Chapter 4- Simultaneous-equation Models

The Simultaneous Dependence of Economic Variables

The application of least squares to a single equation assumes, among others, that
the explanatory variables are truly exogenous, that there is one-way causation
between the dependent variable Y and the explanatory X ’ s. If this is not true, that
is if the X ’s are at the time determined byY . Assumption 6 of OLS is violated [
E( Xu)≠ 0], and the application of this method yields biased and inconsistent
estimates.
If y=f ( X) , but also X =f (Y ) we are not allowed to use a single equation model for
the description of the relationship between Y and X. We must use a multi-equation
model, which would include separate equations in which X and Y would appear as
endogenous variables, although they might appear as explanatory in other
equations of the model. A system describing the joint dependence of variables is
called a system of simultaneous equations.
The bias arising from the application of classical least squares to an equation belong
to a system of simultaneous relations is called simultaneous equations bias. It
originates from the violation of Assumption 6 of OLS, that is it arises from the
dependence of the explanatory variables andu, [ E( Xu)≠ 0].
This creates several problems. Firstly, there arises the problem of identification of
the parameters of individual relationships. Secondly, there arise problems of
estimation. The application of OLS yields biased and inconsistent estimates. One
should therefore choose other estimation methods.

Consequences of Simultaneous Relations

We said that when there is a joint dependence between Y and X, their relationship
cannot be described with a single equation, but with a system of simultaneous
equations. In each relation there are explanatory variables which are endogenous to
the system, that is, they appear as dependent in other equations of the system.
Thus for any particular equation the random variable is not independent of the
explanatory variable(s). Assumption 6 of OLS is not fulfilled [ E( Xu) ≠ 0] and as a
consequence the estimates are both biased and inconsistent.
Assume we have the simple models
Y =b0 +b 1 X +u E ( u )=0 E ( v )=0
X =a0 + a1 Y + a2 Z +v E ( u2 ) =σ 2u E ( v 2) =σ 2v
E ( ui u j ) =0 E ( v i v j )=0 E ( uv )=0
This model is mathematically complete: it contain two equations with two
endogenous variables, X and Y. Z is assumed to be exogenously determined.
Substituting X in the second equation we obtain
a0 +b 0 a 1 a2 a1 u+ v
X =a0 + a1 ( b0 +b 1 X +u ) +a2 Z + v ¿ X= + Z +( )
1−b1 a1 1−b1 a1 1−b1 a1
X and the disturbance term u are related. X is not a truly exogenous variable in the
first equation.
It can be proved that the covariance of X and u is not zero. cov ( Xu)≠ 0.
Proof: By definition the covariance of X and u is

cov ( uX )=E [ { u−E ( u ) }{ X−E ( X ) } ] but E ( u )=0.

Therefore

cov ( Xu ) =E [ u { X−E ( X ) } ]
a0 +b 0 a 1 a2 a1 u+ v
Given that X= + Z +( ) and Z is exogenously determined we
1−b1 a1 1−b1 a1 1−b1 a1
have
a0 +b 0 a 1 a2
E( X )= + Z
1−b1 a1 1−b 1 a 1
Therefore

cov ( uX )=E
[ u
1−b1 a1 0 1 0 2 ] [ u
{ a +a b + a Z +a1 u+ v−( a 0+ a1 b0 +a 2 Z ) } =E 1−b a {a1 u +v }
1 1
]
Since E ( uv )=0 , therefore

1 a1
E ( a 1 u +uv ) = E (u ) ≠ 0
2 2
cov ( uX )=
1−b1 a1 1−b1 a1
As a consequence, if we apply the method of OLS to the first function the estimates
of the coefficients will be biased and inconsistent. The bias is independent of the
sample size. Hence the bias cannot be eliminated by increasing the number of
observations in the sample.

Solution to the Simultaneous-Equation Bias

Since the application of OLS to an equation belonging to a system of simultaneous
equations yields biased and inconsistent estimates, the obvious solution is to apply
other methods of estimation which give better estimates of the parameters. There
are several methods for this purpose. The most common are:
1. The reduced form method or indirect least squares (ILS).
2. The method of instrumental variables (IV).
 3. Two stage least squares (2SLS).
4. Limited information maximum likelihood (LIML).
5. The mixed estimation method.
6. Three-stage least squares (3SLS).
7. Full information maximum likelihood (FIML).
The first five methods are called single-equation methods, because they are applied
to one equation of the system at a time. The three-stage least squares and the full
information maximum likelihood are called systems methods, because they are
applied to all the equations of the system simultaneously. The choice among the
alternative techniques for the estimation of the parameters of a particular model is
a difficult task.

Some Definitions
1. Structural Models
A structural model is a complete system of equations which describe the structure
of the relationships of the economic variables. Structural equations express the
endogenous variables as functions of other endogenous variables, predetermined
variables and disturbances (random variables).
As an illustration we will use the following simple model for a closed economy.
C t=a0 +a 1 Y t +u 1
I t=b0 +b1 Y t + b2 Y t−1 +u2
Y t =C t + I t +Gt
The first equation is a consumption function, the second is an investment function,
the third is a definitional equation. The system is complete in that it contains three
equations in three endogenous variables Y t , C t∧I t . The model contains two
predetermined variables, government expenditure, G, and lagged income Y t −1.
The structural parameters are, in general, propensities, elasticities, or other
parameters of economic theory. A structural parameter expresses the direct effect
of each explanatory variable on the dependent variable. Indirect effects can be
computed only by the solution of the structural system, but not by the individual
structural parameters. Factors not appearing in any function explicitly may have an
indirect influence on the dependent variable of that function. For example, a change
in consumption will affect investment indirectly, through the increase that the
consumption, C, will produce on income, Y, which is a determinant of investment.
The effect of C on I cannot be measured directly by any of the structural
parameters, but it will be taken into account by the simultaneous solution of the
system.
Traditionally the structural parameters are represented by β ’s when they refer to
endogenous variables, and byγ ’s when they are attached to a predetermined
variable. Similarly, endogenous variables are denoted by y ’s while exogenous
variables are represented by x ’s. Using the conventional notations, the structural
system above becomes
y 1=β 13 y 3 +u 1
y 2=β 23 Y 3+ γ 21 x1 +u2
y 3= y1 + y 2 + x 2
Where y 1=C y 2=I y 3=Y
x 1=Y t −1 x 2=G
To obtain the complete table of structural parameters is as follows
y 1 +0 y 2−β 13 y 3 +0 x 1 +0 x 2=u1
0 y 1 + y 2−β 23 Y 3 −γ 21 x 1+ 0 x 2=u 2
− y 1− y 2 + y 3 +0 x 1−x 2=0

Table of structural Table of structural coefficients in

coefficients standard notation
1 0 −a 1 0 0 1 0 −β 13 0 0
0 1 −b 1−b 2 0 0 1 −β 23−γ 21 0
−1 −1 1 0 −1 −1 −1 1 0 −1

The values of the structural parameters may be obtained by using sample

observations on variables of the model and applying an appropriate econometric
method.
2. Reduced Form Models
The reduced form of a structural model is the model in which the endogenous
variables are expressed as a function of the predetermined variables only. The
reduced form is obtained in two ways. The first is to expressed the endogenous
variables directly as functions of the predetermined variables
y i=π i 1 x 1 + π i 2 x 2 +…+ π ik x k + v i
And proceeds with the estimation of the π ’s by applying some appropriate technique
to this expression. In our example of the simple three equation model the reduced
form would be
C t=π 11 Y t −1+ π 12 Gt + v1
I t=π 21 Y t−1 + π 22 Gt + v 2
Y t =π 31 Y t −1+ π 32 Gt +v 3
The second method for obtaining the reduced form of a model is to solve the
structural system of endogenous variables in terms of the predetermined variables,
the structural parameters and the disturbances. The structural system of our
example gives the following reduced form model:
 a1 b2 a1 u1 +a 1 u2−b1 u 1
C t= Y t−1 + Gt +
1−a1−b 1 1−a1 −b1 1−a1−b 1
b 2 (1−a1) b1 u2 +b 1 u1−a1 u 2
I t= Y t−1 + Gt +
1−a1−b 1 1−a1 −b1 1−a1−b 1
b2 1 u1 +u2
Y t= Y t −1+ Gt +
1−a1 −b1 1−a1−b 1 1−a 1−b1
Clearly for the reduced forms to be consistent the following relationships between
the π ’ s and the structural parameters must hold
a1 b2 a1
π 11= π 12=
1−a1−b 1 1−a1−b1
b2 (1−a1 ) b1
π 21= π 22=
1−a1 −b1 1−a1 −b1
b2 1
π 31= π 32=
1−a1−b1 1−a1−b1
It should be clear that there is a definite relationship between the reduced form
parameters and the structural parameters: the π ’ s are functions of the structural
parameters.
Derivation of the reduced form parameters

(a) Substitute C t and I t into the third structural equation

Y t =( a 1 Y t + u1 ) +(b 1 Y t +b2 Y t−1 +u2 )+ Gt
By rearranging we obtain
b2 1 u1 +u2
Y t= Y t −1+ Gt +
1−a1 −b1 1−a1−b 1 1−a 1−b1
This is the reduced form of the third structural equation

(b) Substitute Y t into the consumption function

C t=a1
[ b2
1−a1−b 1
Y t −1 +
1
1−a 1−b1
Gt +
u 1+u 2
1−a1−b1
+u 1
]
¿
a1 b2 a1 u1 +a 1 u2−b1 u 1
C t= Y t−1 + Gt +
1−a1−b 1 1−a1 −b1 1−a1−b 1
This is the reduced form of the consumption function.

(c) Substitute Y t into the investment function

 I t=b1
[ b2
1−a1−b 1
Y t −1 +
1
1−a 1−b1
Gt +
u 1+u 2
1−a1−b1 ]
+ b2 Y t−1 +u2

¿
b 2 (1−a1) b1 u +b u −a u
I t= Y t−1 + Gt + 2 1 1 1 2
1−a1−b 1 1−a1 −b1 1−a1−b 1
This is the reduced form of the investment function.
The reduced form parameters measure the total effect, direct and indirect of a
change in the predetermined variable on the endogenous variables, after taking
into account of the interdependences among the jointly dependent endogenous
variables, while a structural parameter indicates only the direct effect within a
single sector of the economy.
The reduced form coefficients are used for forecasting and policy analysis, since it is
the total effect of a change in the exogenous variables on the dependent variable(s)
that is of interest to the policy maker.
The above two ways of defining the reduced-form model suggests that estimates of
the reduced form coefficients may be obtained in two ways.
Firstly, direct estimation of the reduced-form coefficients. The reduced form
π ' s may be estimated by the method of least-squares-no-restrictions (LSNR).
We express all the endogenous variables as functions of all the predetermined
variables of the system and we apply OLS to these reduced form functions. This
method of obtaining the π ' s is called least squares- no-restrictions (LSNR), because
it does not take into account any information on the structural parameters, that is, it
does not use any restrictions imposed by the form of the structural system.
This method does not require complete knowledge of the structural system. What is
required is knowledge of the predetermined variables appearing in the whole
system.
Secondly, indirect estimation of the reduced-form coefficients: We saw that
there is a definite relationship between the reduced-form coefficients and the
structural parameters. It is thus possible first to obtain estimates of the structural
parameters by any appropriate econometric technique and then substitute these
estimates into the system of parameters’ relationships to obtain (indirectly) values
of the π ' s . This indirect method involves three steps:
I. Solve the system of endogenous variables so that each equation contains only
predetermined explanatory variables. This, as we saw, may be done by
continuous substitutions of variables, until we arrive at the reduced form of all
the equations. This way we obtain the system of parameters’ relations, that is,
to say the system which defines the relations between the π ' s∧the β ' s∧the γ ' s .
II. Obtain estimates of the structural parameters by any appropriate econometric
method.
III. Substitute the estimates β ' s and γ ' s into the system o parameters’ relations to
find the estimates of the reduced form coefficients.

3. Recursive Models
A model is called recursive if its structural equations can be ordered in such a way
that the first includes only predetermined variables in the right-hand side; the
second equation contains predetermined variables and the first endogenous
variable in the right-hand side; and so on. For example
y 1=f (x 1 , x 2 … x k ; u1 )
y 2=f (x 1 , x 2 … x k ; y 1 ; u 2)
y 3=f (x 1 , x 2 … x k ; y 1 , y 2 ; u3 )
¿ so on .
The random variables assumed to be independent.
The special feature of a recursive model is that its equations may be estimated, one
at a time, by OLS without simultaneous-equation bias.

Level of Aggregation – Number of Equations – Number of Variables

The level of aggregation depends on the purpose of the model. For forecasting and
policy analysis the level of aggregation must be chosen carefully. The usefulness of
a model for appraising economic policies depend primarily upon the number of
equations and variables included.
Economic reality is very complex, so that the number of mathematical relationships
which would be required for a complete description of an economic phenomenon
would be large. Systems with large numbers of equations are difficult to handle. So,
if we are to construct an operational econometric model we must simplify the
relationships of the actual world. Simplification is done in several ways:
1. Aggregation of individual consumers.
2. Aggregation of firms into an industry.
3. Aggregation over commodities.
4. Selection of the most important explanatory variables.
5. Simplification of the mathematical form of the function.
The optimal number of equations that should be included in a model should depend
on various considerations: (a) the purpose of the model, (b) data availability, (c)
importance of various sectors and/or various variables, (d) the level of detailed
information in which we are interested, and so on.
With respect to the number of variables we note the following.
 The number of endogenous variables is the same as the number of equations of
the model. However, the number of exogenous variables may be as large as we
choose.
 The more variables we introduce in a model the more complicated it becomes
and the heavier the data requirements and computational difficulties grow.
 If the model is intended for policy analysis it should include as many as possible
policy instruments (that is, variables under the direct control of the government,
like taxes of various item, subsidies of different forms, tariffs and so on) so that
it may be effectively used for the evaluation of the effects of alternative
economic policies.