Abstract

A test for panel structural mean change is developed from the CUSUM of the panel processes. Limiting null distribution and consistency of the test are established. The test is shown to have stable finite sample sizes than the existing test of Horvath and Huskova (2012) based on the squared CUSUM. If the mean changes are not cancelled in that their average is away from zero, the proposed test has better power than the existing test. On the other hand, if the mean changes are nearly cancelled, the existing test has better power. The proposed tests are illustrated by a real data set analysis.

1. Introduction

Structural change problems in panel data models are important issues for economic or financial data analysis because a big macroeconomic policy change or a financial crisis has simultaneous influence on many economic or financial variables. Researches for structural change problems in panel models have been activated due to a vast amount of data in the modern economic world or financial markets. Some studies were made by Bai (1997), Bai et al. (1998) and Han and Park (1989) for change point estimation and testing of multivariate time series models and by Emerson and Kao (2001, 2002) for testing of structural change of a time trend regression in panel data.

6. Conclusion

We have considered a panel mean change test based on the CUSUM of the panel processes. The test has better size performance than the existing test of Horvath and Huskova (2012). Compared with the existing test, the proposed test has better power against non-cancelling changes but has worse power against cancelling changes. Therefore, when we have prior information that the means are shifted to a common direction as in the world wide financial crisis, we may prefer the proposed test to the existing one. When we have the prior information of cancelling changes, the existing test should be preferred. When we do not have information on the direction of mean shifts, we need to consider both of the proposed test and the existing one because none of the two dominates the other one in power performance. It would be better if we have a theoretical explanation on the Monte-Carlo difference between the proposed test SupB and the existing test Sup H. Investigation of the theoretical point may be a good topic of future research.