Eliminating of heteroscedasticity

[Martin Eden]

Hello to everybody,
I faced to following problems when conducting my research:
I have the heteroscedasticity according to test with command -estat imtest- and have not according to Breusch-Pagan test.
I don't know for sure whether I need to correct the model to eliminate possible heteroscedasticity or not. I have panel data, models in three variants - OLS, FE, RE. Please give advice on how to improve the model.
I attach screenshots for clarity

[Carlo Lazzaro]

Martin:
without more substantive details, I'd go -vce(cluster clusterid)- for panel and -robust- for OLS.
That said, I fail to get how -xtreg- can support -estat hettest-.

[Martin Eden]

Martin:
without more substantive details, I'd go -vce(cluster clusterid)- for panel and -robust- for OLS.
That said, I fail to get how -xtreg- can support -estat hettest-.

Carlo, thank you for your response!
Yes, of course. I mean, that I used -estat hettest- for OLS, not for panel data.

But problem have another roots. There is a problem with gaps in my data.
Not that there are a lot of them in general - rather, data for individual IDs are available for different time periods. For example, there are observations from 2000 until 2005 and no one for 2006-2010, after have from 2011-2020 and etc. When I trying to cut my data set and left just one part of period which contain as much as possible observations, the overall number of observations become a small.
As a result, the Hausman test showed that FE and RE are not applicable (prob < 0.05). Thus, the question is, are there any ways to transform the model in order to match the panel data? And otherwise, how correct is it to use OLS for data where there are years, ID and one of the variables is lagged?

Best regards,
Martin

[Carlo Lazzaro]

Martin:
1) as Stata can handle both balanbed and unbalanced panels, there's no need to carving out your dataset. I addition, this approach may end up with making up your original sample size;
2) I would stay with -xtreg-, starting with the -fe- specification.
3) If -hausman- outcome does not reach statistical significance, it means that -re- is the way to go;
4) if you impose non-default standard errors, you should awitch to the community-contributed module -xtoverid- that, being a bit old-fashioned, does not support -fvvarlist- notation (see -xi.- prefix in this respect).

[Martin Eden]

Martin:
1) as Stata can handle both balanbed and unbalanced panels, there's no need to carving out your dataset. I addition, this approach may end up with making up your original sample size;
2) I would stay with -xtreg-, starting with the -fe- specification.
3) If -hausman- outcome does not reach statistical significance, it means that -re- is the way to go;
4) if you impose non-default standard errors, you should awitch to the community-contributed module -xtoverid- that, being a bit old-fashioned, does not support -fvvarlist- notation (see -xi.- prefix in this respect).

Carlo, thanks for your advices
It very helpful for me!

Best regards,
Martin