Which regression command is more appropriate for my dataset? -xtreg,re vce(robust)- or -xtgls,panels(heteroskedastic)

[Luke Silver]

Hi Statalisters,

I am having trouble determining which regression model to use to analyze my data.

I have an unbalanced panel dataset with n=53 and T~68. There are gaps within time variable (measured in days), and the panels don't necessarily share the same time periods (e.g. one panel could cover June 5 - September 5, while another covers October 6 - December 20).

I ran a Hausman test, which indicated that random effects (-xtreg,re-) is preferred over fixed effects (-xtreg,fe-) for this data.
I also ran -lrtest-, which indicated that there is heteroskedasticity between panels.

Based on this information, does anyone have a recommendation for the most appropriate Stata command to use to model this data? From prior reading and research, I believe that -xtreg, re vce(robust)- and -xtgls, panels(heteroskedastic)- would be viable options, but I would welcome any further insight into this matter.

Thanks in advance,
Luke

[Carlo Lazzaro]

Luke:
welcome to the list.
Given that T>N, I would go -xtgls-.

[Luke Silver]

Carlo Lazzaro ,

Thank you for your reply.

I understand that -xtgls- is the preferred model over -xtreg- given that T>N, as you mentioned. However, given the presence of heteroskedasticity and autocorrelation (determined by -xtserial-) I am trying to figure out whether -xtpcse- might be preferred over, say -xtgls,panels(h) corr(ar1) force- (I had to specify "force" because there are gaps within each of my panels).

For context, I just read Beck & Katz (1995) and have been trying to best apply the advice they offer, but would much appreciate your expert opinion.

My two questions are as follows:

1) What is the difference between specifying corr(ar1) vs. corr(psar1) in correcting for autocorrelation?
2) For this data, would you recommend -xtpcse, pairwise hetonly corr(ar1/psar1)- or -xtgls,panels(heteroskedastic) corr(ar1/psar1) force- ?

Thank you for your insight,

Luke

[Luke Silver]

Carlo,

My apologies for failing to follow advice from the FAQ. I have rewritten question 2) below:

2) For this data, would you recommend using:

xtpcse,parwise hetonly corr(ar1/psar1)

or

xtgls,panels(heteroskedastic) corr(ar1/psar1) force

Regards,

Luke

[Carlo Lazzaro]

Luke:
q1) see -help xtpcse- for this question: the main issue there boils down to using AR1 that is panel-specific structure or not;
q2) I woould go as per your first code (-xtpcse-).
That said, I would also consider comparing the outcomes of the two codes and investigate their differences.