransforms
the emainderisturbancesnto phericalisturbances.
n
order o operation-
alizethis ransformation,stimatesf the erial orrelationarametersre
needed. ortheAR(p) model, hese arameterstimatesreeasily btain-
able. However, or heMA(q) process, he stimation ethods more
n-
volved ven
n
conventionalime eriesmodels.Recently, acDonald nd
MacKinnon 14]proposed convenient ethod or stimatingheMA(1)
model,
ut
n
extension
f
this
methodo the
general A(q) process
s dif-
ficult. llah,Vinod, ndSingh 23]proposed new stimation ethod
or
linearmodelswithmoving verage isturbances
f
any order, ut this
method equires onlineareast-squaresegressions.dditionally,
or he r-
ror omponents odel, hismethods furtheromplicatedythepresence
of ndividualffects
hich
re
hard
oseparate
rom he
movingverage
f-
fects.
his
paper roposes simple pproach
o the stimationferror
om-
ponents egression
odel
with
emainder
isturbancesollowing general
MA(q) process.2
n
essence,
e show hat
nly he utocovarianceunction
estimatesf the
omposite
rror erm reneeded
n
thefirst
tep
o
obtain
GLS estimatesf regressionoefficients
n
the
econd tep.
The
computa-
tional teps re as
follows:
i) estimate
he
utocovarianceunctionf the
composite
rror erm
sing
LS
residuals;ii) retrievestimates
f
thegen-
eralMA(q) autocovarianceunctionswell sestimatesfthe arianceom-
ponents;
nd
iii) apply
he
ransformationroposed yBaltagi
nd
Li
[3]
that tilizes
ecursive
ormulas
or he
movingverage rocess
nd
perform
OLS. This
procedure
esults
n
asymptotically
fficientstimates
f
the
e-
gression
oefficientsnd
requires nly east-squaresegressions.
Section reviewshe
Baltagi
ndLi
[3]
transformationor he rror om-
ponent
modelwith
utoregressive
emainder
isturbances,
nd
Section gen-
eralizes he transformation
o the error
omponent
model
with
general
MA(q)
remainder
isturbances.ection
performs
onte-Carlo
xperi-
ments ndconcludes hepaper.
2. THE MODEL
Consider he
followinganel
data
regression odel,
yit
x, to+uit
i
=
1,.
..,9N;
t
=
1,. ..,
T, (1)
whereB s
a
K
x
1
vector f
regressionoefficients,ncluding
he
ntercept.
The disturbancesollow one-wayrroromponent odelseeHsiao [9]),
Uit
=
Ai
+
Pit
(2)
with ndividual
ffects
Ai
i.i.d.
0,
,2)
and
remainder
isturbances
it
hat
follow
stationary-dependentrocess
where
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 4/14
398
BADI H. BALTAGI
AND Ql LI
'Ys
?
<
Isl
c
q
E
(pitPi,
t-s)
=(3
t/0
otherwise,
with
_-
=
Pys*Note that
a standard
MA(q)
process is
equivalent
to a sta-
tionary
-dependent
rocess.
In
fact, the standard
MA(q) process,
vit
=
Cit
+
XIei t-I
+
*.
+
XqEi,t-q,
where
it
-
i.i.d.(0,a,),
is a
special
ase
of
the
tationary
-dependent
rocess efined
n
3). On the
other
and,one
can
also
prove hat
ny
stationary
-dependent
rocess
an be represented
s a
standard
MA(q) process
see Brockwell
nd
Davis
[4]). In what
follows,
we
adopt a general
pproach,
assuming
he
vit's
ave
an arbitrary
erial
corre-
lation
form.
By using
the results
f Baltagi
and
Li [3], we derive
he spec-
traldecompositionf thevariance-covariance atrix fthecomposite rror
and then pply
t specifically
o thegeneral
MA(q) process
given
n (3).3
In
vector
form, quation
1)
can be written
s
y
=
X:
+ u
(4)
with
U
(IN
?&
T)L
+
P
(5)
where
N
is an
identity
matrix
f dimension
N,
eT
is a vector
of 1's of
di-
mension T,
,'
=
(Al,
. . . ,,IN),
and
v'
= (Pl1,
. .
,Y1T,
..
*., N1,
*
.
,PNT).
The variance-covariance
atrix
f u
is
?
2
IN
0
JT
+
g2
IN
0O
Q
=
IN
(&
A,
(6)
where
A
= or
JT
+
o20,
and
E(
vi
)
=
a20
is the variance-covariance
a-
trix
f the remainder
rror
erm
for ach individual 4
=
(Pv1I,
i2,
PiT),
for
=
1,...
,
N. Because
0
is a real
symmetricositive-definite
atrix,
here
exists
T x
T matrix
,
such thatCDC' =
IT.
Therefore,
we
propose
the
following
ransformation.
Definey*=
(IN
( C)Y; X* and u* are similarly efined. hetransformed
error
becomes
U*=
(IN
(D CeT)A
+
(IN
C)V
=
(IN
0
e')/
+
(IN
C)P,
(7)
where '
=
CeT
=
(011,.. T,
Y)
is
a
T x
1
vector,
epending
n the
specific
serial
correlation
rocess.
The
variance-covariance
matrixfor the
trans-
formed
isturbance
*
becomes
*=
IN
0
[CAC']
=
IN
0
*
(8)
withA*
=
CAC. Because C has thepropertyhat CDC'
= IT,
bywriting
JT
=
eTeT
and
noting
hat
CeT
=
e'2,
we
get
from he
definition
f A that
A
=
'2eTeT
+
o2IT
=
2,JT
+
(IT
(9)
whereJ
a=
ece.
This can
be rewritten
s
A*
=
02d2(eTea/d2)
+ or2IT
=
Or2d2JT
J2IT
(10)
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 5/14
ERROR COMPONENTS
WITH MA(q) DISTURBANCES
399
where 2
= e'
e'.
Thisreplaces
T eTeT by ts dempotent
ounterpart
JT= eTeT/d2.
By
following
ansbeek
nd Kapteyn24,25],one also re-
places IT by ET +
JT,
hereET = IT
-J.
From collectingike terms, ne
gets he pectral
ecompositionf A*:
A*
=
2JA
+
2
Ec
(11)
where
2,
u2d2+
u2.
BecauseJT nd
ET are dempotent
atriceshat re
orthogonal
o each
other,
we
have
A*P
=
(uf) JT+ (a2)PET
(12)
where is
an arbitrarycalar.
n particular,
= -1
gives *` (andhence
1*-1)
andp
= I-
gives A*1/2 (and hence E*-1/2). Therefore,
orS
=
(IN
(0
JT)
+
(IN
(?
ET)
=
(IN
(0
IT)
-
6(IN
(092)
(13)
where
= 1-
(/ia). Thus
we can premultiply
he C-transformed
odel
by
UE*-1/2
to make the error pherical.y*
=
u*-l/2y*
and
X**
and u**
are
similarly
efined.
he
typical
lementsf
y**
re given y
T
s=1
Yit
=Y*-
a
(14)
Za
2
This
s a
generalized
ersionfthe
uller ndBattese
5]transformationor
theerroromponent
odelwith n
arbitrary
ariance-covarianceatrix,
E( i', )
=
a 2O,
on
the remainder isturbances. he
OLS regression
n the
(**)
transformedquation
s
equivalent
o theGLS
regression
n the rigi-
nal equation
4).
Equation12) also suggestsaturalstimatorsfthe arianceomponents.
Baltagi
nd
Li
[3] proposed
stimating
2
nd
o,2u
y
=
U*/(IN
0
JT)U*/N
and
2
=
U*J(IN
0
ET)u*/N(T
-
1).
(15)
These re
best
uadratic
nbiased stimators
f a and
2
if
the rue is-
turbances
*
are
known. hetrue esidualsre
generally
otknown.
n
this
case,
one can
replace
*
byu*,
the
OLS residuals n the
*)
transformed
equation.
3. THE GENERALMA(q) PROCESS
In
this ection
e
give
simple ecursiveransformation
or he rrorom-
ponent
model
with emainder
isturbancesollowingn MA(q) process.
Let
ys
=
E(viti,
t-s)
denote the
autocovariance
functionfor
Pit
for s
=
0,1,..
,q.
Let
S
=
ys/-Y
for = 0,1,...
,q, which mplies
hat
O
?
1. Note
that
E
(Pi,
Pis)
s the same
for ll i
=
1,... ,N,
therefore e do not
write he
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 6/14
400 BADI H.
BALTAGI AND
Ql
LI
subscript
in
7y.
The
Appendix ives
standard
rthogonalizinglgorithm
that ransformshe
general A(q)
disturbancesnto
pherical
isturbances
(see Li
[11]).'
With hisnotation, he *) transformationorourmodel,de-
fined
n
7),
is
given ecursively
y
Y7i
Yi,
Yit
=
[yit
-
*t*y*
-b
-
bt,jyi
]I.a
for t
=
2,...,q
Y*
=
[
Yi,
-bt,
t-qYt,
-q]
I
for t
=
q +1,...,T.
(16)
Both
,
and
bt,5
re
n
turn eterminedecursively,s follows:
a,1
=
at= 1-b
-
-b722-r2
1
for
t=2,...,q
atf-
1
-
b
tt
2
for
t=q+
1,...,T. (17)
The recursiveelation or
bt,,
s
given y
(i)
for
<
q,
bt,
=
rtI
bt,,_5s=
[r.
-
bt- s
-t-s-I
b,s-I
bt,s,I
bt,I
/
for s
=
1,...,t-2; (18)
(ii)
for
q
+
1,..
,T
bt,t-q
=
rt,q/atq
b,
t-s=
[r-
bs,s-Ibt1s,1-
-
bs,t-qbt,t-q]
/
for s= 1,...,t-2. (19)
By replacing * by
at
and
replacing
it
y
1
n
(16),
we
can
get
e
=
CT =
(a
,1
C
2,
Ce
T))
as
follows:
1 =
1,
oet
[1
-
b,tbt-
t-.
-bt,IceI]/
/d-
t
=
2,
,q
ot
=
[1
-
bt,t-aoet-I
-
bt,t-qat-q]/v
t= q
+
1,...,
T.
(20)
They.*andX,* are the ame s in 14). The OLS regressionn the **)
transformedquation s equivalento theGLS regression
n theoriginal
equation 4).
The
above
ransformationepends n the utocovarianceunctionf
Pit,
that
s,
-ys(s
0,
1,
.
,q). Therefore, or he
transformationo be feasible,
we must
et
stimates
f
ys.
This
ooks ike difficultroblem ecause t s
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 7/14
ERROR COMPONENTS
WITH
MA(q)
DISTURBANCES 401
not lear
how
one can
separate
he ndividualffect
,i
rom
he
emainder
error erm
it.
We
make
he
followingimplification
hat nables s to
get
consistentstimatesf the
y7's
as well s
u,2
yusing nly east-squarese-
siduals.
Denote by Q5
=
E(uitui,
-s)
the autovariancefunction f
ui,.
Then
we
have
Ur2
+
T,S
=
0,
1, ... .,
q
QS
=
or2
ss
q +
1.
(21)
Qs
can
be consistently
stimated
y5
N T
Qs
=
E(uituiut-s)
E N O0's?
T-1
(22)
i=1I
t=s?
IJV(T
-s)
where he
uit's
enote he
OLS
residuals.
rom
21)
and
22),
we
get onsis-
tent
stimates
f
ys,
2
=a
yo
nd
u2
as follows:
es
=
Qs
-
Q
s
q;
for some r>
q
Uyy2
Q,
for some r
> q. (23)
We eethat
he
stimation
fthe ariance
omponent
2
andthemovingv-
erage arameters
s
s simple, equiringnlyOLS residuals,ndonedoes
notneed o recoverhe
Xt's
f the
MA(q) process.
fter
etting,
Ts
and
"2=
or
=yo one computes
-=
o
for
=1,...
., q. Next, one computes
t
and
bt,s
from
17), 18),
and
19)
and
performs
he
C
transformation
e-
scribed
n
16). Finally,
ne obtains
2
=
l2
from20) andestimate
2
by
c2
=
2
+
2. By using
=
1
-
(4r-y/a,>),
thefinal ransformation
(**)
canbe
performed
s
described
n
14).
The
OLS
estimatef
B
rom he
(**)
transformedquation
s
thus
quivalent
o theGLS estimatefthe
rig-
inal
equation 4).
The advantages
f our
approach
are
by
now evident: 2
-
yo
is
trivially
obtainedfrom
OLS residuals.
This s
because
we
did
not
choose
a,2
as
in
Baltagi nd
Li
[3].
Next,
weestimated
's
atherhanX's.The 's
require
only
inear
east
quares,
whereasheX's
require
onlineareast
quares.
i-
nally,
ur
proposed
stimation
rocedureequiresimple
ecursive
ransfor-
mations hat re
very asy
o
program.
his hould
rove
seful
or
anel
data users.
4. MONTE-CARLORESULTS
This ection
eports
onte-Carloesults
n the
erformance
f
our
proposed
feasible
LS estimatorelative
o
true
GLS.
Weusethe
ollowingimple
e-
gressionquation
Yit
=
a
+
fXit
+
Uit
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 8/14
402
BADI H.
BALTAGI AND Ql Li
where
=
5 and
/3
0.5, nd similarata-generating
rocesss that nNer-
love 16],namely
it
=
0.1 +
0.5Xi,t_1 wit
where
it
s uniformlyistrib-
uted n the nterval-0.5,0.5]. The nitial aluesXi0were hosen s (5 +
10wio). it ,ui vit
with
i
-
IIN(0,o,2)
nd
vit
=
Eit
+
XjEi,t,j
+
.. .
+
XqEi,t-q
with
it
IIN(0,o2).
We focus
n the
MA(1)
and
MA(2)
models
(q
=
1 and 2) in ourexperiments.e fix
he otalvariance cross xperi-
ments o be
0r2
+
o2
=
20.
Note
hat
2
=
(1
+
X2
+
...
+
X2)r2. Next,we
vary
=
or2/(u2
+ ur2)
to take
the
values
0.2,0.5,0.8),respectively.
or
sample ize,
we
choose
N, T)
=
(100,6), 100,12),
250,6),
nd
250,12) e-
causeN>>
T in
a
typicalonsumer
r abor
panel.
or
each
xperiment,
e
calculate
hemean
quare
rror
MSE)
of an estimator
by
MSE(3) =
Z1)2
As
00)2/M,
where
0
= 0.5 s the rue alue f
A3,
is the stimator
of
in
the
th
replication,
nd
M
=
1000s thenumberfreplications.6or
each
replication,
he
following
stimators
f
B
re considered:
i)
the
OLS
estimator;ii)
our
proposed
easible LS estimatorescribed
n
Section
;
(iii)the rueGLS estimator
TGLS;
(iv)
the
onventionalne-way
rror om-
ponents LS estimatorf assuming o
MA(q) process
n the emainder
disturbances;
nd
(v)
the
within
stimatorf/3,
hich
reats
he
<i's
s
fixed ffects.We then ompute he
relative
fficiency
f
/3
y MSE(3)/
MSE(OTGLS).
These elative SE's aredenotedyROLS, RGLS, RCGLS,
andRW correspondingoOLS, ourproposed easibleGLS, the onventional
feasible
GLS,
and the within
stimators,
espectively.
or the
MA(1) case,
we use
r
=
2 in
(23) to get
J2,
Q2, and
for he
MA(2)
case we use
r
=
3
in
(23) to get
2
=
Q
.7 The results re reported
n
Tables
1
and
2
below.
Table
1
gives the Monte-Carlo results
for the MA(1) case:
Pit
=
(it
+
Ei,t-l,
whereX
=
0, 0.5,
and 1.
First,
et us look at the case
of
(N, T)
=
(100,6). Our proposedfeasibleGLS estimator erforms uite
well
even
for
the case whereX
=
0).
Its MSE is at most 5
1 larger
han
that
of the true
GLS for
ll cases
of X and
p
considered. he
OLS
estimator
erforms ery
badlydue to the fact hat t gnores oththe ndividual ffects
,i
ndthese-
rial
correlation
n
the remainder rror
erm
vit.
As
expected,
he
perfor-
manceof OLS deterioratess (i) X ncreases the arger
he erial
orrelation)
for fixed
alue
of
p,
or as
(ii) p increases the arger
he ndividual
ffect)
for fixed alue of X. For theconventional
rror
omponents,
easible
GLS,
which
gnores
he
MA(1) term,
as
the ame
MSE
(up
to
the
econd
decimal
in
relative
fficiency)
s that
f
trueGLS
(when
X
=
0)
for
ll
valuesof
p
con-
sidered.
However,
s
X
departs
rom and
increases,
ts
performance
ete-
riorates.This is because this estimatorgnores
he
serialcorrelation
n
Pit.
For a fixed alue of X (> 0), theperformancef this onventional easible
GLS estimator
mproves
s
p
increases
or
most ases.
This s because
when
p
is
large,
the
ndividual ffect
i
will
dominate he remainder erm ffect
Pft,
iven
hatwe fixed hetotalvariance
cross
experiments.
ut evenwhen
p
=
0.8,
the
efficiency
oss due to
ignoring
heserial orrelation
n
vit
s
quite
largecompared o our proposedfeasibleGLS
estimator. inally, heperfor-
mance of the within
stimator
s
better han
thatof OLS, and a littleworse
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 9/14
ERROR
COMPONENTS
WITH MA(q) DISTURBANCES
403
TABLE 1.
The MA(1)
case
p X ROLS RGLS RCGLS RW
N=
100,
T=6
.2 0 1.19
1.01
1.00
1.06
.2
.5 1.38
1.05
1.27
1.39
.2 1
1.85
1.02
1.64
1.77
.5
0 1.68
1.00
1.00
1.05
.5 .5
1.86
1.01
1.18
1.23
.5
1
2.83
1.01
1.65
1.69
.8
0
4.25
1.04
1.00
1.00
.8 .5 5.78 1.03 1.25 1.27
.8
1
5.78
1.02 1.56
1.58
N= 100,
T= 12
.2
0 1.08
1.01 1.00
1.01
.2
.5 1.19
1.00 1.11
1.12
.2
1
1.66
1.04
1.57 1.59
.5
0 1.44
1.00
1.00
1.01
.5
.5 1.45
1.00
1.14 1.14
.5 1 1.76 1.00 1.42 1.42
.8
0 2.97
1.00
1.00
1.00
.8
.5
2.40
1.01 1.13
1.14
.8
1
2.92
1.03
1.48
1.48
N=
250,
T=
6
.2 0 1.15
1.03
1.00
1.08
.2 .5
1.26 1.02
1.15
1.26
.2
1
1.94
1.01 1.67
1.76
.5
0 1.86 1.00 1.00
1.04
.5 .5 1.99 1.00 1.18 1.22
.5 1
2.47
1.01
1.58
1.64
.8 0 4.82 1.00
1.00
1.00
.8 .5
4.76 1.01 1.21
1.25
.8 1 5.66 1.01
1.64 1.66
N
-
250,
T
=
12
.2 0
1.10
1.00
1.00
1.01
.2 .5
1.18
1.01
1.11
1.12
.2 1 1.62 1.01 1.53 1.54
.5
0
1.49
1.00
1.00
1.00
.5
.5
1.42
1.00
1.15
1.15
.5
1 1.86
1.02
1.47
1.48
.8
0 2.88
1.00 1.00
1.00
.8
.5
2.39
1.00
1.11
1.11
.8
1
3.03
1.02 1.45
1.45
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 10/14
404 BADI H. BALTAGI
AND
Ql LI
TABLE
2.
The
MA(2) case
P
X1
X2
ROLS RGLS
RCGLS
RW
N=
100,
T=
6
.2 0 0
1.19 1.01 1.00 1.06
.2 .5
.5 1.64 1.03
1.46 1.56
.2
1
1
2.46
1.01 2.14
2.27
.5 0 0
1.68
1.01
1.00 1.05
.5 .5
.5 2.22
1.02
1.40 1.44
.5 1 1
3.56 1.02 1.97 2.01
.8 0
0 4.25 1.03
1.00
1.01
.8 .5 .5 5.74 1.03 1.43 1.45
.8
1 1
7.00 1.05 1.94
1.96
N=
100,
T=
12
.2
0
0 1.08 1.01
1.00
1.01
.2
.5 .5
1.33 1.02 1.26 1.27
.2
1
1
2.08 1.03 2.01 2.02
.5 0
0
1.44 1.01 1.00 1.01
.5
.5 .5
1.67 1.02 1.38 1.39
.5 1 1 2.32 1.02 1.96 1.96
.8
0 0
2.75
1.00
1.00 1.00
.8 .5 .5 2.59
1.05 1.26 1.26
.8 1 1
3.63 1.06
2.01
2.01
N=250,
T=6
.2 0
0
1.15 1.02
1.00
1.08
.2
.5
.5 1.59
1.03 1.41
1.51
.2 1
1
2.40
1.01
1.99 2.09
.5 0
0
1.86
1.00
1.00 1.01
.5 .5 .5 2.20 1.02 1.41 1.47
.5 1 1
3.50
1.02 2.02
2.06
.8
0
0
4.99 1.02
1.00
1.01
.8
.5 .5
5.49
1.02
1.45 1.45
.8
1
1
6.89
1.03
2.14
2.18
N=250,
T =
12
.2 0 0
1.10 1.01
1.00
1.01
.2
.5
.5
1.40
1.02
1.33 1.34
.2 1 1 2.01 1.02 1.94 1.97
.5 0
0
1.46 1.00
1.00 1.00
.5 .5
.5 1.63
1.02
1.33
1.33
.5 1 1
2.55 1.01
2.06
2.06
.8
0
0
3.06 1.01
1.00 1.00
.8 .5
.5
2.46
1.00
1.35
1.35
.8
1
1
3.56 1.03
1.94
1.95
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 11/14
ERROR OMPONENTS
WITH
MA(q) DISTURBANCES
405
than hat f theconventional
easible LS estimator. owever,
n most
cases, hewithinstimatorhows big oss
n
efficiencyue
to gnoringhe
serial orrelationn theremaindererm.
We
keep
N fixed
nd double
hevalueof T, that s, (N,T)
=
(100,12).
While
n
general
ll the stimators
mprove
nrelative SEperformance,
he
OLS estimatormproves
he
most.However,
he
rankings
f these
stima-
tors ccording
o
their
elative SE
and
the
eneral attern
bservedbove
remain nchanged. ext,wekeep
T
fixed nd ncreasehe
value fN from
100
o 250. n
general,
heresultsre quite imilar
o
the ase
of
N
=
100.
Note hat he elative SE performancefOLS, conventional
easible
LS,
and thewithinstimatorsid not how ystematicmprovement
s we n-
creased for fixed alue fT. However, urproposed easible LS esti-
mator erforms
ell
ompared
o trueGLS
and ts
relative
fficiency
s
at
most .03
for
ll
cases onsidered.
n
general,
he
ankings
f
these
stima-
tors ccording
o their elative
SE
and
the
eneral attern
bserved
bove
remain
nchanged.
Table
2
gives
he
Monte-Carlo
esults
orthe
MA(2)
case
vit
=
cit
+
XiEi,t-l
+
X2ijt-2
We
varied
I
and
X2
over
,
0.5,
and
1.
As
clear
rom
his
table,
he
onsequences
f gnoringhe erial orrelation
n
theremainder
term remuchmore ronounced
or heMA(2) case. Therelativefficiency
for = 0.8,X1
=2=
1,T= 6,andN= 100, s 7.00forOLS, 1.96for he
within
stimator,
.94 for
he onventionalrror omponent
easible LS
estimatorhat gnoresheMA(2)
process,
nd 1.05for
ur
proposed easi-
ble GLS estimator.
n
general,
ur
proposed stimatorerforms
ell or
ll
experiments,ielding
relative
SE
nolarger
han
Wo
bove hat f true
GLS. The
weak
erformance
f
OLS, within,
ndconventionaleasible LS
warns he esearcherbout he
high rice f gnoringerial orrelation
n
the
remaindererm
ven
forpanels
with
arge
N.
NO TES
1. Lillard ndWillis 12]werehe irsto considern error omponent
odelwith erially
correlated
emainderisturbances
f
heAR(l) type. lternative
ormulationsf he rrorom-
ponentmodel
with
eriallyorrelated
rrors ere onsideredyRevankar21],
Lee
[10],
nd
MaCurdy13]for
he
ingle quation
model, ndbyMagnusndWoodland15]for hemulti-
variate egression odel.
2. Thispaper ssumes hat
he rder ftheMA processs known. owever,fonefollows
themaximumikelihoodoute,ne anuse Hannan ndRissanen
8]type frecursionhich
buildsnthe
chwarz
22]BIC
criterion
t
each tage odeterminehe rder f he
ARMA
pro-
cess.More ecently,hillipsndPloberger19]derivedBayesmodel" estswhichn case of
posterior
dds ead to a
scaling
hat
s
equivalent
o theuse
of a conditionalype
f
Jeffreys
prior ndresultna generalizationftheBIC criterion.hisnew
IC criterionanbe used s
inPhillipsndPloberger20]
for
electingoth he agorder nd
deterministicrend egreen
the lassof ARMA p, q)
models
with
eterministicrend.
3. Wethank eter chmidtor uggesting
his
eneralpproach
o
our
ransformation.
4. Alternativeransformations
or he
MA(l) process
re
given
y
Pesaran
18]
ndBales-
tra
2]. Also,Ansley1]
and
Galbraithnd
Zinde-Walsh
6,7]
for
he
ARMA
p, q) process.
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 12/14
406 BADI H. BALTAGI
AND Ql
LI
5.
Throughout
he
paper, consistency
s forN
-+
oo,
withT finite r T
-*
oo.
6. We also calculate he MSE for
e,
he ntercept. he results re quite similar o thosefor
(,
and are not reported ere to save space.
7. For theMA(q) case, one can also use a linear ombination f
QT's
or
=
q
+.1,.
., T- 1
to estimate
2
. This may givea better stimator f
2
in finite amples,but
should
not make
a differencesymptotically. e performedome Monte-Carlo xperimentsn two alternative
estimators
f
2, one based on
U2'=
t=2
(T
-
)Q/=2
(T
-
T) and another ne based
on
2
2
=
Q2.
Our results how thatforN
=
100, T
=
6, the MSE of
&21
is smaller han that
of
&,22
but the relativeMSE of the corresponding
3's withrespect o trueGLS are virtually
identical.
REFERENCES
1. Ansley,C.F. An algorithm or he exact ikelihood f a mixed utoregressive-movingv-
erage process. Biometrika
6
(1979): 59-65.
2.
Balestra,
P. A
note on the exacttransformationssociated
with
he first-order oving v-
erage process.
Journal
f Econometrics
4
(1980):
381-394.
3.
Baltagi,B.H. & Q.
Li. A
transformationhat
will
circumventhe problem
f
autocorrela-
tion
n
an error
omponentmodel.
Journal
f
Econometrics
8 (1991):
385-393.
4.
Brockwell, .J.
&
R.A. Davis. Time series: Theory nd Methods. New
York:
Springer-
Verlag, 199 .
5. Fuller,W.A. & G.E. Battese.Estimation f linearmodelswith ross-errortructure. our-
nal
of Econometrics (1974): 67-78.
6.
Galbraith,
J.W.
& V. Zinde-Walsh.The GLS transformation atrix nd a semi-recursive
estimator
or
he
inear egression
odelwithARMA
errors. conometric
heory (1992a):
95-111.
7.
Galbraith,
J.W.& V.
Zinde-Walsh. ransforming
he
error-component
odel
for
estima-
tion with
general
ARMA
disturbances. npublished aper,
McGill
University, ontreal,
Quebec, Canada,
1992b.
8.
Hannan, E.J. & J. Rissanen.Recursive stimation f
ARMA
order.Biometrika
9
(1982):
81-94.
9. Hsiao, C. Analysisof Panel Data. Cambridge:CambridgeUniversity ress, 1986.
10.
Lee, Lung-Fei.
Estimation
f autocorrelated rror omponentsmodel
with
panel
data.
Working aper, Department
f
Economics, University
f
Minnesota,
1979.
11. Li, Q. Estimating linear egression odelwith eneralMA(q) disturbances.Working a-
per, University f Guelph, Ontario, Canada, 1992.
12.
Lillard,
L.A.
& R.J. Willis.Dynamic spects f earningmobility.
conometrica
6
(1978):
985-1012.
13.
MaCurdy,
T.A.
The
use
of time eriesprocesses o model the error tructure
f
earnings
in a
longitudinal
ata
analysis.
Journal
f
Econometrics
8
(1982):
83-114.
14. MacDonald, G.M. & J.G. MacKinnon.Convenientmethods or stimationf inear egres-
sion models with
MA(1)
errors.Canadian Journal
f
Economics
18
(1985):
106-116.
15. Magnus, J.R. &
A.D.
Woodland.
On
the maximum
ikelihood
stimation
f
multivariate
regressionmodels ontaining erially
orrelated rror
omponents.
nternational conomic
Review 29
(1988):
707-725.
16. Nerlove,M. Furthervidence n the stimation fdynamic conomic elations rom time-
seriesof crosssections.
Econometrica 9
(1971):
359-382.
17.
Nicholls,D.F.,
A.R.
Pagan
& R.D.
Terrell.
The
estimation nd use of models
with
mov-
ing verage
isturbanceerms:
survey.
nternational
conomic
Review16
1975):
113-134.
18.
Pesaran,
M.H. Exact
maximum
ikelihood stimation f a
regressionquation
with first-
order
moving-average
rror.
Review
of
Economic Studies
40
(1973):
529-536.
19.
Phillips,
.C.B. & W.
Ploberger.
ime
series
modeling
with
Bayesian
frame f reference:
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 13/14
ERROR COMPONENTS
WITH
MA(q)
DISTURBANCES
407
concepts,llustrations
nd
asymptotics.
owles oundation iscussionaper
No. 1038,
1992a.
20. Phillips,.C.B.
& W.Ploberger.osteriordds
estingor unit oot
with ata-based odel
selection. owles oundationiscussionaperNo. 1017, 992b.
21. Revankar, .S.
Error
omponent
odelswith
eriallyorrelatedime ffects. ournal
f
the ndian StatisticalAssociation
17 (1979): 137-160.
22.
Schwarz,
. Estimatinghe imension
f model. nnals fStatistics
(1978): 61-464.
23.
Ullah,A.,
H.D. Vinod R.S.
Singh. stimationf inearmodels
ithmovingverageis-
turbances.Journal f Quantitative
conomics
2
(1986):
137-152.
24. Wansbeek,
. & A. Kapteyn.
simple ay o obtain he pectral
ecompositionf vari-
ance
components
odels orbalanced
ata. Communications
n Statistics
ll
(1982):
2105-2112.
25. Wansbeek,
. & A. Kapteyn. note n spectral
ecompositionnd
maximumikelihood
estimationfANOVAmodels ith alanced ata. tatisticsndProbabilityetters 1983):
213-215.
APPENDIX
This
Appendixderives
standard
rthogonalizing
lgorithm
or
he
generalMA(q)
processdefined n 3). We seek a T x T matrix suchthat
v'
=
Cpiarespherical is-
turbances
with
'
-
(0,0yOIT).
To
keep
the notationmore
compact,
we
drop
the
i
subscript nd
write
t
for
Pit
and
v for
vi.
We define
r,
=
-ys/YO
with
ro
=1),
for
s
=
O,
. .
,
q.
Then
lE')
=
E
=
yo
with
=
Toeplitz
l,1,
l
..
,rp,O,.
.
,0).
De-
fine
wt
Pt/V/y
t
=
1,
. .
.,
T),
thenwe have w
-
(O,Q).
We
wantto find C ma-
trix uch
that
CDC'
=
IT.
The variance-covariance
matrix f w
(i.e., Q)
can be
summarized s follows:
rs
S
IsI
=
0,1 *...
,q
E
cot t-s)
=
0 otherwise (A.1)
with
O
5
I
and
r-s
=
rs.
Let
w*=CW,
we
want
t st=s
0
otherwise.
(A.2)
That s,we wantthetransformedrrors
,*
to be
orthogonal
o
each other
nd have
unitnorm.
We
write
w*
as a
linear ombination
f
wt
and
*
's for
r
< t. Note that
the
wt's
re not
correlated
f
they
re more han
q periods
part.Also,
note that
he
wt
dependonlyon currentndpastvaluesofcoss c t). Therefore,
*
houldhave
the
following
ecursive tructure:
'4
= coi/Vd
=
w[ct-bt,
t-14-I
-
I
-bt,c*1a]/c
(t=2,...,q)
CA)t =
[&)t
-
btI
.
-bt,t-qW
q]/Vat
(t
=
q
+
1,...
., T)
(A.3)
This content downloaded from 92.6.43.63 on Sat, 23 Jan 2016 16:12:21 UTCAll use subject to JSTOR Terms and Conditions
7/25/2019 Baltagi - Estimating Error Component Models With General MA(q) Disturbances
http://slidepdf.com/reader/full/baltagi-estimating-error-component-models-with-general-maq-disturbances 14/14
408
BADI H. BALTAGI ND Ql Li
where t s the
normalizationonstanto
ensure
(w*2)
1 for ll t. One can eas-
ily how hat
t,,
E(wtwc*)
or < t. Weoutlinehe roof ere.
Multiply
4
by
W*
forr
<
tandthen ake hemathematicalxpectation.y using (4C0*o) 0 for
t
*
s
and E(*o2)
= 1
for all s, we have
E(wco*)
=
[E(ct*)
-
t,-
-a
=
0.
Hence,
bt,7
E(wtco*)
for
maxtt
-
q,1]
c r<
t.
By therequirement
hat (c42)
=
1
for ll t, and using hefact
hat
(o2)-
=o
1
for
ll
t,
we
get
a1
=
1
tI-1-[E(wt*
1)2
[E(cotw*
]2
=
I b
2
t 1 b2,
1
~-
-t=
1
-
t-
(t=
2,..
.,q)
=
t
I
-
-[E(wt
-
1)]2
[E(ctw*
)]2 =
1-b-bt2q
(t
=
q
+
1,
.
.,
T).
(A.4)
To determinehe ransformation
ompletely,
e
also need o
know
t,,
E(Cotw4)
for < t. These re determined
ecursivelyy
the
following
ormulas.
Case (i).
For t
c q,
we
have
btj
=
E(wco)
=-E(wtco)/-a =rt
bt,t-, = E(wtw* s) = E[o?tlwt_s -bt-s,t-,-lUw* l-
b
-ts, I
*
I/at-s I
=
ts-
I
-
~
)%f;~
= [rs - bt-s,t-s- bt, -s- I bt-s, bt,,]
Ifd-a
(s
=
1, ..,
t-2). (A.5)
Case
(ii).
For
t
q
+
1,
we
have
bt,t-q
E(WtWt_q)
=
E[wt[Wt-q-
bt-q,t-q-C0t*-q-1
-bt-q,e
fl/
q]
where
Q
=
max(l,t
-
2ql.
Therefore,
bt,t-q
E((tC0t-q)/at-q
=
rq/latq
(A.6)
because
wt
and
wt-r
are
not
orrelated
or
>
q.
For
=
1,
t
2,
we have
bt,t-s
=
E(Wtw*
s) =
E[wt{cwts8
-
*s-
bt-s,t-q?t-q
-bt-s,t-q-
-
IW-
-bt-s,m(0'nJ/
ats]
where
m
=
max{l,t
-
s
-
qJ. Therefore,
bt,
-s
=
[
s
-
bt-s,
-s-
bt,
-s
-
.
-
bt-s,
-qbt,tq]
-q
*.
(A.7)
Therefore,
he
complete
ransformation
f
co*
Cw with
var(w*)
=
IT
is de-
scribed
ecursivelyy quationsA.3)
to
A.7).
Hence,
*
=
V/7_
=
v%CCv
=
can be
obtained
rom
A.3)
with'*and
vt
replacing
and
t,
respectively.