This article is a survey of Jack Kiefer's published works in areas other than design of experiments. The approximately 50 articles can be divided into categories by subject matter. These categories will be discussed individually in the following review. At times Jack worked in several areas simultaneously and ideas from one area sometimes influenced work in another. Some of these influences are noted in the following survey, as the categories are discussed in roughly the order Jack began work in them.

A theory is developed following work by Williams (1952) and Kiefer (1960) for exact treatment designs in one dimension in which the errors are a stationary process. It is shown that the designs which achieve the minimax value of any of a wide class of functionals on the information matrix for estimation of treatment differences have a special property. If the process is autoregressive of order $p$ then a random piece of the design of length $p + 1$ exhibits uncorrelated treatment values. Such designs can be formed using full length cyclic error-correcting codes of a suitable order. A new technique is developed for classifying the ergodic combinatorial structure of exact designs of arbitrary or infinite length. It is shown that all designs are, to $p$th order, generated by a finite number of sequences with finite length. The classification is given explicitly up to order 3. The method is used to find asymptotically optimum designs for different processes. It is also shown that the designs can be achieved to within an arbitrarily good approximation as the realization of an ergodic Markov chain of sufficiently high order.

Consider a sample from a normal population with mean, $\mu$, and variance unknown. Suppose it is desired to test $H_0:\mu \leq \mu_0$ versus $H_1:\mu \geq \mu_1$, with the region $H^I_1:\mu_0 < \mu < \mu_1$ being a (nonempty) indifference zone. It is shown that the usual Student's $t$-test is inadmissible for this problem. An alternative test is proposed. The two sided problem with indifference region is also discussed. By contrast with the above result, the usual Student's $t$-test is admissible here. However the two sided version of the alternative test mentioned above does offer some practical advantages relative to the two sided $t$-test. A 3-decision version of the two sided problem is also discussed. Here the $t$-test is inadmissible, and is dominated by the appropriate version of the alternative test. The results concerning tests are also reformulated as results about confidence procedures.

This paper is about the asymptotic distribution of linear combinations of stratum means in stratified sampling, with and without replacement. Both the number of strata and their size is arbitrary. Lindeberg conditions are shown to guarantee asymptotic normality and consistency of variance estimators. The same conditions also guarantee the validity of the bootstrap approximation for the distribution of the $t$-statistic. Via a bound on the Mallows distance, situations will be identified in which the bootstrap approximation works even though the normal approximation fails. Without proper scaling, the naive bootstrap fails.

We derive invariance principles for processes associated with symmetric statistics of arbitrary order. Using a Poisson sample size, such processes can be viewed as functionals of a Poisson Point Process. Properly normalized, these functionals converge in distribution to functionals of a Gaussian random measure associated with the distribution of the observations. We thus obtain a natural description of the limiting process in terms of multiple Wiener integrals. The results are used to derive asymptotic expansions of processes arising from arbitrary square integrable $U$-statistics.

The errors-in-variables model in which the unobserved true values satisfy multiple linear restrictions is considered. Under the assumptions that the unobservable true values are normally distributed and that an estimator of the covariance matrix of the measurement error is available, the maximum likelihood estimators are derived. The limiting properties of the estimators are obtained for a wide range of assumptions, including the assumption of fixed true values.

In the sequential design of experiments an experimenter performs experiments sequentially to make an eventual inference about the true state of nature. A Bayesian formulation of this problem is considered. The parameter space is assumed finite and there are a finite number of repeatable experiments. Sufficient conditions are given for a procedure to be second order efficient as the sampling costs approach zero. The asymptotic analysis of a related Markov control problem is also presented.

For certain point and interval estimation problems, sequential procedures are proposed for choosing an estimator (from the class of trimmed means with trimming proportion in a specified range) as well as a sample size on which to base the estimator. It is shown that these procedures, which do not require knowledge of the best trimming proportion or the asymptotic variance of the corresponding trimmed mean, are asymptotically efficient with respect to the procedure that uses the best trimmed mean (in the specified class) and the best fixed sample size for that trimmed mean.

Under extremely general patterns of patient-arrival, allocation to treatment and loss to follow-up in (randomized) clinical trial settings, the sequentially computed logrank statistic (Mantel, 1966) is shown (under the null hypothesis of identically distributed lifetimes) to have exactly uncorrelated increments, and is shown via Rebolledo's (1980) martingale invariance principle to satisfy a functional central limit theorem, justifying sequential logrank tests of Jones and Whitehead (1979). Generalizations are made to other two-sample rank tests for censored survival data, and practical applicability to real randomized clinical trials is discussed.

Two samples, $\{X_{ji}, 1 \leq i \leq n(j)\} (j = 1, 2)$ are assumed to be composed of iid random variables with survival functions $(1 - F_j)(1 - H_j)$, where $H$ is the cdf of the "censoring times" and $F$ is the cdf of the "true lifetimes." A unified derivation of the Pitman efficiencies of a class of rank statistics for censored samples is presented. The conditions under which the result holds do not require contiguous alternatives, since convergence to normality is shown to hold uniformly in equicontinuous $(F_1, F_2, H_1, H_2)$ with bounded hazard rates. The uniformity is obtained by studying a convenient joint representation of several counting processes. The results are applied to the translated exponential distributions, a noncontiguous family of alternatives.

Buckley and James (1979) proposed an estimation method for linear regression models with unspecified residual distribution and right-censored response variables. In this paper we consider weak consistency of the regression parameter estimators under regularity conditions which avoid restrictions on the censoring patterns.

Consider a linear model with infinitely many parameters given by $y = \sum^\infty_{i = 1} x_i\theta_i + \varepsilon$ where $\mathbf{x} = (x_1, x_2, \cdots)'$ and $\theta = (\theta_1, \theta_2, \cdots)'$ are infinite dimensional vectors such that $\sum^\infty_{i = 1}x^2_i < \infty$ and $\sum^\infty_{i = 1} \theta^2_i < \infty$. Suppose independent observations $y_1, y_2, \cdots, y_n$ are observed at levels $\mathbf{x}_1, \mathbf{x}_2, \cdots, \mathbf{x}_n$. Under suitable conditions about the error distribution, the set of all bounded linear functionals $T(\theta)$ for which there exists an estimate $\hat{T}_n$ such that $\hat{T}_n \rightarrow T(\theta)$ in probability will be characterized. An application will be extended to the nonparametric regression problem where the response curve $f$ is smooth on the interval [0, 1] in the sense that $f$ has an $(m - 1)$th derivative that is absolutely continuous and $\int^1_0 f^{(m)}(t)^2 dt < \infty$.

We study a wide class of nonparametric regression function estimators including kernel estimators and robust smoothers. Under different assumptions on the kernel and the sequence of bandwidths, we obtain weak uniform consistency rates on a bounded interval. The uniform consistency is shown in a "stochastic design model" and in a "fixed design model".

We study the estimation of a regression function by two classes of estimators, the Nadaraya-Watson Kernel type estimators and the orthogonal polynomial estimators. We obtain sharp pointwise rates of strong consistency by establishing laws of the iterated logarithm for the two classes of estimators. These results parallel those of Hall (1981) on density estimation and extend those of Noda (1976) on strong consistency of kernel regression estimators.

The problem of nonparametric Bayes estimation of a tolerance distribution based on quantal response data has been considered previously with a prior distribution based on the Dirichlet process. In the present article, a broad class of priors is developed for this problem by allowing the hazard function of the tolerance distribution to be a realization of a nonnegative stochastic process with independent increments. This class includes the Dirichlet prior as a special case. In addition, priors over a space of absolutely continuous tolerance distributions, which includes IFR, DFR, and U-shaped failure rate distributions, are constructed by taking the failure rate to be the superposition of two processes with independent increments. Posterior Laplace transforms of the corresponding processes are obtained based on quantal response data with binomial sampling. These posterior Laplace transforms are then used to find Bayes estimates, and examples are given to illustrated the results.

The nonstationary signal extraction problem is to estimate $s_t$ given observations on $z_t = s_t + n_t$ (signal plus noise) when either $s_t$ or $n_t$ or both is nonstationary. Homogeneous or explosive nonstationary time series described by models of the form $\delta(B)z_t = w_t$ where $\delta(B)$ has zeroes on or inside the unit circle and $w_t$ is stationary are considered. For certain cases, approximate solutions to the nonstationary signal extraction problem have been given by Hannan (1967), Sobel (1967), and Cleveland and Tiao (1976). The paper gives exact solutions in the forms of expressions for $E(s_t\mid\{z_t\})$ and $\operatorname{Var}(s_t\mid\{z_t\})$ (assuming normality) under two sets of alternative assumptions regarding the generation of $z_t, s_t$, and $n_t$. Extensions to signal extraction with a finite number of observations, to the nonGaussian case, and to the multivariate case are discussed.

We show that the Sinusoidal Limit Theorem of Slutsky in the Gaussian case is a consequence of the equality of certain higher order crossings.

We consider the problem of the estimation of the total of a stratified finite population. For two levels of prior knowledge about the stratification, we provide Bayes and pseudo-Bayes estimators that make use of this prior knowledge in sensible ways. We then note that admissibility results can be established for these estimators using the techniques of Meeden and Ghosh (1982, 1983) and indicate some possible natural extensions of the present work.

It is shown that large classes of asymptotic relative efficiencies (AREs) are isotonic with respect to various partial orderings on the heaviness of tails of symmetric distributions. The orderings include those of van Zwet (1964), Lawrence (1975), Barlow and Proschan (1975), and a new one that generalizes all three. Characterizations in terms of these orderings are given for many familiar families of distributions with restricted tail and central behavior. By restricting attention to such distributions, finite bounds are obtained for AREs such as that of some robust estimates to the sample mean, which could be unbounded otherwise. Similar results are shown to hold for the approximate Bahadur efficiencies of Kolmogorov-type tests.

The notion of a "simultaneous confidence bound" is redefined by requiring a bound on the expected converge measure (ECM) instead of the coverage probability. This is analogous to a criterion introduced by Spjotvoll for defining simultaneous tests of hypotheses. Bounds which minimize certain width functionals, subject to a bound on the ECM, are characterized. For bounds on a multilinear regression function over an arbitrary subset of Euclidean space, the bounds which minimize weighted average width, among all bounds with prescribed ECM, are expressed in closed form. As a special case, we give a weight function relative to which Scheffe-type bounds are optimal.

A robustified version of the two sample sign test is defined which is insensitive to certain deviations from the assumption of the independence of the observations. These deviations are described in terms of mixing conditions. The asymptotic value of the power function of this robustified sign test is computed on contiguous alternatives possessing the same dependence structure. This entails the calculation of its asymptotic relative efficiencies with respect to some tests which are optimal on these alternatives in the independent case. It becomes apparent that in general the relative performance of two tests heavily depends on the structure of dependence of the observations, i.e. it may either increase or decrease.

The usual approach to estimating regression parameters in the Cox regression model uses the partial likelihood. If the covariates are not time-dependent, the model can be stated in terms of the survival function, which allows one to derive a generalized likelihood containing both regression and survival curve parameters. It is shown that, in the absence of ties, an estimator results which is asymptotically equivalent to the partial likelihood estimator. A joint information matrix leads simply to standard errors for both regression and survival curve parameters which are asymptotically correct.

We consider the problem of determining absolute continuity, and the distribution of the likelihood ratio (Radon-Nikodym derivative) of the measures induced on function space by two infinitely divisible stochastic processes. The results are applied to linear processes, which are shown to be infinitely divisible.

This paper considers estimation and prediction problems invariant under an amenable group that is transitive on the parameter space. It is shown that an invariant confidence or prediction region does not admit super-relevant betting procedures if and only if its level of coverage is defined conditionally given the maximal invariant statistic.

This note is concerned with how to replace assessment of a "true" prior on a nonparametric family of distributions--which is usually infeasible--by assessment of an approximating prior with support in a parametrized subfamily, in such a way that the posterior derived from the parametric model is close to the "true" posterior. In general it is not sufficient that the approximating prior be close to the true prior in the sense of weak convergence, and we characterize the additional aspect of the true prior that must be considered explicitly.

In normal linear regression analysis, many model selection rules proposed from various viewpoints are available. For the information criteria AIC, FPE, $C_p$, PSS and BIC, the asymptotic distribution of the selected model and the asymptotic quadratic risk based on each criterion are explicitly obtained.

Several criteria concerning the choice of kernels for the nonparametric estimation of functions and their derivatives are discussed. A specific optimality criterion is described which applies to kernels of compact support and of different orders of smoothness. The solutions of the corresponding variational problems are explicitly given. Many kernels discussed previously are obtained as special cases.

The question is discussed when of two ordered sets of continuous distribution functions one shows less positive trend than the other. The resulting ordering has applications to power function monotonicity of isotonic rank tests, and is related to certain orderings in parametric models.

Amemiya (1980) derived expressions for the $n^{-2}$-order mean squared errors of the maximum likelihood and the minimum logit chi-squared estimators in the dichotomous logit regression model. He numerically evaluated these expressions for a number of specific examples and in all examples found that the minimum chi-squared estimator has smaller $n^{-2}$-order mean squared error. In this paper, we demonstrate examples in which the maximum likelihood estimator has smaller $n^{-2}$-order mean squared error.