Size And Power Comparison of Cointegration Techniques

Abstract

The ordinary least square (OLS) estimates are inconsistent for non-stationary time series data. However, the OLS estimates are super consistent if the time series variables are cointegrated. Therefore, cointegration has become an important tool for modelling the time series data. To test the cointegration between the time series variables, several cointegration tests are devised in literature. These tests may be classified into two groups: null hypothesis of “cointegration” (CIT) and null hypothesis of “no cointegration” (NCIT). Literature provides evidence for the comparison CIT tests however limited literature is available which studies the size and power properties of NCIT tests. Data is generated using Data generating process (DGP). This study compares single equation static and dynamic cointegration tests in terms of their size and power properties through Monte Carlo simulations which reveal the superiority of Non-linear ARDL to other tests in the study i.e., CRDW, ARDL, and Engle & Granger’s test (EG) for small and medium sample sizes for all the three cases of deterministic part however for large sample sizes with no intercept and trend, ARDL performs better and with trend EG test dominates others.