3.5 Structural VAR and SVEC analysis
3.5.1 Identification structural shocks
According to Breitung et al. [24], if the process yt is I(0), the effects of shocks in the
variables of a given system are most seen in its Wold moving average (MA) represen- tation, yt= ut+Ψ1ut−1+Ψ2ut−2+ ..., where Ψ0= IK, Ψs=Psj=1Ψs−jΦj and
Φj is K × K matrix of coefficients of VAR(p) model, equation 3.1. The coefficients
ψij of the matrices Ψs is interpreted as the response to the impulse of structural
shocks in the system. Note that, obtain forecast error impulse responses is feasible if Σu is diagonal, otherwise, the forecast error impulse responses are not valid, because
the underlying shocks are not likely to occur in isolation if the components of ut are
instantaneously correlated.
In practice, one way to obtain orthogonal innovations is to use Choleski decom- position of the white noise covariance matrix Σu = PP0 such that P is a lower
triangular matrix with positive elements on the main diagonal.
According to Breitung et al. [24] and Lütkepohl [77] there are different procedures used to non-sample information in specifying and then obtaining unique impulse responses. The three most used are associated with the so-called A - model, B - model and AB - model (see also Amisano and Giannini [1]).
The A - model
Lets the structural errors from equation 3.34 have the following MA representation
where the matrices Θj= ΨjA−1 for j = 1, 2, ...; represent the response to structural
shocks εt assumed to be mutually uncorrelated. The relation between structural
innovations and residuals from the VAR model 3.1 are εt= Aut, which is called the
A - model. For this model, the identifying restrictions are imposed on A matrix such that, the product Aut= εt has a diagonal covariance matrix. If it is plausible
that A has a one in the main diagonal, they are necessary K(K − 1)/2 restrictions on A to ensure just identified shocks, consequently identified impulse response.
Summarizing, according to Breitung et al. [24] and Lütkepohl [77], in A - model, the set of innovations εt is modelled in the system of equations as a function of
the residuals such that Aut = εt and the linear restrictions on A can be write in
explicit form as vec(A) = RAγA + rA, where γA represent all unrestrited elements
of A, RA is a suitable matrix with zero and one elements and rA is the vector of
normalized constants.
The B - model
In SVAR model equation 3.34, because the shocks are not observed, some assump- tions are need to identify them. For instance we required that the structural shocks should be orthogonal, and that the structural shock εt are assumed to be related to
the model residuals by linear relations ut= Bεt, where B is a K × K matrix. Note
that we assume that for Aut= Bεt and the matrix A = IK.
Using the relation between model residuals and structural innovations, one can write equation 3.34 as follows:
Ayt=Φ∗1yt−1+ ... +Φ∗pyt−p+ Bεt (3.36)
where εt∼ N (0, IK), Φ∗ are K × K coefficient matrices, B is K × K matrix such
that Σu= BΣεB0. Because the covariance matrix of εt are normalized with one in
diagonal, the covariance matrix of residuals became P
u= BB0, where B is lower
triangular chosen by Choleski decomposition.
According to Lütkepohl and Krätzig [78], the equality ut= Bεt where the struc-
tural shock εt ∼ N (0, IK) is the so called B-model. Due to the symmetry of the
covariance matrix of P
u, we specify only K(K + 1)/2 different equations and need
additional K(K − 1)/2 restrictions to be imposed to identify all K2 elements of B. In summary, for B-model, the identification consists to exclude some linear combinations of the structural shocks by imposing restriction of the form vec(B) = RBγB+ rB, where γB contains unrestricted elements of B, RB are the restricted
3 . 5 . S T RU C T U R A L VA R A N D S V E C A N A LY S I S
The AB model
According to Lütkepohl [77], situation where is possible to consider both types of restrictions, that is, from equation 3.36, Aut = Bεt where εt ∼ N (0, IK), we are
facing a model known as AB - model which relates the reduced form errors ut to
the underlying structural shocks εt.
Identification the structural shocks can be done, imposing restrictions on A and B parameter matrices. Here for K-variables K(K − 1)/2 restrictions are necessary for orthogonalizing the shocks because there are K(K − 1)/2 different instantaneous covariances. Detail of the procedures can be found in Breitung et al. [24] and Galí [50].
For this model, from a previous relation between residuals of the model and structural shock, we get ut= A−1Bεt, and the corresponding covariance matrix is
P
u= A−1BB0A−10. Thus, we have K(K + 1)/2 equations. Meaning that we need
K2 elements for each matrix A and B, additionally 2K2−12K(K + 1) restrictions
to identify all 2K2 elements of A and B at least locally. Even considering that the matrix A has elements one on main diagonal, 2K2− K −12K(K + 1) restrictions are
needed for identification of the AB - model.
In order to avoid the large number of restrictions in most applications is used the special cases just discussed. The A - model where B = IK or the B-model with
A = IK. According to Breitung et al. [24] and Lütkepohl [77], considering the general case AB-model, the constraints are normally normalized and they can be written in the form of linear equations.
vec(A) = RAγA+ rA and vec(B) = RBγB+ rB (3.37)
where RA and RB are possible restricted matrices with zero and ones, γA and γB
are unrestricted vectors with free parameters and finally rA and rB are vectors of
fixed parameters.
Identification SVEC model
If all or some variables in the system are integrated with r cointegrating vectors, which describe the long-run relationships between variables, the VEC model, equation 3.30 is an appropriate model and the corresponding SVEC model is given by,
A∆yt=Π∗yt−1+Γ∗1∆yt−1+ ... +Γ∗p−1∆yt−p+1+ Bεt (3.38)
where Π∗ and Γ∗i are structural form parameters matrices and εt is a structural
error term, B is a K × K matrix of impact of shocks and matrix A models the instantaneous relations among the variables in the system.
According to Pfaff [98], in order to identify the structural form parameters, we need to considers the Beveridge-Nelson (Beveridge and Nelson [10]) moving average representation of the variables yt if they adhere to the VEC model process as in
equation 3.30. yt=Ξ t X i=1 ui | {z } I(1) + ∞ X j=0 Ξ∗jut−j | {z } I(0) +y∗0 (3.39)
where the matrix Ξ is defined as Ξ = β⊥α0⊥(Ik−Pp−1i=1Γi)β⊥
−1
α0⊥, with α⊥ and
β⊥ indicate orthogonal components of α and β respectively. The first term I(1) is
the common trend that drives the system of yt, the second term I(0) it bounded by
the infinite sum because the series ofΞ∗j converges to zero as j → ∞ and y∗0 contains all initial values.
According to Breitung et al. [24] and Pfaff [98], since the long-run effects of shocks is captured on the common trend, this is the centre of modelling Structural VEC model.
Replacing ut in equation 3.39 by A−1Bεt and then, assuming that εt∼ N (0, IK)
and the matrix A = Ik, then the impact of the short-run or transitory orthogonal shocks are obtained by Ξ∗jB and the long-run by ΞB in the same way as in the stationary VAR model.
For example in B - model using the relation ut = Bεt, is necessary at least
K(K − 1)/2 restrictions to identify B. But if rank(ΞB) = K − r, where r is the rank
of the system yt, means that r columns have zeros in the matrix of long-run, which
corresponds to r structural innovations having transitory shocks, the remainder
m = K −r structural innovations have permanent shocks. To identify the permanent
shocks exactly we need m × (m − 1)/2 additional restrictions. Similarly r(r − 1)/2 additional contemporaneous restrictions needed to identify the transitory shocks. Continuing with the assumption that A = IK, we have just enough restrictions to identify B and ΞB. In this case we have a model exactly identified, if we impose additional restrictions on the parameters, so it would be an overidentified model, and the overidentified restriction could be tested.