Scribd
Upload
265 views6 pages

VAR Models With Exogenous Variables

1. VARX (vector autoregression with exogenous variables) models allow for the inclusion of exogenous variables in VAR models. This impacts innovation response analysis. 2. There are weak and strong exogeneity conditions for the exogenous variables. Weak exogeneity means the variables are not Granger caused by the endogenous variables. Strong exogeneity additionally requires the errors to be independent. 3. Tests can be performed to check for weak and strong exogeneity using restrictions on coefficient matrices in an unrestricted VAR model. Innovation response analysis then proceeds based on the appropriate exogeneity condition.

Uploaded by

Abhishek Puri
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
265 views6 pages

VAR Models With Exogenous Variables

1. VARX (vector autoregression with exogenous variables) models allow for the inclusion of exogenous variables in VAR models. This impacts innovation response analysis. 2. There are weak and strong exogeneity conditions for the exogenous variables. Weak exogeneity means the variables are not Granger caused by the endogenous variables. Strong exogeneity additionally requires the errors to be independent. 3. Tests can be performed to check for weak and strong exogeneity using restrictions on coefficient matrices in an unrestricted VAR model. Innovation response analysis then proceeds based on the appropriate exogeneity condition.

Uploaded by

Abhishek Puri
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

VAR models with exogenous variables

Herman J. Bierens February 21, 2004


Abstract In this note I will address the question how to include exogenous variables in a VAR model, and what the consequences are for the innovation response analysis.

VARX models

Consider a VAR model with exogenous variables: Yt = a0 + A1 Yt1 + ... + Ap Ytp +B1 Xt1 + ... + Bq Xtq + Ut , (1)

with probability 1. Next, assuming a VAR model for Xt itself, say

where Yt Rk , Xt Rm is a vector of exogenous variables, a0 Rk is a vector of intercepts, the Aj s are k k coecient matrices, the Bi s are k m coecient matrices, and Ut Rk is the vector of errors. This is an VARX model. The crucial condition for the correctness of this model is that k E Ut |{Ytj } (2) j =1 , {Xti }i=1 = 0 ( R ) Xt = c0 + C1 Xt1 + ... + Cr Xtr + Vt , , { X } = 0 ( Rm ). E Vt |{Ytj } t i j =1 i=1 (3)

Note that model (3) implies that Yt does not Granger-cause Xt , which is a weak form of exogeneity. 1

In principle one could also include a contempaneous Xt in (3): Yt = a0 + A1 Yt1 + ... + Ap Ytp +B0 Xt + B1 Xt1 + ... + Bq Xtq + Ut . (4)

However, if B0 6= O it may be possible that Yt has an indirect impact on Xt via possible mutual dependence of Ut and Vt . If so then (4) and (3) together form a (block-triangular) system of simultaneous equations. However, we can write (4) as a reduced form VARX model by substituting (3) in (4), Yt = a0 + B0 c0 + A1 Yt1 + ... + Ap Ytp +B0 (C1 Xt1 + ... + Cr Xtr ) +B1 Xt1 + ... + Bq Xtq + Ut + B0 Vt , which is of the form (1). Thus, the absence of a contemporaneous Xt in (1) is no loss of generality.

Exogeneity

By allowing some of the coecient matrices Aj , Bj and Cj to be zero matrices we may assume that p = q = r: Yt = a0 + A1 Yt1 + ... + Ap Ytp + B1 Xt1 + ... + Bp Xtp + Ut , Xt = c0 + C1 Xt1 + ... + Cp Xtp + Vt . (5)

Then we can write model (5) as a VAR(p) model with Granger-causality restrictions imposed: Yt Xt + a0 c0 A1 B1 Yt1 + + ... O C1 Xt1 Ap Bp Ytp Ut + . O Cp Xtp Vt

(6)

As said before, the Granger-causality restriction involved is a weak form of exogeneity: 2

Weak exogeneity of Xt : Yt does not Granger-cause Xt .

The usual assumption in VAR analysis is that the errors are i.i.d. normally distributed: Ut 0 11 12 i.i.d. Nk+m , . (7) Vt 0 21 22 A stronger exogeneity condition, in addition to the condition that Yt does not Granger-cause Xt , is now that 12 = O and thus 21 = 012 = O. This implies independence of Ut and Vt :

Strong exogeneity of Xt : Yt does not Granger-cause Xt , and the error vectors Ut and Vt are independent.

Testing weak exogeneity


+ A1 B1 Yt1 + + ... D1 C1 Xt1 Ap Bp Ytp Ut + , Dp Cp Xtp Vt

Setup the VAR model as an unrestricted VAR: Yt Xt a0 c0

(8)

say, and test the joint null hypothesis H0 : D1 = D2 = ... = Dp = 0, 3

using the Wald test option in EasyReg. If you do not reject this weak exogeneity hypothesis, and if you are not interested in the strong exogeneity hypothesis, you may re-estimate the VAR with the Granger causality restrictions imposed, in the form (6), and conduct VAR innovation response analysis, similarly to the example Yt = DIF1[LN[Income Sweden]], Xt = DIF1[LN[nominal GDP]] in the section Granger-causality testing in practice of EasyRegs Guided tour on VAR innovation response analysis.

Testing strong exogeneity

say, where L11 and L22 are lower-triangular matrices, so that L is a lowertriangular matrix. EasyReg provides the option to test the joint signicance of elements of the matrix L. Thus, test the null hypothesis H0 : L21 = O. This hypothesis is equivalent to the hypotheses that 12 = O and 21 = O. You can conduct this test after having re-estimated the VAR in the form (6), but it is better to conduct this test jointly with the test of weak exogeneity, using the unrestricted VAR (8), by testing the joint hypotheses H0 : D1 = D2 = ... = Dp = 0, L21 = O. Of course, if you have rejected the weak exogeneity hypothesis it makes no sense to test the strong exogeneity hypothesis.

The variance matrix in (7) can be written as 0 11 12 L11 O L11 O = = L.L0 , 21 22 L21 L22 L21 L22

Innovation response analysis under the strong exogeneity condition

In structural VAR analysis, the non-structural errors are related to the innovation via a matrix B,1 say. In our case, B11 B12 Ut e1,t Ut = = = et , B Vt B21 B22 Vt e2,t
1

Not related to the matrices Bi in models (1), (6) and (8).

say, where et =

and the errors Ut and Vt are those of the restricted model (6). Now specify B as 11 O B= , O 22 where 11 and 22 are lower-triangular matrices. Then under the strong exogeneity condition Ut Ik+m = Var B Vt 0 11 O 11 O 11 O = O 22 O 22 O 22 0 0 11 O L11 O L11 O 11 O = O 22 O L22 O L22 O 22 0 0 11 L11 L11 11 O = , O 22 L22 L022 022
1 1 which holds for 11 = L 11 and 22 = L22 . Innovation response analysis can now be conducted on the basis of this structural model. Because now 1 Ut 11 e1,t = , (9) 1 Vt 22 e2,t

e1,t e2,t

Nk+m [0, Ik+m ],

where e1,t and e2,t are independent, and because the VAR model is restricted such that Yt does not Granger-cause Xt , a shock in one of the innovations in e1,t will have no eect on Xt and its future values. To see this, write model (6) in VMA() form: Yt Xt
p X Aj = Ik+m O j =1 p X Aj Ik+m O j =1

B1 Cj B1 Cj

!1 L
j

a0 c0

!1

Ut Vt

(10)

1 2

X 11,i 12,i Uti + , 21,i 22,i Vti


i=0

say, where L is now the lag operator, and !1 p X X 11,i 12,i Aj B1 Li = Ik+m Lj . 21,i 22,i O Cj
i=0 j =1

Similarly, Xt can be written in VMA() form as Xt = 2 +


X i=0

2,i Vti .

(11)

Comparing (10) and (11), it follows that 21,i = O and 2,i = 22,i , so that after substituting (9) in (10) the latter becomes Yt Xt = 1 2 1 1 X 11,i e 12 ,i 1 ,t i 11 22 + . 1 O 22,i e2,ti 22
i=0

The innovation responses are now dened by Yt+h Yt+h E ,e e E Xt+h 1,ti 2,ti Xt+h 1 1 11,h 11 12,h 22 e1,t = , h = 0, 1, 2, .... 1 O 22,h 22 e2,t

(12)

Finally, consider the response of Yt and Xt to a unit shock in say the innovation in the rst component of Yt . Thus, replace e1,t in (12) by 1 = (1, 0, 0, ..., 0)0 Rk and e2,t by 0 Rm . Then the resulting innovation responses are Y t + h e1,ti = 1 , e2,ti = 0 E E Xt+h 1 11,h 11 1 = , h = 0, 1, 2, .... 0 Yt+h Xt+h

You might also like