The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
http://clevelandfed.org/research/workpaper/index.cfm Best available copy Working Paper 8303 EXTENSION OF GRANGER CAUSALITY IN MULTIVARIATE TIME SERIES MODELS by Michael L. Bagshaw Working papers of the Federal Reserve Bank of Cleveland are preliminary materials, circulated to stimulate discussion and critical comment. The views stated herein are the author's and not necessarily those of the Federal Reserve Bank of Cleveland or of the Board of Governors of the Federal Reserve System. August 1983 Federal Reserve Bank of Cleveland http://clevelandfed.org/research/workpaper/index.cfm Best available copy EXTENSION OF GRANGER CAUSALITY IN MULTI.VARIATE TIME SERIES MODELS Abstract This paper presents an examp,le of a situation where Granger causality does not exist but an extended definition of causality does. The extended definition of causality is discussed, along with methods to determine its existence in multivariate time series models. Key words: Granger causality, multivariate time series. I. Introduction The concept of Granger causality (Granger 1969) has become widely used in discussing relationships among variables. Some relevant references to Granger causal ity are Sims (1972), Haugh (1972), Pierce (1977), and Pierce and Haugh (1977). Generally, Granger causal ity has been discussed in terms of bivariate models. This paper proposes an extension of Granger causality when more than two variables are used in a multivariate time series model and it is necessary to consider more than one-period-ahead forecasts. Granger causality more appropriately may be called one-period-ahead forecasting ability. Variable z is said to "Granger cause" variable y, with respect to a given information set that includes z and y, if http://clevelandfed.org/research/workpaper/index.cfm Best available copy - 2 - forecasts of present y are more accurate when using past values of z than when not doing so, all other available information being used in either case. The measure of accliracy usually used in the definition o f Granger causality is the mean square error of one-period-ahead forecasts. This idea can be expressed as follows: Let tion set (including at least zt and yt), it At be the given informa- =(AS:s < t), u~(~~B) be the minimum mean square error from forecasting y (one period ahead) given the information set El, and At-zt be the set it without z. Then z is said to Granger cause y if Thus, Granger causality refers to only one-period-ahead forecasts. When forecasting for more than one period ahead, it is necessary to know whether Granger causal i ty would include a1 1 possible causal ity. situations (in terms of forecasting ability). In section I1 is an example of a multivariate model for three variables (x, y, and z). This model demonstrates that while z may not Granger cause y, the two-period-ahead forecasts using z have a smaller forecast error than the forecasts not using z. Thus, Granger causality does not encompass all situations where one would conclude that some type of causality exists (in terms of forecasting ability). In section 11, we also introduce an extension of Granger causality that includes the multiperiod, multivariate situation. 11. Extension of Granger Causality When dealing with only two variables, Pierce (1975) proves that if better L-period-ahead forecasts for any L > 2 are produced by the addition http://clevelandfed.org/research/workpaper/index.cfm Best available copy o f past z, then z must Granger cause y w i t h respect t o t h e s e t (zt, yt). However, as t h e f o l l o w i n g example demonstrates, t h i s i s not necessarily t r u e f o r systems w i t h more than two variables: where t h e a,l a2, and a3 are mutually orthogonal white noise processes 2 w i t h variances u,l 2 2 u2, u3, respectively. In t h i s model, At = (yt, xt, zt). The minimum mean-square-error, one-period-ahead forecasts f o r t h i s model are where Gt - (1) i s the one-period-ahead f o r e c a s t o f w a t time t-1. 2 2 2 forecasts have mean square e r r o r s of u,l These u2, and u3, respectively. (See Tiao and Box (1981) f o r a general discussion o f how t o c a l c u l a t e forecasts from these types o f models and t h e i r mean square errors. ) I f z i s n o t included i n the model, then t h e appropriate b i v a r i a t e model can be derived from t h e model given i n equation (1) by matching t h e variances, covariances, and cross covariances of y and x ' i m p l i e d by t h i s model. T h i s reduced model i s given by where o 2 I 2 = 4:3< + 2 u2. http://clevelandfed.org/research/workpaper/index.cfm Best available copy The minimum mean-square-error, one-period-ahead forecasts from t h i s model are w i t h mean square e r r o r s of U : and 6E30: + u ,; respectively. We thus have shown t h a t z does n o t Granger cause y ( t h e mean-squareforecast e r r o r i s ('2 < '2303 + a ) . U : i n both cases), b u t t h a t z does Granger cause x S i m i l a r l y , we can show t h a t x Granger causes y but n o t z. Also, y does n o t Granger cause x o r z. The c a u s a l i t y chain i s thus an example of i n d i r e c t c a u s a l i t y (Tjostheim 1981) between where - z and y: means Granger causes. When we examine t h e two-period-ahead forecasts, t h e r e s u l t i s different. The two-period-ahead f o r e c a s t s from model 1 are w i t h mean square e r r o r s of 0 : +.4:2~g, U: + 4:3~$, 2 and u3, r e s p e c t i v e l y . (See Tiao and Box (1981) f o r a general discussion of how t o c a l c u l a t e mu1t iple- period f o r e c a s t s and t h e i r mean square e r r o r s . ) For model 3, t h e two-period-ahead f o r e c a s t s are given by http://clevelandfed.org/research/workpaper/index.cfm Best available copy -5- 2 + with mean square forecast errors of :a + dl2(a; 2 respectively. '23 a3) and 0; + ,:a3$ Thus, the mean square forecast error for y two periods ahead is less 2 2 than it is in the model not including in the moiel including z ( a + d12a2) 2 2 2 2 2 2 This illustrates the principle that z (al + d12a2 + $12423a3). even if a variable z does not Granger cause another variable y, z may be useful in forecasting y more than one period ahead. This motivates the following extension of Granger causality. A variable z is said to cause another variable y, with respect to the set (6) U ~ ( ~ ( ILA) ) 4 U ~ ( ~ ( L IA )-) for any L > 0, where y ( ~ )is the L-period-ahead forecast of y at time t-1. We call this type of causality L-period causality, where L is the smallest value so that inequality 6 is true. Thus, Granger causality is one-period causal i ty. The concept of L-period.causality is not the same as the idea that z is related to y with an L-period lag. Consider the model In this model, y is related to z with .a j-period lag, for some j > 1. but z Granger causes y; the value of j is immaterial. This also illustrates the idea that Granger causality does not necessarily involve only one-period lags. . http://clevelandfed.org/research/workpaper/index.cfm Best available copy 111. Determining Multiperiod Causality Given a multivarjate time series model., we wish to determine what patterns of causality are represented by the model. One method of doing this (as demonstrated in the example in section 11) is to compare the mean square forecast errors from the reduced models resulting from deleting one and only one variable with those of the full model f.or different forecast lengths. One advantage in ,doing this is to learn how much the mean square errors change. For example, we saw in section I1 2 2 2 that the two-period-ahead mean square error for y was reduced by 412623a when z was included in the given model. . . Consider the following model for n variables: where K (n-1 x n-1), J (n-1 x l ) , M (1 x n-1), and N (1 x 1) are polynomial matrixes in B (where B is the backshift operator, i .e., Bvt = v t-1 ); and E W is the vector of n-1 variables excluding v; and TI (n-1 x 1) (1 x 1) are the corresponding error terms. If the variable v is omitted from this model, the resulting model is given by (See Quenouil le 1957, p. 43. ) The autoregressive operators wi 11 be given by the right-hand side of equation (7). The moving average operators will have to be dete,rmined by combining the two sources of error n and . Once the submodels are determined, the mean square forecast errors for the submodels for different forecast lengths can be compared with the http://clevelandfed.org/research/workpaper/index.cfm Best available copy corresponding quantities of the full model to determine L-period causality. This method also could be used to determine the effect of Granger causality on the mean square error for one-period-ahead forecasts. One disadvantage of this method is the number of submodels that must be determined. In general, if there are n variables in the model, then we must determine n-1 submodels, each of which has n-1 variables. The determination of these submodels is difficult when there are more than three variables in the model. We now hypothesize an additional method of determining L-period causality. The method is presented without proof, but it is intuitively appealing. The hypothesis is that y is L-period caused by z for some L if there exists a chain of Granger causali3es between z and y. If each Granger causality in the system involves only one lag, then we hypothesize that L is equal to the length of the ~inimum-lengthchain between z and y. However, if some of the lags involved in the Granger causality chain are longer than one, then L may be larger than this minimum, depending on the position of the longer lags. For example, consider the following model: (8) Yt = d12 't-j x t = 423 zt-k + + alt a2t zt = a3t. In section 11, we saw that if j = k = 1, then z two period causes y. In general, it can be shown that for this model z (1 independently of k. + j) period causes y Thus, the value of L depends not only on the length http://clevelandfed.org/research/workpaper/index.cfm Best available copy -8- of the Granger causal chain but also on the lags involved and their location in the chain. In terms of forecasting ability, L-period causality occurs when each variable in the chain is better forecast in the Granger sense using the previous variable. The L-period forecast is thus a better forecast because each variable in the chain is better forecast. That is, the L-period forecast of y depends on the forecast of x, which depends on the forecast of w, ... which depends on lagged z. In the example of equation (l), y is two-period caused by z, because the two-period-ahead forecast of y depends on the one-period-ahead forecast of x, which depends on lagged z. The advantage of this method over the previous method is its ease of use. Determination of Granger causal ity is fairly easy (Tjostheim 1981). Once the Granger causalities are ascertained, it is trivial to determine the chains of causalities. However, this method does not provide an indication of how much the L-period-ahead mean square error is reduced. Thus, we may have L-period causality with no practical significance. This would probably be true when L becomes large. IV. Summary Because Granger causality is determined only in terms of one-period forecasts, and because it is often necessary to forecast more than one period ahead, an extension of Granger causality is necessary in multivariate models.. We have presented an example that illustrates this idea and a proposal for an extension of Granger causality that addresses this problem. This extension involves L-period forecasting abi 1ity. -9- http://clevelandfed.org/research/workpaper/index.cfm Best available copy That is, the ability to forecast a variable for L periods ahead is improved by using another variable versus not using it in the same sense of Granger causality for one period ahead. We have provided methods for determining L-period causality when a model is known. http://clevelandfed.org/research/workpaper/index.cfm Best available copy References Granger, C.W.J. I1InvestigatingCausal Relations by Econometric Models and Cross-Spectral Methods," Econometrica, vol. 37 (1969), pp. 424-38. Haugh, L.D. "The Identification of Time Series Interrelationships with Special Reference to Dynamic Regression," Ph.D. Thesis, Department of Statistics, University of Wisconsin, Madison, 1972. Pierce, D.A. "Forecasting in Dynamic Models with Stochastic Regressors,' Journal of Econometrics, vol. 3 (1975), pp. 349-74. . "Relationships--and the Lack Thereof--Between Economic Time Series, with Special Reference to Money and Interest Rates," Journal of the American Statistical Association, vol. 72 (1977), pp. 11-22. Pierce, D.A., and L.D. Haugh. "Causality in Temporal Systems: Characterizations and a Survey," Journal of Econometrics, vol. 5 Quenouille, M.H. The Analysis of Multiple Time Series, London: E. Griffin and Co., 1957. Sims, C.A. "Money, ,Income, and Causality," American Economic Review, V O ~ .62 (1972), pp. 540-52. http://clevelandfed.org/research/workpaper/index.cfm Best available copy Tiao, G.C., and G.E.P. Box. "Modeling Multiple Time Series with Applications," Journal of the American Statistical Association, vol. 76, no. 376 (1981), pp. 802-16. Tjostheim, D. "Granger-Causal i ty in Mu1 tiple Time Series, " Journal of Econometrics, vol. 17 (1981), pp. 157-76.