I am currently working with an unbalanced panel data set in order to analyse capital structure decisions and determinants. I am using STATA to conduct the analysis. So far I have done the following steps:

1. Fixed-Effect Regression (xtreg)

Nevertheless, the results were mostly insignificant despite tons of empirical evidence in literature and a large data set under analysis.

2. Testing for Heterosced. (xttest3) and Serial Correlation (within the panels) (xtserial)

Result: Heterosced. And Serial Correl.

3. Winsorization oft he data set

4. Fixed-Effect Regression (xtreg) with Clustered Std. Errors.

Improved results, more significant coefficients.

5. GLS-Panel Regression (xtgls) with Hetero and AR(1)

Improved results.

Now, I am not sure which of the models, 4 or 5 I should use.

How can one choose between these two models?

Further, according to Petersen (2009) one should include a Time-Dummy to account for cross-sectional (between panel, over time) correlations. However, this destroys the results. In addition, I am not sure if cross-sect. Correlation exists as I was not able to test for it due to a highly unbalanced sample.

Do I have to include the Time-Dummies? Is there another way to test for cross sect correl instead of XTTEST2 or XTCSD, Pesaran?

Input would be highly appreciated!