Professional Documents
Culture Documents
Errors in Variables in Simultaneous Equation Models
Errors in Variables in Simultaneous Equation Models
Errors in Variables in Simultaneous Equation Models
Jerry A. HAUSMAN
Massachusetts Institute of Technology, Cambridge, MA 02139, USA
The simultaneous equation model is considered when errors in variables are present in the
exogenous variables. By means of a distributional assumption on the exogenous variables, the
system is transformed into an augmented structural model. Full information estimation by
means of maximum-likelihood and instrumental variables is proposed. Depending on the
assumptions on the errors of observation, restrictions may be imposed on the covariance
matrix of the augmented structural residuals.
1. Introduction
The errors in variables model played a large role in the early development of
econometrics. The instrumental variable method of estimation was developed to
deal explicitly with the problem of a ‘right-hand side’ variable in a regression
model being non-orthogonal to the “error” in the regression specification.
However, the outstanding problem with instrumental variables estimators has
always been to specify where the candidates to form instruments come from.
In the usual simultaneous equation specifications this problem does not exist;
the included endogenous variables in an equation are non-orthogonal to the
structural disturbance but the excluded predetermined variables in an identified
system always yield enough additional candidates to form instruments to permit
estimation. This result follows from identification of the structural model under
the usual assumptions. The single-equation errors in variables specification does
not have these convenient instruments to use. Recently Zellner (1970) and
Goldberger (1972) have attempted to surmount this problem by making the
unobserved exogenous variable a (stochastic) linear function of observable
variables. This procedure provides a partial solution to the problem although
in many cases the specification of the linear relationship may be very difficult.
In this paper I investigate the simultaneous equation estimation problem when
errors in variables exist in the exogenous variables. That is, the true structural
*This paper was presented at the Winter meetings of the Econometric Society, 1974. I would
like to thank the discussants, especially P. Dhrymes. I have also benefitted from many conversa-
tions about the errors in variables problem with P.A. Samuelson.
390 J.A. Hausman, Errors in variables
YB+ZT = U,
‘Alternative estimation schemes have been proposed recently by Joreskog (1974) and Geraci
(1974).
J. A. Hausman, Errors in variables 391
Xi containing the r,+si variables whose coefficients are not known a priori to
be zero. Therefore, yi and ui are T vectors, Xi is a TX (ri+si) matrix, and 6i
is a (ri+sJ-dimensional vector. Then stacking the M equations into a system,
y = X6+24, (2.3)
where
z1 = Z:+El, (3.1)
where the components of s1 are i.i.d. with zero mean and constant variance oE,.
Also &I is assumed independent of Z and in particular of zf. Consideration
of errors of this type in the endogenous variables is not interesting since they are
observationally equivalent to the errors in the equations. In fact, the traditional
approach is to assume an exact linear relationship without a structural distur-
bance, and the error in observing the endogenous variable provides the error
term in the usual regression format [see Kendall and Stuart (1971)].
The instrumental variable procedure is usually a single-equation approach;
consider the ith equation assumed to contain zr,
yi = X~6+Ui, (3.2)
where
Xi* = [ Y,z;“Z,],
zi being the matrix Zi with the first column deleted. Now if one replaces the
unobservable zT by the observable z”, , then eq. (3.2) becomes
Yi = YiBi+z”,yil+~i’Y”i+ui-&lYil, (3.3)
where Zi and r”i are the matrix Zi and vector yi with the first column and element
removed, respectively. If ?I is treated as an exogenous variable, inconsistent
estimates result since z”r is not independent of si. However, a consistent estima-
tion procedure proposed by Chernoff and Rubin (1953) and by Sargan (1958)
is to treat z”i as another endogenous variable. Then an instrumental variable
estimator uses instruments Wi to estimate
bi = (w,lx,)-‘w;y,. (3.4)
where Qi is a finite matrix with rank (ri+si) with probability one. Note that all
predetermined variables except z1 are candidates to form instruments; at least
rifsi must be used to insure satisfaction of the rank condition. Chernoff and
Rubin propose a LIML type procedure which is asymptotically equivalent to the
procedure of Sargan which uses all such predetermined variables except z1 to
form the instruments
w; = xi’z(zI’z)-‘z’, (3.61
where 2 denotes the 2 matrix with the first column removed. By construction
and Assumption A.3, the estimates are consistent with asymptotic covariance
matrix
In fact, Sargan shows that this choice of instruments is best asymptotically given
the choice of all candidates to form instruments from Z the matrix of predeter-
mined variables. That is, any choice of a subset of 2 to form instruments will have
a covariance matrix at least as large.
Note that the rank condition on the instruments places an additional restric-
tion on equation i. Qi must have rank ri+si which requires at least rifsi
variables to form the instruments Wi. By the order condition for identification,
K 2 r,+s,. But since z”, cannot be used to form instruments we have only
K- 1 candidates.3 Thus, if equation i is just identified, K = ri+si, then the
instrumental variable procedure cannot be used since Q, will not have full rank.
I will refer to identification when neglecting the presence of exogenous variables
measured with error as conditional identification. Then to have consistent
estimates by the instrumental variable technique, equation i must be conditionally
over-identified. Only then will there be sufficient variables to form the instruments
to insure the rank condition in Qi.
3As Sargan notes, exogenous variables from other equations may be used to form instru-
ments even if they are observed with error so long as their error of observation is uncorrelated
with Us and with E,. Of course, these variables must be correlated with the included exogenous
variables to form proper instruments.
394 J.A. Hausman, Errors in variables
form exists,
Y=ZL!+V, (4.1)
For many cases, especially cross-section data, it is reasonable to assume that all
z,, for t = 1,. . , T, are distributed identically with the same p and A. For time
series specifications, an assumption such as this for variables with a time trend
removed might be plausible. Still, the assumption of a fixed mean and covariance
matrix is quite strong beyond the usual assumptions where only finite second
moments need exist. For instance, the ‘incidental parameter problem’ of errors
in variables [Kendall and Stuart (1971)] disappears since estimates of the constant
mean are trivial to obtain. Then by a well-known theorem on multivariate
normal distributions [cf. Anderson (1958, p. 28)], the regression of the column
vector z1 on the rest of the columns of Z, say 2, gives the conditional expectation
of zl, given Z,
z1 = a,+z:a,+E2,
where the usual independence assumptions of the regression model are satisfied
and e2 N N(0, oJ). Since in our model z1 is the unobservable variable z:, and
we observe z”l by eq. (3.1), the unobservable variable .z: satisfies the stochastic
relationship
Yl = w,+zT~,,+~ly",+~l, (4.5)
Yi = yiPi+ziYi+Ui9 i = 2,. . ., M,
Zl = Z:+E1, (4.6)
2: = ao+Zcr,+EZ,
where in eq. (4.5) 2, denotes Z1 , with the first column deleted, and in eq. (4.6)
z stands for the complete predetermined variable matrix 2 with the first column
deleted. By direct substitution eq. (4.5) and eq. (4.6) are combined to form the
equation system
Yi = yiPt+ziYi+ui, i = 2, . . ., M, (4.7b)
where 2 is now the matrix 2 with a column of ones added to include the constant
if it is not already present, and a = (2) is the combined vector for the unknown
coefficients in eq. (4.6). Note that .?! contains all the predetermined variables
except z:, and 2, contains the included predetermined variables from equation 1
with z: removed. This approach of adding equations for variables measured
with error generalizes so that one additional equation follows from each such
variable measured with error. By Assumption A.1, equation 1 of the structural
%Zellner (1970) considers the case where in eq. (4.4) c2 = 0. Thus the unobservable variable
is known as an exact linear relation of observable variables. While this assumption leads to an
interesting statistical problem, the assumption seems so strong that this case will not be
considered.
396 J.A. Hausman, Errors in variables
where vM+r = s1+s2, note that by assumption the elements of vM+r are
distributed independent normally with zero mean and constant variance.
Furthermore, by Assumption A.3 and eq. (3.1), 2 and v~+~ are contempor-
aneously independent. Therefore a consistent estimator of a is the least-squares
regression
Given & one can form a consistent estimate of z”, from 22, and this conditional
predictor may be substituted into eq. (4.7a). Consistent estimation of eq. (4.7a)
for 6, then proceeds in the usual way by instrumental variables. Note that this
procedure is exactly the instrumental variable procedure outlined in section 3
since
which is the procedure used in eq. (3.6) to form the instruments Wi. Thus no
gain in efficiency is offered by the stochastic specification of the predetermined
variables in eq. (4.2) which lead to the linear structural relationship of eq. (4.4).
However, some structural information has been neglected so now consider the
system in a full-information context.
The equation system (4.7) has the usual format of a structural simultaneous
equation system. The errors are distributed normally, and denoting the errors
as vi for i = 1, . . ., M+ 1 equations,
(4.11)
J.A. Hausman, Errors in variables 391
B ‘Yll
1O
J= I* 1 ’ (4.14)
--_- 0
o***o 1 1
where B is the M-square matrix of the original endogenous variables. Since under
Assumption D. 1 C is unrestricted, it has the form
(4.15)
The first-order conditions for the likelihood functions may be found and an
algorithm specified [e.g., Hausman (1975)]. However, since C is unrestricted,
alternative asymptotically equivalent estimators are easily specified. For instance,
3SLS applied to eq. (4.12) will produce efficient estimates. Alternatively, a full-
‘This assumption relaxes the need to assume that equation 1 is conditionally overidentified
for identification since it leads to restrictions on the covariance matrix.
398 J. A. Hausman, Errors in variables
1
2
(-Jll+Yll GE2 ’ -Y11 gel
1 0
1 :
C= c I * (4.16)
I O
- - - - - -
I---T-----
I -Y11~,, I o...o 1 00 +o,,
As was previously noted, since C now has zero restrictions, the requirement
of conditional overidentification for identification is not needed. Identification
can now be analyzed using the methods of Fisher (1966, ch. 4). The first and last
elements in the last row and column of C put no restrictions on C beyond the
usual symmetry restriction since they are ‘just-identified’, but C~j forj = 2, . . .,
M- 1 are now zero. The likelihood function remains that given in eq. (4.13),
but now the simplification to 3SLS or full-information unstrumental variables
is no longer possible since Cis now restricted. Thus under Assumption D.2,3SLS
is still consistent but no longer attains the Cramer-Rao bound asymptotically
70f course, more efficient estimates ofa are obtained by estimating eq. (4.7~) jointly with the
other equations. Since this equation does not form part of the structural system, one may not
be interested in its coefficients. By the same reasoning, the second equation in Goldberger
(1972, p. 7) may be ignored with no loss in efficiency.
J.A. Hausman, Errors in variables 399
Y, = Y,B,+~~y,,+Z~P1+u,+r,,&,, (5.la)
Yz = Y,P2+~~y,,+Z,y”,+u,+YZ1&2, (5.lb)
_Yi= Y$,+Ziyi+Ui, i = 3,. . ., M, (5.lc)
- = %+a1+a2,
z1 (5.ld)
(5.3)
See Rothenberg (1973) who calculates the change in the CR bound for the special case of a
completely diagonal covariance matrix.
400 J.A. Hausman, Errors in variables
This system of equations still has the standard form of a simultaneous equation
system with the likelihood function given by eq. (4.13). The Jacobian of the
transformation has the form
J= (5.4)
I
~12+YllY21~e2
0
C= c
II .
Y210e=
. (5.5)
I :
__-____------- - 0
YllUQ Y21CE2 o--o aEI+OE2_
This covariance matrix has an even more complicated structure than the co-
variance matrix of eq. (4.16). Not only are there zeros present in the last row and
column but C,,,,, and C,,.,,, must satisfy the nonlinear restriction
Thus again 3SLS no longer attains the CR bound under Assumption A.5. The
nonlinear restriction of eq. (5.6) adds an additional restriction to the covariance
matrix.’ Thus to attain the CR bound, FIML is needed.
‘Note that if eq. (5.lb) were conditionally just identified, identification could be achieved
using the equality of eq. (5.6). For further details of identification see Geraci (1974).
J.A. Hausman, Errors in variables 401
6. Conclusion
In this paper the simultaneous equation model has been enlarged to permit
errors in the exogenous variables. By making an explicit distributional assump-
tion on the relationship of the exogenousvariables, the errors in variables problem
is transformed to a structural specification. The error-ridden variable is then
treated as endogenous, and a just-identified structural equation is added to the
model. In the case of an unrestricted covariance matrix, the 3SLS and FIML
estimators retain their usual properties.
The other case considered is when the error covariance matrix is not un-
restricted. Then only FIML attains the CR bound. Thus the other main con-
clusion of this paper is that the addition of errors in variables to the simultaneous
equation model may lead to restrictions on the structural covariance matrix.
A topic for further research is the practical importance of imposing the restric-
tions. Recent work by Joreskog and Goldberger (1975) and Rothenberg (1973)
does not indicate large gains accrue to imposing zero restrictions on the co-
variance matrix.
References
Anderson, T.W., 1958, An introduction to multivariate statistical analysis (New York).
Berndt, E., B. Hall, R. Hall and J. Hausman, 1974, Estimations and inference in nonlinear
structural models, Annals of Economic and Social Measurement 4.
Chernoff, H. and H. Rubin, 1953, Asymptotic properties of limited-information estimates
under generalized conditions, in: Hood and Koopmans, eds., Studies in econometric
method (New Haven).
Fisher, F., 1966, The identification problem in econometrics (New York).
Geraci, V.J., 1974, Simultaneous equation models with measurement error, University of
Wisconsin SSRI Paper 7407, mimeo.
Goldberger, A.S., 1972, Maximum-likelihood estimation of regressions containing unob-
servable independent variables, International Economic Review 13.
Hausman, J.A., 1975, An instrumental variable approach to full-information estimators for
linear and certain non-linear econometric models, Econometrica 43.
Joreskog, K.G., 1974, A general method for estimating a linear structural equation system,
in: O.D. Duncan and AS. Goldberger, eds., Structural equation models in the social
sciences (New York).
Joreskog, K.G. and A.S. Goldberger, 1975, Estimation of a model with multiple indicators and
multiple causes of a single latent variable, Journal of the American Statistical Society 70.
Kendall, M. and A. Stuart, 1971, Advanced theory of statistics, vol. II (London).
Narayanan, R., 1969, Computation of Zellner-Theil’s three stage least squares estimates,
Econometrica 37.
Rothenburg, T.J., 1973, Efficient estimation with a priori information (New Haven).
Sargan, J.D., 1958, The estimation of economic relationships using instrumental variables,
Econometrica 26.
Zellner, A., 1970, Estimation of regression relationships containing unobservable variables,
International Economic Review 11.