1
Note: IV estimation
The Inconsistency of the LS Estimator
When Cov(x,) is nonzero, the LS estimate of
2 would be inconsistent and the bias is
approximately given in a simple linear model
Cov(x, ) / Var(x)=Corr(x, ) / x
The Inconsistency of the LS Estimator
y 1 2 x e . y E y 2 x E x e
( E y 1 2 E x )
x E x y E y 2 x E x x E x e
2
E x E x y E y 2 E x E x E x E x e
2
cov x, y 2 var x cov x, e 2
b2
cov x, y cov x, e
var x
var x
x, y)
x x y y x x y y / T 1 cov(
x)
var(
x x
x x / T 1
t
If cov x, e 0 then b2
x, y) cov( x, y)
cov(
2
x)
var(
var( x)
If cov x, e 0 , then 2
cov x, y cov x, e
cov x, y
cov x, e
b2
2
2
var x
var x
var x
var x
The Problem of Weak Instrument
The Problem of Weak Instrument
Cov(z,)=0 is never guaranteed. If Cov(z, ) is
nonzero, then the IV estimate of 2 would be
inconsistent and the bias is approximately
Cov(z, ) / Cov(z,x)= Corr(z, ) /[Corr(z,x)x]
If z is a weak instrument (i.e. Corr(z,x) is small), the
2SLS estimate of 2 could be even more biased than the
OLS estimate.
Variance of IV Estimates
Using IV results in less precise estimate than
using OLS (i.e. it is harder to reject null
hypothesis, or less likely to draw conclusion
from data).
Var(2iv)/Var(2ols)=1/r2x,z>1
How much is the return to education?
Return to education
What is the value of Schooling?
Do people really learn useful things in schools?
Do people learn more in college than in high
school?
Should government spend more money on schools?
What level of schools should get more funding?
How hard is it to measure the return to education?
A nave regression
Incomei=Educationi+i
Problems
Measurement Errors (Income, Education)?
Underspecification?
Functional form misspecification?
Endogeneity?
Key Issue
Endogeneity is the key issue
Due to the correlation between education and
omitted variables that affect income (talents,
family environment, )
Due to measurement errors in education
The problem may not be solved by adding more
control variables or using panel data.
The Impact of Years of Education on
Earnings (Angrist and Krueger)
A naive model
Ln(Earnings)= 0+1(years of edu)+u
IV: Quarter of birth
Children born in the fourth quarter (e.g. December 31) enter school at
age 5.75. In contrast, those born in the first quarter (e.g. January 1)
enter school at age 6.75.
Compulsory schooling laws requires students to remain in school until
their 16th birthdays.
Hence, people born in different quarters have different education length
on average. Moreover, the difference is possibly exogenously
determined.
10
11
12
13
Findings
Men born in the first quarter have about onetenth of a year less schooling than men born in
later quarters.
Men born in the first quarter earn about 0.1
percent less than men born in later quarters
14
Findings
15
Men born in the first quarter have about onetenth of a year less schooling than men born in
later quarters.
Men born in the first quarter earn about 0.1
percent less than men born in later quarters
Other examples
Acemoglu et al (2001)
A nave model :
Ln(GDP per capita) = a + b(institution) + u
Instrument : European mortality rates
Frankel and Romer (1999)
A nave model
Ln(GDP per capital) = a + b(Trade Volume) + u
Instrument : Geographic characteristics
16
Natural (Quasi-) Experiment
17
In an ideal experiment we have a treatment group and
a control group. The two groups are the same except
that only the treatment group is affected by some
factor of interest while the control group is not
affected by the factor (both groups can still be
affected by other confounding factors).
A natural experiment is like an experiment. The
difference is that in the latter case we have good
control on which group can be affected by which
factors. In a natural experiment, we can not control
the factors, but we try to find some natural settings in
which the nature has generated two similar groups,
one of which is affected by the factors of interest and
the other group is not.
Natural Experiment
NE required the control and treatment groups
to be matched not only by observed factors,
but also unobserved factors.
Whether unobserved factors are matched is
inferred from theory and other observed
information.
18
Solution with Identical twins
19
Arguments for using twins data
Identical twins should be equally smart; they
should look similar; and they should have the same
family background.
Therefore, if we could find twins with different
education, we can compare their income. The
difference should reflect only the effect of
education but not of omitted variables.
Strictly speaking, we assume there is no
unobserved inter-twin difference that is
correlated with education.
An Empirical Study(Bonjour et al., 2003)
Data: 3,300 same-sex twins in 1999 in U.K.
Detailed medical information
Socioeconomic information
Method
ln(wif)=Sif+aif+if
ln(w1f)-ln(w2f)=WTP(S1f-S2f)+(a1f -a2f)+(1f-2f)
aif indicates ability net of family and genetic
effects
20
21
