This paper investigates by means of Monte Carlo techniques the robustness of the CUSUM and CUSUM-of-squares tests (Brown et al., 1975) to serial correlation, endogeneity and lack of structural invariance. Our findings suggest that these tests perform better in the context of a dynamic model of the ADL type, which is not affected by serial correlation or nonpredetermined regressors even if over-specified. In this case, the empirical sizes of both tests are close to the nominal ones, whether a stationary or a cointegration environment is considered. The CUSUM-of-squares test is to be preferred, as it is very powerful to detect changes in the conditional model parameters, whether or not the variance of the regression error is included in the set of parameters shifting, especially towards the end of the sample.
CUSUM and CUSUM-of-squares tests parameter instability structural invariance marginal and conditional processes ADL model