11

Click here to load reader

Seemingly Unrelated Regressions 2011

Embed Size (px)

Citation preview

Page 1: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 1/11

Seemingly Unrelated Regressions

Anton Parlow

Lab session Econ 710UWM Econ Department

03/04/2011

Anton Parlow Lab session Econ 710 UWM Econ Department ()

Seemingly Unrelated Regressions 03/04/2011 1 / 11

Page 2: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 2/11

Our plan

Overview Sur-Regression.

Stata commands.Example 1.

Example 2.

Anton Parlow Lab session Econ 710 UWM Econ Department ()

Seemingly Unrelated Regressions 03/04/2011 2 / 11

Page 3: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 3/11

Seemingly Unrelated regression

If equations in a system of equations are related by their error-terms,estimating them together using GLS is more efficient compared to OLS.(or the variance is ”more” minimized)

We have: Y i  = X i β i  + i  with i  = 1 . . . L equations

If the error-terms across the equations are linked, so-called

”contemporaneous correlation” s.t. E i 

 j 

= σ2ij I  j 

then Zellner (1962) shows that a GLS-estimator is more efficient.

Anton Parlow Lab session Econ 710 UWM Econ Department ()

Seemingly Unrelated Regressions 03/04/2011 3 / 11

Page 4: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 4/11

Seemingly Unrelated regression continued

Let the above model in matrix notation be: Y  = z δ + e.g. Y  is a nTxL

matrix (and so on)

that δ GLS  = z 

V −1z −1

V −1Y 

and the variance-covariance matrix E  =

= V  with V  = Σ⊗ I T 

For the SUR-model Σ has to be non-diagonal or the off-diagonal termsshould not be zero. If they are zero there is no correlation between the

error-terms.

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 4 / 11

Page 5: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 5/11

Seemingly Unrelated regression continued

This can be tested after the Sur-Regression using a LaGrange-Multipliertest suggested by Breusch Pagan:

λLM  = T 

ni =2

i −1 j =1

r 2ij 

The H 0 = σij  has to be rejected, if we want to use SUR.

Otherwise every equation can be regressed by itself using OLS (there is nogain in using a GLS-estimator)

This is also true (most of the time) if the explanatory variables are thesame across the equations and if the regressors of one equation are asubset of another (Greene, p.343)

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 5 / 11

Page 6: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 6/11

Stata commands

Imagine you have two models:

y1=x1+x2 and y2=x1+x3 using OLS in Stata will be

reg y1 x1 x2

reg y2 x1 x3You suspect they are related by their error-terms → use sureg

sureg (y1 x1 x2) (y2 x1 x3)

in general

sureg (model1) (model2) (model3), options 

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 6 / 11

Page 7: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 7/11

Example 1

from the Stata-website. Load hsb2.dta into Stata

We want to explain test-scores for science and writing for female students:

reg science math female

reg write read femaleand suspect the error-terms are related e.g. because of the school they arein

sureg (science math female) (write read female), corr

Do you see any difference to the OLS-results???

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 7 / 11

Page 8: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 8/11

Example 1 continued

Sur-output → you see the output for the 2 models

Note: Stata adds the BP-test to the output! In this case the error-termsare correlated.

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 8 / 11

Page 9: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 9/11

Example 1 continued

Let us test the effect of type of program prog  on scores. There are threeprograms, a dummy would be nice:tabulate prog, gen(prog)Consider following three OLS-regressions (having the same regressors!):read female prog1 prog3

write female prog1 prog3math female prog1 prog3Still we can assume the errors are related:sureg (read female prog1 prog3) (write female prog1 prog3) (mathfemale prog1 prog3), corr

sames as:sureg (read write math = female prog1 prog3), corrbecause the regressors are the same

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 9 / 11

Page 10: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 10/11

Example 2

Consider a system of cost-share equations derived from a translog costfunction having four inputs (KLEM-model):

S K  = αK  + γ KK lnP K  + γ KLlnP L + γ KE lnP E  + γ KM lnP M 

S L = αL + γ KLlnP K  + γ LLlnP L + γ LE lnP E  + γ LM lnP M 

S E  = αE  + γ KE lnP K  + γ LE lnP L + γ EE lnP E  + γ EM lnP M 

S M  = αM  + γ KM lnP K  + γ LM lnP L + γ EM lnP E  + γ MM lnP M 

We estimate only three share equations because the shares add up to 1.And let us ignore homogeneity and symmetry constraints on that system

for simplicityLoad the bw.dta file

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 10 / 11

Page 11: Seemingly Unrelated Regressions 2011

7/27/2019 Seemingly Unrelated Regressions 2011

http://slidepdf.com/reader/full/seemingly-unrelated-regressions-2011 11/11

Example 2 continued

Use sureg (lk ll le = lpk lpl lpe), corr

and compare it to OLS-estimates for each equation.

What do you see?

Anton Parlow Lab session Econ 710 UWM Econ Department ()Seemingly Unrelated Regressions 03/04/2011 11 / 11