Bootstrap Inferences For Complex Panel Data Models

In this paper,we consider the problem of two kinds of unbalanced panel data model,including one-way error component regression model and two-way error component regression model,we use the parameter bootstrap and the approximate test to test the regression coefficients.According to the assumed distribution of model,we construct the appropriate pivot variable,then we can generate generous subsamples by parameter bootstrap and compute confidence interval and p-value by Monte Carlo simulation.Finally,we can get the coverage probalility and the power,the results of the two methods are compared and analyzed.The simulation results of the one-way error component model show that the PB approaches has more accurate coverage probabilities and more stable performance,in the actual situation,We consider a model with an nonnormal and exponential correlation covariance structure,the PB approaches still have good performance while it's coverage probabilities is basically stable,but the coverage probabilities of approximate test is reduced.In the case of hypothesis test,the PB approaches can control the Type ? error better,while considering the complex model structure,the results of the PB approaches are still stable,but the Type ? error of approximate test increases sharply.The simulation test of the two-way error component model is similar to the one-way model,simulated results show that the PB approaches have higher coverage probabilities and lower Type ? error.Finally,we use a real data example to illustrate our approaches.In summary,the simulation results show that the parameter bootstrap approaches have better performance,and the derivation is easy to understand,but it has a longer length of confidence interval.Therefore,in consideration of the Type ? error,we recommend the parameter bootstrap approaches.