granger causality

Investigating Causal Relations by Econometric Models and Cross-spectral Methods C. W. J. Granger Econometrica, Vol. 37, No. 3. (Aug., 1969), pp. 424-438. Stable URL: http://links.jstor.org/sici?sici=0012-9682%28196908%2937%3A3%3C424%3AICRBEM%3E2.0.CO%3B2-L Econometrica is currently published by The Econometric Society. Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/journals/econosoc.html. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers, and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take advantage of advances in technology. For more information regarding JSTOR, please contact support@jstor.org. http://www.jstor.org Mon Jul 9 21:45:01 2007 Econometrics, Vol. 37, No. 3 (July, 1969) INVESTIGATING CAUSAL RELATIONS BY ECONOMETRIC MODELS AND CROSS-SPECTRAL METHODS There occurs on some occasions a difficulty in deciding the direction of causality between two related variables and also whether or not feedback is occurring. Testable definitions of causality and feedback are proposed and illustrated by use of simple two-variable models. The important problem of apparent instantaneous causality is discussed and it is suggested that the problem often arises due to slowness in recording information or because a sufficiently wide class of possible causal variables has not been used. It can be shown that the cross spectrum between two variables can be decomposed into two parts, each relating to a single causal arm of a feedback situation. Measures of causal lag and causal strength can then be constructed. A generalisation of this result with the partial cross spectrum is suggested. 1. INTRODUCTION THEOBJECT of this paper is to throw light on the relationships between certain classes of econometric models involving feedback and the functions arising in spectral analysis, particularly the cross spectrum and the partial cross spectrum. Causality and feedback are here defined in an explicit and testable fashion. It is shown that in the two-variable case the feedback mechanism can be broken down into two causal relations and that the cross spectrum can be considered as the sum of two cross spectra, each closely connected with one of the causations. The next three sections of the paper briefly introduce those aspects of spectral methods, model building, and causality which are required later. Section 5 presents the results for. the two-variable case and Section 6 generalises these results for three variables. 2. SPECTRAL METHODS If X , is a stationary time series with mean zero, there are two basic spectral representations associated with the series : (i) the Cramer representation, where z,(o) is a complex random process with uncorrelated increments so that (ii) the spectral representation of the covariance sequence CAUSAL RELATIONS 425 If X , has no strictly periodic components, dFx(o) = ,fx(w) d o where fx(w) is the power spectrum of X , . The estimation and interpretation of power spectra have been discussed in [ 4 ] and [S ] .The basic idea underlying the two spectral representations is that the series can be decomposed as a sum (i.e. integral) of uncorrelated components, each associated with a particular frequency. It follows that the variance of the series is equal to the sum of the variances of the components. The power spectrum records the variances of the components as a function of their frequencies and indicates the relative importance of the components in terms of their contribution to the overall variance. If X , and Y , are a pair of stationary time series, so that Y, has the spectrum fy(o) and Cramer representation then the cross spectrum (strictly power cross spectrum) Cr(w) between X , and Y, is a complex function of o and arises both from E[dzx(w) dz,(o)l = 0, w f A , = Cr(o) do, o = A , and eiTWCr(w) ,LL:~ = E [ X t x - , ] = d o .i_". It follows that the relationship between two series can be expressed only in terms of the relationships between corresponding frequency components. Two further functions are defined from the cross spectrum as being more useful for interpreting relationships between variables : (i) the coherence, which is essentially the square of the correlation coefficient between corresponding frequency components of X , and Y , , and (ii) the phase, ,imaginary part of Cr(w) 4(0)= tan- real part of Cr(o) ' which measures the phase difference between corresponding frequency components. When one variable is leading the other, $(o)/o measure the extent of the time lag. Thus, the coherence is used to measure the degree to which two series are related and the phase may be interpreted in terms of time lags. Estimation and interpretation of the coherence and phase function are discussed in [4, Chapters 5 and 61. It is worth noting that 4(o) has been found to be robust under changes in the stationarity assumption [4, Chapter 91. 426 C. W. J . GRANGER If X,, E; , and 2,are three time series, the problem of possibly misleading cor- relation and coherence values between two of them due to the influence on both of the third variable can be overcome by the use of partial cross-spectral methods. The spectral, cross-spectral matrix (fij(w)) = S(w) between the three variables is given by where Lj(w) = fx(w) when i = j = x, = CrXY(w) when i = x, j = y , etc. The partial spectral, cross-spectral matrix between X, and Y, given Z, is found by partitioning S(w) into components : The partitioning lines are between the second and third rows, and second and third columns. The partial spectral matrix is then Interpretation of the components of this matrix is similar to that involving partial correlation coefficients. Thus, the partial cross spectrum can be used to find the relationship between two series once the effect of a third series has been taken into account. The partial coherence and phase are defined directly from the partial cross spectrum as before. Interpretation of all of these functions and generalisations to the n-variable case can be found in [4, Chapter 51. 3. FEEDBACK MODELS Consider initially a stationary random vector X, = {XI,,X,,, . . . ,X,,), each component of which has zero mean. A linear model for such a vector consists of a set of linear equations by which all or a subset of the components of X, are "ex- plained" in terms of present and past values of components of X,. The part not explained by the model may be taken to consist of a white-noise,random vector E,, such that where I is a unit matrix and 0 is a zero matrix. CAUSAL RELATIONS Thus the model may be written as where m may be infinite and the A's are matrices. The completely general model as defined does not have unique matrices A j as an orthogonal transformation. Y, = AX, can be performed which leaves the form of the model the same, where A is the orthogonal matrix, i.e., a square matrix having the property AA' = I. This is seen to be the case as y, = As, is still a white-noise vector. For the model to be determined, sufficient a priori knowledge is required about the values of the coefficients of at least one of the A's, in order for constraints to be set up so that such transformations are not possible. This is the so-called "identification problem" of classical econometrics. In the absence of such a priori constraints, A can always be chosen so that the A. is a triangular matrix, although not uniquely, thus giving a spurious causal-chain appearance to the model. Models for which A, has nonvanishing terms off the main diagonal will be called "models with instantaneous causality." Models for which A, has no nonzero term off the main diagonal will be called "simple causal models." These names will be explained later. Simple causal models are uniquely determined if orthogonal transforms such as A are not possible without changing the basic form of the model. It is possible for a model apparently having instantaneous causality to be trans- formed using an orthogonal A to a simple causal model. These definitions can be illustrated simply in the two variable case. Suppose the variables are X,, Y, . Then the model considered is of the form If bo = co = 0, then this will be a simple causal model. Otherwise it will be a model with instantaneous causality. Whether or not a model involving some group of economic variables can be a simple causal model depends on what one considers to be the speed with which information flows through the economy and also on the sampling period of the data used. It might be true that when quarterly data are used, for example, a simple causal model is not sufficient to explain the relationships between the variables, while for monthly data a simple causal model would be all that is required. Thus, some nonsimple causal models may be constructed not because of the basic pro- perties of the economy being studied but because of the data being used. It has been shown elsewhere [4, Chapter 7; 31 that a simple causal mechanism can appear to be a feedback mechanism if the sampling period for the data is so long that details of causality cannot be picked out. 428 C. W. J. GRANGER 4. CAUSALITY Cross-spectral methods provide a useful way of describing the relationship between two (or more) variables when one is causing the other(s). In many realistic economic situations, however, one suspects that feedback is occurring. In these situations the coherence and phase diagrams become difficult or impossible to interpret, particularly the phase diagram. The problem is how to devise definitions of causality and feedback which permit tests for their existence. Such a definition was proposed in earlier papers [4, Chapter 7; 31. In this section, some of these definitions will be discussed and extended. Although later sections of this paper will use this definition of causality they will not completely depend upon it. Previous papers concerned with causality in econonlic systems [I,6,7,8]have been particu- larly concerned with the problem of determining a causal interpretation of simul- taneous equation systems, usually with instantaneous causality. Feedback is not explicitly discussed. This earlier work has concentrated on the form that the param- eters of the equations should take in order to discern definite causal relationships. The stochastic elements and the natural time ordering of the variables play relative- ly minor roles in the theory. In the alternative theory to be discussed here, the stoch- astic nature of the variables and the direction of the flow of time will be central features. The theory is, in fact, not relevant for nonstochastic variables and will rely entirely on theassumption that the future cannot cause the past. This theory will not, of course, be contradictory to previous work but there appears to be little common ground. Its origins may be found in a suggestion by Wiener [9].The relationship be- tween thedefinition discussed hereand the work ofGood [2]has yet to bedetermined. If A, is a stationary stochastic process, let A, represent the set of past values (A,-j, j = 1,2,. . .,m) and 2, represent the set of past and present values (A,-j, j = 0, 1,. . . ,a}.Further let A(k) represent the set {A,_j, j = k, k + 1,. . . ,a). Denote the optimum, unbiased, least-squares predictor of A, using the set of values B, by P,(AI B). Thus, for instance, P,(XIX) will be the optimum predictor of X, using only past X,. The predictive error series will be denoted by s,(AIB) = A, - P,(AIB). Let a2(AIB) be the variance of &,(AIB). The initial definitions of causality, feedback, and so forth, will be very general in nature. Testable forms will be introduced later. Let U, be all the information in the universe accumulated since time t - 1 and let U, - Y, denote all this information apart from the specified series Y , . We then have the following definitions. DEFINITION1:Causality. If 02(Xl U) < 02(X/ U - Y), we say that Y is causing X, denoted by Y, -X,. We say that Y, is causing X, if we are better able to predict X, using all available information than if the information apart from I:had been used. DEFINITION2 : Feedback. If o2(x1E)< oZ(XIU- Y), a2(y1G) < 02(YIU - X), we say that feedback is occurring, which is denoted Y ,oX,, i.e., feedback is said to occur when X, is causing Y, and also Y, is causing X,. - - - - 429 CAUSAL RELATIONS DEFINITION3 : Instantaneous Causality. If 02 (X IU ,7)< a2 (X1V ) ,we say that instantaneous causality Y , => X , is occurring. In other words, the current value of X , is better "predicted" if the present value of Y,is included in the "prediction" than if it is not. DEFINITION X , , we define the (integer) causality lag m to4 :Causality Lag. If Y, * be the least value of k such that 02 (XIU - Y(k)) < 02 (XIU - Y(k + 1)). Thus, knowing the values Y,-j, j = 0, 1, . . . ,m - 1, will be of no help in improving the prediction of X , . The definitions have assumed that only stationary series are involved. In the nonstationary case, ~ ( ~ 1 8 )etc. will depend on time t and, in general, the existence of causality may alter over time. The definitions can clearly be generalised to be operative for a specified time t. One could then talk of causality existing at this moment of time. Considering nonstationary series, however, takes us further away from testable definitions and this tack will not be discussed further. The one completely unreal aspect of the above definitions is the use of the series U, , representing all available information. The large majority of the information in the universe will be quite irrelevant, i.e., will have no causal consequence. Suppose that all relevant information is numerical in nature and belongs to the vector set of time series Y y = { Y f ,i E D j for some integer set D. Denote the set {i E D , i # j ) by D( j )and {Yf, i E D( j ) )by YP(j), i.e., the full set of relevant information except one particular series. Similarly, wecould leave out more than one series with the obvious notation. The previous definitions can now be used but with U , replaced by Y,and U , - Y, by YD"'. Thus, for example, suppose that the vector set consists only of two series, X , and Y, and that all other information is irrelevant. Then 0 2 ( X I X ) represents the minimum predictive error variance of X , using only past X , and a2 (X IX .7)represents this minimum variance if both past X , and past Y,are used to predict X, . Then Y , is said to cause X , if 0 2 ( X I X )> a 2 ( X IX ,P).The definition of causality is now relative to the set D. If relevant data has not been included in this set, then spurious causality could arise. For instance, if the set D was taken to consist only of the two series X , and Y , ,but in fact there was a third series Z ,which was causing both within the enlarged set D' = ( X , ,I:,Z,),then for the original set D, spurious causality between X , and Y , may be found. This is similar to spurious correlation and partial correlation between sets of data that arise when some other statistical variable of importance has not been included. In practice it will not usually be possible to use completely optimum predictors, unless all sets of series are assumed to be normally distributed, since such optimum predictors may be nonlinear in complicated ways. It seems nitural to use only linear predictors and the above definitions may again be used under this assumption of linearity. Thus, for instance, the best linear predictor of X, using only past X i and past Y,will be of the form where the aj7s and hj's are chosen to minimise a 2 ( X IX ,Y). 430 C. W. J . GRANGER It can be argued that the variance is not the proper criterion to use to measure the closeness of a predictor Pi to the true value Xi .Certainly if some other criteria were used it may be possible to reach different conclusions about whether one series is causing another. The variance does seem to be a natural criterion to use in connection with linear predictors as it is mathematically easy to handle and simple to interpret. If one uses this criterion, a better name might be "causality in mean." The original definition of causality has now been restricted in order to reach a form which can be tested. Whenever the word causality is used in later sections it will be taken to mean "linear causality in mean with respect to a specified set D." It is possible to extend the definitions to the case where a subset of series D* of D is considered to cause X , . This would be the case if 02(XIYD )< 02(XIYD-"*)and then YD" X i . Thus, for instance, one could ask if past X , is causing present X,. Because new concepts are necessary in the consideration of such problems, they will not be discussed here in any detail. It has been pointed out already [3] that instantaneous causality, in which know- ledge of the current value of a series helps in predicting the current value of a second series, can occasionally arise spuriously in certain cases. Suppose Y ,* X , with lag one unit but that the series are sampled every two time units. Then al- though there is no real instantaneous causality, the definitions will appear to suggest that such causality is occurring. This is because certain relevant informa- tion, the missing readings in the data, have not been used. Due to this effect, one might suggest that in many economic situations an apparent instantaneous causality would disappear if the economic variables were recorded at more fre- quent time intervals. The definition of causality used above is based entirely on the predictability of some series, say X, . If some other series Y,contains information in past terms that helps in the prediction of X i and if this information is contained in no other series used in the predictor, then I:is said to cause X, . The flow of time clearly plays a cen- tral role in these definitions. In the author's opinion there is lfttle use in the practice of attempting to discuss causality without introducing time, although philosophers have tried to do so. It also follows from the definitions that a purely deterministic series, that is, a series which can be predicted exactly from its past terms such as a nonstochastic series, cannot be said to have any causal influences other than its own past. This may seem to be contrary to common sense in certain special cases but it is difficult to find a testable alternative definition which could include the deterministic situation. Thus, for instance, if X , = bt and Y , ='c(t + I), then X i can be predicted exactly by b + X i - , or by (blc)I:-,. There seems to be no way of deciding if I: is a causal factor of X , or not. In some cases the notation of the "simplest rule" might be applied. For example, if Xi is some complicated poly- nomial