Behavior in small samples of some tests of non-nested hypotheses in nonstationary regressions and their bootstrap versions (Q2488404)

Three tests are considered testing two non-nested linear regression models: \[ H_i:\;y_t=\beta'_i x_{i,t}+u_{i,t}, \] where \(u_{i,t}\) are stationary \(I(0)\), and \((x_{i,t}, y_t)\) are vector-valued with \(I(1)\) entries. These tests are the J-test of Davidson and McKinnon, the JA-test of Fisher and McAleer, and the bootstrap adjusted J-test. A construction of J-tests is discussed in cases the where regressors are (are not) cointegrated. Monte-Carlo simulations are made for small samples. The authors' conclusion is: ``regardless of the stationarity character of the regressors, if we try to select between non-nested models the best test seems to be the bootstrap-J''.