Universit` a di Pavia

Impulse Response Functions

Eduardo Rossi

VAR VAR(p): yt = (y1t , . . . , yN t ) (L)yt = t (L) = IN 1 L . . . p Lp t = (1t , . . . , N t ) t independent V W N (0, ) The process is stable if det(IN 1 z . . . p z p ) = 0 for |z | 1 On the assumption that the process has been initiated in the innite past (t = 0, 1, 2, . . .) it generates stationary time series that have time-invariant means, variances, and covariances.
Eduardo Rossi c Econometrics 10 2

VAR

Because VAR models represent the correlations among a set of variables, they are often used to analyze certain aspects of the relationships between the variables of interest.

Eduardo Rossi c

Econometrics 10

Granger-Causality Granger (1969) has dened a concept of causality which,under suitable conditions, is fairly easy to deal with in the context of VAR models. Therefore it has become quite popular in recent years. The idea is that a cause cannot come after the eect. If a variable x aects a variable z , the former should help improving the predictions of the latter variable. t is the information set containing all the relevant information in the universe available up to and including period t. zt (h|t ) be the optimal (minimum MSE) h-step predictor of the process zt at origin t, based on the information in t . The corresponding forecast MSE: t (h|t ).
Eduardo Rossi c Econometrics 10 4

Granger-Causality

The process xt is said to cause zt in Grangers sense if t (h|t ) < t (h|t {xs : s t}) for at least one h = 1, 2, . . .

t {xs : s t} is the set containing all the relevant information in the universe except for the information in the past and present of the xt process. If zt can be predicted more eciently if the information in the xt process is taken into account in addition to all other information in the universe, then xt is Granger-causal for zt .

Eduardo Rossi c

Econometrics 10

Instantaneous Causality

If xt : (N 1) causes zt : (M 1) and zt also causes xt the process (z t , xt ) is called a feedback system. We say that there is instantaneous causality between zt and xt if z (1|t xt+1 ) = z (1|t ) In other words, in period t, adding xt+1 to the information set helps to improve the forecast of zt+1 . This concept of causality is really symmetric, that is, if there is instantaneous causality between zt and xt , then there is also instantaneous causality between xt and zt .

Eduardo Rossi c

Econometrics 10

Instantaneous Causality

A possible criticism of the foregoing denitions could relate to the choice of the MSE as a measure of the forecast precision. Of course, the choice of another measure could lead to a dierent denition of causality. Equality of the MSEs will imply equality of the corresponding predictors. In that case a process zt is not Granger-caused by xt if the optimal predictor of zt does not use information from the xt process. This result is intuitively appealing.

Eduardo Rossi c

Econometrics 10

Instantaneous Causality

A more serious practical problem is the choice of the information set t . Usually all the relevant information in the universe is not available to a forecaster and, thus, the optimal predictor given t cannot be determined. Therefore a less demanding denition of causality is often used in practice.

Eduardo Rossi c

Econometrics 10

Instantaneous Causality

Instead of all the information in the universe, only the information in the past and present of the process under study is considered relevant and t is replaced by {zs , xs |s t}. Furthermore, instead of optimal predictors, optimal linear predictors are compared.

Eduardo Rossi c

Econometrics 10

Impulse responses functions Granger-causality may not tell us the complete story about the interactions between the variables of a system. In applied work, it is often of interest to know the response of one variable to an impulse in another variable in a system that involves a number of further variables as well. One would like to investigate the impulse response relationship between two variables in a higher dimensional system. Of course, if there is a reaction of one variable to an impulse in another variable we may call the latter causal for the former. We will study this type of causality by tracing out the eect of an exogenous shock or innovation in one of the variables on some or all of the other variables. This kind of impulse response analysis is called multiplier analysis.
Eduardo Rossi c Econometrics 10 10

Impulse responses functions

For instance, in a system consisting of an ination rate and an interest rate, the eect of an increase in the ination rate may be of interest. In the real world, such an increase may be induced exogenously from outside the system by events like the increase of the oil price in 1973/74 when the OPEC agreed on a joint action to raise prices. Alternatively, an increase or reduction in the interest rate may be administered by the central bank for reasons outside the simple two variable system under study.

Eduardo Rossi c

Econometrics 10

11

Impulse responses functions

Three-variables system: investment y1,t yt = y2,t income y3,t consumption the eect of an innovation in investment. To isolate such an eect, suppose that all three variables assume their mean value prior to time t = 0, yt = , t < 0, and investment increases by one unit in period t = 0, i.e. 10 = 1.

Eduardo Rossi c

Econometrics 10

12

Impulse responses functions

Now we can trace out what happens to the system during periods t = 1, 2, . . . if no further shocks occur, that is, 20 = 30 = 0 and 2 = 0, 3 = 0, . . . We assume that all three variables have mean zero and set c = 0, then yt = 1 yt1 + t

Eduardo Rossi c

Econometrics 10

13

Impulse responses functions

y1,t

.5

y1,t1

1,t

yt = y2,t = .1 .1 .3 y2,t1 + 2,t y3,t 0 .2 .3 y3,t1 3,t Tracing a unit shock in the rst variable in period t = 0 in this system we get y1,0 1 , 0 1 y0 = y2,0 = 2,0 = 0 y3,0 3 , 0 0 y1,1 0.5

y1 = y2,1 = 1 y0 = 0.1 y3,1 0

Eduardo Rossi c Econometrics 10 14

Impulse responses functions

2 y2 = y2,2 = 1 y1 = 1 y0 = 0.06 y3,2 0.02 it turns out that yi = (y1,i , y2,i , y3,i ) is just the rst column of i 1 yi = i 1 ui where u i = (0, 0, . . . , 1, . . . , 0) the elements of i 1 represent the eects of unit shocks in the variables of the system after i periods. They are called impulse responses or dynamic multipliers.

y1,2

0.5

Eduardo Rossi c

Econometrics 10

15

Impulse responses functions

Recall that i 1 = i just the i-th coecient matrix of the MA representation of a VAR(1) process. The MA coecient matrices contain the impulse responses of the system. This result holds more generally for higher order VAR(p) processes as well. V M A () representation:

yt =
i=0

i ti

0 = In

Eduardo Rossi c

Econometrics 10

16

Impulse responses functions

Impulse-response function

yt+n =
i=0

i t+ni yit+n jt

{n }i,j =

the response of yi,t+n to a one-time impulse in yj,t with all other variables dated t or earlier held constant. The response of variable i to a unit shock (forecast error) in variable j is sometimes depicted graphically to get a visual impression of the dynamic interrelationships within the system.

Eduardo Rossi c

Econometrics 10

17

Impulse responses functions If the variables have dierent scales, it is sometimes useful to consider innovations of one standard deviation rather than unit shocks. For instance, instead of tracing an unexpected unit increase in investment in the investment/income/consumption system with V ar (1,t ) = 2.25, one may follow up on a shock of 2.25 = 1.5 units because the standard deviation of 1,t is 1.5. Of course, this is just a matter of rescaling the impulse responses. The impulse responses are zero if one of the variables does not Granger-cause the other variables taken as a group. An innovation in variable k has no eect on the other variables if the former variable does not Granger-cause the set of the remaining variables.
Eduardo Rossi c Econometrics 10 18

The Orthogonalized Impulse-Response Function A problematic assumption in this type of impulse response analysis is that a shock occurs only in one variable at a time. Such an assumption may be reasonable if the shocks in dierent variables are independent. If they are not independent one may argue that the error terms consist of all the inuences and variables that are not directly included in the set of y variables. Thus, in addition to forces that aect all the variables, there may be forces that aect variable 1, say, only. If a shock in the rst variable is due to such forces it may again be reasonable to interpret the i coecients as dynamic responses.

Eduardo Rossi c

Econometrics 10

19

The Orthogonalized Impulse-Response Function

On the other hand, correlation of the error terms may indicate that a shock in one variable is likely to be accompanied by a shock in another variable. In that case, setting all other errors to zero may provide a misleading picture of the actual dynamic relationships between the variables.

Eduardo Rossi c

Econometrics 10

20

The Orthogonalized Impulse-Response Function This is the reason why impulse response analysis is often performed in terms of the MA representation: = PP Cholesky decomposition

P is a lower triangular matrix. From yt = t + 1 t1 + 2 t2 + . . . to yt = 0 wt + 1 wt1 + 2 wt2 + . . . with i = i P wt = P1 t

E [wt wt ] = IN

0 = In

Eduardo Rossi c

Econometrics 10

21

The Orthogonalized Impulse-Response Function It is reasonable to assume that a change in one component of wt has no eect on the other components because the components are orthogonal (uncorrelated). Moreover, the variances of the components are one. Thus, a unit innovation is just an innovation of size one standard deviation. The elements of the i are interpreted as responses of the system to such innovations. {i }jk is assumed to represent the eect on variable j of a unit innovation in the k -th variable that has occurred i periods ago.

Eduardo Rossi c

Econometrics 10

22

The Orthogonalized Impulse-Response Function

To relate these impulse responses to a VAR model, we consider the zero mean VAR(p)process: (L)yt = t This process can be rewritten in such a way that the disturbances of dierent equations are uncorrelated. For this purpose, we choose a decomposition of the white noise covariance matrix = WW is a diagonal matrix with positive diagonal elements and W is a lower triangular matrix with unit diagonal.

Eduardo Rossi c

Econometrics 10

23

The Orthogonalized Impulse-Response Function

This decomposition is obtained from the Choleski decomposition = PP by dening a diagonal matrix D which has the same main diagonal as P and by specifying W = PD1 = DD then from = PP and P = WD = PP = WDD W = WW

Eduardo Rossi c

Econometrics 10

24

The Orthogonalized Impulse-Response Function

Premultiplying the VAR(p) by A W1

Ayt = A 1 yt1 + . . . + Ap ytp + et

where A i = Ai et = At

Eduardo Rossi c

Econometrics 10

25

The Orthogonalized Impulse-Response Function

et has a diagonal var-cov matrix:

E [e t e ] = A E [ ] A = AA t t t

Adding (IN A)yt to both sides gives

(IN A)yt + Ayt = (IN A)yt + A 1 yt1 + . . . + Ap ytp + et y t = A y + A y + . . . + A t t 1 0 1 p ytp + et

Because W is lower triangular with unit diagonal, the same is true for A.

Eduardo Rossi c

Econometrics 10

26

The Orthogonalized Impulse-Response Function Hence, A 0 = (IN A) 0 0 0 12 . .. . = . . . . . N,1 N,2

... ...

0 0 .. .

0 0 . . . . . . 0

. . . N,N 1

is a lower triangular matrix with zero diagonal and, thus, in the representation of the VAR(p) process, the rst equation contains no instantaneous y s on the right-hand side. The second equation may contain y1,t and otherwise lagged y s on the right-hand side. More generally,the k -th equation may contain y1,t , . . . , yk1,t and not yk,t , . . . , yN,t on the right-hand side.
Eduardo Rossi c Econometrics 10 27

The Orthogonalized Impulse-Response Function

Thus, if the VAR(p) with instantaneous eects reects the actual ongoings in the system, yst cannot have an instantaneous impact on ykt for k < s. In the econometrics literature such a system is called a recursive model (Theil (1971)). Wold has advocated these models where the researcher has to specify the instantaneous causal ordering of the variables. This type of causality is therefore sometimes referred to as Wold-causality.

Eduardo Rossi c

Econometrics 10

28

The Orthogonalized Impulse-Response Function

If we trace eit innovations of size one standard error through the system, we just get the impulse responses. This can be seen by solving the system for yt :
(IN A 0 )yt = A1 yt1 + . . . + Ap ytp + et 1 1 1 yt = (IN A A1 yt1 + . . . +(IN A Ap ytp +(IN A et 0) 0) 0)

1 1 Noting that (IN A ) = W = PD shows that the instantaneous 0 eects of one-standard deviation shocks (eit s of size one standard deviation) to the system are represented by the elements of

WD = P = 0 because the diagonal elements of D are just standard deviations of the components of et .

Eduardo Rossi c

Econometrics 10

29

The Orthogonalized Impulse-Response Function

The i may provide response functions that are quite dierent from the i responses. Note that 0 = P is lower triangular and some elements below the diagonal will be nonzero if has nonzero o-diagonal elements. For the investment/income/consumption example 1.5 0 0 0 = P = 0 1 0 0 0.5 0.7 indicates that an income (y2 ) innovation has an immediate impact on consumption (y3 ).

Eduardo Rossi c

Econometrics 10

30

The Orthogonalized Impulse-Response Function If the white noise covariance matrix contains zeros, some components of t are contemporaneously uncorrelated. Suppose,for instance, that 1t is uncorrelated with it , i = 2, . . . , N . In this case, A = W1 and thus A 0 has a block of zeros so that y1 has no instantaneous eect on yi , i = 2, . . . , N . In the example,investment has no instantaneous impact on income and consumption because 2.25 0 0 = 0 1 0.5 0 0.5 0.74 1t is uncorrelated with 2t and 3t . This, of course, is reected in the matrix of instantaneous eects 0 (impact multipliers ).

Eduardo Rossi c

Econometrics 10

31

The Orthogonalized Impulse-Response Function

The fact that 0 is lower triangular shows that the ordering of the variables is of importance, that is, it is important which of the variables is called y1 and which one is called y2 and so on. One problem with this type of impulse response analysis is that the ordering of the variables cannot be determined with statistical methods but has to be specied by the analyst. The ordering has to be such that the rst variable is the only one with a potential immediate impact on all other variables. The second variable may have an immediate impact on the last N 2 components of yt but not on y1t and so on. To establish such an ordering may be a quite dicult exercise in practice.

Eduardo Rossi c

Econometrics 10

32

The Orthogonalized Impulse-Response Function

The choice of the ordering, the Wold causal ordering, may, to a large extent, determine the impulse responses and is therefore critical for the interpretation of the system. For the investment/income/consumption example it may be reasonable to assume that an increase in income has an immediate eect on consumption while increased consumption stimulates the economy and, hence, income with some time lag.

Eduardo Rossi c

Econometrics 10

33

The Orthogonalized Impulse-Response Function

Besides specifying the relevant impulses to a system, there are a number of further problems that render the interpretation of impulse responses dicult. A major limitation of our systems is their potential incompleteness.Although in real economic systems almost everything depends on everything else, we will usually work with low-dimensional VAR systems. All eects of omitted variables are assumed to be in the innovations. If important variables are omitted from the system,this may lead to major distortions in the impulse responses and makes them worthless for structural interpretations.

Eduardo Rossi c

Econometrics 10

34

The Orthogonalized Impulse-Response Function Consider a system yt which is partitioned in vectors zt and xt , with VMA representation zt 1 11 (L) 12 (L) 1t yt = = + xt 2 21 (L) 22 (L) 2t If the zt variables are considered only and the xt variables are omitted from the analysis, we get a prediction error MA representation:

zt = 1 +
i=0

11,i 1ti +
i=1

12,i 2ti

zt = 1 +
i=0

Fi vti

The actual reactions of the zt components to innovations 1t may be given by the 11,i matrices.
Eduardo Rossi c Econometrics 10 35

The Orthogonalized Impulse-Response Function

On the other hand,the Fi or corresponding orthogonalized impulse response are likely to be interpreted as impulse responses if the analyst does not realize that important variables have been omitted. The Fi will be equal to the 11,i , if and only if xt does not Granger-cause zt .

Eduardo Rossi c

Econometrics 10

36

MA Representation The stable VAR(p) has a Wold MA representation: yt = t + 1 t1 + 2 t2 + . . . where s = s j =1 sj j s = 1, 2, . . .

with 0 = IK . The elements of the j are the forecast error impulse responses. Dierent ways to orthogonalize the impulses. Cholesky decomposition of . Such an approach is arbitrary, unless there are special reasons for a recursive structure. Dierent ways to use nonsample information in specifying unique innovations and hence unique i-r.

Eduardo Rossi c

Econometrics 10

37

SVAR

Structural form model Ayt = 1 yt1 + 2 yt2 + . . . + p ytp + Bet The matrix A contains the instantaneous relations between the left-hand-side variables. It has to be invertible.

Eduardo Rossi c

Econometrics 10

38

SVAR

The reduced forms are obtained by premultiplying with A1 yt = A1 1 yt1 + A1 2 yt2 + . . . + A1 p ytp + A1 Bet t = A1 Bet which relates the reduced-form disturbances t to the underlying structural shocks et .

Eduardo Rossi c

Econometrics 10

39

SVAR To identify the structural form parameters, we must place restrictions on the parameter matrices. Even if A = IN the assumption of orthogonal shocks E [e t e t ] = IN is not sucient to achieve identication. For a N -dimensional system,
1) orthogonalizing the shocks because there are N (N potentially 2 dierent instantaneous covariances. N (N 1) 2

restrictions are necessary for

An example of such an identication scheme is the triangular (or recursive) identication suggested by Sims (1980). Such a scheme is also called Wold causal chain system and is often associated with a causal chain from the rst to the last variable in the system.

Eduardo Rossi c

Econometrics 10

40

Identification

Because the i-r functions computed from these models depend on the ordering of the variables, nonrecursive identication schemes that also allow for instantaneous eects of the variables A = IN have been suggested in the literature (Sims (1986), Bernanke (1986)). Restrictions on the long-run eects of some shocks are also used to identify SVAR models (Blanchard and Quah (1989), Gal` (1999), King, Plosser, Stock, and Watson (1991)).

Eduardo Rossi c

Econometrics 10

41

Identification

At = Bet The most popular kind of restrictions: B = IN , the A-model. A = IN , the innovation is t = Bet , the B-model. the AB-model Prior information.

Eduardo Rossi c

Econometrics 10

42