Is xtreg, fe vce(cluster id) sufficient for heteroskedasticity, serial correlation, and cross-sectional dependence? (T=17, N=60)

[Jelly Dang]

I am working with an unbalanced panel dataset of 60 developing countries (id) observed over 17 years.
My model is:
xtreg sdi fin fin_sqr urpop old_age political_violence log_cpi log_trade i.year, fe vce(cluster id) (fin_sqr is the quandaric term of fin)
I want to ensure that my inference is valid in the presence of heteroskedasticity, serial correlation within panels, and potential cross-sectional dependence.

I have run:

• Pesaran CD test → cross-sectional dependence
• xtserial → serial correlation
• xttest3 → heteroskedasticity

Given this structure (N=60, T=17), my question is:

Is it appropriate to rely on xtreg …, fe vce(cluster id) to handle heteroskedasticity, serial correlation, and some degree of cross-sectional dependence?

Or should I prefer:

• xtscc (Driscoll–Kraay),
• or reghdfe with HAC,
• or a CCE estimator (e.g., xtcce)?
• or any other models?

Thank you.

[Carlo Lazzaro]

Jelly:
welcome to this forum.
I would go -xtscc-.

[Jeff Wooldridge]

T = 17 is probably too short to rely on xtscc in all but very simple models. vce(cluster id) does not handle cross-sectional dependence, but you have to have a good reason to allow for such dependence. The tests are very sensitive to, say, unmodeled time-varying coefficients. You will often conclude you have cross-sectional correlation when you don't. If the goal is to model spillover effects then you should do that using spatial methods. But if the variables are defined at the country level, I would just use vce(cluster id).