0% found this document useful (0 votes)
108 views25 pages

Endogeneity: Yusep Suparman

The document discusses various causes of endogeneity in regression analysis, including lagged dependent variables, measurement error, and simultaneity. It also introduces instrumental variable estimation as a method to obtain consistent coefficient estimates when endogeneity is present. Specifically, it describes how instrumental variables can be used to satisfy additional moment conditions needed to identify the coefficients when regressors are correlated with the error term.

Uploaded by

agustinus
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
108 views25 pages

Endogeneity: Yusep Suparman

The document discusses various causes of endogeneity in regression analysis, including lagged dependent variables, measurement error, and simultaneity. It also introduces instrumental variable estimation as a method to obtain consistent coefficient estimates when endogeneity is present. Specifically, it describes how instrumental variables can be used to satisfy additional moment conditions needed to identify the coefficients when regressors are correlated with the error term.

Uploaded by

agustinus
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 25

ENDOGENEITY

Yusep Suparman
ENDOGENEITY
When the error term correlates to one or
more regressors and the OLS estimates can
be used.
THE CAUSES OF ENDOGENNEITY:
LAGGED DEPENDENT VARAIBEL
The model
yt   0  1 xt   2 yt 1   t

Assuming
 t   t 1  t

We have
yt   0  1 xt   2 yt 1   t 1  t

and
yt 1   0  1 xt 1   2 yt  2   t 1
THE CAUSES OF ENDOGENNEITY:
LAGGED DEPENDENT VARAIBEL
Since yt-1 is a function of εt-1 , both are correlated.
Hence as ρ≠0, yt-1 and εt in yt’s regression are correlated too.

Accordingly
E  yt xt , yt 1    0  1 xt   2 yt 1  E   t xt , yt 1 
In which
E   t xt , yt 1   0
OLS does not correspond to the conditional expectation.
OLS is biased and inconsistent.
THE CAUSES OF ENDOGENNEITY:
LAGGED DEPENDENT VARAIBEL
A lagged dependent variable combined with autocorrelation of
the error terms results in endogeneity.

Note
In the present of a lagged dependent variable as a regressor,
Durbin-Watson test for autocorrelation is not valid. One may
use Breusch-Godfrey Lagrange Multiplier test for
autocorrelation.
THE CAUSES OF ENDOGENNEITY:
MEASURMENT ERROR
The Model
yt   0  1wt  t , t ~ iid  0,  2  and E t wt   0
Assume that wt cannot be measured accurately
xt  wt  ut

Hence we have
yt   0  1 xt   t , with  t  t  1ut

Both xt and εt depend on ut, hence they are correlated, the OLS
estimator is inconsistent.
THE CAUSES OF ENDOGENNEITY:
MEASURMENT ERROR

ˆ   2

plim 1  1 1  2 u
2 
  w   u 

The estimator is down-ward biased (attenuation)

Note
Inconsistency of one regression coefficient estimator is usually
carrid over to all other regression coefficient estimators
THE CAUSES OF ENDOGENNEITY:
SIMULTANEITY
The Keynesian model
Ct   0  1Yt   t
Yt  Ct  I t

• Both Ct and Yt are jointly (simultaneously) determined in the


model. They are called endogenous variables.
• The model is call simultaneous equations or a structural
model.
• Since Yt is endogenous hence its covariance with the error
term is no longer zero. OLS estimator will be biased and
inconsistent
THE CAUSES OF ENDOGENNEITY:
SIMULTANEITY

Reduce form: expressing the regression in terms of exogenous


variables.
0 1 1
Yt   It  t
1  1 1  1 1  1
0  1
Ct   It  t
1  1 1  1 1  1
THE CAUSES OF ENDOGENNEITY:
SIMULTANEITY
From the first equation we obtain

1 1 2
cov Yt ,  t   cov  I t ,  t   var   t  
1  1 1  1 1  1

plim ˆ1  1 
cov  Yt ,  t   1  1  1 
 2

var Yt  var  I t    2

assuming
cov  I t ,  t   0
INSTRUMENTAL VARIABLE
ESTIMATOR

Consider
yi  x1i β1  x2i  2   i
It requires
E   i x1i   0 and E   i x2i   0
to associate OLS estimator of the model to conditional
expectation or best linear approximation of yi given xs
INSTRUMENTAL VARIABLE
ESTIMATOR

In the case of x2 associate to components of ε, OLS consistently


estimates conditional expectation of y, but does not consistently
estimate the causal effect of x2 .

“when we interpret the model as a conditional expectation, the


ceteris paribus condition only refers to the included variables,
while for a causal interpretation it is also includes the
unobserved (omitted variables) in the error term.”
INSTRUMENTAL VARIABLE
ESTIMATOR
If
E   i x2i   0

x2 is endogenous, OLS produces biased and inconsistent


estimator.
The OLS’s moment condition

E  y  x ˆ  x ˆ  x   0
i 1i 1 2i 2 2i

Is not longer valid and dropped. The coeficients are no longer


identified
INSTRUMENTAL VARIABLE
ESTIMATOR

At least one additional moment condition is required.


On the availability of an instrument or instrumental variable
(z2) which correlates to x2 but doe not correlate to the error
term, we have an additional moment condition

E  y  x ˆ  x ˆ  z   0
i 1i 1 2i 2 2i
INSTRUMENTAL VARIABLE
ESTIMATOR
The instrumental varaibles estimator can be solved from

  y  x βˆ 
N
1
i 1i 1,iv  x2i ˆ2,iv x1i  0
N i 1

  y  x βˆ 
N
1
i 1i 1,iv  x2i ˆ2,iv z2i  0
N i 1
We obtain
1 N
 N
  x1i   x1i 
βˆ iv    z i xi  z y i i
xi    , z i   
 i 1  i 1  x2i   z2i 
INSTRUMENTAL VARIABLE
ESTIMATOR

  
N βiv  β ~ N 0,   Σ xz Σ zz Σ zx 
ˆ 2 1 1

In practice

   XZ  ZZ  
1 1
cov βˆ iv  ˆ 2
ZX

ˆ 
2 1
N  p 1


y - Xβiv y - Xβˆ iv
ˆ  
GENERAL INSTRUMENTAL VARIABLE
ESTIMATOR

Consider
yi  xi β   i

Where x has K elements. Is there are R instruments in z, the


relevant moment conditions are given by R restrictions

E  i zi   E  yi  xi β  z i   0
INSTRUMENTAL VARIABLE
ESTIMATOR

When R=K, estimator can be obtained from sample moment


conditions
  y  xβˆ  z
N
1
i i iv i 0
N i 1

and
1 N
 N

βˆ iv    z i xi  z y i i
 i 1  i 1
INSTRUMENTAL VARIABLE
ESTIMATOR

When R>K, there are more restrictions than the unknowns. The
estimators are derived by choosing β such that the sample
moment condition

  y  xβˆ  z
N
1
i i iv i 0
N i 1

Here, we minimizing

1 N
 1 N

QN  β   
N

i 1
 yi  xi β  z i  WN

N


i 1
 i i  i
y  xβ z

INSTRUMENTAL VARIABLE
ESTIMATOR

In matrix formulation
1   1  
QN  β    Z  y - Xβ   WN  N Z  y - Xβ  
N 
Differentiating with respect to β gives first order conditions
2XZW Zy  2XZW ZXβˆ  0
N N iv

and
βˆ iv   XZWN ZX  XZWN Zy
1
INSTRUMENTAL VARIABLE
ESTIMATOR

Under the assumptions homoscedastic and non-autocorrelated


error term, the asymptotic covariance matrix of the sample
moment is given by
N
1
 Σ zz   plim
2 2

N
 z z
i 1
i i

Consecquently, the optimal weighting matrix is


1 1
1 N
 1  
WNopt 
N

i 1

zi zi    Z Z 
 N 
INSTRUMENTAL VARIABLE
ESTIMATOR

The resulting IV estimator becomes

 
1
βˆ iv  XZ  ZZ  ZX XZ  ZZ  Zy
1 1

The asymptotic distribution is

  
N βiv  β ~ N 0,   Σ xz Σ zz Σ zx 
ˆ 2 1 1

In practice,

     
1 N
ˆ βˆ iv  ˆ 2 XZ  ZZ  ZX
1 1 2
var ˆ 
2
yi  xi βˆ iv
N i 1
TESTING FOR INSTRUMENTS
VALIDITY

Testing null hypothesis of moment conditions are satisfied by


the model
E  i zi   E  yi  xi β  z i   0
Statistical test

 
 2
1
  N 
 
N N
 2  NQN βˆ iv    ei z i   ˆ  z i zi    ei z i  ~  R2  K
 i 1   i 1   i 1 
GIVE as
TWO-STAGE LEAST SQAURE EST.

Regressors are substituted by their prediction in OLS estimator


ˆ  Z  ZZ 1 ZX
X

 
1
βˆ iv  X
ˆ X
ˆ ˆ y
X
Package:
ivmodel
card.data

You might also like