We discuss frequency properties of Bayes rules, paying special attention to consistency. Some new and fairly natural counterexamples are given, involving nonparametric estimates of location. Even the Dirichlet prior can lead to inconsistent estimates if used too aggressively. Finally, we discuss reasons for Bayesians to be interested in frequency properties of Bayes rules. As a part of the discussion we give a subjective equivalent to consistency and compute the derivative of the map taking priors to posteriors.

In some relatively natural settings, Bayes estimates of location are shown to be inconsistent.

A general asymptotic theory is developed for the maximum likelihood estimator based on a partial likelihood. Conditions are given for consistency and asymptotic normality, and a method is provided for the calculation of the asymptotic efficiency of the estimator. The implications of the general theory are examined in special cases such as inference in stochastic processes, Cox regression models, and AR processes with missing segments.

For an arbitrary one parameter exponential family density it is shown how to construct a mixing distribution (prior) on the parameter in such a way that the resulting mixture distribution is a two (or more) parameter exponential family. Reweighted infinitely divisible distributions are shown to be the parametric mixing distributions for which this occurs. As an illustration conditions are given under which a parametric mixture of negative exponentials is in the exponential family. Properties of the posterior are given, including linearity of the posterior mean in the natural parameter. For the discrete case a class of simply-computed yet fully-efficient least-squares estimators is given. A Poisson example is used to demonstrate the strengths and weaknesses of the approach.

Gaussian Markov distributions are characterised by zeros in the inverse of their covariance matrix and we describe the conditional independencies which follow from a given pattern of zeros. Describing Gaussian distributions with given marginals and solving the likelihood equations with covariance selection models both lead to a problem for which we present two cyclic algorithms. The first generalises a published algorithm for covariance selection whilst the second is analogous to the iterative proportional scaling of contingency tables. A convergence proof is given for these algorithms and this uses the notion of $I$-divergence.

Tests for a null hypothesis whose specification involves an unknown nuisance parameter may be obtained by inverting a bootstrap confidence region for the parameter being tested or by constructing a simulated null distribution for the test statistic. The power of either test against certain alternatives involving the same unknown nuisance parameter can itself be estimated by simulation.

This paper examines the asymptotic properties of compromise estimators. By this we mean an estimation method which compromises between a finite number of sampling situations in a small sample optimal way. We develop the asymptotic theory of such estimators in the location problem and show that under a specific choice of a pair of sampling situations the compromise estimator is asymptotically robust in Huber's sense.

For the canonical problem of estimating a multivariate normal mean under squared-error-loss, this article addresses the problem of selecting a minimax shrinkage estimator when vague or conflicting prior information suggests that more than one estimator from a broad class might be effective. For this situation a new class of alternative estimators, called multiple shrinkage estimators, is proposed. These estimators use the data to emulate the behavior and risk properties of the most effective estimator under consideration. Unbiased estimates of risk and sufficient conditions for minimaxity are provided. Bayesian motivations link this construction to posterior means of mixture priors. To illustrate the theory, minimax multiple shrinkage Stein estimators are constructed which can adaptively shrink the data towards any number of points or subspaces.

A new class of solutions to a general differential inequality often encountered in multiparameter estimation problems is obtained. Using these solutions as guidelines, improved estimators for the scale parameters as well as the natural parameters of independent gamma distributions are obtained for a large class of weighted quadratic losses. The improved estimators have an empirical Bayes interpretation. They also permit an exact analytical representation of the risk improvement. For the ordinary squared-error loss, a larger class of improved estimates is obtained which may allow for incorporation of prior information in choosing an alternative estimate. Numerical results are given which indicate the extent of risk improvement in certain situations.

The estimation problem of the quantiles $\xi + b\sigma$ of an exponential distribution with unknown location-scale parameter $(\xi, \sigma)$ is considered. We establish the admissibility of the traditional (best equivariant) estimator for quadratic loss when $n^{-1} \leq b \leq 1 + n^{-1}$ where $n$ is the sample size. For $b > 1 + n^{-1}$ a class of minimax procedures is found. This class contains generalized Bayes rules and one of them is shown to be admissible within the class of scale-equivariant procedures.

This article presents a family of estimators of the survival function based on right-censored observations which admit the possibility that the censoring variables may not be independent of the true failure variables. This family is obtained by generalizing the self-consistent property (Efron, 1967) of the product limit estimator (Kaplan and Meier, 1958). By assuming a Dirichlet process prior distribution of the observable random vectors, nonparametric Bayesian estimators of the survival curve-which is also a member of this family--are derived under a special loss function. These nonparametric Bayesian estimators generalize results of Susarla and Van Ryzin (1976), who impose a Dirichlet process prior on the failure survival function without considering any prior distribution of the censoring variables. Large sample properties of this family of nonparametric Bayesian estimators are also derived.

The paper discusses statistics that can be used to test whether two failure times, say $X_1$ and $X_2$, are independent. The two variables are subject to right censoring so that what is observed is $Y_i = \min(X_i, Z_i)$ and $\delta_i = I(X_i = Y_i)$, where $(Z_1, Z_2)$ are censoring times independent of $(X_1, X_2)$. Statistics that generalize the Spearman rank correlation and the log-rank correlation are considered, as well as general linear rank statistics. The Chernoff-Savage approach is adopted to show that suitably standardized versions of these statistics are asymptotically normal under both fixed and converging alternatives.

In the course of conducting a clinical trial to compare $K (\geq 2)$ treatments, it is often desirable to balance the trial with respect to the assignments of patients to treatments. On the other hand, some form of randomization of treatment assignments is essential for reducing experimental bias. In this article, the large-sample approximation to the null distribution of $K$-sample randomization tests generated from a broad class of restricted randomization rules is derived. The implication of this result for conditional inference is also discussed.

This paper shows that for sequential sampling without recall from a Dirichlet process, there exists a finite bound beyond which one will cease sampling with probability 1. The result is valid for Dirichlet processes defined on closed bounded intervals of the real line.

Two treatments that yield Bernoulli outcomes are available in a clinical trial. One success probability is known. A probability distribution reflects opinion about the other success rate. $N$ patients are to be treated, with $N$ possibly unknown, in a multistage trial. The goal is to maximize the total number of successes on the $N$ patients. Optimal lengths for each stage and optimal treatment allocations are found for two-stage trials with $N$ known. When $N$ is unknown the problem is shown to be equivalent to that of discounting future observations. Optimal stage lengths and treatment allocations are characterized for distributions on $N$ that yield regular discount sequences. This class of distributions includes the geometric family, which is given special consideration. It is shown that if the number of stages in the trial is fixed and if the distribution on $N$ yields a regular discount sequence, then it is optimal to use the known treatment in the last stage only. This extends the work of Berry and Fristedt (1979).

In a recent article Bechhofer and Kulkarni proposed a class of closed adaptive sequential procedures for selecting that one of $k \geq 2$ Bernoulli populations with the largest single-trial success probability. These sequential procedures which take no more than $n$ observations from any one of the $k$ populations achieve the same probability of a correct selection as does a single-stage procedure which takes exactly $n$ observations from every one of the $k$ populations. In addition, they often require substantially less than a total of $kn$ observations to terminate sampling. Amongst other problems, Bechhofer and Kulkarni considered the problem of devising a procedure within this class which minimizes the expected total number of observations to terminate sampling. For their proposed procedure they cited several optimality properties for the case $k = 2$ and conjectured additional optimality properties for the case $k \geq 3$. In this article we use a new method of proof to establish stronger results than those cited by Bechhofer and Kulkarni for the case $k = 2$, and prove stronger results than those conjectured for $k \geq 3$. We also describe a new procedure for $k \geq 3$ and prove that it minimizes the expected total number of observations to terminate sampling when all of the success probabilities are small.

The autoregressive orders selected by the criterion autoregressive transfer function (CAT) of Parzen (1974), a new version, CAT$^\ast$, of CAT introduced by Parzen (1977), and the CAT$_2$ criterion of Bhansali (1985) are shown to be asymptotically efficient in the sense defined by Shibata (1980, 1981). A generalization of the penalty function considered by Shibata (1980) is introduced. The order selected by the CAT$_\alpha$ criterion of Bhansali (1985), with any fixed $\alpha > 1$, is asymptotically efficient with respect to this generalized penalty function.

The size of error is investigated when the log-likelihood of multivariate generalized linear models is approximated by a quadratic function. The nonquadratic tail is characterized by analyzing the cubic part of the log-likelihood. In a local analysis simple bounds for that part can be expressed in terms of expectations of the related random variables for arbitrary sample size $N$. Additionally global error bounds are given for the univariate case.

Since their introduction by Dwass (1957) and Barnard (1963), Monte Carlo tests have attracted considerable attention. The aim of this paper is to give a unified approach that covers the case of an arbitrary null distribution in order to study the statistical properties of Monte Carlo tests under the null hypothesis and under the alternative. For finite samples we obtain bounds for the power of the Monte Carlo test with the original test that allow determination of the required simulation effort. Furthermore the concept of asymptotic (resp. local asymptotic) relative Pitman efficiency (ARPE, resp. LARPE) is adapted to Monte Carlo tests for the study of their asymptotic behaviour. The normal limit case is investigated in more detail, leading to explicit formulas for ARPE and LARPE.

We consider the following problem arising in robust estimation theory: Find the maximum asymptotic variance of a trimmed mean used to estimate an unknown location parameter when the error distribution is subject to asymmetric contamination. The model for the error distribution is $F = (1 - \varepsilon)F_0 + \varepsilon G$, where $F_0$ is a known distribution symmetric about $0, \varepsilon$ is fixed proportion of contamination, and $G$ is an unknown and possibly asymmetric distribution. We prove, under the assumption that $F_0$ has a symmetric unimodal density function $f_0$, that the maximal asymptotic variance is obtained when $G$ places mass 1 at either $+\infty$ or $-\infty$. The key idea of the proof is first to maximize the asymptotic variance subject to the side conditions $F(a) = \alpha$ and $F(b) = 1 - \alpha$ when $a$ and $b$ are given.

Gong and Samaniego (1981) define pseudo maximum likelihood estimation and derive the asymptotic distribution of the resulting estimates. This note gives a simpler and more elegant expression for the asymptotic variance of a pseudo maximum likelihood estimate.