The panel dimensions are N=170 and T=21 and the estimated model is a dynamic panel estimated using xtlsdvc (corrected least squares dummy variables).

Some of the variables are I(0) or I(1) depending on the country. My guess is that the error term must be stationary for all the panels. The panel is highly heterogenous and N is 170 whereas the model specification must be the same for all the panels which makes it difficult to find a specification (in terms of differences or levels) that fits all the panels. With any specification I have tried there is always a certain proportion of countries for which the error terms are I(1) instead of I(0). Since these countries are problematic a solution would be to remove them but the estimator tends to be unbiased as N increases.

With this panel dimensions, is it possible to allow non-stationarity in the error term for until a certain percentage of the panels? Or the totality of them must be I(0)?

How relevant are stochastic trends in a panel with T<<N dimensions?

Thank you so much.