Using weighted least squares on Heteroskedasticity

[Dawud Sikander]

I used the Breusch-Pagan to check for Heteroskedasticity which I found (1st attachment). Then I used weighted least squares to mitigate it. I did this by trial and error and used the code [aw=1/experience]. I have attached the results (2nd attachment) . Does the 2nd image still show Heteroskedasticity?

Also I am running two regressions this one(the images attached) is for return to education for white ethnicity in the UK and my other regression which does not have any heterskedascity is for returns to education for ethnic minorities in the UK. Although I do not have any Heteroskedasticity in the ethnic minorities regression do I still need to use weighted least squares. As both regressions run the same variables and I don't want to create any bias in either one.

I am new to stata and any responses will be very appreciated

[Carlo Lazzaro]

Dawud:
please start sharing what you typed and what Stata gave you back via CODE delimiters (as per FAQ, that also remind listers that screenshots are deprecated,mainly because it is impossible to work/elaborate on them). Thanks.

[Dawud Sikander]

Dawud:
please start sharing what you typed and what Stata gave you back via CODE delimiters (as per FAQ, that also remind listers that screenshots are deprecated,mainly because it is impossible to work/elaborate on them). Thanks.

Hello,
Thank you for letting me know I will do that.

[Joao Santos Silva]

Dear Dawud Sikander,

I have some comments and suggestions on this:

1 - You should use robust standard errors both in the OLS and WLS regressions.

2 - You have a different number of observations in the two equations. I presume that is because experience is zero for some individuals. This suggests that your weights are not very sensible.

3 - As an alternative, I suggest you run a Poisson regression of grsswk (not it log) on the same regressors. Again, you should use robust standard errors.

Best wishes,

Joao

[Dawud Sikander]

Dear Dawud Sikander,

I have some comments and suggestions on this:

1 - You should use robust standard errors both in the OLS and WLS regressions.

2 - You have a different number of observations in the two equations. I presume that is because experience is zero for some individuals. This suggests that your weights are not very sensible.

3 - As an alternative, I suggest you run a Poisson regression of grsswk (not it log) on the same regressors. Again, you should use robust standard errors.

Best wishes,

Joao

Hello,
Thank you very much for your reply

Should I change my weights the code i used to change them was

[aw=1/experience]

. I used the weights to fix my heteroskedasticty. Do you suggest I change these weights in order to fix the observations?
Also I am unfamiliar with robust standard errors is that another way of testing heteroskedascity? Also I have a a basic understanding of Poisson regression where you predict the count?
I am sorry for my lack of knowledge

[Joao Santos Silva]

Dear Dawud Sikander,

You should really learn about robust standard errors; please get a good textbook and study that section.

Best wishes,

Joao

[Uyi Erhabor]

Dear Stata users,

Please could someone let me know how to compute a WLS fixed effects regression model with robust standard errors to address heteroskedasticity in a panel data set. I have managed to compute the robust standard errors but I'm not sure how to get the WLS estimates. I have tried to but I received the following error:


xtreg DiffMeanHourlyPercent responsible_person_changes [weight = yhat4_inv], vce(robust) fe
(analytic weights assumed)

'weight must be constant within CompanyNumber'


Kind regards,

Uyi Erhabor
(Stata 16.1 SE)

[Joro Kolev]

Either make the weights to be constant within panels, or if you insist on having time varying weights, do the time demeaning manually and use -regress- to fit the fixed effects regression.