Scribd
Upload
39 views

3.3 Omitted Variable Bias: U X X y

1) When a valid explanatory variable is omitted from a regression model, it results in omitted variable bias. The OLS estimates will be biased. 2) The direction and size of the bias depends on the correlation between the omitted variable and the included variables, as well as the true parameter estimate for the omitted variable. 3) Under the Gauss-Markov assumptions, including homoskedasticity, the variance of the OLS slope estimates is inversely related to R-squared from regressing the explanatory variable on the other included variables and the total sample variation of that explanatory variable.

Uploaded by

Juan Tello
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
39 views

3.3 Omitted Variable Bias: U X X y

1) When a valid explanatory variable is omitted from a regression model, it results in omitted variable bias. The OLS estimates will be biased. 2) The direction and size of the bias depends on the correlation between the omitted variable and the included variables, as well as the true parameter estimate for the omitted variable. 3) Under the Gauss-Markov assumptions, including homoskedasticity, the variance of the OLS slope estimates is inversely related to R-squared from regressing the explanatory variable on the other included variables and the total sample variation of that explanatory variable.

Uploaded by

Juan Tello
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 18

3.

3 Omitted Variable Bias


-When a valid variable is excluded, we
UNDERSPECIFY THE MODEL and OLS estimates
are biased
-Consider the true population model:


(3.40)
2 2 1 1 0
u x x y + + + = | | |
-Assume this satisfies all 4 assumptions and that
we are concerned with x
1

-if we exclude x
2
, our estimation becomes:


1 1 0
~ ~
x y | | + =
3.3 Omitted Variable Bias
-From (3.23) we know that:


(3.43)
~

~
2 1 1
o | | | + =
-where Bhats come from regressing y on ALL xs
and deltatilde comes from regressing x
2
on x
1

-since deltatilde depends on independent
variables, it is considered fixed
-we also know from Theorem 3.1 that Bhats are
unbiased estimators, therefore:


(3.45)
~
)
~
(
2 1 1
o | | | + = E
3.3 Omitted Variable Bias
-From this we can calculate Btildes bias:


(3.46)
~

)
~
( )
~
(
2 1 1 1
o | | | | = = E Bias
-this bias is often called OMITTED VARIABLE
BIAS
-From this equation, B
1
tilde is unbiased in two
cases:
1) B
2
=0; x
2
has no impact on y in the true model
2) deltatilde=0


3.3 Deltatilde=0
-deltatilde is equal to the covariance of x
1
and x
2

over the variance of x
1
, all in the sample
-deltatilde is equal to zero only if x
1
and x
2
are
uncorrelated
-therefore if they are uncorrelated, B
1
hat is
unbiased
-it is also unbiased if we can show that:


) ( ) | (
2 1 2
x E x x E =
3.3 Omitted Variable Bias
-As B
1
hats bias depends on B
2
and deltatilde, the
following table summarizes the possible biases:


Corr(x
1
,x
2
)>0 Corr(x
1
,x
2
)<0
B
2
hat>0 Positive Bias Negative Bias
B
2
hat<0 Negative Bias Positive Bias
3.3 Omitted Variable Bias
-the SIZE of the bias is also important, as a small
bias may not be cause for concern
-therefore the SIZE of B
2
and deltatilde are
important
-although B
2
is unknown, theory can give us a
good idea about its sign
-likewise, the direction of correlation between
x
1
and x
2
can be guessed through theory
-a positive (negative) bias indicates that given
random sampling, on average your estimates
will be too large (small)


3.3 Example
Take the true regression:
(ie) u ove E
2 1 0
+ + + = L xperience Pasta | | |
Where pasta taste depends on experience
making pasta and love
-While we can measure years of experience, we
cant measure love, so we find that:


(ie) E 4 . 0 3 . 5

xperience sta a P + =
What is the bias?


3.3 Example
(ie) E 4 . 0 3 . 5

xperience sta a P + =
We know that the true B
2
should be positive; love
improves cooking
We can also support a positive correlation
between experience and love, if you love
someone you spend time cooking for them
Therefore B
1
hat will have a positive bias
However, since the correlation between
experience and love is small, the bias will
likewise be small


3.3 Bias Notes
-It is important to realize that the direction of
bias is ON AVERAGE
-a positive bias on average may
underestimate in a given sample
If


1 1
)
~
( | | > E
There is an UPWARD BIAS
If


1 1
)
~
( | | < E
There is a DOWNWARD BIAS
And B
1
tilde is BIASED TOWARDS ZERO if it is
closer to zero than B
1



3.3 General Omitted Bias
Deriving the direction of omitted variable bias
with more independent variables is more difficult
-Note that correlation between any explanatory
variable and the error causes ALL OLS estimates
to be biased.
-Consider the true and estimated models:
(3.50) x
~
x
~ ~
~
(3.49) x x x
2 2 1 1 0
3 3 2 2 1 1 0
| | |
| | | |
+ + =
+ + + + =
y
u y
x
3
is omitted and correlated with x
1
but not x
2
Both B
1
tilde and B
2
tilde will always be biased
unless x
1
and x
2
are uncorrelated


3.3 General Omitted Bias
Since our x values can be pairwise correlated, it
is hard to derive the bias for our OLS estimates
-If we assume that x
1
and x
2
are uncorrelated,
we can analyze B
1
tildes bias without x
2
having an
effect, similar to our 2 variable regression:

+ =
2
1
1
3
1
1
3 1 1
) (
) (
)
~
(
x x
x x x
E
i
i i
| | |
With this formula similar to (3.45), the previous
table can be used to determine bias
-Note that much uncorrelation is needed to
determine bias


3.4 The Variance of OLS Estimators
-We now know the expected value, or central
tendency, of the OLS estimators
-Next we need information on how much spread
OLS has in its sampling distribution

-To calculate variance, we impose a
HOMOSKEDASTICITY (constant error
variance) assumption in order to
1) Simplify variance formulas
2) Give OLS an important efficiency property
Assumption MLR.5
(Homoskedasticity)
The error u has the same variance
given any values of the explanatory
variables. In other words,


2
2 1
) ,..., , | ( o =
k
x x x u Var
Assumption MLR.5 Notes
-MLR. 5 assumes that the variance of the error
term, u, is the SAME for ANY combination of
explanatory variables
-If ANY explanatory variable affects the errors
variance, HETEROSKEDASTICITY is present
-The above five assumptions are called the
GAUSS-MARKOV ASSUMPTIONS
-As listed above, they apply only to cross-
sectional data with random sampling
-time series and panel data analysis require
more complicated, related assumptions


Assumption MLR.5 Notes
If we let X represent all x variables, combining
assumptions 1 through 4 give us:
k k 2 2 1 1 0
x ... x x X) | (y | | | | + + + + = E
Or as an example:


ove ime I E L T X) | (U
3 2 1 0
| | | | + + + =
MLR. 5 can be simplified to:


2
) | ( o = X y Var
Or for example:


2
) | ( o = X U Var
3.4 MLR.4 vs. MLR.5
Assumption MRL. 4 says that the expected value
of y, given X, is linear in the parameters but it
certainly depends on x
1
, x
2
,.,x
k
.

Assumption MLR. 5 says that the variance of y,
given X, does not depend on the values of
the independent variables.
(bold added)


Theorem 3.2
(Sampling Variances of the
OLS Slope Estimators)
Under assumptions MLR. 1 through MRL. 5,
conditional on the sample values of the
independent variables,


(3.51)
) 1 (
)

(
2
2
j j
j
R SST
Var

=
o
|
For j= 1, 2,,k, where R
j
2
is the R-squared from
regressing x
j
on all other independent variables
(and including an intercept) and SST is the total
sample variation in x
j
:


2
) (

=
j
ij j
x x SST
Theorem 3.2 Notes

Note that all FIVE Gauss-Markov assumptions
were needed for this theorem
Homoskedasticity (MLR. 5) wasnt needed to
prove OLS bias

The size of Var(B
j
hat) is very important
-a large variance leads to larger confidence
intervals and less accurate hypothesis tests

You might also like