Panel cross-sectional dependence test: 'xtcd' vs 'xtcd2'

[Robert Akunga]

I'm very new to the stata forum and had an urgent query on my panel data. I am working with panel data with T>N, and when testing for cross section dependence within the variables using Pesaran (2004) CD test, 'xtcd' i get a p-value of 0.000, thus inferring cross-sectional dependence. However when I test using the Pesaran (2015) CD test for weak cross-sectional dependence, 'xtcd2 [variable], noestimation', i get a p-value of 0.000, thus inferring rejection of weak cross-sectional dependence. I am getting the opposite in both tests, kindly assist to shade light. Thanks.

[JanDitzen]

Hi Robert,
both the xtcd and xtcd2 have implicitly the same hypothesis and this of weak cross sectional dependence. the alternative for both tests is strong cross sectional dependence.

In the Pesaran (2004) paper the hypothesis is of cross-sectional independence. However (and please note that the 2004 paper is only a working paper and was never published anywhere) this restriction is very strong and unlikely to be met in large panels. Therefore the null hypothesis Pesaran (2015 - Econometric Reviews) is of weak dependence. So the difference between the two papers is that the hypothesis in the published paper is less restrictive.

Weak dependence means, that in the limit (i.e. N and T -> infinity) cross sectional dependence disappears and therefore the estimator is consistent. Strong dependence on the other hand poses real problems (omitted variable bias etc.). Thus in a finite panel it is sufficient to test weak vs. strong cross sectional dependence, knowing that weak dependence disappears if N and T go to infinity.

Therefore the p-value in your tests imply that you have strong cross sectional dependence.

[Robert Akunga]

Hi Jan,
Thanks. If that's the case, I go ahead and test for unit roots (Pesaran CADF) and find both I(0) and I(1) variables and decide to run an ARDL regression. I test for cointegration using Westurland (including a constant, trend and bootstrap and find no cointegration). Do you just run the pmg estimator afterward for both short run and long run or the cce estimator?