Problem with correcting a heteroskedasticity and serial correlation on panel data with fixed effect

[proud pm]

Hello Statalists,

I have a balanced panel data of 131 countries over 20 years with fixed effect (using xtreg) and I found serial correlations (using xtcsd) and heteroskedasticity (using xttest3) in the model. I tried correcting the model by using

xtreg <depvar> <indepvars>, fe vce(cluster panelid)

but it turned out all of independent variables are not significant (which were significant before). So, I would like to know what should I do about this or there is any other method to fix those two problems rather than cluster? Any help would be much appreciated

Thanks

[Jack Grover]

Hello.
the option "vce(cluster panelid)" takes into account the within error correlation of the cluster comprehensive of the various ID (in your case). In general, if sigma_u is small (small variability in the residuals within groups), the option cluster is not feasible (it can be that this is your case (?)).
You may perhaps use the cluster option controlling for different clusters (e.g., north-south-central countries -a geographical cluster).
Alternatively, you can try to use the vce(robust) option, which does not take into account clusters but it controls for heteroskedasticity among residuals.
The key point here, is to understand the behavior of the residuals' variance.

[proud pm]

Hello.
the option "vce(cluster panelid)" takes into account the within error correlation of the cluster comprehensive of the various ID (in your case). In general, if sigma_u is small (small variability in the residuals within groups), the option cluster is not feasible (it can be that this is your case (?)).
You may perhaps use the cluster option controlling for different clusters (e.g., north-south-central countries -a geographical cluster).
Alternatively, you can try to use the vce(robust) option, which does not take into account clusters but it controls for heteroskedasticity among residuals.
The key point here, is to understand the behavior of the residuals' variance.

Thank you so much for your help,

The sigma_u is actually small. I just tried vce(robust) option and it gave the exact same result with the cluster option.

[Jack Grover]

Did you go through the ANOVA test statistics? Is the F test statistic significant?

Alternatively, you may want to try different estimators, such as the newey-west two-step GMM estimator, which produces estimates robust to heteroskedasticity and autocorrelation (https://www.stata.com/manuals13/rgmm.pdf).

[Carlo Lazzaro]

Proud PM (just in case, please see: https://www.statalist.org/forums/help#realnames. Thanks):
- the options -robust- and -cluster- do the very same job under -xtreg- (and unlike -regress-): hence, no wonder you got identical results;
- you do not explain why you went -fe- vs -re-;
- if you actually detected heteroskedasticity and/or autocorrelation and you stick with default standard errors, the statistical significance (which, by the way, should not be the ultimate goal of your research) of your coefficients is unreliable.

[proud pm]

Did you go through the ANOVA test statistics? Is the F test statistic significant?

Alternatively, you may want to try different estimators, such as the newey-west two-step GMM estimator, which produces estimates robust to heteroskedasticity and autocorrelation (https://www.stata.com/manuals13/rgmm.pdf).

Yes, I did. The F test is significant.
Thank you so much, I will try different estimators.

[proud pm]

Proud PM (just in case, please see: https://www.statalist.org/forums/help#realnames. Thanks):
- the options -robust- and -cluster- do the very same job under -xtreg- (and unlike -regress-): hence, no wonder you got identical results;
- you do not explain why you went -fe- vs -re-;
- if you actually detected heteroskedasticity and/or autocorrelation and you stick with default standard errors, the statistical significance (which, by the way, should not be the ultimate goal of your research) of your coefficients is unreliable.

Thank you so much for your response,

First, I used pvar but the results is not significant. Then, I was suggested to go for fe/re.

[Carlo Lazzaro]

How did you select -fe- vs -re- specification?

[proud pm]

How did you select -fe- vs -re- specification?

I used Hausman test

[Carlo Lazzaro]

Did you invoke -cluster()- standard errors after .hausman- outcome?