Bootstrapping error component model (Q1861604)

The one-way error component model \(y_{it}=\alpha+x_{it}\beta+\mu_i+\varepsilon_{it}\), \(i=1,\dots,n\), \(t=1,\dots,T\), is considered, where \(y_{it}\) and \(x_{it}\) are observable variables, \(\mu_i\sim(0,\sigma_\mu^2)\), \(\varepsilon_{it}\sim(0,\sigma_\varepsilon^2)\), \(E\varepsilon_{it},\varepsilon_{js}=0\), \(E\mu_i\mu_j=0\) and \(E\mu_i\varepsilon_{jt}=0\). Generalized least squares estimators for the parameters are described. The authors discuss different bootstrap-type procedures for tests of hypotheses on the variance ratios \(\sigma^2_\mu/(\sigma_\mu^2+\sigma_\varepsilon^2)\). These procedures are: direct nonparametric resampling, nonparametric resampling of the error components, parametric resampling of the error components, and nonparametric resampling of the orthogonalized residuals. The sizes and power of the obtained tests are compared via simulations.