Panel data with small sample and autocorrelation

[Rafael Leao]

Hi all. How are you? This is my second post here.

I'm trying to run a panel regression with a small sample, with N=22 (countries) and T=18 (years).

. xtset id year
       panel variable:  id (strongly balanced)
        time variable:  year, 2001 to 2018
                delta:  1 unit

In my data autocorrelation seems to be an issue.

Wooldridge test for autocorrelation in panel data
H0: no first-order autocorrelation
    F(  1,      21) =      5.272
           Prob > F =      0.0321

In such a case, as I'm not able to increase N or T, which regression should I be concerned to solve issues with such a sample?

Reading this forum I found that -xtscc- with FE and Lag would help, but I'm not sure, since T doesn't seem to be big enough. If -xtscc- is good enough in this case, is it ok to use differences in variables too?
Or should I stick to -xtreg- FE robust to solve it? In any chance is it possible to go for a -xtabond2- controlling for the proliferation of instruments?

If there is any other option to model it I would appreciate it very much to be aware of. Also, if I need to provide more info about my data, just let me know.

Thanks in advance you all.

Best regards.

[Carlo Lazzaro]

Leorari:
welcome to this forum.
I would go -xtreg, fe robust- or -xtreg, fe vce(cluster panelid)-.
They both call cluster-robust standard errors and. as such, do the very same job (a word of caution: the same is not true for -regress-).

[Rafael Leao]

Dear Carlo, thank you for your answer.

I understand that robust and -xtreg, fe robust- or vce(cluster) should address the issue of autocorrel? Right? Or do I have to include lagged dependent variable?

Also, may I use variables in difference if needed?

And any kind of postestimation tests should be highlighted concerning a small sample?

Thank you.
Best regards!

[Carlo Lazzaro]

Rafael:
1) -xtreg, fe robust- or -xtreg, fe vce(cluster panelid) deal with heteroskedasticity and/or autocorrelation. No need to go -xtabond- (provided you want to remain within the boundaries of the static panel data regression).
2) -D.- independednt variables are allowed;
3) the usual test: are those for heteroskedasticity, autocorrelation (that you can skip if you invoke non-default standard errors) and misspecification of the functional form of the regressand.

What above holds for -xtreg,re-, too.