Вы находитесь на странице: 1из 13

Regression Analysis Tutorial 207

LECTURE / DISCUSSION

Instrumental Variables

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 208

Instrumental Variables Estimation

Problem: error is correlated with explanatory variable.

Miles

. .
.
. . .
. . .
. . . . . . .
. . . . . .
. . . .
. . .. .
.
true

Operating
cost

Miles = " + $(op. cost) + g

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 209

Solution #1: Include the variable that causes


the correlation, if possible.
(Miles) = " + $(op. cost) + 2(distance to work) + g

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 210

Solution #2: Instrumental variables estimation


Yn = " + $Xn + gn

Corr(X,g) ú 0

Find a variable, Z, that is


C correlated with X
C not correlated with g

Z is called an instrument.

j z ny n cov (z,y)
ˆ
IV ' '
j z nx n cov (z,x)

Example: Miles = " + $(op. cost) + g


Instrument: gas price

cov (gas price , miles)


ˆ
IV '
cov (gas price , op. cost)

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 211

Why Does IV Work?

β
Y X
θ

cov (zy)
E( ˆ IV) ' ' '
cov (zx)

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 212

When IV is Biased

β
Y X

θ
λ
Z

cov (zy) %
E( ˆ IV) ' ' ' %
cov (zx)

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 213

Why Is IV Unbiased?

yn' xn%gn

j z n y n ' j z n x n %j z n g n

E(j znyn)' E(j znxn)%E(j zngn)


' E(j znxn)

So that zn isolates the covariation between yn and xn .


This leads to

zy
ˆ 'j n n
IV
j z nx n

L Efficiency is greater with instruments that are more


highly correlated with X , while still uncorrelated with
the error terms.

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 214

Can Apply IV Through 2SLS

Yn = " + $Xn + gn

instrument Zn

Step 1: Run OLS regression

Xn = 8 + 2Zn + µn

Get predicted X̂ n ' ˆ % ˆ Zn

Step 2: Run OLS regression using X̂


(
Yn ' % X̂ n % gn

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 215

Example

Miles = " + $(op. cost) + g


instrument: gas price

Step 1: Regress

(op. cost) = 8 + 2(gas price) + µ

Calculate
(op. cost) ' ˆ % ˆ (gas price)

Step 2: Regress

Miles = " + $(op. cost) + g*

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 216

2SLS Interpretation of IV

Yn = " + $Xn + gn

Step 1: X is decomposed into a part that is uncorrelated


with g and a part that is correlated with g .

Xn = 8 + 2Zn + µn
________ ____
| |
| |
uncorrelated correlated
with g with g

Step 2: Use the portion of X that is uncorrelated with g


in estimating the original equation.

(
Yn ' % X̂ % gn

X̂ n ' ˆ % ˆ zn

(
gn ' gn % µ̂ n

The part of X that was correlated with g gets moved


into the error.

The part of X that was uncorrelated with g stays as an


explanatory variable.

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 217

2SLS method also shows that many instruments can be used


in IV.

(Miles) = " + $(op. cost) + g

Instruments: gas price


relative prices for large and small cars

Not instruments: income


family size

Note: Income and family size can serve as instruments if


they also enter the original regression as explanatory
variables.

Miles = " + $(op. cost) + 2(income) + 8(family size) + g

Instruments: gas price


relative prices for large and small cars
income
family size

Because income and family size enter directly, they are not
part of g and hence are not correlated with g .

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 218

Equivalence of IV and 2SLS

2SLS

Step 1:
x n ' z n % gn

cov (zx)
ˆ
OLS '
var (z)
x̂ n ' ˆ OLSzn

Step 2:
yn ' x̂ n % gn

cov (x̂y)
ˆ
2SLS '
var (x̂)

cov ( ˆ OLSzy) ˆ
OLS cov (zy)
ˆ
2SLS ' '
var ( ˆ OLSz) ˆ2
OLS var (z)

cov (zy) cov (zy)


' '
ˆ cov (zx)
OLS var (z) @ var (z)
var (z)
cov (zy)
' ' ˆ IV
cov (zx)

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999


Regression Analysis Tutorial 219

TSP Commands

Three ways to estimate

Y = " + $X + g with instrument Z

1. 2SLS explicitly as two OLS regressions


olsq x c,z;
genr xhat = @fit;
olsq y c,xhat;

2. 2SLS implicitly

2sls(inst = (c,z)) y c,x;

3. IV
inst y c,x invr c,z;

Econometrics Laboratory C University of California at Berkeley C 22-26 March 1999

Вам также может понравиться