When does bootstrap work! Asymptotic results and simulations (Q1189002): Difference between revisions

A large number of papers have appeared in the last decade on the bootstrap procedure. This monograph is designed to study and to clarify in a unified approach the conditions under which the bootstrap yields a satisfactory approximation of the unknown distribution of a statistic. After a short abstract description of the bootstrap procedure the author investigates the application of this resampling method in three important contexts: estimation of smooth functionals, nonparametric curve estimation, and linear models. A central point of the discussion is the question of equivalence of the consistency of bootstrap and the asymptotic normality of the considered functional. Whereas in the case of linear functionals this equivalence holds and therefore also the validity of bootstrap is proved, it is not possible to formulate a general statement for arbitrary functionals. This fact is demonstrated in examples concerning the application of bootstrap in nonparametric curve estimation. So it is shown that the distribution of the \(L^ 2\)-distance between a nonparametric kernel regression estimator and a parametric regression estimator is not estimated consistently by bootstrap. This statistic is proposed as a goodness-of-fit test statistic of a parametric regression model, and for an approximate determination of critical values an alternative resampling procedure, the so-called wild bootstrap, is presented. For the comparison of different data sets on the basis of nonparametric density estimators bootstrap tests and bootstrap confidence regions are constructed by a suitable choice of the statistical functional. Furthermore, using results about the number of modes of a kernel density estimate, the author considers bootstrap tests on the number of modes of a density. The equivalence of asymptotic normality and consistency of bootstrap is connected with the more general question whether bootstrap should be applied in models where classical asymptotic distribution approximations are available. To answer this question the author proves for arbitrary smooth functionals the higher-order accuracy of bootstrap estimates. Moreover, for the linear model with nonrandom design and i.i.d. error variables it is shown that bootstrapping the least squares estimates and the \(M\)-estimates works under essentially weaker conditions than those which are necessary for the classical normal approximation. Whereas this example shows that bootstrap can be applied in cases where the normal approximation breaks down, it is stated in a further chapter that in the more complex high dimensional linear random design model bootstrap works not always as satisfactorily as in the case of a nonrandom design. The book gives a good insight into basic concepts of bootstrap methods and inspires the reader to a further study of this modern field of statistical research. It is recommended to researchers in mathematical statistics.