[Carlo Lazzaro]

Guest:
if you only have two panels, use -panels(heteroskedastic)- if you detect heteroskedastcity.

[Guest]

Guest:
if you only have two panels, use -panels(heteroskedastic)- if you detect heteroskedastcity.

Sorry for the confusion.

My thesis consists of the cross-sectional analysis (all the countries within the sample, its the one this forum entry is about) and the case study (only two countries, the one the forum entry "Warning: variance matrix is nonsymmetric or highly singular" for robust random effects model is about)

So for the case study (2 countries), I use "panels(heteroskedastic)", right?

And for the cross sectional analysis (all countries), I cannot use "panels(correlated)" due to some missing values, so can I just use "panels(heteroskedastic)" instead?

[Carlo Lazzaro]

Guest:
1) if you have a cross-sectional dataset, I read it as you have many countries observed only once. If that were the case, you do not have a panel data, as T=1. Therefore, if your regressand is continous, you shuould go -regress-, possibly with -robust- standard errors (that in -regress- deals with heteroskedasticity only);
2) correct.

[Guest]

Guest:
1) if you have a cross-sectional dataset, I read it as you have many countries observed only once. If that were the case, you do not have a panel data, as T=1. Therefore, if your regressand is continous, you shuould go -regress-, possibly with -robust- standard errors (that in -regress- deals with heteroskedasticity only);
2) correct.

Sorry again for the confusion, cross-sectional is not the right word then. By cross-sectional I meant all the countries within the dataset, I need to find an accurate word for that then.

So yes, I have panel data for the "all countries" analysis (1960-2018 for 31 countries, T>N) and for the case study (two countries) analysis (1960-2018 for 2 countries, which is T>N as well).

So I have to use "xtgls" with "panels(heteroskedastic)" for both analysis then?

Thanks again for all your patience.

[Carlo Lazzaro]

Guest:
thanks for clarifying (and please be precise with econometric terms in your dissertation!).
You have two long panels (ie, T>N).
Both deserve -xtgls-.
The one with many panels should benefit from -panels(correlated)- if applicable (if not, explain why in your dissertation). As a second best, use "panels(heteroskedastic)" if you detect heteroskedastcity.
The one woth 2 panels only needs "panels(heteroskedastic)".

[Guest]

Guest:
thanks for clarifying (and please be precise with econometric terms in your dissertation!).
You have two long panels (ie, T>N).
Both deserve -xtgls-.
The one with many panels should benefit from -panels(correlated)- if applicable (if not, explain why in your dissertation). As a second best, use "panels(heteroskedastic)" if you detect heteroskedastcity.
The one woth 2 panels only needs "panels(heteroskedastic)".

I understand everything now. Sorry if most of my questions were not very smart.

Thanks so much Carlo, your help has been invaluable.

[Jeff Wooldridge]

My suggestion is not to use xtgls as it does not allow for country fixed effects, which is of primary concern in these studies. I would use the user-written command -xtscc- which allows for the longer T that Carlo is referring to. It essentially allows any N. If you choose the fe option with xtscc then you're doing fixed effects but getting a serial correlation and heteroskedasticity-robust standard error.

I would estimate the equation separately for Latin American countries. You might need to center your variables about averages before squaring and cubing them but I'd need to think more about this.

[Guest]

My suggestion is not to use xtgls as it does not allow for country fixed effects, which is of primary concern in these studies. I would use the user-written command -xtscc- which allows for the longer T that Carlo is referring to. It essentially allows any N. If you choose the fe option with xtscc then you're doing fixed effects but getting a serial correlation and heteroskedasticity-robust standard error.

I would estimate the equation separately for Latin American countries. You might need to center your variables about averages before squaring and cubing them but I'd need to think more about this.

Thanks for your suggestion.

But doesnt "i.country" with xtgls capture the country fixed effects already?
Moreover, I would like to perform a random effects model as well to see the comparison between fe and re.