Moneta Chlass Entner Hoyerto the ‘right’ rotation of the VAR model, that is the rotation compatible both with thecontemporaneous causal structure of the variable and the structure of the <strong>in</strong>novationterm. Let us consider a matrix B 0 = I − Γ 0 . If the system is normalized such that thematrix Γ 0 has all the elements of the pr<strong>in</strong>cipal diagonal equal to one (which can be donestraightforwardly), the diagonal elements of B 0 will be equal to zero. We can write:Y t = B 0 Y t + Γ 1 Y t−1 + ... + Γ p Y t−p + ε t (6)from which we see that B 0 (and thus Γ 0 ) determ<strong>in</strong>es <strong>in</strong> which form the values of a variableY i,t will be dependent on the contemporaneous value of another variable Y j,t . The‘right’ rotation will also be the one which makes ε t a vector of authentic <strong>in</strong>novationterms, which are expected to be <strong>in</strong>dependent (not only over time, but also contemporaneously)sources or shocks.In the literature, different methods have been proposed to identify the SVAR model(4) on the basis of the estimation of the VAR model (5). Notice that there are moreunobserved parameters <strong>in</strong> (4), whose number amounts to k 2 (p + 1), than parametersthat can be estimated from (5), which are k 2 p + k(k + 1)/2, so one has to impose atleast k(k − 1)/2 restrictions on the system. One solution to this problem is to get arotation of (5) such that the covariance matrix of the SVAR residuals Σ ε is diagonal,us<strong>in</strong>g the Cholesky factorization of the estimated residuals Σ u . That is, let P be thelower-triangular Cholesky factorization of Σ u (i.e. Σ u = PP ′ ), let D be a k × k diagonalmatrix with the same diagonal as P, and let Γ 0 = DP −1 . By pre-multiply<strong>in</strong>g (5) byΓ 0 , it turns out that Σ ε = E[Γ 0 u t u ′ t Γ′ 0 ] = DD′ , which is diagonal. A problem withthis method is that P changes if the order<strong>in</strong>g of the variables (Y 1t ,...,Y kt ) ′ <strong>in</strong> Y t and,consequently, the order of residuals <strong>in</strong> Σ u , changes. S<strong>in</strong>ce researchers who estimate aSVAR are often exclusively <strong>in</strong>terested on track<strong>in</strong>g down the effect of a structural shockε it on the variables Y 1,t ,...,Y k,t over time (impulse response functions), Sims (1981)suggested <strong>in</strong>vestigat<strong>in</strong>g to what extent the impulse response functions rema<strong>in</strong> robustunder changes of the order of variables.Popular alternatives to the Cholesky identification scheme are based either on theuse of a priori, theory-based, restrictions or on the use of long-run restrictions. Theformer solution consists <strong>in</strong> impos<strong>in</strong>g economically plausible constra<strong>in</strong>ts on the contemporaneous<strong>in</strong>teractions among variables (Blanchard and Watson, 1986; Bernanke,1986) and has the drawback of ultimately depend<strong>in</strong>g on the a priori reliability of economictheory, similarly to the Cowles Commission approach. The second solution isbased on the assumptions that certa<strong>in</strong> economic shocks have long-run effect to othervariables, but do not <strong>in</strong>fluence <strong>in</strong> the long-run the level of other variables (see Shapiroand Watson, 1988; Blanchard and Quah, 1989; K<strong>in</strong>g et al., 1991). This approach hasbeen criticized as not be<strong>in</strong>g very reliable unless strong a priori restrictions are imposed(see Faust and Leeper, 1997).In the rest of the paper, we first present a method, based on the graphical causalmodel framework, to identify the SVAR (section 2). This method is based on conditional<strong>in</strong>dependence tests among the estimated residuals of the VAR estimated model.Such tests rely on the assumption that the shocks affect<strong>in</strong>g the model are Gaussian.108
Causal Search <strong>in</strong> SVARWe then relax the Gaussianity assumption and present a method to identify the SVARmodel based on <strong>in</strong>dependent component analysis (section 3). Here the ma<strong>in</strong> assumptionis that shocks are non-Gaussian and <strong>in</strong>dependent. F<strong>in</strong>ally (section 4), we explorethe possibility of extend<strong>in</strong>g the framework for causal <strong>in</strong>ference to a nonparametric sett<strong>in</strong>g.In section 5 we wrap up the discussion and conclude by formulat<strong>in</strong>g some openquestions.2. SVAR identification via graphical causal models2.1. BackgroundA data-driven approach to identify the structural VAR is based on the analysis of theestimated residuals û t . Notice that when a basic VAR model is estimated (equation 3),the <strong>in</strong>formation about contemporaneous causal dependence is <strong>in</strong>corporated exclusively<strong>in</strong> the residuals (be<strong>in</strong>g not modeled among the variables). Graphical causal models,as orig<strong>in</strong>ally developed by Pearl (2000) and Spirtes et al. (2000), represent an efficientmethod to recover, at least <strong>in</strong> part, the contemporaneous causal structure mov<strong>in</strong>g fromthe analysis of the conditional <strong>in</strong>dependencies among the estimated residuals. Once thecontemporaneous causal structure is recovered, the estimation of the lagged autoregressivecoefficients permits us to identify the complete SVAR model.This approach was <strong>in</strong>itiated by Swanson and Granger (1997), who proposed to testwhether a particular causal order of the VAR is <strong>in</strong> accord with the data by test<strong>in</strong>g all thepartial correlations of order one among error terms and check<strong>in</strong>g whether some partialcorrelations are vanish<strong>in</strong>g. Reale and Wilson (2001), Bessler and Lee (2002), Demiralpand Hoover (2003), and Moneta (2008) extended the approach by us<strong>in</strong>g the partialcorrelations of the VAR residuals as <strong>in</strong>put to graphical causal model search algorithms.In graphical causal models, the structural model is represented as a causal graph (aDirected Acyclic Graph if the presence of causal loops is excluded), <strong>in</strong> which each noderepresents a random variable and each edge a causal dependence. Furthermore, a setof assumptions or ‘rules of <strong>in</strong>ference’ are formulated, which regulate the relationshipbetween causal and probabilistic dependencies: the causal Markov and the faithfulnessconditions (Spirtes et al., 2000). The former restricts the jo<strong>in</strong>t probability distributionof modeled variables: each variable is <strong>in</strong>dependent of its graphical non-descendantsconditional on its graphical parents. The latter makes causal discovery possible: all ofthe conditional <strong>in</strong>dependence relations among the modeled variables follow from thecausal Markov condition. Thus, for example, if the causal structure is represented asY 1t → Y 2t → Y t,3 , it follows from the Markov condition that Y 1,t ⊥ Y 3,t |Y 2,t . If, on theother hand, the only (conditional) <strong>in</strong>dependence relation among Y 1,t ,Y 2,t ,Y 3,t is Y 1,t ⊥Y 3,t , it follows from the faithfulness condition that Y 1,t → Y 3,t