You are on page 1of 20

# Regression Analysis Tutorial 228

LECTURE / DISCUSSION

Simultaneous Equations

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 229

Simultaneous Equations

Example 1

Gas
consumption
OLS estimate

. .
. true
. . .
. .
. . .
. .
. .
. . .
Thermostat
. setting

## Problem: Because µ is correlated with g , and µ affects

thermostat setting, thermostat setting is
correlated with g .

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 230

Example 2

## Price elasticity of telecommunications demand:

MOU ' % (price) % (Nemply) % g
Price ' % (MOU) % (region dum.) % µ

MOU
OLS
.
.
.
.
.
.
. . .
.
. true
. Price

Demand

## Problem: Because µ is correlated with g , and µ affects

price, price is correlated with g .

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 231

Example 3

## Aggregate demand and supply:

Demand: Q = " + \$P + g
Supply: P = 8 + 2Q + µ

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 232

M ' % P % N % g
P ' % M % R % µ

Endogenous: M, P
Exogenous: R, N

## Step 1: Regress endogenous variables against all the

exogenous variables.

M = a + bR + cN + e
P = d + fR + gN + u
Get M̂ , P̂

## Step 2: Regress original equations, replacing endogenous

explanatory variables M and P with predicted
values M̂ and P̂ .

M ' % P̂ % N % g(

P ' % M̂ % R % µ (

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 233

Or apply IV directly.

Regress M = " + \$P + 2N + g
with instruments N and R
and P = 8 + NM + 0R + µ
with instruments N and R .

TSP commands:

## 2sls(inst = (c,n,r)) m c,p,n ;

2sls(inst = (c,n,r)) p c,m,r ;
or
inst m c,p,n invr c,n,r ;
inst p c,m,r invr c,n,r ;

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 234

Reduced-Form Equation

## Structural model: causal relationships

M = " + \$P + 2N + g
P = 8 + NM + 0R + µ

## Structural parameters: ", \$, 2, 8, N, 0 .

Reduced-form
equations: endogenous variables expressed as a
function of exogenous variables.

M ' % ( % M % R % µ) % N % g
M ' % % M % R % N % µ % g

(1 & )M ' % % R % N % µ % g

% µ % g
M ' % R % N %
1& 1& 1& 1&
| | | |
| | | |
M= a + bR + cN + e

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 235

Similarly,

% g % µ
P ' % R % N %
1& 1& 1& 1&
| | | |
| | | |
P= d + fR + gN + u

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 236

## 1. Reduced-form equations are estimated in first step of

2SLS. Structural equations are estimated in second
step of 2SLS.

## 2. Reduced-form equations can be estimated by OLS

because exogenous variables are uncorrelated with
errors.

## 3. There is a relation between the reduced-form

parameters and the structural parameters.

For example:

c '
1&

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 237

## Reduced form parameters give the full effect of a change in

an exogenous variable.

M = " + \$P + 2N + g
P = 8 + NM + 0R + µ

N rises by 1 unit:

M increases by 2
P increases by N2
M increases by \$[N2]
P increases by N[\$N2] = \$N22
M increases by \$[\$N22] = \$2N22
P increases by N[\$2N22] = \$2N32
M increases by \$[\$2N32] = \$3N32
.
.
.

Total effect on M

4
% % ( ) 2
% ( ) 3
% ... ' j ( )t
t'0

'
1&

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 238

Total effect on P

% 2
% 2 3
' % ( ) % ( )2
4
' j( )t '
t'0 1&

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 239

## Sometimes, structural parameters can be calculated from

reduced-form parameters.

% %
a ' d '
1& 1&

b ' f '
1& 1&

c ' g '
1& 1&

b g
' ' etc.
f c

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 240

Example:

if â ' 4 d̂ ' 8
b̂ ' 3 f̂ ' 2
ĉ ' 7 ĝ ' 3.5

## then ˆ ' & 8. ˆ ' 6

ˆ ' 1.5 ˆ ' 0.50
ˆ ' 1.75 ˆ ' 0.50

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 241

## I. Add an exogenous variable:

M ' % P % N % I % g
P ' % M % R % µ
7 structural parameters

Reduced-form equations:

M = a + bN + cI + dR + e

P = f + gN + hI + RR + u

8 reduced-form parameters

## 2SLS finds structural parameters that best fit the

reduced form parameters.

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 242

## II. Omit an exogenous variable:

M ' % P % g
P ' % M % R % µ
5 structural parameters

Reduced-form equations

M = a + bR + e

P = c + dR + u

4 reduced-form parameters

solutions.

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 243

## Step 1: M̂ ' â % b̂R

P̂ ' ĉ % d̂R

Step 2: M ' % P̂ % g

P ' % M̂ % R % µ
' % â % b̂R % R % µ
' ( % â) % ( b̂ % )R % µ

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 244

Identification

## Name Estimation method

Just identified NRFP = NSP 2SLS = ILS
Over identified NRFP > NSP 2SLS
Under identified NRFP < NSP Cannot estimate

## NRFP: Number of reduced form parameters.

NSP: Number of structural parameters.

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 245

Efficiency

## But 2SLS ignores information contained in correlation

between errors. So, 2SLS is not efficient.

M ' % P % N % g
P ' % M % R % µ
g, µ correlated

## If we knew µ was high, we would know that g is

probably also high and hence M is higher than predicted
from P and N only.

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 246

and µ̂ .

## Step 3: Re-estimate structural equations with ĝ and µ̂

included as explanatory variables.

M ' % P̂ % N % µ̂ % g((

P ' % M̂ % R % ĝ % µ ((

## Because g and µ are correlated, µ̂ provides

information for explaining M and ĝ provides
information for explaining P . Including this
information makes the estimates better.

## Econometrics Laboratory C University of California at Berkeley C 23-27 March 1998

Regression Analysis Tutorial 247

Special Case:
Seemingly Unrelated Equations
Yn ' % X n % gn

Wn ' % Sn % µ n

## No endogenous explanatory variables.

But: gn and µn are correlated.

## Example: Yn = temperature in San Francisco.

Wn = temperature in Monterey.

## 3SLS can be done in two steps, because reduced-form

equations are the same as the structural equations.

## Step 1: Estimate equations by OLS.

Get residuals ĝ n and µ̂ n .

## Step 2: Re-estimate equations including residuals as

explanatory variables:
(
Yn ' % Xn % µ̂ n % gn
(
Wn ' % S n % ĝ n % µ n