2.8 Structural Analysis with VAR Models 43, of all the Information in the universe, only mnation in the past and present of the procoss tinder study is considered relevant and 12 is replaced ly < #}, Furthermore, instoad of optimal prodictors, optimal linear predictors are compared. In other words, =(h|$2,) is replaced by the linear minimum MSE f-step predictor based on the inforn dhl \ {r, tion in {=,.t6l9 < ¢} amd #}) is replaced by the linear minimum MSE h-step predictor \s < ¢}. In the following, when the terms “Granger-causality” nstantancous causality” are used, these restrictive assumptions are i used ifnot otherwise noted. Characterizati n of Granger-Causality Tn onder to determine the Granger-cusal relationships between the variables of the K-cimensional VAR process ys, suppose it has tho canonical MA rep resentation n=e¢Ftei=psAln =o ean) where u; is a white noise process with nonsingular covariance matrix Zy. Suppose that yy consists of the M-dimensional process 2, and the (K — M)- imensional process ¢ and the MA representation is partitioned accordingly, afm], [Pu G2(L))] fe l |- [a ]+ [ont wel “| 235) Using the prediction formula (2.2.10), the optimal L-step forecast of & based on ues se(H{vels <0) = [Lar = Olu) (236) = 1 DO aness i+ DOs, Hence the forecast error is — 2A fasls S tb) = ayer. (23.7) ‘As mentioned in Section 2.1.3, a subprocess of @ stationary process also has a prediction error MA representation. Thus, +S Fas, 238) where Fy = Ty and the last expression is a prediction orror MA reprosenta- tion. Thus, the optimal I-step predictor based on 2 only is M2 Stable Vector Autoregressive Processes ated $8) = 11+ Five (238) and the corresponding forecast ervor is Sepa — He f2.l8 < th) = test (2.3.10) Consequently, the predictors (2.3.6) and (23.9) are identical if and only if vy = ung for all & Tn other words, equality of the predictors is equivalent to = having the MA representation a= Hat Dasa Prate mnt Paes tases Uniqueness of the canonical MA reprosentation implios that 124 = 0 for i= 1,2,.... Hence, we get the following propositi Dy. and Proposition 2.2 (Chamctertzation of Granger-Noneausaity) Let gy be a VAR process as in (23.4)/(2.35) with canonical MA. operator (2). Then HCMUlS SO) = 2e(AM Gale SO) Hr, for t= 1,2,. (23.1) Because we have just used the MA representation (2.3.4) and not its finite order VAR form, the proposition is not only valid for VAR processes but ‘more generally for processes having a canonical NIA representation stich as (2.3.4), From (2.2.10) it isobvious that equality of the I-step predictors implies equality of the A-step predictors for h = 2,3,.... Hence, the proposition provides a necessary and sufficient condition for ry being not Granger-cansal for =, that 4s, 2 is not Granger-caused by x; if and only if P42, = 0 for = 1,2,.... Thus, Granger-noneausality can be checked easily by looking at the MA representation of yz. Becatse we are mostly: concerned with VAR processes, it is worth noting that for a stationary, stable VAR(p) process -[2]+[4e te] Ea] + [4 dee] [Ez] + [2]. ea.) H238 Structural Analysis with VAR Models 45 the condition (2.2.11) is satisfied if and only if Aig = 0 fori P. This result follows from the recursios the inverse of in (2.1.22) or, alternatively, beens [243 oto] [ 2(L)~! o | A Oag(LPPry(L)Oyy(L)-* Pea(L)* ‘Thus, we have the following result. Corollary 2.2.1 If y is a stable VAR(p) process as in (2.3.12) with nonsingular white noise covariance matrix Sy, then a(hl{unl SO) = a(hfaslo SY), k= 1,2... Aga 0 for im 1. d. (2313) Alternatively, xi(al(asls St) = achlfals St), = 1,2... Aga O for im 1p. (2314) ‘This corollary implics that noncausalities can be determined by just look- ing at the VAR representation of the system, For instance, for the example process (2.1.14), fe) (243])- 2 2= (dans Yue)’ doos not Granger-caus = = yyy bocanse Ayga = 0 if the coefficient’ matrix is partitioned according to (2.8.12). On the other hand, z+ Granger-causes 2%. To give thls discussion economic content Tet us assuune that the variables in the system are rates of change of investment (yt), in- comie (yz), and consumption (ys). With these specifieations, the previous discussion shows that investment Grangerseauses the consumption income system whereas the converse is not true. It is also easy to check that con sumption causes the income/investment system and vice versa, Note that so 452 Stable Vector Autoregressive Processes far we have defined Granger-causality only in terns of two groups of vati- les. Thatofore, at this stage, we cannot talk about the Granger-eausal r= lationship between consumption and income in the three-cimensional invest ment income/consumption system. Let us asstime that the variables in the example VAR(2) process (2.1.19), [as}eo+ [2s] [ter ]+L 2s 0] [te represont the inflation rato (y;), and some intorest rato (yz). Using Corollary 2.2.1, ibis immodiatoly obvions that inflation enuses tho intorost rato and vice versa, Hence the system is a feedback system. In the following. we will refer to (21.15) as the inflation interest rate system, i] Charactorization of Instantaneous Causality In order to study the concept of instantaneous causality in the framework of the MA process (2.3.5), it is usefil to rewrite that representation, Note thot the positive definite symmetric matrix %, can be written as the product, = PP", whero P is a lower triangular nonsingular matrix with positive dingonal elements (see Apponclix A.9.3), ‘Thus, (2.3.5) ean he represented as ” ne DHPP us et Dewi (2.3.15) where Op — BP aml wy — Phar io white noise with covariance matibs By = PAE MPY = Ik, (2.3.16) Because the white noise errors wy have uncorrelated components, they are led orthogonal residuals or innovations. joning the representation (2.8.15) according to the partitioning of = (2%)! gives Li [i] [ete of, I [en oa] [Sus Jee Hence, Fa mnt Orrottigsr + Ornawya + Praauns t and Bey = He + OaiatOrs omayyr F Oa e + so. Wae +2.8 Structural Analysis with VAR Models 47 ‘The optimal L-step predictor of 2; based on {us| < t} and, in addition, on 2441, i equal to the Lstep predictor of sr, based on (wal < OU feesa}s that is, aallltvsla S6)U (20) (Itt, = Ce, .105,.) 18 $8 Ufo) = Onpwres +riKuls 1. Consequently, it makes sense to define more refined concepts of causality which refer explicitly to the forecast horizon. For instance, ye may be called h-step noncausal for =. (yezny%e) for h = 1,2. if the #step ahead forecasts of z; cannot be improved for 7 < h by taking into account the information in past and present ge. Now the original concept of Granger-causality corresponds to infinite-step causality ‘The corresponding restrictions of multi-step causality on the VAR coetti- cents have been considered by Dufour & Renault (1998). Unlike in the biva ate setting explored earlier, now nonlinear restrictions on the VAR cooficients 50 2 Stable Vector Autoregressive Processes ‘are obtained which make it mote difficult to check for h-step causality ifthe information sot is expanded by additional variables ‘To state the restrictions formally, let A be defined as in the VAR(1) rep= resentation (2:18), let J = [lj 10: >>: 0) be a (AC x Kp) matrix as before and define AD) = JA? and a) = veo( AY). Dufour & Renault (1908) show that in the process (2.8.22), yey) i and only if Ro =0 for 5 hy (2323) and tA (9) HE and only if Ra) =0 for 5 PK, +1 (2324), Here the restriction tmatrix Ais such that Rvec{ 4a, Azyh, that is it collects tho elements of the sevon each of the coefficient matrices. ‘As an example consider again the dimensional VAR(1) process (2.1.14). te-step causality or nonceusality from gor to yie We need to check elements of the coefficient matrix and its second power 500 250 0 A= faa a], ag=| ‘06 o7 az]. 023 02 08 15 Ay] = veelAsyis- ++» lode in tho first row of Clearly, wae/qy¥is holds because 42,1 = @ and also the restrictions for to (jie ate satisfied in this case because the (1,2)-th element of AF is also zero. Tn contrast, wie? holds, while yie/(qysie does not, becatse ‘the lower left-hand element of A? io nonnero, Notice that the definition and chatacterizations of multistep causality are given for the first two sets of ssubvectors with the third one containing the extra variables, For applving the definition and results in the present example, the variables may just be rearranged acconlingly mn addition to these extensions related to increasing the information set, there are also other problems which may make it difficult to interpret Granger cansal tolations even in a bivariate setting, Lot usd ‘er of an inflation interest rate system, For example, it may make «differ- cance whether the information set contains monthly, quarterly or antmal date Ta quarterly system is considered and no easality is found from the iter= Cat rate to inflation it does not follow that a corresponding monthly interest rate has no impact on the monthly inflation rate, In other words, the interest rate may Granger-eause inflation in a monthly system even if ie doos not quarterly system, Furthermore, putting seasonally adjusted variables in the information st is not the same as using unadjusted variables, Consequontly, if Granger-causality is found for the seasonally adjusted variables, itis still possible that in the ac- tual seasonal system the interest rate is not Granger-causal for inflation. Sime ilar comments apply in the prosonce of measurement errors. Finally, causality uss some of th23 Structural alysis with VAR Models. 51 analyses are usually basod on estimated rathor than known systems. Addi- tional problems result in that case, We will returm to them in the next chapter. ‘The previous critical remarks are meant to caution the reader and multiple ‘me series analyst against overinterpreting the evidence from a YAR model. Still, causality analyses are useful tools in practice if these critical points are opt in mind, At the very least, a Granger-catsality analysis tells the analyst wwhethor a set of variables contains usoful information for improving the pro: dictions of another set of variables, Further discussions of eautsality istes snd ‘many further references may be found in Geweke (1984) and Granger (1982). 2.3.2 Impulse Response Analysis In the previous subsection, we have seen that 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 vatiable to fan impulse in another variable in a system that involves « number of further variables as woll. Thus, one would like to investigato the impulse response relationship between two variables in a higher dimensional system. Of course, if there isa reaction of ope Variable to an impulse in another variable we may call the latter causal for the former. In this subsection, we will study this type of causality by tracing out the elfect of an exogenous shock or innovation it one of the variables on some oF all of the other variables. This kind of impulse response analysis is sometimes called malfiphier analysis. For instance, in « system consisting of an inflation rate and an interest rate, the offect of an Imerease in the inflation 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 pricss. Altematively, 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. Responses to Forecast Errors Suppose the effect of an innovation in investment in a system containing Investment. (94), iecme (y), ancl consumption (ys) is of interest. To isolate such an effect, suppose that all three variables assume theit mean value prior totime £= 0, ye =p for €< 0, and investinent increases by one unit in period t= 0, that is, 1, — 1. Now we can trace out what happens to the system during periods t= 1,2,... if no further shocks occur, that is, w2,0 tu = 0, uy = 0,.... Because we are not interested int the me ‘ jm such an exercise but just in the variations of the variables around their tmeans, we assume that all three variables have mean zero and sat » — 0 in (2.1.14). Hence, ye = Anyes +; oF, more precisely, 52 2 Stable Vector Autoregressive Processes [E)-(EESEIE] a ‘Tracing a unit shock in the first variable in period m= [ms] tno in this system we get - Continuing the procedure, it turns ont that: ys = (Ws. Yi)" is Just the first column of Aj. An analogous line of arguments shows that a unit shock in mo (use) at € = 0, after 7 periods, results in a vector yi which is just the second (third) column of Aj. Thus, the elements of Al represent the effects of tit shocks in the variables of the system after i periods. Therefore they are called impulse responses or dynamic nai Recall that A; = ® is just the i-th coofficient matrix of the MA rep- resentation of VAR(1) process. Consequently, the MA coefficient matrices contain the impulse responses of the system. This result holds more gene ally for higher order VAR(p) processes as well. To sce this, suppose that yp is a stationary VAR(p) process as in (2.1.1) with » = 0. This process has corresponding VAR(I) process Y; = A¥s1 + U, as in (2.1.8) with » = 0. ‘Under the assumptions of the previous example, ys = 0 for t < 0, tu = 0 for (> 0 and yp = uo is a A-dimensional unit vector ex, say, with a one as the Fth coordinate and zoros eliewhere. It follows that ¥y = (¢4,0,---.0)" and ¥, = A'Ys. Hence, the impulse responses are the elements of the upper left-hand ({ x) block of A‘. "This matrix, however, was shown to be the ith cocflicient matrix #; of the MA representation (2.1.17) of yuo ies, Be = FAT = [lm £0: --- 10) a(K x Kp) matrix. In othor words, dye the k-th element of @, represents the reaction of the j-th variable of the system to & unit shock in variable &, # petiods ago, provided, of course, the effect is, not contaminated by other shocks to the system. Because the uy are just the step ahead forecast errors of the VAR process, the shocks considered here tay be regarded as forecast errors and the impubso responses are sometimes referred to as forerast error impulse responses. The response of variable j to a unit shock (forecast error) in variable ‘is sometimes depicted grophically to get a visual impression of the dy- namic intorrelationships within the system. Impulse responses of the invest tment income/consumption system aro plotted in Figure 25 andl the dynamicwith VAR Models | | | gt fore teem | Fig. 2.5. Impulse responses of the investment ncomo/consumption systom (im pase — response), responses of the inflation/interest rate system are depicted in Figure 2.6. For iustance, in the latter igure an inflation ianovation is seen wo induce the ine terest rate to increase for two poriods an then it tapers off to zero. In both, systems the effect of a unit shock in any of the variables dies away quite rapidly dia to the stability of the eystoms. Ifthe variables have different scales, it is Sometimes useful to consider in- ovations of one standard deviation rather than unit shocks. For instance, instoad of tracing an unexpected unit increase in investment in the ine westinent/ineome/consumption system with white noise covariance matrix (2.1.32), one may follow up on a shock of V2.25 = 15 units because the standard deviation of ue is 1.5. OF course, this is just a matter of rescaling the impulse responses. Ih Figures 25 and 2.6, it suffices to choose the units a the vertical axes equal to the standard deviations of the residuals corre: sponding to the variables whose effects ate considered, Such a rescaling may sometimes give a hettor picture ofthe dynamic relationships hecause the a\- crage sie of the innovations occurring in a system depends on their standard doviation. 54 2 Stable Vector Autoregressive Processes fr) intation + inttation fer interest ~ inflation 4 nr so #15, intlation - interest f24<) interest ~ interost d ge Ts So oF Fig. 2.6. Impulso responses of the sponse). flation/intorost rato system (impulse — ro. It follows from Proposition 2.2 that the impulse responses are zero if one of the variables does not Granger-cause the other variables taken as a group. Moro precisely, an innovation in vnriable E bass no effet em the other variables Hf the former variable does not Granger-cause the set of the remaining: vari- ables, As we have mentioned previously, in opplied work it is often of foremoet interest whether one variable has an impact on a specific other variable. That is, one would like to know whether, for some k # j, dj, = 0 for i= 1,2,..-- If the dja represent the actual renctions of variable j to a unit shock in variable &, we may call the latter noncausal forthe j-th variable if oj; = 0 for i — 1,2,.... In ordor to chock the lattor condition, it i not necassary to ‘compute infinitoly many ®, matrices. The following proposition shows that it sntfices to check the first p(X — 1) #, matrices. Proposition 2.4 (Zero Impuise Responses) Ify, is a K-dimensional stable VAR(p) process, then, for j # k, y= 0 fori=1,2, is equivalent to23 Structural Analysis with VAR Models 55 Ong20 fords tyes AK 1) . In other words, the proposition asserts that for a K-dimensionsl, station- ary, stable VAR(p), if the first pC ~ p responses of variable J to an impulse ‘m variable k are zero, all the following respouses must also be zero. For it~ stance, in the investiient/income/constmption VAR(1) system, because the responses of investment for the next two periods alter a consumption impulse fare zero, we Imow that investment will not react at all to such an impulse. Note, that in « VAR(I) system of dim 2, it does not suffice to check, say, the upper right-hand oon 1 coefficient matrix in omer to determine whether the last variable is noncansal for the fist vari able, Notice that Proposition 2.4 is related to the conditions for multi-step ‘causality in (231.28) and (2.3.24). In general, the conditions are not identical, wecause the two concepts differ, Proposition 2.4 will be helpful whet pulse response relations is discussed in the next chapter, Wo vi now prove the proposition, Proof of Proposition 2. Returning to the lag operator notation of Section 2.1 BL) = (Gja(L))0 = ACL) where A(E)® = (Ajy(E)) ie is the adjoint of A(L) = fx AL ~ (cee Appendix. Ai.1), Obviously, e(L) = 0 is equivalent to Aye(Z) = 0. Brom the definition ofa cofacter of n mati in Appendix. A.3, iis easy to see that Aj, (L) has degree not greater than pK =p. Defining 4(L) = [det A(L)}~!, wo got for k # j, OL) = onal + dyn al? +> = Anh) (AjeaL to + Aje px pL YL + Lt). we have A(E)%4 /dot(A(L)), baa Ane and 6ua= Ane SA — ford 1, & with Aji = 0 for > pip. Consequently, Ajai = 0 ford = 1,-..,pA—p, bs equivalent to dj = O for i= 1,2,...,pK—p, which proves the proposition. il Sometimes interest centers on the accumulated effect ever several o nore periods of a shock in one variable, This effect may be determine by summing up the MA coefficient matrices. For instance, the Ath colwman of Bq i= Sig Gy contains the accumulated responses over n periods to a 5B 2 Stable Veotor Autorogressive Procasies unit shock in the A- variable of the system, These quantities are some: times called mth interim multipliers. The total accumulated effects for all future periods are obtained by sumraing up all the MA coefficient matrices. Yu = CRy Ar is semetines called the matrix of long-run effects or total ‘multipliers. Bocauso the MA operator (2) is the inverse of the VAR operate tk ~ Ay2?, the long-nun effects are easily obtained as Way =O) = x = Ay == Ay) (2.3.26) As an example, accumulated responses for the investment /income/consump- tion aystem are depicted in Figure 2.7. Similarly, interim and total multipliers Of the inflation/interest rate system are shown in Figure 28. vig erm rmet—| fig = 7 || 34 + + Fig: 2.7. Accumulated and long-run responses of the investment sumption system (impulse — response). Responses to Orthogonal Impulses A problematic assumption in this type of impulse response anal ‘4 shock occurs only in one variable at a time. Such an assumption may be2.8 Structural Analysis with VAR Models 57 | Fig. 24 (impulse — response). reasonable ifthe shocks in different variables are independent. If they are not independent one may argue that the error terms consist of all the influences and variables that are not directly included in the set of y variables, Thus, in dition to foroos that affoct all the variables, thoro maybe forces that affoct variable 1, say, only. Ifa shock in the first variable fs due to stich forces it aay aguin be reasonable to interpret the &, codficients as dymamie responses On the other hand, correlation of the error tems may indicate that a shock in one variable is Bkely ta be accompanied by a shock in another variable. In that case, seting all other residuals to zero may provide a misleading picture ‘of tho actual dynamic relationships between the varihles. For example, in the investment/income/consumption system, the white nolse or innovation ‘covariance matrix is given in (2.1.38), 225 0 0 o 10 0 5 7 Obviously, there is a quite strong positive correlation between wa, and tay, the residuals of the income and constmption equations, respectively. Conse 58 2 Stable Vector Autoregressive Processes ‘quently, # shock in income may be accompanied Ly a shock in consumption in the same poried. Therefore, forcing the conssimption innovation to zero when ‘the effect of an income shock is traced, as in the previous analysis, may in fact obscure the actual relation between the variables, ‘This is the reason why impulse response analysis is often performed in terms of the MA representation (23.15), w= SPs (2.3.27) (ty,,-- 2x4)! are uncorrelated variance, Sy = Ix. The mean term is dropped again because it is of no inter- est in the present analysis. Recall that the representation (2.3.27) is obtained by decomposing Dy as Dy = PP!, where P is a lower triangular matrix, and defining 8; = %; P and w, = Py. In (23.27) it is reasonable to assume that ‘a change in one component of w; has no effect on the other components be- ‘cause the components are orthogonal (uncorrelated). Moreover, the variances of the components are one. Thus, a unit innovation is just an innovation of size cone standard deviation. The elements of the 6; are interpreted as responses Of the system to such innovations. More precisely, the jA-th element of Oj is assumed to represent the effect on variable j of a unit innovation in the k-th variable that has occurred # periods ago. ‘To relate those impulse responses to a VAR model, we consider the zero. mean VAR(p) process em Anya bot Apyinp (2.3.28) This process ean be rowritton in such a way that the residuals of different ‘equations are uncorrelated. For this purpose, we choose a decomposition of the white noise covariance matrix I, = WELW", where £; is a diagonal matrix ‘with positive diagonal elements and W is lower triangular matnix with unit, Aiagonal. This decomposition is obtained from the Choleski decomposition 2, = PP’ by defining a diagonal matrix D which has the same main diagonal fas P and by specifying W = PD~! and ©, = DD! Prenultiplying (2.3.28) by A= W-* gives Aye = Ajyaa to + Agutp bet, (2.3.29) whore Ay = Ady i covariance matrix, (ents 52We)! = Aug has diagonal ass and ce Be = Eleee}) = AB(wu,)A’ = ADA’. Adding (Iq, Aye to both sides of (2.3.29) gives (2.3.30)2.8 Structural Analyvis with VAR Modo 50 where Aj s= Ly — A. Because W is lower triangular with unit diagonal, the same is true for A. Hence, oo o 0 By 0 0 0 Mste-A= I LAK Ow Dee 0 {sa lower triangular matrix with zero diagonal and, thus, in the representation (23330) of our VAR(p) process the fist equation contains no istantancons 17 on tho righthand side, The sceond equation may contain yur and other- wise Inggod y's on the right-hand side. More generally, the k-th eeptation may contain piss. thie andl HOt tues. .+5 Ke o0 the tightehand side, ‘Thus, if (23.30) rstects the neuaal ongoings in the system, ay cannot have an insta taneous impact on yar for k-< s, In the economettics literature such a system called a recursive model (seo Theil (1071, Section 9.6)). Herman Wold has advoeated these models where the resoarcher has to specify the instantancons “vansal’ ordering of the variables. This type of cansality is therefore sometimes referred to as Wolt-oansality. If we trace ey, innovations of size one standard error through the system (23.80), we just get the © impulse responses. This ‘ean be scon by solving the systom (2.3.30) for yes w= (le A) Aya 0k (lg = A Ap lis = Ms) tee Noting that (Iq — Aj)~! = W = PD-* shows that the instantaneous effects of one-standard deviation shocks (cj¢'s of size one standard deviation) to the system are represented by the elements of WD = P = @p because the diagonal elements of D ate just standard dovintions of the components of ¢-. The 8, may then be obtained by tracing these effects through the system, ‘The 6's may provide response functions that are quite different from the 4, responses. For the example VAR(1) aystem (2.8.25) with © as in (2.1.33) we get a 0 S-P-} 0 10), aa ap wl a sh esa 3m 0 e020 064 090 055.105 60 2 Stable Vector Autoregressive Procasses fon responses are depicted in Figure 2.9, Ak though they are similar to those givon in Figure 2.5, thero is an obvious difference in tho response of consumption to an income innovation. While ‘constmption responds with a time lag of one period in Figure 2.5, there is an fnstantaneous effect in Figure 2.9. 2 [ecg t= eee os (orm emnmon hone or ig ema ewmmoin | fo econ mmetar | [ogy comm seman Fig. 2.9. Orthogonalized impulse responses of the investment incon system (impulse — response) sfeonsumpt Note that © ~ Pis lower triangular and some elements below the diagonal will be nonzero if £y has nonzero off-diagonal elements. For instance, for the investment /income consumption example y indicates that an income (y) innovation has an immediate impact on consumption (ys). Ifthe white noise covariance matrix 2%, contains Zeros, some components of ue = (uti, txt) are contemperancousty uncorrslated. Suppose, for instance, that wy¢ is uncor= rolated with a for i= 2,..2,4. In this caso, A=W"! and, thus, Aj has « Mock of zeros s0 that ya has no instantanconts effect on yy, #= 2..... In the ‘example, investment has no instantaneous impacton income and eonsumption Ddecause2.3 Structural Analysis with VAR Models 61 Zu and, hence, uy¢ is uncorrelated with uz, ancl ny. This, of courso, is reflected in the matrix of instantaneous effects Op given in (2.5.51). Because the elements of P =p roprosont the immediate responses of the systom to unit innovations they are sometimes called impact multipliers. In onder to determine whether there is no response at all of one vatiable to sn impulse in one of tho othor variables, it sufficos to consider the first pK — p response coefficients and the immediate effect. This result is stated formally in the next proposition where Oj. denotes the jk-th element of 6, Proposition 2.5 (Zero Orthogonalized Impulse Responses) Ify is a K-dimensional stable VAR(p) process, then, for j # k, a, es =0 for i=0,1, equivalent to 0.1... p(K ~ 1). ‘The proof of this result is analogous to that of Proposition 2.4 and is left as an exercise (see Problem 2.2), The fact that 8) is lowor triangular shows that the ordering of the vari- ables is of importance. that is. itis important which of the variables is called sn and which one is called yy 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 specified by the analyst. The order ing has to be such that the first variable is the only one with a potential jediate impact on all other variables. The second variable may have an immediate impact on the last K’— 2 components of ye bitt not on gue and so on. To establish such an ordering may be a quite difficult exercise in practice The choice of tho ordoring, the Wold causal ordering, may, to a large extent, determine the impulse responses and is therefore critical for the interpreta- tion of the eystom. Currently we are dealing with known eystems only. In this situation, assuming that the ordering is known may not be a great restric- tion. Por the investment/income/consumption example it may be reasonable to assume that an increase in income has an immediate effect on consumption while increased consumption stimulates the economy and, hence, income with some time lag. Our interpretation of orthogonalized impulse responses is based on the rep- resentation (2.3.80) and the impulses are viewed as changes in the observed variables. Sometimes it is more plausible to focus on impulses which cannot bo associated easily with changes in a specific observed variable within the 62 2 Stable Veotor Autorogromive Processes systom. In that case, it may be more logical to base the interpretation on the MA representation which decomposes the variables in contributions of the wie innovations. If these innovations can be associated with a specific impulse to the system, the orthogonalized impulse responses reflect the reac- tons of the variables to such possibly unobserved innovations. In that case, a specific impulse or shock to the system ean have an instantaneous impact on several variables while some other impulse may only effect on one specific variable and may effect the ot! ly some dolay. By decomposing ¥,, = PP’ with some non-triangular P matrix, it is also possible that all shocks have instantaneous effects on all observed ‘variables of the system, In this kind of interpretation, finding the decomposi- tion matrix P and, hence, the innovations a, which actually can be associated with shocks of Interest, is often a dificult part of the analysis, We will provide 4 more in-dopth discussion of the related problems in Chapter 9 which deals with structural VAR models. Critique of Impulse Response Analysis Besides specifying the relevant impulses to a system, there are a number of farther problems that render the interpretation of impulse responses difficult. ‘We have mentioned some of them in the context of Granger-causality. A major limitation of our systoms is their potontial incomplotoncss. Although in real economic systems almost everything depends on everything else, we will ust~ ally work with low-dimensional VAR systems. All effects of omitted variables are assumned to be in the innovations. [f important variables are omitted from the system, this may lead to major distortions in the impulse responses and makes thom worthless for structural interprotations. The systom may still be useful for prediction, though. ‘To ace the related problems more clearly, consider a system ye which is partitioned in vectors 2 and 2 as in (2.3.5). Ifthe 2 variables are considered only aud the 2, variables are omitted from the analysis, we get a system c= nt Senet Stans sin (2.3.8). The actual reactions ofthe = components to innovations we may be given by the 1 matrices. On the other hand, the For corresponding orthogonalized ‘impulse responses” are likely to be interpreted as impulse responses if the analyst: does not realize that important variables have been mitted. As we have seen in Section 2.3.1, the F, will be equal to the #1, if and only if 1 does not Granger-cas Further problems related to the interpretation of the MA coefficients as dynamic multipliers or impulse rosponsos rst from asnremont errors and2.8 Structural Analysis with VAR Models 63 the use of seasonally adjusted or temporally and/or contemporancously ag- sgrogated variables, A detailed account of the aggregation problem is given by: Liitkopohl (1987). We will discuss these probloms in more detail in Chepter 11 in tho contoxt of mor goneral models. Thes» problems soverely limit the interpretability of the MA coefficients of a VAR system as impulse responses. 1 further possibility to interprot VAR models will be In the next subsccti considered, 2.3.3 Forecast Error Variance Decomposition Ifthe innovations which actually drive the syste can be identified, a farther tool for interpreting VAR models is available. Suppose a recursive identifi cation scheme is available so that the MA representation (2.3.15) with or thogoual white noise imiovations may be considered, In the context of the represerttion went Dewi (2.2.39) with Sy = Jig, the error of the optimal hstep forocast is tern vel) = SS eiugs-s = Serr wns Denoting the m-th ekanent of 8, by Bas ts befur,the step forecast error of te jth component ofp visa yell) = SO jnenivanecbeet irae = LO noweren + pana eer): (2.3.35) ‘Thus, the forecast error of the j-th component potentially consists of all the Imovations wie,.+.,txe OF course, some of the @mn,¢ may be zero so that some components may not appear in (2.3.85). Because the wy,'s are uncom related and havo unit variances, the MSE of y¢(lt) is E(yseyn — yell) K LG a0t + Fenn) Thorfore, 612 Stabe Vector Autoregresive Process met Gay + Ba 42+ Bas = SG 800)? (23.36) is sometimes interpreted as the contribution of innovations in variable & to the forecast error variance or MSE of the Irstep forecast of viable j. Here ‘4 is the Keth clan of 1. Dividing (233.86) by toa MSElu.(0l] = 3° 0 O35 gine winn = TGP cea)"/MSElyye(0)) (eas7 which is the proportion of the h-stop forecast error variance of variable j, ac. counted for by wy, innovations. If wa, can be associated with variable i 4ji,4 represents the proportion of the h.step forecast error variance accounted for Dy innovations inv o is decom. ‘posed into ecmponents accounted for by innovations in the different variables of the system. From (2.3.44), the A-step forecast MSE matrix is seen to be USE| (M)] = D> 0.6; = Da, ‘The diagonal elements of this matrix are the MSEs of the yy, variables which nay be used in (2.8.87) Tor the investment/income/consamption example, forecast exzor variance lecompesitions of all three variables are given in Table 2.1, For instance, about 66% of the I-step forecast ervor variance of consumption is accounted ble k. Thereby, the forecast error var Bylh) For long torm forocasts, 57.5% and 42.3% of the error variance is accounted for by consumption and income innovations, respectively. For sny forecast horizon, investinent innovations contribute less than 1% to the forecast error variance of consumption. Moreover, only small fractions (less than 10%) of the forveast error variances of income are accounted for by innovations in the other variables of the system. TI 1 of analysis is some innovation accounting. From Proposition 2.5, it obvious that for a stationary, stable, K-dimen- sional VAR(p) process y, all forscast error variance proportions of variable Jy accounted for by innovations in variable &, will be zero if aay, = 0 for h = pK p+. Inthis context itis perhaps worth pointing ont the relationship Detween Granger-causality and forecast error variance components, For that purpose we consider a bivariate system w= (1,1) first. In suck a system, if 2, does not Granger-cause 27, the proportions of forecast error variances of 2 nes called