Bandwidth selection in nonparametric density estimation under dependence: a simulation study (Q1965935)

For kernel density estimators (with Gaussian and Epanechnikov kernels) there are considered methods of data driven bandwidth selection: cross-validation (pseudo-likelihood, least squares, biased and smoothed); plug-in method; non-smoothed and smoothed bootstrap; and double kernel method. The simulated data were drawn from an AR(1) model \(X_t=\rho X_{t-1}+ e_t\), where \(e_t\) are i.i.d. with normal, exponential, double-exponential and Rademacher distributions. A mixture of two normal AR(1) processes and \(X_t=2^{-1} \sum_{k=0}^{j-1} e_{t+k}^2\) are also considered. The mean integrated square and absolute errors (MISE, MIAE) and the mean uniform absolute error (MUAE) are used as measures of the estimators' efficiency. The authors concluded that the smoothed cross-validation and plunging methods present a generally good performance.