GLS in Eviews 9 confusing!

[omid_ie]

I am trying to run GLS on this OLS (attached file)

Code: Select all

ls miles c income age kids

(eq1 in attached file)

As there is Heteroskedasticity, I want to run GLS given that var(u_i)=sigma^2*income

Now I want to replicate GLS Both with codes and menu provided by Eviews 9.0. But I got problem!


so I tried these codes:

Code: Select all

scalar sigma = eq1.@se genr income2=income/(sigma*((income)^(0.5))) genr kids2=kids/(sigma*((income)^(0.5))) genr miles2=miles/(sigma*((income)^(0.5))) genr age2=age/(sigma*((income)^(0.5))) ls miles2 c income2 age2 kids2

(wls2 in attached file)


then I tried this from menu:
in Equation Estimation window > miles c income age kids
then Options:
type: inverse std. dev.
weight sires: 1/(sigma*income^0.5)
scaling: eviews default
(eviewswls in attached file)


but as you can see the results for constant value and income coefficient are different in both ways. I might made a mistake in both codes and proper selection of menu.
My only concern is running GLS with scaling var(u_i)=sigma^2*income.

Thanks in advance

[EViews Glenn]

You need to scale the intercept as well.

Code: Select all

equation eq1.ls miles c income age kids scalar sigma = eq1.@se genr income2=income/(sigma*((income)^(0.5))) genr kids2=kids/(sigma*((income)^(0.5))) genr miles2=miles/(sigma*((income)^(0.5))) genr age2=age/(sigma*((income)^(0.5))) genr c2 = 1/(sigma*((income)^(0.5))) equation eq2.ls miles2 c2 income2 age2 kids2 equation eq3.ls(w=1/(sigma*income^0.5)) miles c income age kids

Note that the definitions of the summary statistics differ when you run the model in the two different ways.

[omid_ie]

Thanks... Now It works!

Now I have a question!

If I run BP test on equ2 (

Code: Select all

equation eq2.ls miles2 c2 income2 age2 kids2

)

then I will get LM=15.65 showing there is still Heteroskedasticity!

What is conclusion? Is it correct to say:
var(u_i)=sigma^2*income

above scaling is not correct?
Is there any other way to come up with solution about Heteroskedasticity with GLS? ( I mean not using ARCH models)

[EViews Glenn]

The implication of the test result is that the weighted model using your specified weights still exhibits heteroskedasticity.
You can come up with an alternative model for the variance.
Alternately, you can ignore the heteroskedasticity in estimation and use heteroskedasticity robust standard errors to obtain proper inference.