Scribd
Upload
31 views

Instrumental Variables & 2SLS: y + X + X + - . - X + U X + Z+ X + - . - X + V

1. Instrumental variables (IV) estimation is used when a model has endogenous variables, or variables that are correlated with the error term. IV addresses omitted variable bias. 2. A valid instrumental variable must be correlated with the endogenous explanatory variable but uncorrelated with the error term. 3. Two-stage least squares (2SLS) is a common approach to IV estimation that first estimates the endogenous variables based on the instruments, then uses those predicted values in the main regression equation.

Uploaded by

Sahrish Jaleel Shaikh
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
31 views

Instrumental Variables & 2SLS: y + X + X + - . - X + U X + Z+ X + - . - X + V

1. Instrumental variables (IV) estimation is used when a model has endogenous variables, or variables that are correlated with the error term. IV addresses omitted variable bias. 2. A valid instrumental variable must be correlated with the endogenous explanatory variable but uncorrelated with the error term. 3. Two-stage least squares (2SLS) is a common approach to IV estimation that first estimates the endogenous variables based on the instruments, then uses those predicted values in the main regression equation.

Uploaded by

Sahrish Jaleel Shaikh
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/ 21

Instrumental Variables & 2SLS

y =
0
+
1
x
1
+
2
x
2
+ . . .
k
x
k
+ u
x
1
=
0
+
1
z +
2
x
2
+ . . .
k
x
k
+ v
Why Use Instrumental Variables?
Instrumental Variables (IV) estimation is
used when your model has endogenous xs
That is whene!er "o!(x,u) # $
Thus IV %an be used to address the
&roblem o' omitted !ariable bias
(mitted Variable )ias (re%a&)
( )
* 2 * *
* * $
2 2 * * $
+ +
then
+ +
+
estimate
but we
as gi!en is model true the Su&&ose



+ =
+ + =
+ + + =
E
u x y
u x x y
Summary o' ,ire%tion o' )ias
(re%a&)
"orr(x
1
, x
2
) - $ "orr(x
1
, x
2
) . $

2
- $
/ositi!e bias 0egati!e bias

2
. $
0egati!e bias /ositi!e bias
/ro1y Variables
What i' model is miss&e%i'ied be%ause no
data is a!ailable on an im&ortant x !ariable?
It may be &ossible to a!oid omitted
!ariable bias by using a &ro1y !ariable
2 &ro1y !ariable must be related to the
unobser!able !ariable
What i' a suitable &ro1y !ariable is also not
a!ailable? 3ay use IV or 2SLS 2&&roa%h4
What Is an Instrumental Variable?
In order 'or a !ariable z to ser!e as a !alid
instrument 'or x the 'ollowing must be true
The instrument must be e1ogenous
That is "o!(z,u) 5 $
The instrument must be %orrelated with the
endogenous !ariable x
That is "o!(z,x) # $
3ore on Valid Instruments
We ha!e to use %ommon sense and
e%onomi% theory to de%ide i' it ma6es sense
to assume "o!(z,u) 5 $
We %an test i' "o!(z,x) # $
7ust testing 8
$
9
1
5 $ in x =
0
+
1
z + v
Sometimes re'er to this regression as the
'irst:stage regression
IV ;stimation in the Sim&le
<egression "ase
=or y =
0
+
1
x + u and gi!en our
assum&tions
"o!(z,y) 5
1
"o!(z,x) > "o!(z,u) so

1
5 "o!(z,y) ? "o!(z,x)
Then the IV estimator 'or
1
is
( ) ( )
( )( )



=
x x z z
y y z z
i i
i i
*
@

The ;''e%t o' /oor Instruments


What i' our assum&tion that "o!(z,u) 5 $ is 'alse?
The IV estimator will be in%onsistent too
"an %om&are asym&toti% bias in (LS and IV
/re'er IV i' "orr(z,u)?"orr(z,x) . "orr(x,u)
x
u
x
u
u x Corr
x z Corr
u z Corr


+ =
+ =
) (
+
&lim 9 (LS
) (
) (
@
&lim 9 IV
* *
* *
In'eren%e with IV ;stimation
The homos6edasti%ity assum&tion in this %ase is
;(u
2
|z) 5
2
5 Var(u)
2s in the (LS %ase gi!en the asym&toti%
!arian%e we %an estimate the standard error
( )
2 ? *
2

2
*
)
@
(
@
z x x
R SST
se

=
IV !ersus (LS estimation
Standard error in IV %ase di''ers 'rom (LS
only in the R
2
'rom regressing x on z
Sin%e R
2
. * IV standard errors are larger
8owe!er IV is %onsistent while (LS is
in%onsistent when "o!(x,u) # $
The stronger the %orrelation between z and
x the smaller the IV standard errors
IV ;stimation in the 3ulti&le
<egression "ase
IV estimation %an be e1tended to the
multi&le regression %ase
"all the model we are interested in
estimating the stru%tural model
(ur &roblem is that one or more o' the
!ariables are endogenous
We need an instrument 'or ea%h
endogenous !ariable
3ulti&le <egression IV (%ont)
Write the stru%tural model as y
1
=
0
+
1
y
2

+
2
z
1
+ u
1
where y
2
is endogenous and z
1

is e1ogenous
Let z
2
be the instrument so "o!(z
2
,u
1
) 5 $
and
y
2
=
0
+
1
z
1
+
2
z
2
+ v
2
where
2
# $
This redu%ed 'orm eAuation regresses the
endogenous !ariable on all e1ogenous ones
Two Stage Least SAuares (2SLS)
Its &ossible to ha!e multi&le instruments
"onsider our original stru%tural model and
let y
2
=
0
+
1
z
1
+
2
z
2
+
3
z
3
+ v
2
8ere were assuming that both z
2
and z
3
are
!alid instruments B they do not a&&ear in
the stru%tural model and are un%orrelated
with the stru%tural error term u
1
)est Instrument
"ould use either z
2
or z
3
as an instrument
The best instrument is a linear %ombination
o' all o' the e1ogenous !ariables y
2
* =
0
+

1
z
1
+
2
z
2
+
3
z
3

We %an estimate y
2
* by regressing y
2
on z
1
,
z
2
and z
3
B %an %all this the 'irst stage
I' then substitute
2
'or y
2
in the stru%tural
model get same %oe''i%ient as IV
3ore on 2SLS
While the %oe''i%ients are the same the
standard errors 'rom doing 2SLS by hand
are in%orre%t so let Stata do it 'or you
3ethod e1tends to multi&le endogenous
!ariables B need to be sure that we ha!e at
least as many e1%luded e1ogenous !ariables
(instruments) as there are endogenous
!ariables in the stru%tural eAuation
Testing 'or ;ndogeneity
Sin%e (LS is &re'erred to IV i' we do not
ha!e an endogeneity &roblem then wed
li6e to be able to test 'or endogeneity
I' we do not ha!e endogeneity both (LS
and IV are %onsistent
Idea o' 8ausman test is to see i' the
estimates 'rom (LS and IV are di''erent
Testing 'or ;ndogeneity (%ont)
While its a good idea to see i' IV and (LS
ha!e di''erent im&li%ations its easier to use
a regression test 'or endogeneity
I' y
2
is endogenous then v
2
('rom the
redu%ed 'orm eAuation) and u
1
'rom the
stru%tural model will be %orrelated
The test is based on this obser!ation
Testing 'or ;ndogeneity (%ont)
Sa!e the residuals 'rom the 'irst stage
In%lude the residual in the stru%tural
eAuation (whi%h o' %ourse has y
2
in it)
I' the %oe''i%ient on the residual is
statisti%ally di''erent 'rom Cero reDe%t the
null o' e1ogeneity
I' multi&le endogenous !ariables Dointly
test the residuals 'rom ea%h 'irst stage
Testing (!eridenti'ying
<estri%tions
I' there is Dust one instrument 'or our
endogenous !ariable we %ant test whether
the instrument is un%orrelated with the error
We say the model is Dust identi'ied
I' we ha!e multi&le instruments it is
&ossible to test the o!eridenti'ying
restri%tions B to see i' some o' the
instruments are %orrelated with the error
The (!erI, Test
;stimate the stru%tural model using IV and
obtain the residuals
<egress the residuals on all the e1ogenous
!ariables and obtain the R
2
to 'orm nR
2
Under the null that all instruments are
un%orrelated with the error L3 +
A
2
where
q is the number o' e1tra instruments

You might also like