TESTING FOR SERIAL-CORRELATION BY VARIABLE ADDITION IN DYNAMIC-MODELS ESTIMATED BY INSTRUMENTAL VARIABLES

Instrumental variable tests for serial correlation can be carried out by adding lagged residuals from initial estimation to the regressors of the model under scrutiny, and then checking their joint significance. It is shown that asymptotically valid tests are obtained if the lagged residuals are also added to the initial instrument set. Monte Carlo evidence suggests that useful improvements in finite sample behaviour under null and alternative hypotheses can be produced when the instrument set is extended to include the relevant lagged residuals. Links with other tests are discussed and a modification allowing for conditional heteroskedasticity is described.