Evaluation Of Causality Methods And Tests For Panel Data
Author: Abida Naurin

Causality is the most important concept which is tested frequently in social sciences. Unfortunately, it is not easily detected from observational studies. It is difficult for a researcher to differentiate between cause and effect, which is even more challenging in economics. Researchers have been applied various approaches to test causality, i.e., Regression discontinuity design (1960), Cross correlation-based methodology (1994), Difference in differences approach (2005), Granger causality test (1969), Error correction mechanism (1986), Toda and Yamamoto method (1995) for cross-sectional and time-series data. These usual regression methodologies rely on normality and linearity assumptions not supported by the data used in the analysis and may lead to unreliable results. Therefore, the policies built on poor tests for causality remain unreliable. Some causality techniques are sensitive to distributional assumptions and specification issues. In the literature, no study compares causality methods and tests for all types of data and panel data that can help choose an appropriate testing methodology. The current study evaluates the size and power based on Monte Carlo Simulations of various causality methods for panel data under whole alternative hypotheses for all possible causal combinations. This study also modifies the Sims test (1972) and Final Prediction Error (FPE) test (1981) for the Panel dataset. Comparison of Panel Causality Tests (PCT) has been used with different model specifications (stationary series with drift only, with drift and trend) for different sample sizes (small, medium, and large) under this study. Monte Carlo results reveal that the Granger Non-Causality (GC) test by Dumitrescu and Hurlin (2012) has the least size distortion from nominal compared to size distortion of the Sims test and FPE test for all sample sizes of cross-section units. However, the GC test’s power attainment is much better than the other two tests at all alternatives and all sample sizes. Among the Sims test and FPE tests, the former gains lower power at all alternatives than the latter one corresponding to small, medium, and large cross-section units and thus identified as the worst performer. A similar pattern has been observed for almost all tests at different sample sizes; medium sample size (i.e., T=50) and large sample size (i.e., T=200). Based on the comparison of size and power analysis of the PCT, this study concludes that the GC test is a point optimal and performs better at all causal combinations and panel dimensions, whether drift only or both drift and trend have been taken into account. On the other hand, the Sims test with its lowest power gain at all causal combinations and panel dimensions is the worst performer test. However, the FPE test having a power curve between the better and worst performer is graded as the average performer test. We investigate the government and household spending nexus on education in Pakistan for the applied application of our proposed PCT testing procedure. The result is fascinating and useful for policymakers. It indicates that the causality clearly runs from the intensity of government spending on education to the corresponding household intensity, but the effect is only direct. Supervisor:- Dr. Ahsan Ul Haq Co- Supervisor:- Dr. Asad Zaman

Meta Data

Keywords : Causality, Econometrics, Granger Non-Causality Test, Monte Carlo Simulations, Panel Data
Supervisor: Ahsan ul Haq Satti

Related Thesis​