Confusion about cross sectional dependence

[Alberto Martinelli]

Hi, I have a lot of confusion about cross sectional dependence in panel data, particularly I don't understand what is done to account for this problem when we have N>T. In the case of heteroskedasticity and serial correlation what is usually done is to cluster standard errors, however, from what I have learned, this doesn't solve the issue of cross sectional dependence. I have heard that to account for this problem some people use Driscoll-Kraay standard errors with the command xtscc, but I have seen in other posts that this can be done only when the T dimension of the data is sufficiently large (I don't have clear what is 'sufficiently large' either, if you could explain it would be great). Moreover, some people say that cross sectional dependence is a problem only in some specific cases and most of the times it is possible to ignore it, I am really confused... if someone could help me in this regard I would be so grateful. By the way sorry if I posted this question in another section of the forum initially, it is my first time using it.

[Carlo Lazzaro]

Alberto:
while -xtscc- is the way to go in your case, caution should be paid when dealing with N>T panel dataset (as highlighted in -xtscc- help file reported below):

Description

xtscc produces Driscoll and Kraay (1998) standard errors for coefficients estimated by pooled OLS/WLS or fixed-effects (within) regression. depvar is the dependent variable and varlist is an optional list of explanatory variables.

The error structure is assumed to be heteroskedastic, autocorrelated up to some lag and possibly correlated between the groups (panels). These standard errors are robust to general forms of cross-sectional (spatial) and temporal dependence when thetime dimension becomes large. Because this nonparametric technique of estimating standard errors places no restrictions on the limiting behavior of the number of panels, the size of the cross-sectional dimension in finite samples does not constitute
a constraint on feasibility -- even if the number of panels is much larger than T. Nevertheless, because the estimator is based on an asymptotic theory, one should be somewhat cautious with applying this estimator to panels that contain a largecross-section but only a short time dimension.

That said, in your cse I would compare the DK standard errors with their default counterparts and inspect the difference.

[Alberto Martinelli]

Thanks Carlo, I am a bit hesitant about using xtscc because T=20 and I feel like it might be too small. Using year fixed effects can somehow attenuate the cross sectional dependence problem in your opinion?

[Carlo Lazzaro]

Aberto:
20 years is the T of the -xtscc example (see related helpfile).

[Alberto Martinelli]

Ok, thanks again Carlo.