Professional Documents
Culture Documents
Abstract
Regression coefficients are interpreted by a counterfactual experiment. For
simultaneous equations this experiment can be implemented if the coefficients
are identified, and throws some light on the role of instruments and the method
of indirect least squares. This paper discusses another counterfactual
experiment in the vector autoregressive model in order to interpret the
coefficients of an identified cointegrating relation. The dynamics of the model
is used to implement a long-run change by changing the current values. The
counterfactual experiment can be conducted precisely when the cointegrating
relation is identified.
I. Introduction
The purpose of this note is to discuss the interpretation of cointegrating
coefficients as elasticities. Regression coefficients are usually interpreted by a
counterfactual experiment or thought experiment, which does not make
immediate sense in the cointegrated vector autoregressive (VAR) model,
because we do not have the dichotomy of endogenous and exogenous
variables.
*I would like to thank Giuseppe Bertola, Martin Browning, Tony Hall, Henrik Hansen, Jim
Heckman, David Hendry, Andrea Ichino, Alison Major, Adrian Pagan, and Tom Rothenberg for
useful comments. The support of the Danish Social Sciences Research Council and the ESF-EMM
network is gratefully acknowledged.
JEL Classification numbers: C32, C30.
93
Blackwell Publishing Ltd, 2005. Published by Blackwell Publishing Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK
and 350 Main Street, Malden, MA 02148, USA.
94 Bulletin
Example 1
Consider first the univariate regression model
yt ¼ c1 x1t þ c2 x2t þ et
where we assume that the errors et are independent of the regressors xt. The
usual interpretation of the regression coefficient c1, say, is that c1 is the causal
effect of x1 on y, keeping x2 fixed. Thus, in order to interpret the coefficient c1
we conduct the thought experiment of changing the system by changing x1,
without changing x2, and observing the change in y. It is implicit in the
thought experiment that there is no other relation between y and x, and that we
can disregard relations (or correlation) between the xs and finally manipulate x
without changing e (independence between x and e).
Note that because
Eðyt jxt Þ ¼ c1 x1t þ c2 x2t
it follows that
@
Eðyt jxt Þ ¼ c1
@x1t
which is a convenient formulation of the role of c1 as a partial derivative
and the thought experiment above describes the definition of a partial
derivative.
Example 2
As another example, we consider a system of simultaneous equations,
constructed to illustrate various situations
y1t /2 y2t ¼ c11 x1t þ e1t
y2t /3 y3t ¼ c22 x2t þ e2t ð1Þ
y2t /4 y3t ¼ c31 x1t þ c32 x2t þ e3t :
y1h y1 /2 ðy2h y2 Þ ¼ 0
y2h y2 /3 ðy3h y3 Þ ¼ c22 h
y2h y2 /4 ðy3h y3 Þ ¼ c32 h:
The solution is
y1h y1 ¼ /2 ðy2h y2 Þ
hð/4 c22 /3 c32 Þ
y2h y2 ¼
ð/4 /3 Þ
hðc22 c32 Þ
y3h y3 ¼ :
ð/4 /3 Þ
Hence, by choosing an increment of h ¼ (/4 ) /3)/(/4c22 ) /3c32) to x2 we find
that the effect on y2 is y2h y2 ¼ 1 and that on y1 is y1h y1 ¼ /2 , as required.
Thus we can conduct the thought experiment of using the instrument x2 to
produce a change of 1 in y2, disregarding the change in y3, thereby producing
the required change of /2 in y1. Note the role of the exogenous variable x2 as
instrument to produce the required effect, and in a sense implement the
2 ‘design of experiment’ of Haavelmo (1944).
The idea of influencing y1 via y2, using x2 as an instrument, and first
discussing the effect of x2 on y2 (regressing y2 on x2), and then using the
effect h(/4c22 ) /3c32)/(/4 ) /3) (fitted value ^y2 ) to measure the effect on y1
(regression of y1 on ^y2 Þ is clearly the idea behind the two-stage least squares
(2SLS) by Theil (1953a,b, 1971), although in this particular just-identified
case, it is just indirect least squares, which was used by Tinbergen (1930).
Note also that the condition for identification of the first equation is that
/4c22 ) c32/3 6¼ 0. Thus, in the construction of the thought experiment, we
apply the assumption that the three equations determine the endogenous
variables as functions of the exogenous (/3 6¼ /4), and the assumption that
the first equation is identified (/4c22 ) c32/3 6¼ 0).
Another way of expressing the interpretation of /2, is to note that from
equation (1) it follows that
Eðy1t jx1t ; x2t Þ ¼ /2 Eðy2t jx1t ; x2t Þ þ c11 x1t :
Here E(y1t|x1t, x2t) is considered a function of (x1t, x2t) through E(y2t|x1t, x2t)
and x1t, so that
@Eðy1t jx1t ; x2t Þ
/2 ¼ :
@Eðy2t jx1t ; x2t Þ
The counterfactual experiment then shows how one could, in principle,
implement the change in E(y2t|x1t, x2t) keeping x1t fixed, using x2t as an
instrument. Note that if x2t had a non-zero coefficient in the equation for y1t,
the expectation would be a function of x1t and x2t as well as of E(y2t|x1t, x2t),
and the counterfactual experiment could not be performed, as one cannot keep
x1t and x2t fixed and change E(y2t|x1t, x2t). In this case, however, the parameter
/2 would not be uniquely identified in the system, although it would still have
a meaning according to the original ‘design of experiment’.
where we have left out deterministic terms to simplify the notation. Here, a
and b are n · r matrices of full rank. We assume that the process is I(1) so that
Xk 1
0
a In C b 6¼ 0
? i ?
i¼1
we define
C ¼ b? ða0? Cb? Þ1 a0?
and find from the solution to the equilibrium correction model in the
companion form [see equation (8) in the Appendix] the long-run value x1|t as
a function of current values (xt, . . . , xt)k+1) as given by
!
X
k1
x1jt ¼ lim Eðxtþh jxt ; . . . ; xtkþ1 Þ ¼ C xt Ci xti : ð3Þ
h!1
i¼1
If, however, k 2 sp(b’) is the desired effect in the long run, then we write
k ¼ b’w, for some w 2 Rn)r, and find that a simultaneous change to all
current values by k will lead to the long-run change
which again is approximately Nn(0, tR) with the singular long-run covariance
matrix R ¼ CXC¢. Because b¢C ¼ 0, we find from
0 0
b1 R11 R12 0
¼
b2 R21 R22 0
that
b01 R12 þ b02 R22 ¼ 0:
If x2t does not cointegrate, then R22 has full rank and we find
c0 ¼ b10 0 1
1 b2 ¼ R12 R22
long-run values (see Proposition 1). We apply this to consider a different type
of thought experiment, and use current values to implement a change in long-
run values. This thought experiment gives the possibility of implementing
changes in the long run needed for interpretation of the coefficients instead of,
as before, using a change in exogenous variables to produce a change in
endogenous variables. We first consider two examples and then give a general
result.
Example 3
Suppose the dimension n ¼ 3, and
b0 xt ¼ b1 x1t þ b2 x2t þ b3 x3t ¼ ut
is the only cointegrating relation in model (2), so that ut is stationary, whereas,
in general, the components xit are non-stationary. If b1 6¼ 0, we identify the
coefficients by solving for x1t, and find
x1t ¼ c2 x2t þ c3 x3t þ vt ð4Þ
with ci ¼ )bi/b1. We would like to interpret the coefficient c2 as the long-run
effect of x2 on x1, keeping x3 fixed; i.e. we would like to change x2 by 1, and x3
by 0, and observe a long-run change (c2) in x1. Because the relation is a long-
run relation, i.e. can be formulated as b¢x1|t ¼ 0, it is natural to think of these
changes as those in the long-run values and the idea is to accomplish these
changes by a suitable change in the current values, as formulated in
Proposition 1.
We therefore introduce a change to xt so that in the long run x2 is shifted by
1 and x3 is kept constant, while x1 is shifted by c2; i.e. we want to change
(x1, x2, x3)1|t to (x1, x2, x3)1|t + (c2, 1, 0). Now notice that the vector k ¼
(c2, 1, 0)¢ is orthogonal to b ¼ (1, )c2, )c3)¢:
k 0 b ¼ ðc2 ; 1; 0Þb ¼ c2 c2 ¼ 0
which means that the desired long-run change is a vector in the space spanned
by the columns of C, i.e. sp(b’). In general, we can ask the question if it is
possible to find a vector h and shift xt to xt + h so that the long-run value is
shifted from x1|t to x1|t + k, for a given vector k in sp(b’). This is exactly the
content of Proposition 1, where the relation between initial change and long-
run change is given, and it is seen that we can choose h ¼ Ck, and shift xt to
xt + Ck.
We use this result to interpret the coefficient c2 in the identified
cointegrating relation (4) as the effect of x2 on x1, in the sense that a
change of xt by C(c2, 1, 0)¢ will lead to a long-run change of the system
where x11|t is changed to x11|t + c2, x21|t to x21|t + 1, and x31|t is kept
Example 4
Consider next a cointegrated system with four variables and two cointegrating
relations, (b1, b2), which we identify by solving them for x1 and x2
x1t ¼ c13 x3t þ c14 x4t þ u1t
x2t ¼ c23 x3t þ c24 x4t þ u2t :
We want to implement a long-run change in the system, so that we can
interpret c13 as the long-run effect of x3 on x1 keeping x4 fixed. Because x2
does not enter the first relation, we consider a long-run change of the form
k 0 ðkÞ ¼ ðc13 ; k; 1; 0Þ
for some k, where the value of k is at our disposal, and we choose it so that
k(k) is orthogonal to b1 and b2. Note that
k 0 ðkÞb1 ¼ ðc13 ; k; 1; 0Þð1; 0; c13 ; c14 Þ0 ¼ 0
k 0 ðkÞb2 ¼ ðc13 ; k; 1; 0Þð0; 1; c23 ; c24 Þ0 ¼ k c23 :
Thus, by the choice k ¼ c23 we make k(c23) orthogonal to both cointegrat-
ing vectors. Hence we can achieve the long-run change k ¼ k(c23) by
3 giving a change to xt of the form Ck, which by Proposition 1 will give a
long-run change of CCk ¼ k. Note that the long-run change also changes x2,
and it is the change of x2 that achieves that k(k) is orthogonal to both b
vectors. Thus, the role of the excluded variable x2 is that of an
instrument needed to produce the required effect on other variables
namely that the point stays on the attractor set, where the equations are
satisfied.
We finally show in Proposition 2 the general result that the identifying
restrictions define the instruments, that allow the interpretation of an identified
coefficient in a cointegrating relation. We construct a thought experiment that
achieves the required counterfactual change in the long-run value by suitable
changes in the current values.
We consider the model (2) for the cointegrated I(1) process Xt with r
cointegrating vectors. Let b1 be one of them of the form
b01 ¼ ð1; b21 ; . . . ; bs1 ; 0; . . . ; 0Þ
that is normalized on b11 ¼ 1, and restricted by R¢b1 ¼ 0, where
R0 ¼ ð0ðnsÞs ; Ins Þ:
b~1 ¼ ðb2 ; . . . ; br Þv
and
b1 ¼ ð1 b~11 Þb1 þ b~1 :
IV. Conclusion
We have discussed interpretation of regression coefficients in simultaneous
equations and in coefficients in cointegrating relations. Whereas interpretation
of individual (autonomous) equations or cointegrating relations is in terms of a
counterfactual experiment that allows one to vary some variables and
keep others fixed, we discuss here how to implement such a change in terms
of another counterfactual experiment, when the equations form part of a
system of simultaneous equations. This experiment consists, for regression
coefficients, of manipulating the exogenous variables to achieve the relevant
effect on the endogenous variables. For cointegrating coefficients we
manipulate the current values in order to implement the required effect on
the long-run values. This last counterfactual experiment can be conducted
precisely when the coefficient is identified.
References
Angrist, J. D., Imbens, G. W. and Rubin, D. B. (1996). ‘Identification of causal effects using
instrumental variables’, Journal of American Statistical Association, Vol. 91, pp. 444–472.
Engle, R. F. and Granger, C. W. J. (1987). ‘Co-integration and error correction: representation,
estimation and testing’, Econometrica, Vol. 55, pp. 251–276.
Haavelmo, T. (1944). ‘The probability approach in econometrics’, Econometrica, Vol. 12, Issue
Supplement, pp. 1–115.
with errors
~et ¼ ðe0t ; 0; . . . ; 0Þ0
and parameters
0 1 0 1
a C1 Ck1 b In 0
B0 In 0 C B 0 In 0 C
B C B C
a ¼ B ..
~ .. .. C; b~ ¼ B .. .. .. C ð7Þ
@. . . A @. . . A
0 0 In 0 0 In
so that
0
ab~ ~xt1 þ ~et
D~xt ¼ ~
with solution
0 X
h1
0
ab~ Þh~xt þ
~xtþh ¼ ðIkn þ ~ ðIkn þ ~ab~ Þi~etþhi : ð8Þ
i¼0
Hence
0
Eð~xtþh j~xt Þ ¼ ðIkn þ ~ab~ Þh~xt : ð9Þ
The I(1) condition (Johansen, 1996, Theorem 4.2) is equivalent to
0
jeigðIrþðk1Þn þ b~ ~aÞj < 1
0 h
which implies that ðIrþðk1Þn þ b~ ~
aÞ ! 0. Applying this we find, using the
well-known relations
0 0 0
ab~ Þh ~
ðIkn þ ~ ab~ Þh1 ðIkn þ ~ab~ Þ~a
a ¼ ðIkn þ ~
0 0
ab~ Þh1 ~aðIrþðk1Þn þ b~ ~aÞ
¼ ðIkn þ ~
0
aðIrþðk1Þn þ b~ ~aÞh
¼ ¼ ~
and
0 0 0
ab~ Þh b~? ¼ ðIkn þ ~
ðIkn þ ~ ab~ Þh1 ðIkn þ ~ab~ Þb~?
0
ab~ Þh1 b~? ¼ ¼ b~?
¼ ðIkn þ ~
that
0 0 0 0
ab~ Þh~xt ¼ ðIkn þ ~
ðIkn þ ~ a0? b~? Þ1 ~a0? þ ~aðb~ ~aÞ1 b~ ~xt
ab~ Þh ½b~? ð~
0 0 0
a0? b~? Þ1 ~
¼ b~? ð~ a0?~xt þ ~aðIrþðk1Þn þ b~ ~aÞh ðb~ ~aÞ1 b~ ~xt
! b~ ð~a0 b ~ Þ1 ~a0 ~xt ¼ C~~ xt ¼ ~x1jt :
? ? ? ?
The limit ~x1jt is called the long-run value of ~xtþh starting at ~xt .
It is seen that
0
a0? ¼ a0? ðIn ; C1 ; . . . ; Ck1 Þ;
~ b~? ¼ b0? ðIn ; . . . ; In Þ
which shows that
1
C~ ¼ b~? ~
a0? b~? a0? ¼ ðIn ; . . . ; In Þ0 CðIn ; C1 ; . . . ; Ck1 Þ :
~ ð10Þ