Econometrics:

Specification Errors

Burcu Eke

UC3M
Introduction

I So far, we have focused on the models that satisfy the

assumptions of linear regression model (LRM) and
therefore have nice properties and interpretations.

I The model specification of the LRM includes choosing

Choosing the independent variables, and thus the omitted
variables
Choosing the functional form

I Some very important questions are:

What would happen if we used the LRM when the
assumptions are not met, i.e., when it is not appropriate?
What are the properties of the OLS estimators under a
specification error?

2
Introduction

I Specification errors in which we are interested:

1. Inclusion of irrelevant variables.
2. Omission of relevant variables.
3. Measurement errors in variables.

3
Omitting a Relevant Variable

I Recall the full and reduced models we have talked about

when we first introduced multiple linear regression.

I Now, lets revisit this issue from another perspective:

Let’s consider the multiple linear regression model:
Y = β0 + β1 X1 + β2 X2 + ε
For some reason such as unavailability of the data on X2 ,
we construct a regression model without the variable X2
variable
Y = γ0 + γ1 X1 + ε1

4
Omitting a Relevant Variable

I In econometrics, this issue is known as “omitting a relevant

variable”, if β2 6= 0, and this is a type of misspecification
The big question then is what is the effect of omitting a
relevant variable?
The answer is given in the equation below.

C(X1 , X2 )
γ1 = β1 + β2 (1)
V (X1 )

The equation (1) is called the rule of omitted variable,

which shows that the slope of the reduced model is a linear
combination of β1 and β2 (the two slopes of the full model)

5
Omitting a Relevant Variable

I Then, by omitting X2 , its effect becomes part of the error

term in the reduced model:
ε1 = ε + β2 X2

I This implies:

E[ε1 |X1 ] = E [ε + β2 X2 |X1 ]

= E [ε|X1 ] + E [β2 X2 |X1 ]
= E [E [ε|X1 , X2 ] |X1 ] + β2 E [X2 |X1 ]
= β2 E [X2 |X1 ]

I Hence,

E[Y |X1 ] = γ0 + γ1 X1 + β2 E [X2 |X1 ]

6
Omitting a Relevant Variable

I In general, whenever we have E[ε|X1 , . . . , Xk ] 6= 0 for

multiple linear model
Y = β0 + β1 X1 + . . . + βk Xk + ε
It means that the model is misspecified

7
Omitting a Relevant Variable: The Properties
of OLS Estimators

I Now, let’s focus on the properties of the OLS estimators

under this specification error.

I Consider the simplest multiple linear regression, and a

simple regression using only one of the variables
Let the MLR model be Y = β0 + β1 X1 + β2 X2 + ε, with the
estimated model Ŷ = β̂0 + β̂1 X1 + β̂2 X2
Let the SLR model be Y = γ0 + γ1 X1 + ε, with the
estimated model Ŷ = γ̂0 + γ̂1 X1

8
Omitting a Relevant Variable: The Properties
of OLS Estimators

I Then we know that γ̂1 = β̂1 + β̂2 δ̂1 , where δ̂1 is the OLS
estimator of the slope of L (X2 |X1 ) = δ0 + δ1 X1

I We also know that

E[β̂1 ] = β1 and p lim β̂1 = β1
E[γ̂1 ] = γ1 and p lim γ̂1 = γ1

9
Omitting a Relevant Variable: The Properties
of OLS Estimators

I Using all these information, we have

h i h i
E [γ̂1 |X1 , X2 ] = E β̂1 |X1 , X2 +E β̂2 δ̂1 |X1 , X2 = β1 +β2 δ̂1

I This implies

E[γ̂1 ] = β1 + β2 E[δ̂1 ]
p lim(γ̂1 ) = β1 + β2 δ1

I In general:
γ̂1 won’t be appropriate if we want to make inference about
β1  
Furthermore, it is easy to show that V (γ̂1 ) ≤ V β̂1

10
Omitting a Relevant Variable: The Properties
of OLS Estimators
I Whenever X2 is a “relevant variable” (that is,
β2 6= 0),γ̂1 ⇒ inconsistent and biased estimator of β1 ⇒
bias will not disappear no matter how big the sample size is

σ2   σ2
I V (γ̂1 ) = P and V β̂1 = P 2 , that is
i x1i i x1i (1 − R1 )
V (γ̂1 ) ≤ V β̂1 , but V̂ (γ̂1 ) is a biased estimator for the
variance of β̂1

I That is, the “omission of relevant variables” in the analysis

generates inconsistency and bias in estimating the effects of
variables, though a reduction in the variance of the
estimator.

I How about hypothesis test?

11 Are they valid?
Omitting a Relevant Variable: The Properties
of OLS Estimators

I In other words, the coefficient of X1 in the regression that

incorrectly omits X2 :
does not capture the ceteris paribus effect of X1 on Y (sice
when X1 changes, so does X2 )
captures the effect on Y coming from a change in X1 plus
the effect of X1 on X2 (which DOES have an effect on Y )

I We can summarize the bias in estimating β1 when X2 is

incorrectly omitted as:
C(X1 , X2 ) > 0 C(X1 , X2 ) < 0
β2 > 0 + −
β2 < 0 − +

12
Including an Irrelevant Variable

I X2 is an “irrelevant” variable, that is, β2 = 0

I Now, then, the true model is Y = γ0 + γ1 X1 + ε1 and we

estimate
Y = β0 + β1 X1 + β2 X2 + ε, thinking that X2 is relevant

I In this case, ε = ε1 − β2 X2 = ε1 since β2 = 0 in the true

model

I This implies:

E[ε|X1 , X2 ] = E [ε1 − β2 X2 |X1 , X2 ]

= E [ε1 |X1 ]

13
Including an Irrelevant Variable: The
Properties of OLS Estimators

I Then, β̂1 is a consistent and unbiased estimator of γ1

and it has less variance than V (β̂1 ) ≥ V (γ̂1 ), which is also
unbiased and consistent.

I That is, the “inclusion of irrelevant variables” in the

analysis, does not affect the consistency of the estimated
effect of the variables.

I Intuition: The true population value of the coefficient of an

irrelevant variable is 0, so by including this variable, the
coefficient estimators for the other variables are not
affected in the limit.

14
Including an Irrelevant Variable: The
Properties of OLS Estimators

I However, estimated β’s will be generally inefficient, that

is, their variances will be greater than those of the true
model
Intuition: The higher the correlation between the irrelevant
and relevant variables, the greater the variance of the
estimated coefficient for the relevant variables.
This means that when irrelevant variables are included, we
do not have the problem of inconsistency of the OLS
estimator (and hence including an irrelevant variable is a
less serious problem)
Nevertheless,the inefficiency problem can generate a serious
consequences when testing hypotheses of type H0 : βj = 0
due to the lost of power, so we might infer that they are not
relevant variables when they truly are (type II error)

15
Including an Irrelevant Variable: The
Properties of OLS Estimators

I In practice, Is it possible to know which model is the

appropriate one?
Strictly: No.
In general the best approach is to include only the variables
that, based on economic theory, affects the dependent
variable, and are not accounted for other variables in the
model
Then we can gather evidence for or against the ”relevance”
or ”irrelevance” of one or more variables through the testing
of hypotheses.

16
Example 1: The impact of tobacco consumption
on cancer

I Assume that we information on existence of cancer in two

groups of people: smokers and nonsmokers

I Additionally, let’s assume that smokers are more likely to

engage in more physical activities which reduces the
likelihood of cancer. But we don’t observe these activities

I Therefore the impact of smoking on cancer will be

overestimated because the tobacco consumption may
decrease the level of physical activity.

17
Example 1: The impact of tobacco consumption
on cancer

I Formally, Ci = β0 + β1 Fi + β2E Ji + ε, where,

Ci as the measure of cancer for individual i
Fi is a dummy variable that takes a value of 1 is the
individual i is a smoker and 0 otherwise
EJi is a measure of physical activity, i.e., exercise
Let the true values be β1 > 0, β2 < 0

18
Example 1: The impact of tobacco consumption
on cancer

I Additionally, Ci = δ0 + δ1 Fi + ε1 , with δ1 < 0

I Therefore, when we run the simple regression of Ci on Fi ,

we will have
Ci = γ0 + γ1 Fi + ε2 ,
Then we have γ̂1 = β̂1 + β̂2 δ̂1

19
Example 2: The impact of tobacco on wages

I Additionally to the impact on health outcomes, does

smoking have economic consequences?

I Smokers might face lower wages than non-smokers:

If they were less productive due to “cigarette breaks”;
If smoking has an impact on health outcomes, smokers
would be more likely ask for sick-leaves;
If the firm would discriminate against smokers; and so on.

20
Example 2: The impact of tobacco on wages

I We have representative data for individuals 30 years old for

the US. Levine, Gustafson and Velenchik (1997) estimated
a wage equation using the following variables:
Y = ln(wage)
F = a dummy variable that takes a value of 1 for smokers
and 0, otherwise.
ED = Years of education

I We must take into consideration that smokers have, on

average, less education than non-smokers (thus, education
is negatively correlated with smoking)

21
Example 2: The impact of tobacco on wages

I Two specifications considered are:

Omitting education: Ŷi = −0.176Fi with sβ̂1 = 0.021
Including education: Ŷi = −0.080Fi + 0.070EDi with
sβ̂1 = 0.021 and sβ̂2 = 0.004

I By not including education in the regression we

overestimated the impact of smoking.

22
Measurement Error

I So far, we have assumed that Y and Xj ’s are accurate, i.e.,

there are no measurement errors

I Sometimes we have data that has measurement errors, or

that there are no data available on the variable of interest.

I For example, according to the life cycle models, the

consumption depends on the permanent income which we
cannot measure without error, or we have data on the
reported annual income, but not the real annual income.

23
Measurement Error

I Measurement error occurs when we cannot accurately

measure the magnitude of the variable of interest. We are
interested in the effects of such errors on our OLS
estimators.

I There are two types of measurement error:

1. Measurement error in the dependent variable, Y
2. Measurement error in the independent variable, X.

24
Measurement Error in Y

I Consider the following model:

Y ∗ = β0 + β1 X + ε (2)

I E[ε|X] = 0 ⇒ E[Y ∗ |X] = L(Y ∗ |X) = β0 + β1 X

I E(ε) = 0, C(X, ε) = 0 and β0 = E[Y ∗ ] − β1 E[X],

C(X, Y ∗ )
β1 =
V (X)

25
Measurement Error in Y

I Assume that the Model (2), Y ∗ has measurement error, so

that Y = Y ∗ + v0 , where v0 is the measurement error in Y ∗

I Then this model becomes Y = β0 + β1 X + ε + v0 :

Y = β0 + β1 X + (ε + v0 ) (3)
| {z }
u

I u is the composite error term

26
Measurement Error in Y

1 P
n i xi yi
I Recall that β̂1 = 1 P 2 , yi = Yi − Ȳ , and xi = Xi − X̄
n i xi

1X
p lim x i yi
n→∞ n
i C(X, Y )
p lim β̂1 = =
n→∞ 1 X V (X)
p lim x2i
n→∞ n
i
C(X, Y ∗ + v0 ) C(X, Y ∗ ) + C(X, v0 )
= =
V (X) V (X)
C(X, v0 )
= β1 + ⇒ consistent if C(X, v0 ) = 0
V (X)

27
Measurement Error in Y

I For β̂0
 
p lim β̂0 = p lim Ȳ − β̂1 X̄
n→∞ n→∞
∗
= E[Y + v0 ] − E[X]p lim β̂1
n→∞

I Therefore, β̂0 and β̂1 are consistent if:

C(X, v0 ) = 0 and
E[v0 ] = 0.
I The variances are now larger under measurement error
V (Y ∗ |X) = V (ε|X) = σ 2
V (Y |X) = V (Y ∗ + v0 |X) = σ 2 + σv20 , assuming that
C (ε, v0 |X) = 0

28
Measurement Error in X

I Consider the following model:

Y = β0 + β1 X ∗ + ε (4)

I E[ε|X ∗ ] = 0 ⇒ E[Y |X ∗ ] = L(Y |X ∗ ) = β0 + β1 X ∗

I E(ε) = 0, C(X ∗ , ε) = 0 and β0 = E[Y ] − β1 E[X ∗ ],

C(X ∗ , Y )
β1 =
V (X ∗ )

29
Measurement Error in X

I Assume that the Model (4), X ∗ has measurement error, so

that X = X ∗ + v1 , where v1 is the measurement error in X ∗

I Then this model becomes Y = β0 + β1 (X − v1 ) + ε:

Y = β0 + β1 X + (ε − β1 v1 ) (5)
| {z }
u1

I E[v1 ] = 0, and C(X, ε) = 0: ε is correlated neither with X

nor with X ∗ , therefore nor with v1 .

I E[Y |X ∗ , X] = E[Y |X ∗ ] ⇒ Given X ∗ , X does not contain

any additional relevant information.

30
Measurement Error in X

I Assume the classical ME assumptions hold, i.e.,

C(X ∗ , v1 ) = 0 and C(v1 , ε) = 0

1X
p lim xi yi
n→∞ n
i C(X, Y )
p lim β̂1 = =
n→∞ 1X 2 V (X)
p lim xi
n→∞ n
i
0
z }| {
C(X ∗ − v1 , Y ) C(X ∗ , Y ) + C(Y, v0 )
= =
V (X ∗ − v1 ) V (X ∗ ) + V (v1 )
C(X ∗ , Y )/V (X ∗ )
=
(V (X ∗ ) + V (v1 )) /V (X ∗ )
β1
= 6= β1
1 + V (v1 )/V (X ∗ )
31
Measurement Error in X

I The asymptotic bias then:

  β1
p lim β̂1 = β1 = − β1
n→∞ 1 + V (v1 )/V (X ∗ )
V (v1 )
= −β1
V (v1 ) + V (X ∗ )

I What happens to the bias if V (X ∗ ) is relatively much

larger than V (v1 )? What happens otherwise?

32
Measurement Error in X

I In a multiple regression model, the overall measurement

error in an explanatory variable produces inconsistency of
all the estimators, i.e., all β̂j ’s are inconsistent. In a
multiple regression model with ME in only one of the Xj ’s
and this error is not correlated with either wrongly
measured, say Xm or with the remaining Xj ’s (m 6= j):
There is a bias in β̂m (underestimate in absolute value), and
it is inconsistent.
For the remaining β̂j ’s (m 6= j) they are generally
inconsistent, although it is not easy to know the directions
and magnitudes of the biases and inconsistency.
Only in the unlikely event that Xm is orthogonal to the
remaining Xj ’s (m 6= j), β̂j ’s are consistent.

33
Example 3: The impact of family income on
college performance

I We want to see if the family income has an effect on the

grade point average in the college. Since it is not clear that
family income has a direct effect on academic performance,
the recommended strategy would be to include this variable
as a regressor and test whether its coefficient is significant.

I CAL = β0 + β1 I ∗ + β2 PRE + β3 SEL + ε, where

CAL = Average grade in college,
I ∗ = Family income,
PRE = Average grade prior college entrance,
SEL = Average grade on the admission exam.

34
Example 3: The impact of family income on
college performance

I Suppose the data are obtained by surveying students.

There may be errors in the declared family incomes, so
I = I ∗ + v1 .
I Even if we assume that the measurement error, v1 , is
uncorrelated neither with I ∗ nor with the rest of the
explanatory variables (PRE, SEL), the estimators obtained
by using I instead of the true value I ∗ will be inconsistent.
I Specifically, we will underestimate |β1 |. Therefore, when
testing the significance of β1 , we will be more likely to do
not reject (DNR) H0
I In this example, it is difficult to determine the magnitude
and direction of the bias and the inconsistency for the
estimators of β2 and β3 .
35