st: Choosing standard error form to correct for serial correlation in panel data

[Nishant Tharani <nishant.tharani@gmail.com>]

Dear Statalist-ers,


Thank you for your consideration of my question. My question is about the
appropriate choice of standard errors to use in a panel data context to
correct for serial correlation. I'm performing a fixed effects regression
(xtreg, fe) with T=14 and N=13, and the dataset is strongly balanced. I've
detected serial correlation using the 'xtserial' test.

My choice at the moment is between clustering and using Newey-West standard
errors, both of which correct for correlation within a panel over time if I
understand correctly. From what I've gathered, clustering doesn't require
the specification of any autocorrelation structure in particular, and is
ideal for panel data in which N tends to infinity (N > T). Newey-West
errors, on the other hand, require one to specify the number of maximum
lags of autocorrelation in the model, and are used more for T > N. I'm
unsure as to which is the appropriate choice given that T and N are roughly
equal in my model. I'm tempted to use the Newey-West standard errors as my
N doesn't seem large enough for clustering to be consistent - would this be
correct? If so, is there a standard number of lags that would be
appropriate for annual data?

Another option is to report results with both - I've tested my model with
both types and the results are qualitatively the same.

Many thanks,
Nishant Tharani

(Using Stata 11.1 on Windows / OSX)
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/