Structural VARs Structural Representation Consider the structural VAR (SVAR) model y1t = 10 b12y2t + 11y1t1 + 12y2t1 + 1t y2t

y2t = 20 b21y1t + 21y1t1 + 22y2t1 + 2t where

1t 2t

iid

0 0

2 1 0

0 2 2

!!

Remarks:

1t and 2t are called structural errors In general, cov(y2t, 1t) 6= 0 and cov(y1t, 2t) 6= 0 All variables are endogenous - OLS is not appropriate!

In matrix form, the model becomes

"

= or

"

1 b12 b21 1 10 20
#

"

#"

y1t y2t

11 12 21 22

#"

y1t1 y2t1

"

1t 2t

Byt = 0 + 1yt1 + t ! 2 0 1 E[t0 ] = D = t 0 2 2

In lag operator notation, the SVAR is

B(L)yt = 0 + t, B(L) = B 1L.

Reduced Form Representation Solve for yt in terms of yt1 and t :

yt = B1 0 + B11yt1 + B1t = a0 + A1yt1 + ut a0 = B1 0, A1 = B11, ut = B1t

or

Note that
1 = 1 B

A(L)yt = a0 + ut A(L) = I2 A1L

"

The reduced form errors ut are linear combinations of the structural errors t and have covariance matrix E[utu0 ] = B1E[t0 ]B10 t t 1DB10 = B = . Remark: Parameters of RF may be estimated by OLS equation by equation

1 b12 b21 1

, = det(B) = 1b12b21

Identication Issues Without some restrictions, the parameters in the SVAR are not identied. That is, given values of the reduced form parameters a0, A1 and , it is not possible to uniquely solve for the structural parameters B, 0, 1 and D. 10 structural parameters and 9 reduced form parameters Order condition requires at least 1 restriction on the SVAR parameters Typical identifying restrictions include Zero (exclusion) restrictions on the elements of B; e.g., b12 = 0. Linear restrictions on the elements of B; e.g., b12 + b21 = 1.

MA Representations Wold representation Multiplying both sides of reduced form by A(L)1 = (I2 A1L)1 to give

yt = + (L)ut (L) = (I2 A1L)1

=
k=0

= A(1)1a0 E[utu0 ] = t
Remark: Wold representation may be estimated using RF VAR estimates

k Lk , 0 = I2, k = Ak 1

Structural moving average (SMA) representation SMA of yt is based on an innite moving average of the structural innovations t. Using ut = B1t in the Wold form gives

yt = + (L)B1t = + (L)t (L) =

k=0 X

k Lk

= (L)B1 = B1 + 1B1L + That is,

k = k B1 = Ak B1, k = 0, 1, . . . . 1 0 = B1 6= I2

Example: SMA for bivariate system

"

y1t y2t

Notes

# (0) (0) " 1 12 1t = + 11 (0) (0) 2 2t 21 22 # (1) (1) " 12 1t1 + 11 + (1) (1) 2t1 21 22 " #

0 = B1 6= I2. 0 captures initial impacts of structural shocks, and determines the contemporaneous correlation between y1t and y2t.
(k)

Elements of the k matrices, ij , give the dynamic multipliers or impulse responses of y1t and y2t to changes in the structural errors 1t and 2t.

Impulse Response Functions Consider the SMA representation at time t + s

" # (0) (0) " 1 12 1t+s = + 11 + (0) (0) 2 2t+s 21 22 # (s) (s) " 12 1t + . + 11 (s) (s) 2t 21 22 " #

y1t+s y2t+s

The structural dynamic multipliers are

y1t+s (s) y (s) = 11 , 1t+s = 12 1t 2t y2t+s (s) y2t+s (s) = 21 , = 22 1t 2t The structural impulse response functions (IRFs) are the plots of ij vs. s for i, j = 1, 2. These plots summarize how unit impulses of the structural shocks at time t impact the level of y at time t + s for dierent values of s. Stationarity of yt implies lim = 0, i, j = 1, 2 s ij
(s) (s)

The long-run cumulative impact of the structural shocks is captured by

P (s) P (s) 11(1) 12(1) (1) = = Ps=0 11 Ps=0 12 (s) (s) 21(1) 22(1) s=0 21 s=0 22 " # 11(L) 12(L) (L) = 21(L) 22(L) P (s) s P (s) s L s=0 12 L = Ps=0 11 (s)Ls P (s)Ls s=0 21 s=0 22 " #

Digression: Dynamic Regression Models In the SVAR every variable is engodenous. Suppose, for example, y2t is strictly exogenous which implies b21 = 0 and 21 = 0. Then, the rst equation is an ADL(1,1) y1t = + y1t1 + 0y2t + 1y2t1 + 1t cov(y2t, 1t) = 0 In lag operator notation the equation becomes (L)y1t = + (L)y2t + 1t (L) = 1 L, (L) = 0 + 1L The second equation is an AR(1) model for y2t y2t = c + y2t1 + 2t Stationarity now only requires || < 1 and || < 1.

The rst equation may then be solved for y1t as a function of y2t and 1t y1t = + (L)1(L)y2t + (L)11t (1) = + (L)y2t + (L)t = (1) (L) = (L)1(L), (L) = (L)1 Since y2t is exogenous, we have two sources of shocks. Note: there can be four types of dynamic multipliers : y1t+s y1t+s y1t+s y2t+s , , , y2t 2t 1t 2t

The short-run dyamic multipliers with respect to y2t and 1t are y1t+s y1t = = ,s y2t y2ts y1t y1t+s = = s 1t 1ts In the steady state or long-run equilibrium all variables are constant
y1 = + (L)y2 = + (1)y2 c y2 = 1 + 1 (1) = (1)1(1) = 0 1 The long-run impact of a change in y2 on y1 is then X y1t+s y1 0 + 1 = (1) = 1 = y2 s=0 y2t

Identication issues In some applications, identication of the parameters of the SVAR is achieved through restrictions on the parameters of the SMA representation. Identication through contemporaneous restrictions Suppose that 2t has no contemporaneous impact on y1t. (0) Then 12 = 0 and
(0) 0 0 = 11 (0) (0) . 21 22 Since 0 = B1 then " # (0) 1 0 1 b12 11 = (0) (0) 1 b21 21 22 b12 = 0 (0) Hence, assuming 12 = 0 in the SMA representation is equivalent to assuming b12 = 0 in the SVAR representa

tion.

Identication through long-run restrictions Suppose 2t has no long-run cumulative impact on y1t. Then
X (s) 12 = 0 12(1) = s=0 " # 11(1) 0 (1) = . 21(1) 22(1)

This type of long-run restriction places nonlinear restrictions on the coecients of the SVAR since

(1) = (1)B1 = A(1)1B1 = (I2 B11)1B1

Estimation Issues In order to compute the structural IRFs, the parameters of the SMA representation need to be estimated. Since

(L) = (L)B1 (L) = A(L)1 = (I2 A1L)1

the estimation of the elements in (L) can often be broken down into steps: A1 is estimated from the reduced form VAR.
c Given A1, the matrices in (L) can be estimated c c using k = Ak . 1

B is estimated from the identied SVAR.

Given B and k , the estimates of k , k = 0, 1, . . . , c are given by k = k B1.

Forecast Error Variance Decompositions Idea: determine the proportion of the variability of the errors in forecasting y1 and y2 at time t + s based on information available at time t that is due to variability in the structural shocks 1 and 2 between times t and t + s. To derive the FEVD, start with the Wold representation for yt+s

yt+s = + ut+s + 1ut+s1 + +s1ut+1 + sut + s+1ut1 + .

The best linear forecast of yt+s based on information available at time t is

yt+s|t = + sut + s+1ut1 +

and the forecast error is

yt+s yt+s|t = ut+s + 1ut+s1 + + s1ut+1.

Using

t = B1ut, k = k B1
The forecast error in terms of the structural shocks is

yt+s yt+s|t = B1t+s + 1B1t+s1 +

= 0t+s + 1t+s1 + + s1t+1 The forecast errors equation by equation are
"
(0) y1t+s y1t+s|t 11 = (0) y2t+s y2t+s|t 21 (s1) (s1) 11 12 + (s1) (s1) 21 22

+ s1B1t+1

# (0) " 12 1t+s + (0) 2t+s 22 # " 1t+1 2t+1

For the rst equation y1t+s y1t+s|t = 11 1t+s + + 11

(0) (0) (s1)

1t+1 2t+1

+12 2t+s + + 12

(s1)

Since it is assumed that t i.i.d. (0, D) where D is diagonal, the variance of the forecast error in may be decomposed as
2 2! (0) (s1) 11 + + 11 = 2 1 2 2! (0) (s1) + 2 12 + + 12 . 2

var(y1t+s y1t+s|t) = 2(s) 1

The proportion of 2(s) due to shocks in 1 is then 1 2 1 1,1(s) =

2 2! (0) (s1) 11 + + 11

2(s) 1

the proportion of 2(s) due to shocks in 2 is 1 2 2 1,2(s) =

2 2! (0) (s1) 12 + + 12

2(s) 1

The forecast error variance decompositions (FEVDs) for y2t+s are 2 1 2,1(s) = 2 2 2,2(s) = where
2 2! (0) (s1) 21 + + 21 = 2 1 2 2! (0) (s1) + 2 22 + + 22 . 2 2 2! (0) (s1) 21 + + 21 2 2! (0) (s1) 22 + + 22

2(s) 2

2(s) 2

var(y2t+s y2t+s|t) = 2(s) 2

Identication Using Recursive Causal Orderings Consider the bivariate SVAR. We need at least one restriction on the parameters for identication. Suppose b12 = 0 so that B is lower triangular. That is,

B =

"

1 0 b21 1
"

B1 = 0 =

1 0 b21 1

The SVAR model becomes the recursive model y1t = 10 + 11y1t1 + 12y2t1 + 1t y2t = 20 b21y1t + 21y1t1 + 22y2t1 + 2t The recursive model imposes the restriction that the value y2t does not have a contemporaneous eect on y1t. Since b21 6= 0 a priori we allow for the possibility that y1t has a contemporaneous eect on y2t.

The reduced form VAR errors ut = B1t become

ut =
=

" "

u1t u2t

"

1t 2t b211t

1 0 b21 1
#

#"

1t 2t

Claim: The restriction b12 = 0 is sucient to just identify b21 and, hence, just identify B.

To establish this result, we show how b21 can be uniquely identied from the elements of the reduced form covariance matrix . Note
"

2 12 1 12 2 2

= =

" "

2 b21 2 1 1 2 2 + b2 2 b21 1 2 21 1

1 0 b21 1

#"

2 1 0

0 2 2

#" #

1 b21 0 1

Then, we can solve for b21 via b21 = 12 = 2 , 1 2 1 where = 12/ 12 is the correlation between u1 and u2. Notice that b21 6= 0 provided 6= 0.

Estimation Procedure 1. Estimate the reduced form VAR by OLS equation by equation:
c b b yt = a0 + A1yt1 + ut b 2. Estimate b21 and B from :
T 1 X b b bt = utu0 T t=1

3. Estimate SMA from estimates of a0, A1 and B:

b b b21 = 12 , b 2 " 1 # 1 0 b B = . b b21 1 c b b = + (L)t b b b = BB0. c b = a0(I2A1)1

yt b c k c D

c b = Ak B1, k = 0, 1, . . . 1

Remark: Above procedure is numerically equivalent to estimating the triangular system by OLS equation by equation: y1t = 10 + 11y1t1 + 12y2t1 + 1t y2t = 20 b21y1t + 21y1t1 + 22y2t1 + 2t Why? Since cov(1t, 2) = 0 by assumption, cov(y1t, 2t) = 0

Recovering the SMA representation using the Choleski Factorization of . Claim: The SVAR representation based on a recursive causal ordering may be computed using the Choleski factorization of the reduced form covariance matrix . Recall, the Choleski factorization of the positive semidenite matrix is given by

= PP0 " # p11 0 P = p21 p22

A closely related factorization obtained from the Choleski factorization is the triangular factorization

= TT0 " # " # 1 0 1 0 T= , = , t21 1 0 2 i 0, i = 1, 2.

Consider the reduced form VAR

yt = a0 + A1yt1 + ut, = E[utu0 ] t = TT0

Construct a pseudo SVAR model by premultiplying by T1 :

T1yt = T1a0 + T1A1yt1 + T1ut

or

Byt = 0 + 1yt1 + t
where

B = T1, 0 = T1a0, 1 = T1A1, t = T1ut.

The pseudo structural errors t have a diagonal covariance matrix E[t0 ] = T1E[utu0 ]T10 t t = T1T10 = T1TT0T10 = . In the pseudo SVAR, 1 0 = T1 = b21 1 b12 = 0, b21 = t21
" # " #

B =

1 0 t21 1

Ordering of Variables The identication of the SVAR using the triangular factorization depends on the ordering of the variables in yt. In the above analysis, it is assumed that yt = (y1t, y2t)0 so that y1t comes rst in the ordering of the variables. When the triangular factorization is conducted and the pseudo SVAR is computed the structural B matrix is

B = T1 =
b12 = 0

"

1 0 b21 1

If the ordering of the variables is reversed, yt = (y2t, y1t)0, then the recursive causal ordering of the SVAR is reversed and the structural B matrix becomes

B = T1 =
b21 = 0

"

1 0 b12 1

Sensitivity Analysis Ordering of the variables in yt determines the recursive causal structure of the SVAR, This identication assumption is not testable Sensitivity analysis is often performed to determine how the structural analysis based on the IRFs and FEVDs are inuenced by the assumed causal ordering. This sensitivity analysis is based on estimating the SVAR for dierent orderings of the variables. If the IRFs and FEVDs change considerably for different orderings of the variables in yt then it is clear that the assumed recursive causal structure heavily inuences the structural inference.

Residual Analysis One way to determine if the assumed causal ordering inuences the structural inferences is to look at the resid ual covariance matrix from the estimated reduced form VAR. If this covariance matrix is close to being diagonal then the estimated value of B will be close to diagonal and so the ordering of the variables will not inuence the structural inference.