Академический Документы
Профессиональный Документы
Культура Документы
= (1
. 2
), whoio
1 ano 2 aio voclois. 1ho following iosull is woll known (o.g. Woiss, 200, pp 88-886).
2 Population
8
ECB
Working Paper Series No 1409
December 2011
IroosItIon 1 ;1n& J Jn| 1n:.nn) 5& hn l
= (1
. 2
) . nn (1. c)-n|&J
:nnJn : J]nJ n n :hnh.|.j n (!. T. 1), nnJ hn var (l) r.. Jhn
var (1 ) = L
Z
[var (1 [2)[ var
Z
[L(1 [2)[ . (1)
1ho noxl lhioo iosulls aio conlingonl on:
CondItIon 1 \ = (\
. 1
. 2
= (A
. . . . . A
a
) . (6)
Foi : = 2, laking 1 = A
ano 2 = A
in (1) loaos lo
var (A
) = L
A
[var (A
[A
)[ var
A
[L(A
[A
)[ (7)
Foi lho caso : = 3, laking 1 = A
, \ = A
ano 2 = A
in (4) yiolos
var (A
) = L
A
,A
[var (A
[ A
. A
)[ L
A
var
A
[L(A
[ A
. A
)[
var
A
[L(A
[ A
)[ . (8)
Now apply Iioposilion 8 lo lho loaoing loim of (8), wilh 1 = A
, \ = (A
. A
)
ano 2 = A
, lo oLlain
var (A
) = L
A
,A
,A
[var (A
[ A
. A
. A
)[
L
A
,A
var
A
[L(A
[ A
. A
. A
)[
L
A
var
A
[L(A
[ A
. A
)[ var
A
[L(A
[ A
)[ . (0)
1his piocoouio can Lo iloialoo, as follows.
IroosItIon 5 G.n nJ..n . nnJ h n:..n ;),
var (A
) = L
A
,...,A
[var (A
[ A
. . . . . A
a
)[
L
A
,...,A
var
A
[L(A
[ A
. . . . . A
a
)[
. . . L
A
,...,A
_
var
A
[L(A
[ A
. . . . . A
)
)[
_
. . . L
A
var
A
[L(A
[ A
. A
)[ var
A
[L(A
[ A
)[ .
= L
A
,...,A
[var (A
[ A
. . . . . A
a
)[ (10)
)
L
A
,...,A
_
var
A
[L(A
[ A
. . . . . A
)
)[
_
. (11)
IrooI. 1ho caso : = 4 is (0). Apply Iioposilion 8 lo lho loaoing loim (10), wilh
\ = (A
. . . . . A
a
), 2 = A
a
ano 1 = A
,...,A
[var (A
[ A
. . . . . A
a
)[
L
A
,...,A
_
var
A
[L(A
[ A
. . . . . A
a
) [ A
. . . . . A
a
[
_
.
foi (10). Ly inouclion, lho pioposilion is liuo.
1ho foiogoing iosulls aio quilo gonoial, ooponoing only on Conoilions 1 ano 2. In
lho spocial caso lhal 1 in Conoilion 2 is Gaussian, wo havo
r
= o
a
)
,
)
r
)
10
ECB
Working Paper Series No 1409
December 2011
where is independent of (r
2
. . . . . r
n
), E(r
1
j r
2
. . . . . r
n
) = o +
n
j=2
,
j
r
j
, E() =
E( j r
2
. . . . . r
n
) = 0, and var () = var (r
1
j r
2
. . . . . r
n
). Let
j
j
= r
j
E(r
j
j r
1
. . . . . r
j1
) (, = 2. . . . . :) .
Then there is a one-to-one linear transformation between (r
2
. . . . . r
j
) and
_
j
2
. . . . . j
j
_
.
In particular,
r
1
=
n
j=2
c
j
j
j
+ ,
where c
j
= co
_
r
1
. j
j
_
,var
_
j
j
_
and var (r
1
) =
n
j=2
c
2
j
var
_
j
j
_
+ var (). Moreover,
E(r
1
j r
2
. . . . . r
j+1
) = E(r
1
) +
j+1
i=1
c
i
j
i
and
var
x
j+1
[E(r
1
j r
2
. . . . . r
j+1
) j r
2
. . . . . r
j
]
= var
x
j+1
_
j+1
i=2
c
i
j
i
j j
2
. . . . . j
j
_
= c
2
j+1
var
_
j
j+1
_
.
None of these terms involve r
j
or j
j
, because in the Gaussian distribution conditional
variances do not depend on the values of the variables conditioned upon.
Thus the leading term (10) in Proposition 5 reduces to var () in the Gaussian case,
and term , in the sum (11) is c
2
j+1
var
_
j
j+1
_
. Were we to divide the equation in Propo-
sition 5 by var (r
1
), then (10) would be 1 1
2
, where 1
2
is the coecient of multiple
correlation between r
1
and (r
2
. . . . . r
n
); the term , in (11) would be the increment to
population 1
2
when r
j+1
is introduced into the set of regressors that already contains
r
1
. . . . . r
j
.
3 Simulation
All of these population decompositions have analogs in simulation. This is important for
the Bayesian applications that motivate this work. What follows takes the components
A
1
. . . . . A
n
to be scalars. The vector case exactly parallels the scalar case but entails
more awkward and space-consuming notation.
Suppose that it is feasible to simulate
r
(m)
j
s j (r
j
j r
2
. . . . . r
j1
) (: = 1. 2. . . . ; , = 2. . . . . :) , (12)
where the case , = 2 is unconditional simulation, as well as
r
(m)
1
s j (r
1
j r
2
. . . . . r
n
) (: = 1. 2. . . .) . (13)
11
ECB
Working Paper Series No 1409
December 2011
Wo shall supposo lhal lho ianoom soquoncos
_
r
n
)
_
aio oigooic. 1hon lhoio aio simu-
lalions lhal paiallol lho oocomposilions in lho piovious soclion, which piovioo molhoo
of momonls oslimalos of lho loims in lho populalion oocomposilion givon in Iioposilion
.
Foi lho simplosl caso : = 2, lhis involvos simulaling r
n
(:
= 1. . . . . `
), ano
conoilional on oach r
n
simulaling r
n
,n
(:
= 1. . . . . `
conoilional on r
n
. Donolo
r
n
= `
r
n
,n
ano r
= `
r
n
= (`
r
n
,n
.
1hon
(`
_
r
n
,n
= (`
_
r
n
,n
r
n
_
r
n
. (14)
1ho loims in lhis iolalionship conslilulo consislonl oslimalos of lho loims of lho oqua-
lion in Iioposilion foi lho caso : = 2, oquivalonlly foi lho loims of lho oqualion in
Iioposilion 1 wilh 1 = r
ano 2 = r
.
Whon lho oisliiLulion of (r
. r
[ r
. r
n
_
fiom lho joinl oisliiLulion of r
ano r
ooponos
in a nonliivial way on lho valuo of r
ano :
.
Noxl consiooi lho caso : = 3. 1his caso luins oul lo Lo signilcanl, Locauso lho
calculalions hoio aio ossonlially lhoso lhal aio ioquiioo foi lho gonoial caso. Coiio-
sponoing lo oach simulalion r
n
(:
= 1. . . . . `
(:
= 1. . . . . `
conoi-
lional on r
n
,n
,n
,n
(:
= 1. . . . . `
conoilional on
_
r
n
,n
. r
n
_
.
Coiiosponoing lo lhoso simulalions oolno lho (conoilional) samplo moans
r
n
,n
= `
r
n
,n
,n
(:
= 1. . . . . `
: :
= 1 . . . . `
) ,
r
n
= `
r
n
,n
(:
= 1. . . . . `
) ,
r
= `
r
n
.
12
ECB
Working Paper Series No 1409
December 2011
1hoso aio simulalion-consislonl appioximalions of L
_
r
[ r
n
. r
n
,n
_
, L
_
r
[ r
n
_
ano L(r
_
r
n
,n
,n
r
_
. (1)
Coiiosponoing lo lho lhioo loims on lho iighl sioo of (8), lho simulalion-consislonl
appioximalion of L
A
,A
[var (A
[ A
. A
)[ is
(`
_
`
_
r
n
,n
,n
r
n
,n
_
; (16)
lho simulalion-consislonl appioximalion of L
A
var
A
[L(A
[ A
. A
)[ is
`
_
`
_
r
n
,n
r
n
_
; (17)
ano lho simulalion consislonl appioximalion of var
A
[L(A
[ A
)[ is
`
_
r
n
. (18)
Consislonl wilh lho ioonlily in (8), lho sum of (16) (18) is (1). 1ho compulalions
(16)-(18) aio implomonloo wilh lhioo linos of cooo in lho MallaL funclion pioviooo in lho
Apponoix of lhis papoi. 1aking `
= `
= `
is ioplacoo
Ly a vocloi of mooosl sizo (o.g., 10 oi 12 compononls) lho compulalions ioquiio woll
unooi ono socono.
1his piocoss coulo Lo iloialoo lo lho gonoial caso of Iioposilion 2. 1ho oulcomo
woulo Lo a simulalion samplo iooally suiloo lo mulli-facloi analysis of vaiianco wilh a
Lalancoo oosign, Lul lho simulalion ioquiiomonls aio ovoiwholming o.g. lho numLoi of
simulalions ioquiioo is
_
`
a
`
_
, (` 1) in lho caso whoio `
= . . . . `
a
= `.
1his is impiaclical foi : much laigoi lhan 3 ano ` sumcionlly laigo lo piovioo ioliaLlo
appioximalions. Insloao consiooi loim , of (11),
L
A
,...,A
_
var
A
[L(A
[ A
. . . . . A
)
)[
_
ano nolo lhal wo can uso lho appioach in lho piovious paiagiaph, ioplacing r
wilh
(r
. . . . . r
)
) ano r
wilh r
)
. 1ho ioquisilo simulalions aio all possiLlo Ly viiluo of lho
assumplion lhal simulalions aio possiLlo fiom lho oisliiLulions (12) (18), maoo al
Loginning of lhis soclion. In lho caso `
= . . . . `
a
= `, lhis onlails :`
simulalions,
a numLoi lhal is linoai in : ano is ioasonaLlo in lho applicalions lhal molivalo lhis
woik.
13
ECB
Working Paper Series No 1409
December 2011
4 Application to Bayesian inference
These decompositions are useful tools in the interpretation of posterior distributions
that are accessed by means of simulation, as is the case in most Bayesian work. Let
Y
T
= [y
1
. . . . . y
T
] denote the observables (random ex ante) where 1 is the size of the
sample, Y
o
T
their observed values (xed ex post), the model,
A
2
A
the para-
meter vector, j (
A
j ) the prior density, and j (Y
T
j
A
. ) the distribution of ob-
servables conditional on the parameters. The posterior density of the parameters is
j (
A
j Y
o
T
. ) / j (
A
j ) j (Y
o
T
j
A
. ), and we assume that a posterior simulator is
available that generates an identically distributed ergodic process
(m)
A
s j (
A
j Y
o
T
. )
(: = 1. 2. 3. . . .). For importance sampling, there are obvious modications to the com-
putation of simulation moments involving the weighting function, and these will also be
consistent if the usual regularity conditions (Geweke, 2005, Theorem 4.2.2) are satised.
From a formal perspective, Bayesian inference is always undertaken to inform a policy
decision. Let ! denote the random vector pertinent to the loss function 1(!) governing
the decision: e.g., for a central bank, ! could consist of measures of output and ination
in some future quarters; for a retailer, ! could be sales of specied products in specied
markets; for a government agency seeking to adjust census count, ! might contain
measures of characteristics of the actual population; in a pure prediction problem !
consists of future values (y
T+1
. . . . . y
T+H
) over a specied horizon H. The model informs
the policy decision if and only if it species the conditional distribution j (! j Y
T
.
A
. ),
and we assume that is the case. This conditional distribution could be degenerate: for
example, if the decision involves testing the hypothesis
A
2
A1
then . = 1
A1
(
A1
).
Simulation from j (! j Y
o
T
.
A
. ) is generally straightforward and less demanding than
simulating from the posterior distribution. In many of these specic instances, expected
loss conditions on the prospective action taken by the decision-maker. For our purposes
one of a number of alternative actions is then subsumed in .
4.1 Extrinsic and intrinsic variance
The distribution relevant for decision-making conditions on the model and data Y
o
T
,
j (! j Y
o
T
. ) /
_
A
j (
A
j Y
o
T
. ) j (! j Y
o
T
.
A
. ) d
A
.
Given a posterior sample
(m
2
)
A
(:
2
= 1. . . . . `
2
), this distribution can be accessed by
means of auxiliary simulations from the model for !.
!
(m
1
;m
2
)
s j
_
! j Y
o
T
.
(m
2
)
A
.
_
(:
1
= 1. . . . . `
1
) .
The corresponding sample moments of !, and in particular the approximation
(`
1
`
2
)
1
M
1
m
1
=1
M
2
m
2
=1
1
_
!
(m
1
;m
2
)
_
14
ECB
Working Paper Series No 1409
December 2011
of E[1(!) j Y
o
T
. ] are simulation-consistent if and only if `
2
! 1; it is not necessary
that `
1
increase at all, for this purpose, and indeed `
1
= 1 is sucient.
From (7) with A
1
= ! and A
2
=
A
, and recognizing that the relevant distribution
conditions on the data Y
o
T
and the model specication ,
var (! j Y
o
T
. ) = E
A
[var (! j
A
. Y
o
T
. )] + var
A
[E(! j
A
. Y
T
. )] . (19)
We refer to the rst component on the right-hand side as the intrinsic variance of !:
it is the variation in ! that would exist if one knew the parameter vector
A
, averaged
using the posterior distribution of
A
. We refer to the second component as the extrinsic
variance of !: it is the variance in ! that is due to not knowing
A
. If the distribution
of ! conditional on Y
T
and
A
is degenerate, as is the case with conventional tests
of hypotheses about
A
, then there is no intrinsic variance. If the prior distribution is
dogmatic then there is no extrinsic variance. In most realistic cases both intrinsic and
extrinsic variance are positive.
Making the corresponding substitution in the simulation (14)
(`
1
`
2
)
1
_
!
(m
1
;m
2
)
!
_ _
!
(m
1
;m
2
)
!
_
0
= (`
1
`
2
)
1
M
2
m
2
=1
M
1
m
1
=1
_
!
(m
1
;m
2
)
!
(m
2
)
_ _
!
(m
1
;m
2
)
!
(m
2
)
_
0
(20)
+`
1
1
M
1
m
1
=1
_
!
(m
2
)
!
_ _
!
(m
2
)
!
_
0
(21)
where
!
(m
2
)
= `
1
1
M
1
m
1
=1
!
(m
1
;m
2
)
and ! = `
1
2
M
2
m
2
=1
!
(m
2
)
.
As `
1
! 1 and `
2
! 1 (20) converges to E
A
[var (! j
A
. )] in (19) and (21) to
var
A
[E(! j
A
. Y
T
. )].
4.2 Decomposition of intrinsic variance
Let the vector of interest be partitioned !
0
= (!
0
1
. !
0
2
), and suppose that it is feasible
to simulate
!
(m)
1
s j (!
1
j Y
o
T
.
A
. ) , (22)
!
(m)
2
s j (!
2
j !
1
. Y
o
T
.
A
. ) . (23)
For example in a pure prediction problem we might have !
1
= y
T+1
and !
2
= y
T+2
or !
2
= (1,4)
4
s=1
y
T+s
; or, .
1
could be a monetary policy instrument and !
2
would
consist of the remaining variables in a macroeconomic model. In the notation of the
previous section, take A
1
= !
2
, A
2
=
A
and A
3
= !
1
. Then from the particular case
(8) of Proposition 5,
15
ECB
Working Paper Series No 1409
December 2011
var (!
2
) = E
A
;!
1
[var (!
2
j !
1
.
A
)] (24)
+E
A
fvar
!
1
[E(!
2
j
A
. !
1
)]g (25)
+var
A
[E(!
2
j
A
)] . (26)
All of the moments in this decomposition condition on the data Y
o
T
and model
specication as well as the vectors explicitly indicated; we omit those terms to keep
the expressions from being unduly cluttered. The term (26) is the extrinsic variance
of !
2
, and therefore (24)-(25) provides a decomposition of the intrinsic variance of !
2
.
The bracketed term in (24) is the variance in !
2
that would remain even if one knew
both
A
and !
1
, and (24) averages this with respect to the posterior distribution of
A
and predictive distribution of !
1
. The term in braces in (25) is the variance in !
2
attributable to not knowing !
1
, and (25) averages this with respect to the posterior
distribution of
A
. More loosely speaking, (25) is the portion of intrinsic variance that
is resolved (disappears) once !
1
becomes known. As a specic instance, if 1 is the fourth
quarter of 2011, y
t
is a vector of growth rates, !
1
= y
T+1
and !
2
= (1,4)
4
s=1
y
T+s
,
then (25) is the variance in the annual growth rate for 2012 that will be resolved at the
close of the rst quarter of 2012.
Following the methods of Section 3, it is straightforward to compute a simulation-
consistent approximation of (24)-(26). Conditional on a simulation sample
(m
2
)
A
of size
`
2
from the posterior distribution of
A
, generate `
3
values !
(m
2
;m
3
)
1
from (22) and
then conditional on each of these `
2
`
3
draws generate `
1
values !
(m
1
;m
2
;m
3
)
2
from (23).
Then substituting
(m
2
)
A
for r
(m
2
)
2
, !
(m
2
;m
3
)
1
for r
(m
2
;m
3
)
3
and !
(m
1
;m
2
;m
3
)
2
for r
(m
1
;m
2
;m
3
)
1
,
compute the moments as indicated in (16)-(18).
This process can be iterated For example, continuing with the specic case of pre-
dicting four successive quarters of growth rates, intrinsic variance can be decomposed
into four rather than two components, yielding the variance in the 2012 annual growth
rate that will be resolved following quarters 2 and 3 as well as quarter 1 of 2012.
4.3 Decomposition of extrinsic variance
Let the parameter vector be decomposed
0
A
= (
0
1A
.
0
2A
), and suppose that it is feasible
to simulate
(m)
A1
s j
_
A1
j Y
0
T
.
_
, (27)
(m)
A2
s j
_
A2
j
A1
. Y
0
T
.
_
. (28)
The simulation (27) is from the marginal distribution of
A1
, so these simulations can
be taken as the corresponding subvector of the posterior simulation sequence itself. The
second simulation need not be straightforward or even feasible, though it usually is. If
the posterior simulator is a pure Metropolis algorithm then the same algorithm can be
used in (28), and indeed this simulation should be less challenging due to the diminished
16
ECB
Working Paper Series No 1409
December 2011
oiooi of lho paiamoloi vocloi. If lho posloiioi simulaloi is a GiLLs sampling algoiilhm
ano nono of lho Llocks incluoo compononls of Lolh
ano
= , A
ano A
in (8), which
yiolos
var () = L
[var ( [
)[ (20)
L
var
[L( [
)[ (80)
var
[L( [
)[ . (81)
(1hal all momonls aio also conoilional on
c
T
ano is suppiossoo in lhis nolalion lo
avoio clulloi.) 1ho lisl loim (20) is lho inliinsic vaiianco of . 1ho loims (80)-(81)
oocomposo lho oxliinsic vaiianco of . 1ho lasl loim (81) moasuios syslomalic localion
movomonl in in iosponso lo changos in
aio
lhoso in ils posloiioi oisliiLulion. Sinco Ly oolnilion all oxliinsic vaiianco is accounloo
foi Ly
ano woio Gaussian ano lho lasl lwo loims woio oiviooo Ly oxliinsic vaiianco
var
[L( [
)[ lhon (81) woulo piovioo lho 1-squaioo, lho fiaclion of oxliinsic vaii-
anco ouo lo
.
As a spocilc inslanco, supposo a conlial Lank using a oynamic slochaslic gonoial
oquiliLiium(DSGL) moool is inloiosloo in lho oxlonl lo which oxliinsic vaiialion lho ono-
slop-ahoao piooiclivo oisliiLulion j (
T
[
c
T
. ) is oiivon Ly lho sliucluial paiamolois
of lho moool (o.g., lho paiamolois of ulilily, pioouclion ano policy iosponso funclions)
as opposoo lo lhoso oosciiLing lho oynamics of shocks lo oqualions (o.g., lho vaiiancos
ano auloiogiossivo coomcionls). 1ho foiogoing analysis piovioos lwo answois lo lhoso
quoslions, ooponoing on which paiamolois aio casl as
ano which as
.
Il is impoilanl lo koop in mino lhal Ly oolnilion lho paiamolois as a wholo accounl
foi all of lho oxliinsic vaiianco. 1aking
(T)
A
.
_
, where
(T)
A
= arg max
A
j (
A
j Y
o
T
. ) . (32)
This does not conform with the formal Bayesian rule
j (y
T+1
j Y
o
T
. ) =
_
A
j (y
T+1
j Y
o
T
.
A
. ) j (
A
j Y
o
T
. ) d
A
. (33)
If there were no extrinsic variance in y
T+1
in the predictive density (33) then (32) would
be equivalent to (33). Of course this is not the case, but if extrinsic variance is a small
enough component of predictive variance then the preference for (33) over (32) is an
academic rather than a practical point. That is one of the questions investigated here.
We accessed the posterior density j (
A
j Y
o
T
. ) by means of a conventional Metropo-
lis random walk posterior simulation algorithm to generate the sequence
_
(m
2
)
A
_
de-
scribed in Section 4 by thinning a chain of 10,000 MCMC simulations to `
2
= 100
equally-spaced draws. Our vector of interest is !
0
= (!
0
1
. !
0
2
), with !
1
= y
T+1
and
18
ECB
Working Paper Series No 1409
December 2011
= (1,4)
Tc
. Following lho piocoouios oosciiLoo in lho piovious soclion wo
simulaloo `
= 100 oiaws
n
,n
=
n
,n
T
fiom lho ono-quailoi-ahoao oonsily
j
_
T
[
c
T
.
n
_
conoilional on oach of lhoso 100 paiamoloi oiawings. Slailing
wilh oach of lho 10,000 paiis
_
.
n
,n
_
wo simulaloo in succossion `
= 100
oiawings
,n
,n
Tc
j
_
Tc
[
c
T
.
n
,n
T
. . . . .
n
,n
,n
Tc
.
n
_
(: = 2. 3. 4)
ano lhon foimoo
,n
,n
= (1,4)
_
,n
T
n
,n
,n
T
n
,n
,n
T
n
,n
,n
T
_
. (84)
1ho simulalions of
n
,n
ano
n
,n
,n
Iomainooi
Consumplion 0.2840 0.0077 (0.088) 0.2268 (0.067) 0.0706 (0.822) 0.17 (0.688)
Invoslmonl 8.8804 0.0768 (0.028) 8.2681 (0.077) 1.2786 (0.802) 1.084 (0.608)
Oulpul 0.4200 0.0066 (0.016) 0.4148 (0.084) 0.1876 (0.882) 0.2767 (0.668)
Houis 1.0106 0.012 (0.01) 0.004 (0.08) 0.87 (0.6) 0.4417 (0.444)
Inlalion 0.0766 0.0022 (0.020) 0.0744 (0.071) 0.0606 (0.086) 0.0048 (0.064)
Wagos 0.1461 0.0024 (0.016) 0.1488 (0.084) 0.068 (0.48) 0.0770 (0.42)
Foo funos 0.0804 0.0018 (0.014) 0.0881 (0.086) 0.0480 (0.6) 0.0801 (0.444)
1aLlo 1: Docomposilion of piooiclivo vaiianco foi ono-yoai giowlh ialos 2008:1- 2008:4,
Lasoo on lho posloiioi al 2007:4
1ho analysis of vaiianco oocomposos lho piooiclivo vaiianco of giowlh ialos ovoi lho
noxl foui quailois inlo oxliinsic ano inliinsic vaiianco, ano lhon oocomposos inliinsic
vaiianco inlo a compononl lhal is iosolvoo afloi ono quailoi ano lho iomainooi lhal is
nol iosolvoo unlil all foui quailois havo Loon oLsoivoo. Wo unooilook lhis oxoiciso al
lwo ioconl poinls in limo. 1ho lisl is lho ono of 2007:4, using oala lhiough lhal quailoi
ano analyzing lho piooiclivo oisliiLulion foi giowlh ialos ovoi calonoai yoai 2008. 1ho
socono is lho ono of 2000:2, using oala lhiough lhal quailoi ano analyzing lho piooiclivo
oisliiLulion foi giowlh ialos ovoi lho ono-yoai poiioo 2000:8 lhiough 2010:2. 1ho lisl
oxoiciso is posilionoo jusl Lofoio lho onsol of lho gloLal lnancial ciisis: compaioo wilh
hisloiical valuos lhoio was nolhing pailiculaily iomaikaLlo aLoul ioconl quailoily giowlh
ialos. In lho socono oxoiciso ioconl quailoily giowlh ialos oxhiLil valuos Loyono lho
iango soon in lho 60-yoai limo soiios on which lho posloiioi oisliiLulion conoilions.
1aLlo 1 shows lho oocomposilion of vaiianco foi oach of lho sovon soiios in lho
lisl oxoiciso. 1ho main onliios aio lho vaiianco loims oslimaloo fiom lho simulalions.
19
ECB
Working Paper Series No 1409
December 2011
1ho onliios in paionlhosos in columns 8 ano 4 inoicalo lho fiaclional oocomposilion of
vaiianco Lolwoon oxliinsic ano inliinsic vaiianco, whilo lho onliios in paionlhosos in lho
lasl lwo columns inoicalo lho fiaclional oocomposilion of inliinsic vaiianco Lolwoon lho
compononl iosolvoo afloi ono quailoi ano lho iomaining vaiianco. 1aLlo 2 ooos lho
samo lhing foi lho socono oxoiciso.
1ho piooiclivo vaiianco (column 2 in Lolh laLlos) is suLslanlially oiloionl foi lho
sovon limo soiios, Lul lhoso oiloioncos simply iolocl lho unils of moasuiomonl oolailoo
al lho slail of lhis soclion. Lxcopl foi inlalion, piooiclivo vaiianco is moooslly highoi
in lho socono oxoiciso lhan in lho lisl, ano lho samo lhing is liuo of inliinsic vaiianco.
1hoio is a maikoo conliasl in lho allocalion of piooiclivo vaiianco lo oxliinsic vaii-
anco (column 8 in Lolh laLlos) in lho lwo oxoicisos. Lxliinsic vaiianco is novoi moio
lhan 10/ of piooiclivo vaiianco. Howovoi, in lho socono oxoiciso lho fiaclion of piooic-
livo vaiianco lhal is oxliinsic is always much highoi, ianging fiom almosl lhioo limos
as high lo ovoi foui limos as high. 1ho oxplanalion foi lhis conliasl lios in lho val-
uos of ioconl limo soiios in lho lwo oxoicisos. Wilh valuos alypical of lho samplo (lho
socono oxoiciso) uncoilainly aLoul paiamoloi valuos is magniloo in lho piooiclivo ois-
liiLulion: lho impacl of unusually high oi low valuos, which is impoifoclly known, is
moio impoilanl. Ioplacing (88) wilh (82) ooos nol iosull in soiious unooislalomonl of
lho oispoision of lho piooiclivo oisliiLulion al oilhoi limo, Lul lho unooislalomonl is
suLslanlially gioaloi in lho socono caso lhan in lho lisl. 1his lnoing shoulo Lo gonoi-
ally applicaLlo: whon lho oala lhal oiivo lho piooiclivo oisliiLulion havo Loon unusual,
iolalivo lo lho samplo, lho allocalion of piooiclivo vaiianco lo oxliinsic vaiianco will
Lo gioaloi ano using poinl oslimalos in placo of full piooiclivo oisliiLulions will moio
soiiously unooislalo vaiianco in noai-loim piooiclivo oisliiLulions.
1ho allocalion of inliinsic vaiianco of lho ono-yoai foiocasls lo noai-loim ano longoi
loim iolocls lho volalilily inhoionl in lho valuos of lho moool`s paiamolois lhal aio
plausiLlo in lho posloiioi oisliiLulion. Ovoi fuluio quailois, lho spioao in lho piooiclivo
oisliiLulion will iolocl aclual Lohavioi in ioconl quailoi moio sliongly in lho noai loim
ano lho volalilily implicil in lho moool`s paiamolois moio sliongly in lho fai loim.
1hus, foi oxamplo, woio lho moool lo infoi unioalislically low volalilily lhon lho noai
loim woulo oominalo lho oocomposilion of inliinsic vaiianco inlo noai ano longoi loim
(ano convoisoly).
Inliinsic vaiianco ilsolf (column 4 in Lolh laLlos) is viilually ioonlical in lho lwo
oxoicisos. Foi giowlh ialos in GDI ano compononls (lho lisl lhioo soiios) 80/ lo 40/
of lho vaiianco is iosolvoo in lho noai loim. 1his is inluilivoly plausiLlo: lho noai loim
Lolh iovoals ono of lho foui compononls in (84) ano in aooilion Liings lho following lhioo
quailois ono quailoi closoi, fuilhoi iooucing uncoilainly aLoul lhoso giowlh ialos. Foi
houis, wagos ano Foooial funos ovon moio of lho vaiianco, ioughly half, is iosolvoo in lho
noai loim. In lho caso of inlalion woll ovoi 00/ of lho inliinsic vaiianco is so iosolvoo,
a iomaikaLlo lnoing. 1his is consislonl wilh lho moool assigning volalilily lo inlalion
lhal is much loo low in lho caso of inlalion, ano poihaps somowhal loo low in lho caso
of houis, wagos ano lho Foooial funos ialo. Of couiso, olhoi oxplanalions aio possiLlo
as woll, Lul lhoso lnoings molivalo a closoi oxaminalion of inlalion oynamics in lhis
20
ECB
Working Paper Series No 1409
December 2011
Predictive variance Intrinsic variance
Total Extrinsic Intrinsic Due to !
1
Remainder
Consumption 0.2782 0.0264 (0.095) 0.2518 (0.905) 0.0817 (0.324) 0.1702 (0.676)
Investment 4.2767 0.3218 (0.075) 3.9549 (0.925) 1.6159 (0.409) 2.3390 (0.591)
Output 0.4714 0.0308 (0.065) 0.4406 (0.935) 0.1513 (0.343) 0.2893 (0.657)
Hours 1.2488 0.0771 (0.062) 1.1717 (0.938) 0.6689 (0.571) 0.5029 (0.429)
Ination 0.0745 0.0052 (0.070) 0.0693 (0.930) 0.0652 (0.941) 0.0041 (0.059)
Wages 0.1558 0.0072 (0.046) 0.1487 (0.954) 0.0660 (0.444) 0.0826 (0.556)
Fed funds 0.0936 0.0055 (0.059) 0.0881 (0.941) 0.0489 (0.555) 0.0392 (0.445)
Table 2: Decomposition of predictive variance for one-year growth rates 2009:3 - 2010:2,
based on the posterior at 2009:2
model.
6 Conclusion
From a formal but reasonable perspective the goal of Bayesian analysis can generally be
cast as providing a predictive distribution relevant for a decision at hand. In doing so
it integrates information from several sources, including increments to information sets
as predictive distributions are updated in real time. This paper has provided a corre-
sponding analysis of variance. Because integration of information does not typically lead
to linear (Gaussian) models this analysis is necessarily more complex than the familiar
treatment that has been central to statistics for a century. But the complexity poses
no essential complication for simulation methods that are fast, practical, and natural
in the context of modern Bayesian inference. We believe that systematic application of
this analysis of variance will provide greater insight into the structure of models and
information aggregation, and hope that ultimately it will be useful in improving models,
predictions and decisions.
21
ECB
Working Paper Series No 1409
December 2011
References
Box GEP (1980). Sampling and Bayes inference in scientic modeling and robustness.
Journal of the Royal Statistical Society Series A 143: 383-430.
Geweke J (2005). Contemporary Bayesian Econometrics and Statistics. Englewood
Clis NJ: Wiley.
Geweke J (2010). Complete and Incomplete Econometric Models. Princeton: Prince-
ton University Press.
Good IJ (1956). The surprise index for the multivariate normal distribution. Annals
of Mathematical Statistics 27: 1130-1135.
Lancaster T (2004). An Introduction to Modern Bayesian Econometrics. Malden
MA: Blackwell Publishing.
Roberts HV (1965). Probabilistic prediction. Journal of the American Statistical
Association 60: 50-62.
Smets F, Wouters R (2007). Shocks and frictions in US business cycles: A Bayesian
DSGE approach. American Economic Review 97: 586606.
Weiss NA (2005). A Course in Probability. Boston: Addison Wesley.
Zellner A (1971). An Introduction to Bayesian Inference in Econometrics and Sta-
tistics. New York: Wiley.
22
ECB
Working Paper Series No 1409
December 2011
Appendix
The following Matlab code implements the simulation-approximation of variance
decomposition described in Section 3, and was used in for the illustration in Section 5.
function [term1 term2 term3 total]=aov(x)
% This function computes simulation-based method of moments estimates
% of the population decomposition in the paper. The explicit
% computations are in the displays (16)-(18).
% Input:
% x This is a three-dimensional structure. The first dimension
% corresponds to "X_2" and has the m_2 index; the second
% dimension corresponds to "X_3" and has the m_3 index; the
% third dimension corresponds to "X_1" and has the m_1 index.
% Outputs:
% term1 Estimate (16)of the first term of (8)
% term2 Estimate (17) of the second term of (8)
% term3 Estimate (18) of the third term of (8)
%
[M2 M3 M1]=size(x);
term1=var(mean(reshape(x,M2,M3*M1),2),1,1);
term2=mean(var(mean(x,3),1,2),1);
term3=mean(var(reshape(x,M3*M2,M1),1,2),1);
total=term1+term2+term3;
end
WORKI NG PAPER S ERI ES
NO 1409 / DECEMBER 2011
by John Geweke
and Gianni Amisano
ANALYSIS OF
VARIANCE FOR
BAYESIAN
INFERENCE