You are on page 1of 17

Hausman Tests Overidentifying Restrictions Tests

Modelos Lineares
Variveis Instrumentais: Testes de Especicao
Cristine Campos de Xavier Pinto
CEDEPLAR/UFMG
Maio/2010
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
This lecture covers some useful tests for some of the
underlying hypothesis of OLS and IV.
We will see 2 type of tests: Hausman specication tests and
test for overidentifying restrictions.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
Hausman suggested a general class of diagnostic tests based
on comparison of two estimators,

and

for the same
parameter
0
.
Under the null hypothesis both of the estimators are
_
N
consistent so that
_
N
_

d
A (0, V)
Under departures from the model in the null hypothesis, the
estimators diverge in probability
p lim

,= 0
The Hausman specication test statistics takes the quadratic
form
HS = N
_

_/

V

D
_

_
where

V

D
is a consistent estimator of a generalized inverse of
V
D
.
There is one special case of hausman test that use 2SLS and
OLS.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
Consider the following partitioned regression model
Y
i
= X
1i

1
+X
2i

2
+
i
Consider a model where K
1
< K explanatory variables are
exogenous, and the last K K
1
variables are endogenous. .
We have a vector of instruments Z = [X
1
, W] . We assume
that the rank condition is satised, L _ K and
rank[M
ZX
] = K.
The null hypothesis is exogeneity:
H
0
: E
_
X
/

_
= 0
If the null hypothesis is true, the OLS estimator is relatively
ecient 2SLS estimator of
0
.
In the null hypothesis is false, the OLS estimator is
inconsistent. However, the 2SLS estimator that uses the
vector of instruments (exogenous variables) is a consistent
estimator of
0
.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
Hausman suggested a test based on the contrast

2SLS

OLS
.
Because under the null,

OLS
is ecient relative to

2SLS
Var
_

2SLS

OLS
_
= Var
_

2SLS
_
Var
_

OLS
_
and we can estimate this variance.
Under homoskedasticity
HS
=
_

2SLS

OLS
_/
_
_
X
/
P
Z
X
_
1

_
X
/
X
_
1
_

2SLS

OLS
_

2

d
A
2
K
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
There is a regression-based form of the Hausman test, that is
asymptotically equivalent to the original form of the Hausman
test.
Lets assume that K K
1
= 1. We have only one endogenous
explanatory variable.
To get the regression-based test, lets write the linear
projection of X
2
on Z
X
2
= Z
2
+v
E
_
Z
/
v
_
= 0
Z is a valid vector of instruments if assumptions IV1 and IV2
hold
IV1 :E[Z
i

i
] = 0
IV2 :E
_
Z
i
X
/
i
_
= M
ZX
where rank [M
ZX
] = K.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
Since is uncorrelated with Z, X
2
is only endogenous if and
only if
E[v] ,= 0
Write the linear projection of onto v
= v +
where
=
E[v]
E[v
2
]
, E[v
i

i
] = 0, and E[Z
i

i
] = 0
In this case, X
2
is exogenous only if = 0.
Plugging in the linear model for into our original model, we
have
Y = X
1i

1
+X
2i

2
+ v
i
+
i
where by construction
i
is uncorrelated with X
1i
, X
2i
and v
i
.
We need to test: H
0
: = 0.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
The problem is that v
i
is not observed.
However we can estimate
2
by an OLS regression of X
2
on Z
and v will be the residuals of this regression.
If we replace v by v, we have the equation
Y = X
1i

1
+X
2i

2
+ v
i
+error
i
where the parameters of this model can be consistently
estimated by OLS.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
The usual t statistics is a valid test for H
0
: = 0 provided
that the homoskedasticity assumption is satised under H
0
(E
_

Z, X
2

=
2
)
If we do not have homoskedasticity, we need to use t statistics
with a variance estimator that is robust under the presence of
heteroskedasticity
You are going to show in your problem set that the OLS
estimator of this equation,

is identical to the 2SLS
estimator.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
We can extend the regression-based Hausman test to the case
of several potentially endogenous explanatory variables. Lets
get back to the case where X
2
is vector K K
1
x1.
We assume IV1 and IV2.
In this case, the model in the rst step is
X
2
= Z
2
+v
where
2
is Lx(K K
1
) and v is a vector of population errors.
We can obtain the vector of OLS residuals,v.
We estimate the model
Y = X
1i

1
+X
2i

2
+v
i
+error
i
We do a F-test of H
0
: = 0 (under homoskedasticity)
If homoskedasticity is violated, we can use the
heteroskedasticity-robust Wald test.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
When we have more instruments than we need to identify an
equation, we can test whether the additional instruments are
valid, in the sense that they are uncorrelated with .
Null Hypothesis:
H
0
: m(Y, X, Z, ) = E
__
Y
i
X
/
i

_
Z
i

= 0
Suppose we partitioned the vector of the moments into two:
m(Y, X, Z, ) =
_
m
1
(Y, X, Z, )
m
2
(Y, X, Z, )
_
where m
1
(Y, X, Z, ) = E[(Y
i
X
/
i
) Z
1i
] is Kx1 and
m
2
(Y, X, Z, ) =E[(Y
i
X
/
i
) Z
2i
]is (L K)x1.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
Augmented model
m(Y, X, Z, ) =
_
m
1
(Y, X, Z, )
m
2
(Y, X, Z, ) +
_
where is a (L K) additional parameters.
In this augmented model, we want to test the hypothesis that
H
0
: = 0
.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
Test Statistics:
DM = 2N
_
S
N
_

N
, 0
_
S
N
_

N
,
N
_
_
= 2NS
N
_

N
_
d
A
2
LK
where
S
N
() = [m
N
()]
/
A
N
m
N
()
with
m
N
() =
1
N
N

i =1
_
Y
i
X
/
i

_
Z
i
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
There is an equivalent regression-based procedure.
Consider a overidentifying model:
Y
i
= X
1i

1
+X
2i

2
+
i
Consider a model where K
1
< K explanatory variables are
exogenous, and the last K K
1
variables are endogenous.
We have a vector of instruments Z = [X
1
, W] , with L > K.
The null hypothesis is that: H
0
: E[
i
Z
i
] = 0
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
We estimate a regression by 2SLS using all the instruments,
and obtain the residuals
i
.
Run a regression of
i
on a constant and the vector of
instruments Z , including all the instruments.
Obtain the R
2
of this regression,
LM = NR
2
d
A
2
LK
If we reject the null, we should not use your instruments.
If we fail to reject the null, we have evidence that the overall
set of instruments should be used.
This test is valid under homoskedasticity.
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
In a heteroskedasticity-version of this test, we rst need to get
the tted values of a regression of each element of X
2
on Z,

X
2
.
Now, let Z
2
be any 1xL K subset of W. I can be any subset.
Regress each element of Z
2
onto
_
X
1
,

X
2
_
and get the
residuals
2
.
2
is 1xL K .
Run a regression of 1 on
i

2
and get the SSR
S = N SSR
d
A
2
LK
Cristine Campos de Xavier Pinto Institute
Modelos Lineares
Hausman Tests Overidentifying Restrictions Tests
References
Rudd: 22
Hayashi: 3
Wooldridge: 6
Cristine Campos de Xavier Pinto Institute
Modelos Lineares

You might also like