Robust asymptotic statistics

Publication date 1 August 1994

Published in Springer Series in Statistics (Search for Journal in Brave)

log likelihoods optimal influence curves optimally robust functionals robust regression stable constructions von Mises functionals

Mathematics Subject Classification ID

Introductory exposition (textbooks, tutorial papers, etc.) pertaining to statistics (62-01) Research exposition (monographs, survey articles) pertaining to statistics (62-02) Robustness and adaptive procedures (parametric inference) (62F35) Asymptotic properties of nonparametric inference (62G20) Nonparametric robustness (62G35)

Review text

This book is about optimally robust functionals and their unbiased estimators and tests. Functionals extend the parameter of the assumed ideal center model to neighborhoods of this model that contain the actual distribution. The two principal questions are (F): Which functional to choose? and (P): Which statistical procedure to use for the selected functional? Using a local asymptotic framework, we deal with both problems by linking up nonparametric statistical optimality with infinitesimal robustness criteria. Thus, seemingly separate developments in robust statistics are presented in a unifying way. Instead of principles, we prefer objective criteria that are applicable to a wide class of functionals. Moreover, we look more closely into the nonparametric optimality theorems, whose proofs completely rely on the parametric results applied to one least favorable local alternative. It is true that these abstract theorems are geared to, and are to some extent justified by, the utmost generality. As a side condition on mathematical abstraction, however, elementary procedures should still be dealt with according to original intentions. But for the class of asymptotically linear estimators, it turns out that the asymptotic statistical optimality results, which -- in the parametric case -- generalize the Cramér-Rao bound for asymptotic variance, are -- in the nonparametric setup with full neighborhoods -- dominated by bias alone: To be able to trace the proofs of the nonparametric optimality theorems, and because a sufficiently lucid and elementary reference on asymptotic statistics (in particular, on the asymptotic minimax theorem) has been lacking, it proved necessary to develop a self-contained treatment of the main statistical optimality results. Our presentation is fairly simple (assuming only weak convergence), so that students with a basic knowledge of probability and statistics can understand the proofs. As such, the corresponding Chapters 2, 3, 4, and 6 (supported by Appendix A) should be interesting also outside robust statistics. Chapter 1, a nonstatistical extra, merely supplies some asymptotically normal procedures via differentiation. In the sequel, the nonparametric optimality results are used in just two ways: First, as unbiasedness statements, and secondly, to quantify the costs at which any functional can be estimated or tested. On this basis, we then determine that functional which, subject to certain robustness constraints, is easiest to estimate or test. As robustness side conditions, bounds are imposed on the infinitesimal oscillation (estimation), respectively, the inclusion of given capacity balls into nonparametric hypotheses is required (testing). It is at this instance where the classical robustness theories on bias and variance may be invoked to complement nonparametric unbiasedness by some further, nontrivial, optimality. Optimality results for general parameters are contained in Chapter 5. Chapter 7 derives those for linear regression, which serves as a prototype example of a structured model, where additional variants of neighborhoods and optimality problems arise. The conditional centering constraint occurring in this context takes values in some infinite-dimensional function space, and makes it necessary to explicitly derive a suitable Lagrange multiplier theorem [Appendix B]. We must admit that the monolithic type of optimality known from parametric statistics cannot be achieved in robust statistics. The solutions depend on the choice of neighborhoods, and one hardly knows precisely what sort of neighborhood one would `like to be robust over'. But the variety of robust solutions at least constitute a complete class. The book has no pretensions of treating all developments in robustness, nonparametrics, or asymptotic statistics. We only consider the local, infinitesimal, approach (that is, no qualitative robustness, no breakdown aspects), and only the case of independent observations. The restricted selection and comprehensive treatment of basic issues, however, defines the framework for further refinements, extensions, and specializations.

Recommendations

Cited in

(60)

This page was built for publication: Robust asymptotic statistics

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q1329887)