A quasi-Bayesian model selection approach is employed to detect the number and dates of structural changes in China's GDP and labour productivity data. It is shown that the predictive likelihood information criterion is valid only among models with well-behaved residuals.