An Estimation Method for Panel Data Model with Heteroscedasticity, Serial and Spatial Correlations

Edokpayi A. A, Abiodun A. A

Panel Data, heteroscedasticty, serial correlation, spatial Dependence, kernel Estimators

The kernel based nonparametric HAC estimation methods have been suggested as alternative to PCSE for panel data with heteroscedasticity, serial and spatial correlations, But the two commonly used kernel functions –the truncated and Bartlett functions for kernel based HAC estimations are too restrictive and that they yield negative bias and that such bias could be substantial in finite samples In this study, the error structure of the PCSE was modified with the introduction of two new kernel functions -.the Parzen kernel and Turkey-Hannings kernel functions and a non-linear weight was defined for them. Using simulated data for varied levels of heteroscedasticity, serial and spatial correlations, varying spatial weight matrix specification and different cross-sectional and time dimensions, the performances of these two new kernel functions were compared with Bartlett kernel ,the truncated kernel functions and the PCSE’ Using absolute bias (AB), residual variance (RVar) and the root mean squares error (RMSE) as assessment criteria, the performances of these estimators were determined The results from the study showed that the kernel based nonparametric approach performed better than the PCSE and that the Bartlett kernel, the Turkey-Hanning kernel and the Parzen performed better in the presence of heteroscedasticity, serial and spatial correlations . However, the Tukey-Hannings kernel was generally more preferred for small sample sizes, narrow spatial weight matrix specifications and for short panels N>T), while the Parzen kernel estimator performed better for long panels (N<T) and wider spatial weight matrix specifications .The Bartlett kernel functions however, performed better than the Turkey-Hanning and Parzen kernels generally for large sample sizes, wider spatial weight matrix specifications. From the results the study concludes that the performances of the different estimators were generally influenced by the type of panel data, the size of the cross-sectional and time dimensions and the spatial weight matrix specifications.

https://dx.doi.org/10.31873/IJEAS.11.05.04