The McGraw-Hill Series Economics essentials of economics brue, McConnell, and Flynn Essentials of Economics

neo-classical linear regression model (NLRM),
10
in contrast to the CLRM, where the
X
’s are treated as fixed or nonrandom. For discussion purposes, we will call the former the
stochastic regressor model
and the latter the 
fixed regressor model.
9
James Davidson, 
Econometric Theory,
Blackwell Publishers, U.K., 2000, p. 10.
10
A term due to Arthur S. Goldberger, 
A Course in Econometrics,
Harvard University Press, Cambridge,
MA, 1991, p. 264.
ASSUMPTION 3
Zero Mean Value of Disturbance
u
i
:
Given the value of 
X
i
,
the mean, or expected,
value of the random disturbance term 
u
i
is zero. Symbolically, we have
E
(
u
i
|
X
i
) 
=
0
(3.2.1)
Or, if 
X
is nonstochastic,
E
(
u
i
) 
=
0
Assumption 3 states that the mean value of
u
i
conditional upon the given
X
i
is zero.
Geometrically, this assumption can be pictured as in Figure 3.3, which shows a few values
of the variable
X
and the
Y
populations associated with each of them. As shown, each
Y
X
1
X
2
X
3
X
4
X
Y
Mean
Y
i
=
β
β
1
+
2
X
i
PRF:
+
u
i
–
u
i
FIGURE 3.3
Conditional
distribution of the
disturbances 
u
i
.
guj75772_ch03.qxd 23/08/2008 02:34 PM Page 63

---

64
Part One
Single-Equation Regression Models
population corresponding to a given
X
is distributed around its mean value (shown by the cir-
cled points on the PRF), with some
Y
values above the mean and some below it. The distances
above and below the mean values are nothing but the
u
i
. Equation 3.2.1 requires that the
average or mean value of these deviations corresponding to any given
X
should be zero.
This assumption should not be difficult to comprehend in view of the discussion in
Section 2.4 (see Eq. [2.4.5]). Assumption 3 simply says that the factors not explicitly
included in the model, and therefore subsumed in 
u
i
,
do not systematically affect the mean
value of 
Y
; in other words, the positive 
u
i
values cancel out the negative 
u
i
values so that
their average or mean effect on 
Y
is zero.
11
In passing, note that the assumption 
E
(
u
i
|
X
i
) 
=
0 implies that 
E
(
Y
i
|
X
i
) = 
β
1
+ 
β
2
X
i
.
(Why?) Therefore, the two assumption are equivalent.
It is important to point out that Assumption 3 implies that there is no 
specification bias
or 
specification error
in the model used in empirical analysis. In other words, the regres-
sion model is correctly specified. Leaving out important explanatory variables, including
unnecessary variables, or choosing the wrong functional form of the relationship between
the 
Y
and 
X
variables are some examples of specification error. We will discuss this topic in
considerable detail in Chapter 13.
Note also that if the conditional mean of one random variable given another random
variable is zero, the covariance between the two variables is zero and hence the two vari-
ables are uncorrelated. Assumption 3 therefore implies that 
X
i
and 
u
i
are uncorrelated.
12
The reason for assuming that the disturbance term 
u
and the explanatory variable(s) 
X
are uncorrelated is simple. When we expressed the PRF as in Eq. (2.4.2), we assumed that
X
and 
u
(which represent the influence of all omitted variables) have separate (and additive)
influences on 
Y
. But if 
X
and 
u
are correlated, it is not possible to assess their individual
effects on 
Y
. Thus, if 
X
and 
u
are positively correlated, 
X
increases when 
u
increases and
decreases when 
u
decreases. Similarly, if 
X
and 
u
are negatively correlated, 
X
increases
when 
u
decreases and decreases when 
u
increases. In situations like this it is quite possible
that the error term actually includes some variables that should have been included as
additional regressors in the model. This is why Assumption 3 is another way of stating that
there is no specification error in the chosen regression model.
11
For a more technical reason why Assumption 3 is necessary see E. Malinvaud, 
Statistical Methods of
Econometrics,
Rand McNally, Chicago, 1966, p. 75. See also Exercise 3.3.
12
The converse, however, is not true because correlation is a measure of linear association only. That
is, even if 
X
i
and 
u
i
are uncorrelated, the conditional mean of 
u
i
given 
X
i
may not be zero. However, if
X
i
and 
u
i
are correlated, 
E
(
u
i
|
X
i
) must be nonzero, violating Assumption 3. We owe this point to Stock
and Watson. See James H. Stock and Mark W. Watson, 
Introduction to Econometrics,
Addison-Wesley,
Boston, 2003, pp. 104–105.