# Variance Decomposition

**DOI:**https://doi.org/10.1057/978-1-349-95189-5_2274

## Abstract

Variance decomposition is a classical statistical method in multivariate analysis for uncovering simplifying structures in a large set of variables (for example, Anderson 2003). For example, *factor analysis or principal components* are tools that are in widespread use. Factor analytic methods have, for instance, been used extensively in economic forecasting (see for example, Forni et al. 2000; Stock and Watson 2002). In macroeconomic analysis the term ‘variance decomposition’ or, more precisely, ‘forecast error variance decomposition’ is used more narrowly for a specific tool for interpreting the relations between variables described by vector autoregressive (VAR) models. These models were advocated by Sims (1980) and used since then by many economists and econometricians as alternatives to classical simultaneous equations models. Sims criticized the way the latter models were specified, and questioned in particular the exogeneity assumptions common in simultaneous equations modelling.

### Keywords

Bayesian methods Bootstrap Choleski decompositions Cointegrated variables Cointegration Factor analysis Forecasting Least squares Maximum likelihood Multivariate analysis Principal components Simultaneous equations models Structural vector autoregressions Variance decomposition Vector autoregressionsVariance decomposition is a classical statistical method in multivariate analysis for uncovering simplifying structures in a large set of variables (for example, Anderson 2003). For example, *factor analysis or principal components* are tools that are in widespread use. Factor analytic methods have, for instance, been used extensively in economic forecasting (see for example, Forni et al. 2000; Stock and Watson 2002). In macroeconomic analysis the term ‘variance decomposition’ or, more precisely, ‘forecast error variance decomposition’ is used more narrowly for a specific tool for interpreting the relations between variables described by vector autoregressive (VAR) models. These models were advocated by Sims (1980) and used since then by many economists and econometricians as alternatives to classical simultaneous equations models. Sims criticized the way the latter models were specified, and questioned in particular the exogeneity assumptions common in simultaneous equations modelling.

*y*

_{t}= (

*y*

_{1t}, …,

*y*

_{Kt})

^{′}(the prime denotes the transpose) is a vector of

*K*observed variables of interest, the

*A*

_{i}’s are (

*K*×

*K*) parameter matrices,

*p*is the lag order and

*u*

_{t}is a zero mean error process which is assumed to be white noise, that is,

*E*(

*u*

_{t}) = 0, the covariance matrix, \( E\left({u}_t{u}_t^{\prime}\right)={\Sigma}_u \), is time invariant and the

*u*

_{t}’s are serially uncorrelated or independent. Here deterministic terms such as constants, seasonal dummies or polynomial trends are neglected because they are of no interest in the following. In the VAR model (1) all the variables are a priori endogenous. It is usually difficult to disentangle the relations between the variables directly from the coefficient matrices. Therefore it is useful to have special tools which help with the interpretation of VAR models. Forecast error variance decompositions are such tools. They are presented in the following.

*h*steps ahead forecast or briefly

*h*-step forecast at origin t can be obtained from (1) recursively for

*h*= 1, 2, …, as

*y*

_{t + j|t}=

*y*

_{t + j}for

*j*≤ 0. The forecast error turns out to be

_{h}. Here the Φ

_{i}’s are the coefficient matrices of the power series expansion \( {\left({I}_K-{A}_1z-\cdots -{A}_p{z}^p\right)}^{-1}={I}_K+{\sum}_{i=1}^{\infty }{\Phi}_i{z}^i. \) Note that the inverse exists in a neighbourhood of

*z*= 0 even if the VAR process contains integrated and cointegrated variables. (For an introductory exposition of forecasting VARs, see Lütkepohl 2005.)

*u*

_{t}can be decomposed in instantaneously uncorrelated innovations with economically meaningful interpretation, say,

*u*

_{t}=

*Bε*

_{t}with

*ε*

_{t}~ (0,

*I*

_{K}), then Σ

_{u}=

*BB*

^{′}and the forecast error variance can be written as \( {\Sigma}_h={\sum}_{i=0}^{h-1}{\Theta}_i{\Theta}_i^{\prime } \), where Θ

_{0}=

*B*and Θ

_{i}= Φ

_{i}

*B*;

*i*= 1, 2, …. Denoting the (

*n*,

*m*)th element of Θ

_{j}by

*θ*

_{nm, j}, the forecast error variance of the kth element of the forecast error vector is seen to be

*j*th innovation to the

*h*-step forecast error variance of variable

*k*. Dividing the term by \( {\sigma}_k^2(h) \) gives the percentage contribution of innovation

*j*to the

*h*-step forecast error variance of variable

*k*. This quantity is denoted by

*ω*

_{kj, h}in the following. The

*ω*

_{kj, h},

*j*= 1, … ,

*K*, decompose the

*h*-step ahead forecast error variance of variable

*k*in the contributions of the

*ε*

_{t}innovations. They were proposed by Sims (1980) and are often reported and interpreted for various forecast horizons.

For such an interpretation to make sense it is important to have economically meaningful innovations. In other words, a suitable transformation matrix *B* for the reduced form residuals has to be found. Clearly, *B* has to satisfy Σ_{u} = *BB*^{′}. These relations do not uniquely determine *B*, however. Thus, restrictions from subject matter theory are needed to obtain a unique *B* matrix and, hence, unique innovations *ε*_{t}. A number of different possible sets of restrictions and approaches for specifying restrictions have been proposed in the literature in the framework of *structural VAR models*. A popular example is the choice of a lower-triangular matrix *B* obtained by a Choleski decomposition of Σ_{u} (for example, Sims 1980). Such a choice amounts to setting up a system in recursive form where shocks in the first variable have potentially instantaneous effects also on all the other variables, shocks to the second variable can also affect the third to last variable instantaneously, and so on. In recursive systems it may be possible to associate the innovations with variables, that is, the *j*th component of *ε*_{t}is primarily viewed as a shock to the *j*th observed variable. Generally, the innovations can also be associated with unobserved variables, factors or forces and they may be named accordingly. For example, Blanchard and Quah (1989) consider a bivariate model for output and the unemployment rate, and they investigate effects of supply and demand shocks. Generally, if economically meaningful innovations can be found, forecast error variance decompositions provide information about the relative importance of different shocks for the variables described by the VAR model.

Estimation of reduced form and structural form parameters of VAR processes is usually done by least squares, maximum likelihood or Bayesian methods. Estimates of the forecast error variance components, *ω*_{kj, h}, are then obtained from the VAR parameter estimates. Suppose the VAR coefficients are contained in a vector *α*, then *ω*_{kj, h} is a function of *α*, *ω*_{kj, h} = *ω*_{kj, h}(*a*). Denoting the estimator of *α* by \( \widehat{\alpha} \), *ω*_{kj, h} may be estimated as \( {\widehat{\omega}}_{kj,h}={\sigma}_{kj,h}\left(\widehat{\alpha}\right) \). If \( \widehat{\alpha} \) is asymptotically normal, that is, \( \sqrt{T}\left(\widehat{\alpha}-\alpha \right)\underrightarrow{d}\mathcal{N}\left(0,{\Sigma}_{\widehat{\alpha}}\right) \), then, under general conditions, \( {\widehat{\omega}}_{kj,h} \) is also asymptotically normally distributed, \( \sqrt{T}\left({\widehat{\omega}}_{kj,h}-{\omega}_{kj,h}\right)\underrightarrow{d}\mathcal{N}\left(0,{\sigma}_{kj,h}^2=\frac{\partial {\omega}_{kj,h}}{\partial {\alpha}^{\prime }}{\Sigma}_{\widehat{\alpha}}\frac{\partial {\omega}_{kj,h}}{\partial \alpha}\right), \) provided the variance of the asymptotic distribution is non-zero. Here ∂*ω*_{kj, h}/∂*a* denotes the vector of first-order partial derivatives of *ω*_{kj, h} with respect to the elements of *α* (see Lütkepohl 1990, for the specific form of the partial derivatives). Unfortunately, \( {\sigma}_{kj,h}^2 \) is zero even for cases of particular interest, for example, if *ω*_{kj, h} = 0 and, hence, the *j*th innovation does not contribute to the *h*-step forecast error variance of variable *k* (see Lütkepohl 2005, Sect. 3.7.1, for a more detailed discussion). The problem can also not easily be solved by using bootstrap techniques (cf. Benkwitz et al. 2000). Thus, standard statistical techniques such as setting up confidence intervals are problematic for the forecast error variance components. They can at best give rough indications of sampling uncertainty. The estimated *ω*_{kj, h}’s are perhaps best viewed as descriptive statistics.

## See Also

### Bibliography

- Anderson, T. 2003.
*An introduction to multivariate statistical analysis*. 3rd ed. New York: John Wiley.Google Scholar - Benkwitz, A., H. Lütkepohl, and M. Neumann. 2000. Problems related to bootstrapping impulse responses of autoregressive processes.
*Econometric Reviews*19: 69–103.CrossRefGoogle Scholar - Blanchard, O., and D. Quah. 1989. The dynamic effects of aggregate demand and supply disturbances.
*American Economic Review*79: 655–673.Google Scholar - Forni, M., M. Hallin, M. Lippi, and L. Reichlin. 2000. The generalized dynamic factor model: Identification and estimation.
*Review of Economics and Statistics*82: 540–552.CrossRefGoogle Scholar - Lütkepohl, H. 1990. Asymptotic distributions of impulse response functions and forecast error variance decompositions of vector autoregressive models.
*Review of Economics and Statistics*72: 116–125.CrossRefGoogle Scholar - Lütkepohl, H. 2005.
*New introduction to multiple time series analysis*. Berlin: Springer-Verlag.CrossRefGoogle Scholar - Sims, C. 1980. Macroeconomics and reality.
*Econometrica*48: 1–48.CrossRefGoogle Scholar - Stock, J., and M. Watson. 2002. Forecasting using principal components from a large number of predictors.
*Journal of the American Statistical Association*97: 1167–1179.CrossRefGoogle Scholar