Scribd
Upload
18 views

Rec 4 20

This document discusses methods for dealing with heteroskedasticity in regression models, including generalized least squares (GLS) and feasible generalized least squares (FGLS). It describes how to implement GLS if the variance model is known, and how to use FGLS if it is unknown by estimating it from the data. Two tests for detecting heteroskedasticity are presented: the Breusch-Pagan test and the White test. The document recommends using White standard errors as a robust approach if heteroskedasticity is suspected.

Uploaded by

Malek Blr
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
18 views

Rec 4 20

This document discusses methods for dealing with heteroskedasticity in regression models, including generalized least squares (GLS) and feasible generalized least squares (FGLS). It describes how to implement GLS if the variance model is known, and how to use FGLS if it is unknown by estimating it from the data. Two tests for detecting heteroskedasticity are presented: the Breusch-Pagan test and the White test. The document recommends using White standard errors as a robust approach if heteroskedasticity is suspected.

Uploaded by

Malek Blr
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

More on FGLS and How to Test for Heteroskedasticity

Generalized Least Squares and Feasible GLS

Here’s a little more detail on how to perform GLS/FGLS:


Suppose you have a standard multivariate model such as
y i   0   1 x 1i   2 x 2i   i
which satisfies all of the standard assumptions except Var i |x 1i , x 2i  ≠  2 . Instead,
you are told (or you guess) that
Var i |x 1i , x 2i    2 gx 1i , x 2i    2  0   1 |x 1i |   2 |x 2i |
This is definitely not the only way to model the standard errors. For example, you
can use interactions, higher powers of the x’s, etc. There are two issues to watch
out for:

1. All Variances must be positive (the reason for using the exponential
function above)
2. You need to believe in your model! In most cases, there is no strong
theory as to what the heteroskedasticity should look like. This is why many
empirical economists just "punt" and just use robust standard errors.

If you are confident in your model and you know  0 ,  1 , and  2 , then you can just
use generalized least squares. Instead of running the model
y i   0   1 x 1i   2 x 2i   i

you need to divide all of the variables by gx 1i , x 2i  and regress


yi 0  1 x 1i  2 x 2i i
   
gx 1i , x 2i  gx 1i , x 2i  gx 1i , x 2i  gx 1i , x 2i  gx 1i , x 2i 

Now, we have eliminated the heteroskedasticity, since

Var i  1 Var i    2
gx 1i , x 2i  gx 1i , x 2i 

This procedure is relatively easy, but not feasible since we almost never know the
parameters of the variance equation,  0 ,  1 , and  2 . Hence the need for Feasible
Generalized Least Squares.

Cite as: James Berry, course materials for 14.32 Econometrics, Spring 2007. MIT OpenCourseWare (http://ocw.mit.edu/),
Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].
The general idea of FGLS is to use the estimated ̂ i from the standard OLS
equation, then estimate ̂ 0 , ̂ 1 , and ̂ 2 using another OLS regression. The key
here is that even though OLS standard errors are wrong, the estimates are still
consistent. This means that we can use the estimated errors as consistent
estimators of the actual standard errors.
FGLS is a straightforward process, but there are more steps:

1. Run the OLS Regression y i   0   1 x 1i   2 x 2i   i to get the estimated


residuals ̂ i .
2. Run the regression ̂ 2i  gx 1i , x 2i , in our case we have
̂ 2i   0   1 |x 1i |   2 |x 2i |  

The fitted values from this regression, call them ̂ are our estimated gx 1i , x 2i 
(up to a multiplicative constant, which doesn’t matter).
3. We can now use a similar formula as before
yi 0  1 x 1i  2 x 2i i
   
gx 1i , x 2i  gx 1i , x 2i  gx 1i , x 2i  gx 1i , x 2i  gx 1i , x 2i 
i
Computing the actual variance of the new residual is tricky, but suffice to
gx 1i ,x 2i 

say that in large samples

Var i  Var i  1 Var i    2


gx 1i , x 2i  gx 1i , x 2i 
gx 1i , x 2i 
which is exactly the same as in GLS, as if we really knew the parameters of
gx 1i , x 2i !

How to check for Heteroskedasticity

Here are two ways to test for heteroskedasticity, and they use the same concepts
as we use above: they test whether the squared estimated residuals are related to
the x’s.

The Breusch-Pagan Test

1. Run the OLS Regression y i   0   1 x 1i   2 x 2i   i to get the estimated


residuals.

Cite as: James Berry, course materials for 14.32 Econometrics, Spring 2007. MIT OpenCourseWare (http://ocw.mit.edu/),
Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].
2. Run the OLS regression of the estimated residuals on the independent
variables, that is,

̂ 2i   0   1 x 1i   2 x 2i  

3. Construct and F-test of the joint hypothesis that  1  0 and  2  0, as


you would in any other OLS situation.

The White Test

1. Run the OLS Regression y i   0   1 x 1i   2 x 2i   i to get the estimated


residuals
2. Run the OLS regression of the estimated residuals on the independent
variables and interactions of the independent variables, that is,
̂ 2i   0   1 x 1i   2 x 2i   3 x 1i x 2i  

3. Construct and F-test of the joint hypothesis that  1  0,  2  0 and


 3  0 , as you would in any other OLS situation.

Q: Which of these two tests is better?

SAS can automatically compute a variant of the White test using the SPEC option
in the model statement. (It uses a Chi-squared LM-type test instead of an F-test to
check for joint significance, but it doesn’t really matter which one you use.)

Punting with White Standard Errors

Finally, recall that if we decide to "punt" and assume general heteroskedasticity


just in case it’s there, we simply use the formula
∑ x̃ 1i û 2i
Var̂  
s 2x̃ 1  2

I put x̃ in there because you need to partial out other variables when doing a
multivariate regression.

To do this numerically in SAS, you need to use the option ACOV in the MODEL

Cite as: James Berry, course materials for 14.32 Econometrics, Spring 2007. MIT OpenCourseWare (http://ocw.mit.edu/),
Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].
statement. To get the standard errors of each variable, take the square root of
each diagonal element of the matrix that turns up in the output. Check out the
output on the next page using a standard earnings-schooling regression with both
the ACOV and SPEC options:

proc reg dataone;


model rearningsed_comp exp / acov spec;

This was done with Kreuger’s 1984 data from problem set 5.

Cite as: James Berry, course materials for 14.32 Econometrics, Spring 2007. MIT OpenCourseWare (http://ocw.mit.edu/),
Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

You might also like