you have to options with the xtreg,, fe to account for serial correlation and hetroska., either to cluster your variables across your panel id, using cluster (panel id) or use robust option. Both way are fine and will get you consistent and efficient estimates. both xtregar and xtgls works only if you have auto correlation of order 1, have you tested this ?. I would suggest going with xtreg robust and clustered options as a start.

Sherif Maher Hassan Thank you very much! Yes i test for serial correlation and I have this problem in model. on the attached capture you can see that signification of coefficients are very low when I use cluster id SEs. However, signification of coefficients are high when I use -xtpcse id, c(psar1)- . Is this last model better? As I know I cann't use cluster id when I have cross-sectional dependence. is this true? thank in advance