An Analysis of the Performance of Structural Change Tests in OLS Regressions
|Speaker:||Brendan McCabe, University of Liverpool|
|Date:||Friday 12 October 2012|
|Location:||Matrix Lecture Theatre, Building One|
In this paper, we analyse the relationship between the timing of structural breaks and the ability of tests to detect that timing. We consider tests for structural breaks in Gaussian OLS regressions in a multiple decision theory framework. The optimal tests derived are forms of cumulative sum procedures based on OLS residuals. They are derived from Bayes rules which maximise the probability of finding the correct change points, should they exist. The results are exact and valid in small samples. The standard OLS CUSUM procedures are found to be optimal for changes in the regrsssion coefficients but they have an implicit prior over the set of posible break points which can have a significant impact on the ability of the CUSUM to find the timing of the break. Approximate optimal tests are also derived which allows any prior probability to be combined with any desired pattern of structural breaks. Monte Carlo simulation experiments are used to look at the finite sample performances of the derived optimal tests in the mean shift, derterministic trend and stochastic trend regression models.