Comparison of panel unit root tests under cross sectional dependence

For cross-sectionally dependent panels, various tests are proposed according to all possible combinations of ways of de-factoring (projection, subtraction), methods of combining results from different panel units (averaging, pooling), and estimation schemes for the unit root parameters (ordinary least squares and others). A Monte Carlo experiment is conducted to compare finite sample sizes and powers of the proposed tests. It shows that subtraction is a better policy than projection in terms of size performance; for projection, pooling yields very wild-sized tests; for projection, ordinary least squares (OLS) estimation yields lower-powered tests than other estimation methods; for subtraction, averaging is slightly better than pooling in terms of size. New tests based on subtraction, averaging and OLS estimation emerged as best test in terms of both size and power.