Bootstrap Tests for Structural Breaks When the Regressors and Error Term are Nonstationary

This paper considers tests for structural breaks in linear models when the regressors and the serially dependent error process are unstable. The set of models contains various economic circumstances such as the structural breaks in the regressors and/or the error variance, and a linear trend model with I(0)/I(1) error. We show that the existing heteroscedasticity robust tests and the fixed regressor bootstrap method of Hansen (2000) have severe size distortion problem even in the asymptotics. We suggest a method which combines the fixed regressor bootstrap and the sieve-wild bootstrap method to nonparametrically approximate the serially dependent unstable error process. The suggested method is shown to asymptotically replicates the true distribution of the existing tests under various circumstances. Monte Carlo experiments show significant improvements both in the size and the power properties. Once the size is controlled by the bootstrap, Wald type tests have better power properties relative to LM type tests.