Examples are given to illustrate the "linear" and "general" empirical Bayes approaches to estimation. A final example concerns testing the null hypothesis that a treatment has had no effect.

A unified treatment is given of the classical additive models for complete factorial experiments and of multiplicative models and Lancaster-additive models for multi-dimensional contingency tables. The models are characterised by properties of being simplest subject to having a prescribed set of marginals. It is shown that, by using averaging operators and the notion of a generalised interaction, the interaction properties of these models can be derived very simply.

The asymptotic behaviour of symmetric statistics of arbitrary order is studied. As an application we describe all limit distributions of square integrable $U$-statistics. We use as a tool a randomization of the sample size. A sample of Poisson size $N_\lambda$ with $EN_\lambda = \lambda$ can be interpreted as a Poisson point process with intensity $\lambda$, and randomized symmetric statistics are its functionals. As $\lambda \rightarrow \infty$, the probability distribution of these functionals tend to the distribution of multiple Wiener integrals. This can be considered as a stronger form of the following well-known fact: properly normalized, a Poisson point process with intensity $\lambda$ approaches a Gaussian random measure, as $\lambda \rightarrow \infty$.

Let $W$ be a $(k + r)$-dimensional random (column) vector with distribution $F^W_\xi$ belonging to a $(k + r)$-parameter exponential family, $\Pi = \{F^W_\xi: \xi \in \Omega \subset R^{k+r}\}$. Let $\mu$ be a $\sigma$-finite measure which dominates $\Pi$ such that for all $\xi \in \Omega, F^W_\xi$ has a density, with respect to $\mu$, of the form $f^W(w:\xi) = h(w) \exp\{\xi'w + c(\xi)\}$. Consider the partitions $W' = (U', T')$ and $\xi' = (\theta', \nu')$, where $U, \theta \in R^k$ and $T, \nu \in R^r$. It is proven that the conditional covariance matrix of $U$ given $T = t$ does not depend on $t$ for almost all values of $t$ if and only if there exists a unique measurable vector-valued function $g(T) = (g_1(T), \cdots, g_k(T))'$, such that the random vector $Z = U - g(T)$ is stochastically independent of $T$ under any member of $\Pi$. Furthermore, the distribution of $Z$ is shown to constitute a $k$-parameter exponential family with $\theta$ as the vector of natural parameters. Further results are obtained and exemplified.

Suppose an exponential model $\mathscr{M}$ is partitioned into submodels, all of the same parametric dimension. If each of these models corresponds to a linear hypothesis about the canonical parameter and also to a linear hypothesis about the mean parameter, we speak of the partitioning of $\mathscr{M}$ as an affine dual foliation. We study those cases where the parameter sets defining the hypotheses are parallel either in the canonical space or in the mean space, and obtain various characterisations and properties of these cases. It is shown, inter alia, that canonical parallelism and mean parallelism are related to likelihood independence of $\theta_1$ and $\tau_2$ (and hence to $S$-ancillarity and $S$-sufficiency), respectively to stochastic independence of $\hat{\theta}_1$ and $\hat{\tau}_2$. Here $(\theta_1, \tau_2)$ denotes a mixed parametrisation of $\mathscr{M}$ and $\hat{\theta}_1$ and $\hat{\tau}_2$ are the maximum likelihood estimators of $\theta_1$ and $\tau_2$. Also, the two types of parallelism are characterised in terms of observed and expected information. Mean parallelism is closely related to the concept of reproductivity of exponential models that forms the subject of a separate paper. A number of requisite general results for exponential families are established, and these are also of some independent interest.

Consider a full and steep exponential model $\mathscr{M}$ with model function $a(\theta)b(x)\exp\{\theta \cdot t(x)\}$ and a sample $x_1, \cdots, x_n$ from $\mathscr{M}$. Let $\bar{t} = \{t(x_1) + \cdots + t(x_n)\}/n$ and let $\bar{t} = (\bar{t}_1, \bar{t}_2)$ be a partition of the canonical statistic $\bar{t}$. We say that $\mathscr{M}$ is reproductive in $t_2$ if there exists a function $H$ independent of $n$ such that for every $n$ the marginal model for $\bar{t}_2$ is exponential with $n\theta$ as canonical parameter and $(H(\bar{t}_2), \bar{t}_2)$ as canonical statistic. Furthermore we call $\mathscr{M}$ strongly reproductive if these marginal models are all contained in that for $n = 1$. Conditions for these properties to hold are discussed. Reproductive exponential models are shown to allow of a decomposition theorem analogous to the standard decomposition theorem for $\chi^2$-distributed quadratic forms in normal variates. A number of new exponential models are adduced that illustrate the concepts and also seem of some independent interest. In particular, a combination of the inverse Gaussian distributions and the Gaussian distributions is discussed in detail.

Geometric analysis of the mixture likelihood set for univariate exponential family densities yields results which tie the number and location of support points for the nonparametric maximum likelihood estimator of the mixing distribution to sign changes in certain integrated polynomials. One corollary is a very general uniqueness theorem for the estimator.

This paper presents a sufficient condition for second order efficiency of an estimator. The condition is easily checked in the case of minimum contrast estimators. The $\alpha^\ast$-minimum contrast estimator is defined and proved to be second order efficient for every $\alpha, 0 < \alpha < 1$. The Fisher scoring method is also considered in the light of second order efficiency. It is shown that a contrast function is associated with the second order tensor and the affine connection. This fact leads us to prove the above assertions in the differential geometric framework due to Amari.

The theory and application of the Cox (1972) failure time regression model has, almost without exception, assumed an exponential form for the dependence of the hazard function on regression variables. Other regression forms may be more natural or descriptive in some applications. For example, a linear relative risk regression model provides a convenient framework for studying epidemiologic risk factor interactions. This note uses the counting process formulation of Andersen and Gill (1982) to develop asymptotic distribution theory for a class of intensity function regression models in which the usual exponential regression form is relaxed to an arbitrary non-negative twice differentiable form. Some stability and regularity conditions, beyond those of Andersen and Gill, are required to show the consistency of the observed information matrix, which in general need not be positive semidefinite.

Let $X$ have a $p$-variate normal distribution with unknown mean $\theta$ and identity covariance matrix. The following transformed version of a control problem (Zaman, 1981) is considered: estimate $\theta$ by $d$ subject to incurring a loss $L(d, \theta) = (\theta^t d - 1)^2$. The comparison of decision rules in terms of expected loss is reduced to the study of differential inequalities. Results establishing the minimaxity of a large class of estimators are obtained. Special attention is given to the proposition of admissible, generalized Bayes rules which dominate the uniform prior, generalized Bayes controller when $p \geq 5$.

Given $X_1, X_2, \cdots$, i.i.d. with mean $\mu$ and variance $\sigma^2$, suppose that at stage $n$ one wishes to estimate $\mu$ by the sample mean $\bar{X}_n$, subject to the loss function $L_n = A(\bar{X}_n - \mu)^2 + n, A > 0$. If $\sigma$ is known, the optimal fixed sample size $n_0 \approx A^{1/2}\sigma$ can be used, with corresponding risk $R_{n_0}$, but if $\sigma$ is unknown there is no fixed sample size procedure that will achieve the risk $R_{n_0}$. For the sequential estimation procedure with stopping rule $T = \inf\{n \geq n_A: n^{-1} \sum^n_1 (X_i - \bar{X}_n)^2 \leq A^{-1} n^2\}$, the second order approximation of Woodroofe (1977) to the risk $R_T$ for normal $X_i$ is extended to the distribution-free case. Specifically, if the $X_i$ have finite moments of order greater than eight and are non-lattice, under certain conditions on the delay $n_A$ it is shown that the regret $R_T - R_{n_0} = c + o(1)$ as $A \rightarrow \infty$, where $c$ depends on the first four moments of the distribution of the $X_i$. For the lattice case, bounds of the form $c_1 + o(1) \leq R_T - R_{n_0} \leq c_2 + o(1)$ are obtained, where the $c_j$ are $c \pm 3$. It follows from these approximations that the regret can take arbitrarily large negative values as the distribution of the $X_i$ varies, in contrast to previous results for normal and gamma cases.

The concepts of canonical correlations and canonical components are familiar ideas in multivariate statistics. In this paper we extend these notions to stationary time series with a view to determining the most predictable aspect of the future of a time series. We relate properties of the canonical description of a time series to well known structural properties of the series such as (i) rational spectra (i.e., ARMA series), (ii) strong mixing, (iii) absolute regularity, etc.

This paper continues an investigation into the canonical correlations and canonical components of the past and future of a stationary Gaussian time series which were introduced in Jewell and Bloomfield (1983). Bounds for the maximum canonical correlation are provided under specified conditions on the spectrum of the series. A computational scheme is described for estimating the canonical correlations and components and the procedure is illustrated on the well-known sunspot number series.

A unified treatment of the consistency properties of the ordinary least squares estimates in an autoregressive fitting of time series from nonstationary or stationary autoregressive moving average models is given. For a given model, the orders of autoregressions which produce consistent estimates are obtained and the limiting values, hence the biases, of the estimates of other autoregressions are investigated.

A stepwise testing procedure using Lagrangian multiplier tests is developed to determine the order of an ARMA-process. The strong consistency of this procedure under slightly weaker assumptions than in Hannan (1980) (proof of the consistency of the order estimators obtained via BIC) is proved.

A Bayesian treatment of the quantal bioassay design problem is given. It is assumed that the potency curve is a Dirichlet random distribution $F$ with parameter $\alpha(t) = MF_0(t)$, and that $n_1, \cdots, n_L$ animals are treated at drug levels $t_1, \cdots, t_L$ respectively. The optimal design levels $t_1, \cdots, t_L$ that minimize the Bayes risk for weighted integrated quadratic loss functions are found in the following cases: (i) $L = 1$ and the weight function arbitrary; (ii) uniform prior guess, uniform weight and two animals treated; and (iii) uniform weight and $L$ arbitrary, but $M \rightarrow 0$. These results disprove a conjecture of Antoniak.

Let $f$ be a density on $R^d$, and let $f_n$ be the kernel estimate of $f$, $f_n(x) = (nh^d)^{-1} \sum^n_{i=1} K((x - X_i)/h)$ where $h = h_n$ is a sequence of positive numbers, and $K$ is an absolutely integrable function with $\int K(x) dx = 1$. Let $J_n = \int |f_n(x) - f(x)| dx$. We show that when $\lim_nh = 0$ and $\lim_nnh^d = \infty$, then for every $\varepsilon > 0$ there exist constants $r, n_0 > 0$ such that $P(J_n \geq \varepsilon) \leq \exp(-rn), n \geq n_0$. Also, when $J_n \rightarrow 0$ in probability as $n \rightarrow \infty$ and $K$ is a density, then $\lim_nh = 0$ and $\lim_nnh^d = \infty$.

The coverage of a multinomial random sample is the sum of the probabilities of the observed classes. A normal limit law is rigorously proved for Good's (1953) coverage estimator. The result is valid under very general conditions and all terms except the coverage itself are observable. Nevertheless the implied confidence intervals are not much wider than those developed under restrictive assumptions such as in the classical occupancy problem. The asymptotic variance is somewhat unexpected. The proof utilizes a method of Holst (1979).

We derive the joint asymptotic distribution of the $k$ midranges formed by averaging the $i$th smallest normalized order statistic with the $i$th largest normalized order statistic, $i = 1, \cdots, k$. We then derive the distribution of the maximum midrange among these $k$ extreme midranges and the limiting distribution of this maximum as $k \rightarrow \infty$. These results imply that, even in infinite samples, different distributions in the class of symmetric, unimodal distributions with tails that die at least as fast as a double exponential distribution may have different maximum likelihood estimates for the location parameter. We also discuss the application of these results to a test of symmetry suggested by Wilk and Gnanadesikan (1968).

This paper discusses a class of minimum distance Cramer-Von Mises type estimators of the slope parameter in a linear regression model. These estimators are obtained by minimizing an integral of squared difference between weighted empiricals of the residuals and their expectations with respect to a large class of integrating measures. The estimator corresponding to the weights proportional to the design variable is shown to be asymptotically efficient within the class at a given error distribution. The paper also discusses the asymptotic null distribution of a class of minimum Cramer-Von Mises type goodness-of-fit test statistics.

It is well known that the Kolmogorov-Smirnov (K-S) test exhibits poor sensitivity to deviations from the hypothesized distribution that occur in the tails. A modified version of the K-S test is introduced that is more sensitive than the K-S test to deviations in the tails. The finite and infinite sample distribution along with the consistency properties of the proposed test are studied. Tables of critical values are provided for two versions of the test (one sensitive to heavy tail alternatives and one sensitive to light tail alternatives) and the finite sample properties of these two versions of the test are investigated.

In this paper the multivariate normal linear patterned mean and covariance matrix testing and estimation problem is studied in the presence of missing data for general one-population hypotheses. The Newton-Raphson, Method of Scoring and EM algorithms are given for finding the maximum likelihood estimates. The asymptotic joint distribution of the maximum likelihood estimates under null and alternative hypotheses are derived along with the form of the likelihood ratio statistic and its asymptotically Chi squared null and asymptotically normal nonnull distributions. The distributions of the maximum likelihood estimates and nonnull distributions of the likelihood ratio tests are derived using the standard multivariate and univariate delta method respectively, and may be evaluated at a parameter point under the alternative hypothesis parameter space or at a parameter point in a parameter space that contains the null and alternative hypothesis parameter spaces. New results for these problems in the presence of complete data as well as known results (Szatrowski, 1979) are special cases of the results of this paper.

The main purpose of this paper is the development of an asymptotic theory of systematic sampling from a stochastic population. The superpopulation model assumed is that the population arises from a second-order stationary process. A comparison among the multiple random start systematic sampling schemes is made in terms of the limiting expected variance of the sample mean. Asymptotic normality of the systematic sampling mean is obtained, both unconditionally and conditionally on the given population. Finally, the asymptotic behavior of confidence intervals based on two distinct variance estimators is studied.

This paper deals with techniques for finding $\Phi$-optimal designs for weighing $v$ objects in $b$ weighings using a spring balance. The optimality functions considered encompass a large class of functions. Results are applied to find $A$-, $D$- and $E$-optimal designs and the optimal designs obtained are seen to be related to certain types of well known block designs.

The affine type of distributions on the real line are represented as sequences of distributions of maximal invariants on spheres. It is shown that such a representation characterizes the affine type. A consistency condition is introduced, and it is shown that any sequence of maximal invariant distributions satisfying the condition is generated by some affine type on $\mathbf{R}$.

In some problems of directional data, finite mixtures of simple distributions have been proposed as statistical models. In this paper we show that for a wide class of directional distributions, all such mixtures are identifiable.

By convolution smoothing of the empirical hazards, a kernel estimate of the hazard function from censored data is obtained. Small and large sample expressions for the mean and the variance of the estimator are given. Conditions for asymptotic normality are investigated using the Hajek projection method.

In a recent paper (Tanner and Wong, 1983b), a family of data-based nonparametric hazard estimators was introduced. Several of these estimators were studied in an extensive simulation experiment. The estimator which allows for variable bandwidth was found to have a superior performance. In this note, sufficient conditions for the variable kernel estimator to be strongly consistent are presented.

We study the asymptotic accuracy of the bootstrap approximation to the distribution of a $k$-sample studentized mean.

We introduce and describe orthogonal series methods for estimating the density of qualitative, quantitative or mixed data. The techniques are completely nonparametric in character, and so may be used in situations where parametric models are difficult to construct. Just this situation arises in the context of mixed--both qualitative and quantitative--data, where there are few parametric models.

Let $p_\nu(t, \delta)$ be the density at $t$ of a noncentral $t$-variable with $\nu$ degrees of freedom and noncentrality parameter $\delta$. It is proved that for any $d > 0$ and fixed $t, p_\nu(t, \delta + d)/p_\nu(t, \delta)$ is a strictly decreasing function of $\delta$.

Suppose $x(t)$ is a vector stochastic process generated by a first order differential equation and $f(t)$ is a linear combination of the elements of $x(t)$. Functionals of $x(t)$ are observed with noise. We obtain the smoothness properties of the best linear unbiased estimate of $f(t)$, and those of its derivatives that exist. In addition we obtain the smoothness properties of their mean squared errors.