Professional Documents
Culture Documents
+
1
]
1
1
]
1
+
1
]
1
1
]
1
1
]
1
zt
yt
t
t
t
t
z
y
c c
c c
b
b
z
y
b
b
1
1
22 21
12 11
2(
1(
21
12
1
1
,ore simpl-:
(.)
t t t
X BX + +
1 1 (
Structural VAR (SVAR) or the Primitive
System
/o normali#e the 012 vector, we need to m"ltipl- the e3"ation b- inverse 4:
t t t
B X B B BX B
1
1 1
1
(
1 1
+ + , th"s:
(5)
t t t
e X A A X + +
1 1 (
VAR in standard form (unstructured VAR=UVAR).
or:
(6)
1
]
1
+
1
]
1
1
]
1
+
1
]
1
1
]
1
t
t
t
t
t
t
e
e
z
y
a a
a a
a
a
z
y
2
1
1
1
22 21
12 11
2(
1(
/hese error terms are composites of the str"ct"ral innovations from the primitive s-stem.
7hat are their characteristics8moments9
t t
B e
1
where
1
B
a
B
B
1
T
B
B
:) (
1
1
]
1
1
1
) 1 (
1
21
12
12 21
b
b
b b
VAR 1
: B
'cofactor of 4 and
T
B:) ( 'transpose.
/h"s
(;)
1
]
1
1
]
1
1
]
1
zt
yt
t
t
b
b
b b e
e
1
1
) 1 (
1
21
12
12 21 2
1
<r
zt yt
t
b
e
12
1
where 12 21
1 b b
zt yt
t
b
e
21
2
s ' are white noise, th"s es are ) , ( (
2
i
:
( ) (
it
e E
2
2 2
12
2
2
2 2
12
2
2
1 1
) (
) ( ) (
+
z y zt yt
t t
b b E
e E e Var
is time independent, and the same is tr"e for
) (
2t
e Var
.
4"t covariances are not #ero:
$ovar (
) ( )= )( >(
) ( ) , (
2
2
21
2
12
2
21 12
2 1 2 1
y z yt zt zt yt
t t t t
b b b b E
e e E e e
.
2o the shoc&s in a standard VAR are correlated. /he onl- wa- to remove the correlation and
ma&e the covar'( is if we ass"me that the contemporaneo"s effects are #ero:
(
21 12
b b
.
/he var8covar matri! of the VAR shoc&s:
1
]
1
2
2 21
12
2
1
.
Identification
7e can estimate (6) with <02, since the R12 consists of predetermined variables and the
error terms are white noise. /he errors are seriall- "ncorrelated b"t correlated across
e3"ations. Altho"?h 2@R co"ld be "sed in these cases, here we do not need it since all the
R12 variables are identical, so there is no efficienc- ?ain in "sin? 2@R over <02. 4"t we
cannot "se <02 to estimate the 2VAR beca"se of contemporaneo"s effects, which are
correlated with the s ' (str"ct"ral innovations).
<"r ?oal:
/o see how a str"ct"ral innovation it
+
1
]
1
1
]
1
+
1
]
1
1
]
1
1
]
1
zt
yt
t
t
t
t
z
y
c c
c c
b
b
z
y b
1
1
22 21
12 11
2(
1( 12
1 (
1
1
]
1
1
1
) 1 (
1
21
12
12 21
1
b
b
b b
B
1
]
1
1 (
1
12
b
.
1ence the VAR s-stem in standard form can be written:
(C)
1
]
1
+
1
]
1
1
]
1
+
1
]
1
1
]
1
zt
zt yt
t
t
t
t
b
z
y
c c
c b c c b c
b
b b b
z
y
12
1
1
22 21
22 12 12 21 12 11
2(
2( 12 1(
) ( ) (
*f we match the coefficients in (C) with the estimates in (6)
1
]
1
+
1
]
1
1
]
1
+
1
]
1
1
]
1
t
t
t
t
t
t
e
e
z
y
a a
a a
a
a
z
y
2
1
1
1
22 21
12 11
2(
1(
, we can e!tract the coefficients of the 2VAR:
2( 12 1( 1(
b b b a
2( 2(
b a
z y
e b e
12 1
21 12 11 11
c b c a
21 21
c a
z
e e
2
22 12 12 12
c b c a
22 22
c a
12
Cov '
2
2
21
2
12
) (
+
y z
b b
'
2
12 z
b
Impulse response functions
7e want to trace o"t the time path of the effect of str"ct"ral shoc&s on the dependent
variables of the model. For this, we first need to transform the VAR into a V,A
representation.
Rewrite the @VAR more compactl-:
L A I
e
L A I
A
X e X A A X
t
t t t t
1 1
(
1 1 (
) 5 (
+ +
VAR +
First, consider the first component on the R12:
12 21 22 11
2(
1(
22 12
21 22
22 21
12 11
(
22 21
12 11
1
( 1
(
1
1
1
(
) 1 )( 1 (
1
1
1
1
1
1
) (
) (
a a a a
a
a
a a
a a
a a
a a
A
a a
a a
A I
A A I
A A I
A I
A
a
1
]
1
1
]
1
1
]
1
1
]
1
1
]
1
+
+
z
y
a a a a
a a a a
2( 22 1( 12
2( 21 1( 22
) 1 (
) 1 (
1
2tabilit- re3"ires that the roots of
L A I
1
1
]
1
i t
i t
i
i i i t
i t
e
e
a a
a a
e A
L A I
e
, 2
, 1
(
22 21
12 11
( 1
1
7e can th"s write the VAR as a V,A with the standard VARs error terms.
(A)
1
]
1
1
]
1
+
1
]
1
1
]
1
i t
i t
i
i
A
t
t
e
e
a a
a a
z
y
z
y
i
, 2
, 1
(
22 21
12 11
4"t these are composite errors consistin? of the str"ct"ral innovations. 7e m"st th"s replace
the es with the s ' from (;)
t t
b
b
e
1
]
1
1
1
1
21
12
(Aa)
1
]
1
1
]
1
+
1
]
1
1
]
1
i t z
i t y
i
i
t
t
i
b
b
b b
A
z
y
z
y
,
,
(
21
12
21 12
1
1
1
1
]
1
1
]
1
+
1
]
1
i t z
i t y
i
i i
i i
i
z
y
,
,
) (
22
) (
21
) (
12
) (
11
(
i t
i i
X
+
(
.
Impact multipliers
/he- trace the impact effect of a one "nit chan?e in a str"ct"ral innovation. D!: find the
impact effect of t z,
on t
y
and t
z
:
) ( (
12
,
t z
t
d
dy
) ( (
22
,
t z
t
d
dz
) 1 (
22
,
1
+
t z
t
d
dz
VAR .
Eote that this is the same effect on t
y
and t
z
of a str"ct"ral innovation one period a?o:
) 1 (
12
1 ,
t z
t
d
dy
) 1 (
22
1 ,
t z
t
d
dz
,
) 1 (
12
,
) 2 (
12
,
Cumulated effect is the s"m over *R f"nctions:
n
i
i
( 12
) ( .
Lon!run cumulated effect:
n
lim
n
i
i
( 12
) (
*n practice we cannot calc"late these effects since the 2VAR is "nderidentified. 2o we m"st
impose additional restrictions on the VAR to identif- the imp"lse responses.
*f we "se the $holes&- decomposition and ass"me that - does not have a contemporaneo"s
effect on #, then
(
12
b
. /h"s the error str"ct"re becomes lower trian?"lar:
(1()
1
]
1
1
]
1
1
]
1
zt
yt
t
t
b
e
e
1 (
1
12
2
1
/he y
shoc& doesnt affect # directl- b"t it affects it indirectl- thro"?h its la??ed effect in
VAR.
"raner Causality: *f the # shoc& affects e1, e2 and the - shoc& doesnt affect e2 b"t it
affects e1, then # is ca"sall- prior to -.
!"ample#
$alc"late the imp"lse response f"nctions on
F G
t
y
,
F G
t
z
of a "nit chan?e in # shoc&
) (
zt
b
b
#E #E
Cov
Altho"?h this information is s"fficient to calc"late the imp"lse responses in this simple
model, we can e!tract all of the coefficients of the primitive s-stem as follows:
VAR 5
(
2( 1(
a a (
2(
b
( (
1( 2( 12 1(
b b b b
; . (
22 22
c a and 2 . (
21 21
c a
From
21 12 11 11
; . ( c b c a we ?et 5. . (
11
c .
From
21 12 11 11
; . ( c b c a
and +6 . ( ) ; . ( ( C . ( ; . (
12 12 22 12 12 12
+ c c c b c a and 5. . (
11
c
2"bstit"te 12
b
into (1() to ?et:
zt yt t
e C . (
1
+
zt t
e
2
A 1 "nit zt
1
,
t z
t
d
dz
+
n
i
t z
i t
d
dy
... ;5 . ( C6 . ( 1
(
,
+ + +
+
n
i
t z
i t
d
dz
1
]
1
+
1
]
1
1
]
1
i t z
i t y
i
i
i
t
t
t
i
b
b
b b
A
z
y
z
y
$
,
,
(
21
12
21 12
1
1
1
1
]
1
1
]
1
+
1
]
1
i t z
i t y
i
i i
i i
i
z
y
,
,
) (
22
) (
21
) (
12
) (
11
(
or
i t
i i t
X $
+
(
.
7e want to calc"late the n%period forecast error of ! in order to find that of sa-, -.
2tart from 1 period:
...
1 2 1 1 ( 1
+ + + +
+ + t t t t
X $
...
1 2 1 1
+ + +
+ t t t t
X $ E
1%period forecast error 1 ( 1 1 + + +
t t t
E$ $
Lroceed in the same wa- and ?et 2%period forecast error:
1 1 2 ( 2 2 + + + +
+
t t t t
E$ $
+%period forecast error:
1 2 2 1 + ( + + + + + + +
+ +
t t t t t
E$ $
n%period forecast error:
+ + + + + + +
+ + + +
1
( 1 1 2 2 1 1 (
...
n
i i n t t n n t n t n t n t n t
E$ $
Eow consider -, the first element of the ! matri!. *ts n%step%ahead forecast error is:
VAR C
) ... (
1 , 1 , 11 1 , 1 , 11 , ( , 11 + + + + +
+ + +
t y n n t y n t y n t n t
Ey y
) ... (
1 , 1 , 21 1 , 1 , 21 , ( , 21 + + +
+ + + +
t z n n t z n t z
/he variance of its n%step%ahead forecast error is:
+ + + +
time over %ecreases
shoc& o'n to due
iance of proportion
n
y n y
var
1 , 11
2
1 , 11
2
( , 11
2 2 2
,
) ... (
time over "ro's
shoc& z a to due
iance of proportion
n
z
var
1 , 21
2
1 , 21
2
( , 21
2 2
) ... ( + + +
*f z
dof ' ) (
2
p h n n'Mvariables, h'Mma! chosen
la?s, p'order of the VAR.
Eot a ?ood statistics to "se if there is a 3"asi%"nit root (re3"ires hi?h order ,A
coefficients to be ().
A"tocorrelation 0, /est.
VAR 1(
E"ll h-pothesis:
no a"tocorrelation "p to la? h.
0, statistics distrib"ted
2
dof '
2
n .
Eormalit- tests
,"ltivariate version of the Rar3"e 4era tests. *t compares the +
rd
and .
th
moments
(s&ewness and &"rtosis) to those from a normal distrib"tion. ,"st specif- a
factori#ation of the resid"als. $hoices in Dviews:
$holes&-: the statistics will depend on the orderin? of the variables.
Noorni& and 1ansen (A.) S*nverse 2QR/ of resid"al correlation matri!:
invariant to the orderin? of variables and the scale of the variables in the
s-stem.
@r#"a (A;)% *nverse 2QR/ of resid"al covariance matri!: same advanta?e as
Noornic& and 1ansen, b"t better.
Factori#ation from 2VAR (later: need to have estimated an 2VAR)
.. Kran?er $a"salit-
*n a two%variable VAR(p)/he process
F G
t
z
does not K%ca"se
F G
t
y
if all coefficients in
( ) (
12
L A
(or a Joint test of
( ) ( ... ) 2 ( ) 1 (
21 21 21
p a a a
at all la?s is not reJected).
/his concept involves the effect of past val"es of # on the c"rrent val"e of -. 2o it answers
the 3"estion whether past and c"rrent val"es of # help predict the f"t"re val"e of -.
*t is different from e!o?eneit- tests, which loo& at whether the c"rrent val"es of #
e!plains c"rrent and f"t"re val"es of -.
*n a n%variable VAR(p), )loc*+e"o&eneity ('bloc&%K%ca"salit-) test loo&s at whether the
la?s of an- variables K%ca"se an- other variable in the s-stem. Po" can test this "sin? the
0R test in (1().
Application
$reate a bivariate VAR(1) and appl- the tests to ?et the best specification of the model.
7or&file:DENDR2Q@AR/DR0P.wf
%Kenerate the rate of ?rowth of mone- s"ppl-
m'lo?(,1E2At)%lo?(,1E2At%1)
%Kenerate the rate of LL* inflation:
*nf'lo?(LL*t)%lo?(LL*t%1)
%Kenerate seasonal d"mmies for each 3"arter of the -ear:
Ni'Tseas(i), where i'1,2,+
or: enders3"artd"mmies.pr?
smpl @all
inf=log(ppi)-log(ppi(-1))
VAR 11
m=log(m1nsa)-log(m1nsa(-1))
for !j=1 to 4
series d{!j}=@seas({!j})
next
%$hec& whether m and inf are *(()
Eow we can create o"r bivariate VAR(1):
Dndo?eno"s variables: m, inf
D!o?eno"s variables: constant, . seasonal d"mmies
Dstimate an "nrestricted VAR
1. /est the la? len?th
/he se3"ential 0R statistics indicates 5 la?s. Also confirmed b- info criteria.
View%0a? 2tr"ct"re%0a? len?th criteria%la?s to incl"de >C=%<U
VAR Lag rder !ele"tion #riteria
$ndogeno%s &aria'les( )*+ ,
$xogeno%s &aria'les( # -1 -. -/
!ample( 101231 .22.34
)n"l%ded o'ser&ations( 112
Lag LogL LR +4$ A)# !# 53
2 0.672202 *A /78.e-29 -11749611 -117///98 -1174.819
1 00/71648 1.97//20 1711e-29 -1.7.620/ -1.7242.0: -1.7166.9
. 12227/91 1.76410. 1788e-29 -1.7/2461 -117006.4 -1.716090
/ 121270.2 10761194 174/e-29 -1.7/9182 -1.722.11 -1.7./241
4 12197129 1/7.062. 17/9e-29 -1.74.1/8 -11701826 -1.7./024:
8 12./7612 127..//4: 17/8e-29: -1.7441/6: -117029.. -1.7..698
1 12.87029 /7081.19 17/9e-29 -1.74./98 -11792991 -1.716412
6 12.97611 8726.614 1742e-29 -1.742086 -11761611 -1.71.911
9 12//7419 9714./.1 17/0e-29 -1.74166/ -11714904 -1.712888
: indi"ates lag order sele"ted '; t<e "riterion
LR( se=%ential modified LR test statisti" (ea"< test at 8> le&el)
+4$( +inal predi"tion error
A)#( A?ai?e information "riterion
!#( !"<@arA information "riterion
53( 5annan-3%inn information "riterion
Po" ma- have priors and want to test for la? len?th -o"rself "sin? a 0R test. 2"ppose we
start with 12 la?s and compare it with C la?s.
$alc"late the determinant of the resid"al covariance matri!:
VAR 12
Dviews ?ives it at the bottom of the estimation
,
_
t t t
p T
' V V
1
det
V
with p'M parameters per e3"ation in the VAR. /he "nadJ"sted
i?nores p.
$stimate @it< 1. lags t<e %nrestri"ted VAR
Ve"tor A%toregression $stimates
-ate( 24B1.B26 Cime( 1.(1.
!ample (adj%sted)( 101/3. .22.31
)n"l%ded o'ser&ations( 181 after adj%stments
!tandard errors in ( ) D t-statis E
-eterminant resid
"o&arian"e (dof adj7) 1712$-29
-eterminant resid
"o&arian"e 6741$-20
Log li?eli<ood 12167820
A?ai?e information
"riterion -1.7/.624
!"<@arA "riterion -117./...
Dstimate with C la?s over the same VAR over the same sample:
-eterminant resid
"o&arian"e (dof adj7) 1712$-29
-eterminant resid
"o&arian"e 9741$-20
Log li?eli<ood 12//7419
A?ai?e information
"riterion -1.74166/
!"<@arA "riterion -11714904
Co do t<e "omparison properl;F @e m%st %se t<e same sample of 1. lags (101/=. .22.=1)
-eterminant resid
"o&arian"e (dof adj7) 1714$-29
-eterminant resid
"o&arian"e 9718$-20
Log li?eli<ood 122874..
A?ai?e information
"riterion -1.7/66.1
!"<@arA "riterion -117808.2
+orm t<e LR test statisti"s( ) ( ) ) ln )(ln (
2
( m T L)
u r
(m+,parameteres in each
e(uation of the unrestricted system-constants.
2
. M n ns restrictio (
, n+,variables)
VAR 1+
0R'>156%(.B(2!12))=>ln(9718$-20)-ln(6741$-20)E =10792 G "<is=r((1.-9)x4)="<is=r(11)= /4
-o not reje"t t<e n%ll7
2. /est the si?nificance of the d"mmies "sin? the same 0R test.
+. Nia?nostic tests of the resid"als
View%Resid"al tests%
Portmanteau test#
VAR Resid%al 4ortmantea% Cests for A%to"orrelations
52( no resid%al a%to"orrelations %p to lag <
!ample( 101/3. .22.34
)n"l%ded o'ser&ations( 181
Lags 3-!tat 4ro'7 Adj 3-!tat 4ro'7 df
1 272///11 *A: 272//866 *A: *A:
. 27/11.66 *A: 27/.2116 *A: *A:
/ 17191./9 *A: 17.26198 *A: *A:
4 17680211 *A: 17608299 *A: *A:
8 /7.18180 *A: /7.00/01 *A: *A:
1 /76/1089 *A: /794.216 *A: *A:
6 47/66006 *A: 4781/... *A: *A:
9 87.008/4 *A: 8749486. *A: *A:
0 87968262 27.296 17208/48 2710.1 4 (14791)
12 07984/00 27.684 127/46./ 27.418 9 (.170)
11 167./001 271429 197.0/29 271261 1.
1. 19791282 27.691 10700482 27..28 11
:C<e test is &alid onl; for lags larger t<an t<e VAR lag order7
df is degrees of freedom for (approximate) "<i-s=%are distri'%tion
*ot reje"t t<e n%ll7
LM test
VAR Resid%al !erial #orrelation L,
Cests
52( no serial "orrelation at lag order <
!ample( 101/3. .22.34
)n"l%ded o'ser&ations( 181
Lags L,-!tat 4ro'
1 .7/.62.9 271680
. 47911900 27/219
/ 187/212. 272241
VAR 1.
4 87480/91 27.4//
8 07.61611 272846
1 .74..11. 271898
6 /7164/0/ 278.01
9 .72018.. 276190
0 276.60.1 270469
12 4718011/ 27/.41
11 976611.. 272161
1. 17028.91 2768/.
4ro's from "<i-s=%are @it< 4 df7
#<is=r(4)=14791
,ostl- not reJect the n"ll.
,ormality (est
VAR Resid%al *ormalit; Cests
rt<ogonaliAation( Resid%al #orrelation (-oorni?-5ansen)
52( resid%als are m%lti&ariate normal
!ample( 101/3. .22.34
)n"l%ded o'ser&ations( 181
#omponent !?e@ness #<i-s= df 4ro'7
1 -2726122/ 2714./61 1 276280
. 272966/4 27.161/. 1 27141.
Hoint 27/80820 . 279/88
#omponent I%rtosis #<i-s= df 4ro'7
1 /7121299 /79/.192 1 27282/
. 171081/8 .17421/2 1 272222
Hoint .87.//09 . 272222
#omponent Har=%e-Jera df 4ro'7
1 /7068286 . 271/62
. .171194/ . 272222
Hoint .8780/40 4 272222
C<e n%ll is a joint test of 'ot< t<e s?e@ness and t<e ?%rtosis7
VAR 15
*ormalit; not reje"ted for inf '%t reje"ted for m d%e to ?%rtosis pro'lem7
*s this somethin? we sho"ld worr- abo"t9 in principle reJection of normal distrib"tion invalidates the
test statistics. 4"t meas"res of s&ewness are fo"nd to be not informative in small samples (4ai, E?
4oston $olle?e 7L 115, 2((1).
.. -ran&er causality
View%la? str"ct"re%K%ca"salit-8bloc& e!o?eneit- test
VAR Kranger #a%salit;BJlo"? $xogeneit; Lald Cests
!ample( 101/3. .22.34
)n"l%ded o'ser&ations( 181
-ependent &aria'le( )*+
$x"l%ded #<i-s= df 4ro'7
, 67126888 8 27.1.9
All 67126888 8 27.1.9
-ependent &aria'le( ,
$x"l%ded #<i-s= df 4ro'7
)*+ 167084.2 8 2722/2
All 167084.2 8 2722/2
#<is=r(8)=11768
*t tests bilaterall- whether the la?s of the e!cl"ded variable affect the endo?eno"s variable.
/he n"ll: the la??ed coefficients are si?nificantl- different than (.
All: Joint test that the la?s of all other variables affect the endo?eno"s variable.
D!: on top panel, first row shows if la??ed variables of , are si?nificantl- different than (,
the second row shows if la??ed variables of all variables other than *EF are #ero (in o"r case
both tests are identical since we onl- have two variables).
For both the n"ll is reJected, tho"?h there is some evidence abo"t effect of inf on m at 1( H
si?nificance level.
VAR 16