In the case of the exponential family of distributions it is well known that the use of a prior distribution belonging to the natural conjugate family substantially simplifies the analysis whilst often being realistic in applications. The present paper explores the related idea of a conjugate family of utilities, and generally the notion of choosing a utility structure that is suitably "matched" to the probability structure and, at the same time, realistic in application.

Ever since Stein's result, that the sample mean vector $$ of a $$ dimensional normal distribution is an inadmissible estimator of its expectation $$, statisticians have searched for uniformly better (minimax) estimators. Unbiased estimators are derived here of the risk of arbitrary orthogonally-invariant and scale-invariant estimators of $$ when the dispersion matrix $$ of $$ is unknown and must be estimated. Stein obtained this result earlier for known $$. Minimax conditions which are weaker than any yet published are derived by finding all estimators whose unbiased estimate of risk is bounded uniformly by $$, the risk of $$. One sequence of risk functions and risk estimates applies simultaneously to the various assumptions about $$, resulting in a unified theory for these situations.

The problem of estimating several normal mean vectors in an empirical Bayes situation is considered. In this case, it reduces to the problem of estimating the inverse of a covariance matrix in the standard multivariate normal situation using a particular loss function. Estimators which dominate any constant multiple of the inverse sample covariance matrix are presented. These estimators work by shrinking the sample eigenvalues toward a central value, in much the same way as the James-Stein estimator for a mean vector shrinks the maximum likelihood estimators toward a common value. These covariance estimators then lead to a class of multivariate estimators of the mean, each of which dominates the maximum likelihood estimator.

Let $$, have density $$ with respect to Lebesgue measure. It is desired to estimate $$ under the loss $$. Assuming the problem has a minimax risk $$, an estimator is defined to be tail minimax if its risk is no larger than $$ outside some compact set. Under quite general conditions on $$ and $$, sufficient conditions for an estimator to be tail minimax are given. A class of good tail minimax estimators is then developed and compared with the best invariant estimator.

Let $$ be a sample from an $$-variate distribution which is spherically symmetric up to an affine transformation. This paper deals with the robust estimation of the location vector $$ and scatter matrix $$ by means of "$$-estimators," defined as solutions of the system: $$ and $$, where $$. Existence and uniqueness of solutions of this system are proved under general assumptions about the functions $$ and $$. Then the estimators are shown to be consistent and asymptotically normal. The breakdown bound and the influence function are calculated, showing some weaknesses of the estimates for high dimensionality. An algorithm for the numerical calculation of the estimators is described. Finally, numerical values of asymptotic variances, and Monte Carlo small-sample results are exhibited.

Huber's theory of robust estimation of a location parameter is adapted to obtain estimators that are robust against a class of asymmetric departures from normality. Let $$ be a distribution function that is governed by the standard normal density on the set $$ and is otherwise arbitrary. Let $$ be a random sample from $$, where $$ is the unknown location parameter. If $$ is in a class of continuous skew-symmetric functions $$ which vanish outside a certain set $$, then the estimator $$, obtained by solving $$ by Newton's method with the sample median as starting value, is a consistent estimator of $$. Also $$ is asymptotically normal. For a model of symmetric contamination of the normal center of $$, an asymptotic minimax variance problem is solved for the optimal $$. The solution has the form $$ for $$ for $$, and $$ for $$. The results are extended to include an unknown scale parameter in the model.

Two algorithms, the $$-process and the $$-process, are presented which can be effective tools for determining the estimable contrasts in a classification model. Both algorithms operate on the incidence matrix of the model as opposed to the design matrix. If the model is partitioned as $$, the $$-process produces a spanning set for the estimable $$-contrasts (i.e., contrasts involving only $$ parameters) whenever the set of distinct $$ vectors is linearly independent. If the distinct $$ vectors are dependent, the $$-process is still useful and often simplifies the problem of obtaining a spanning set for the estimable $$-contrasts. After the $$-process has been applied in a case when the distinct $$ vectors are dependent, the $$-process produces a spanning set for the estimable $$-contrasts provided a partition $$, can be made such that the sets of distinct $$ vectors and distinct $$ vectors are both linearly independent. As examples, the $$-process is used to investigate the additive two-way model; and the $$-process and $$-process together are used to investigate an additive three-way model, a two-way model with interaction, and a Graeco-Latin square model.

Asymptotic expansions are established for the power of distribution free tests in the one-sample problem. These expansions are then used to obtain deficiencies in the sense of Hodges and Lehmann (1970) for distribution free tests with respect to their parametric competitors and for the estimators of location associated with these tests.

Without assumptions on the underlying distributions we prove the asymptotic normality of averaged scores rank statistics under all product distributions which are contiguous to the null hypothesis, and find a very simple form of the centering constants. In the one-sided two-sample and trend situations this enables us to show that monotonicity of the scores generating function is equivalent to the asymptotic unbiasedness of the corresponding averaged scores rank test. For such asymptotically unbiased tests we prove simple necessary and sufficient conditions for having bounds for their asymptotic relative efficiency under all contiguous alternatives of the model. As a by-product, we get the local asymptotic superiority of averaged scores rank tests to the associated tests with randomized ranks not only for shift but for general alternatives. In addition we prove that every one-sided averaged scores rank test is asymptotically most powerful (asymptotically equivalent to likelihood ratio test) for a suitable nonparametric subclass of alternatives, provided the test and the associated subclass of alternatives are generated by a nondecreasing, square-integrable function defined on the unit interval.

As a "robust" alternative to the least squares estimates for a regression parameter, Koul (1969) proposed new estimates based on signed-rank statistics. To find out their asymptotic distribution Koul proved that under quite general assumptions, the signed-rank statistics of Wilcoxon type are asymptotically linear in the sense that they are uniformly approximable by linear forms in the regression parameter. More general results have been obtained by Van Eeden (1972) in a paper which is an analog to Jureckova's paper (1969) dealing with linear rank statistics. In 1972 the author proved that the statistic used to define the Hodges-Lehmann estimate for a location parameter is asymptotically linear in a stronger sense, the result being to Koul's theorem what the central limit theorem is to the weak law of large numbers. For the general linear regression model with one parameter the signed-rank statistics are proved to be linear in a strong sense, that is, the differences between the statistics and the linear forms mentioned above, properly normalized, converge weakly to linear processes. Results in this direction for linear rank statistics have been obtained by Jureckova (1973). As an application of the theorems presented here, one can construct new estimates for the squared $$-norm of the underlying density, and this in much the same way as in Antille (1974). It is also possible to get more information about the asymptotic behavior of the linearized versions, proposed by Kraft and Van Eeden (1972).

The subject of this research is the maximum probability assignable to closed subintervals of a closed, bounded, nondegenerate interval by distributions on that interval whose first three moments are specified. This maximum probability is explicitely displayed as a function of both the moments and the subintervals. The ready application of these results is illustrated by numerical examples.

Kendall's rank correlation statistic $$ sgn $$ is well known to be asymptotically normally distributed under the null hypothesis of independence as the sample size $$. In the note it is shown that this assertion can be obtained easily from the recurrence formula $$ for the probability distribution $$ of $$ (see Kendall (1970), e.g.). This recurrence formula implies that $$ has the same distribution as a sum of $$ well defined independent random variables to which the Lyapunov criterion applies.

Let $$ and $$ be two unbounded random variables. Then two necessary conditions are proved concerning the structure of the bivariate distribution function of $$ and $$ when it is expanded in the orthonormal polynomials of its marginal distributions. The first condition concerns the shrinking of the polynomial representation into a diagonal form, and the second is a generalization of the Sarmanov-Bratoeva theorem.

The problem of estimating the mean of a $$-variate $$ normal distribution is considered. It is assumed that the covariance matrix $$ is known and that the loss function is quadratic. A class of minimax estimators is given, out of which admissible minimax estimators are developed.

Arnold (1973) studied testing problems where the covariance matrix is assumed to have the generalized correlation structure under both hypotheses. That paper showed how to transform such problems to "products" of unpatterned problems. This paper extends those results to testing problems where the covariance matrix is assumed to have Geisser's (1963) generalization of the pattern of compound symmetry. We prove theorems indicating how to transform such problems to products of unpatterned problems. These results are then applied to three problems: 1. a general problem where both the mean vectors and covariance matrix are patterned (this problem is general enough to include both the multivariate analysis of variance (MANOVA) and classification problems.); 2. the MANOVA problem when only the covariance matrix is patterned; 3. a problem arising only when the covariance matrix is patterned. In this paper we only show how to transform such problems to products of unpatterned problems that have been studied, since in Arnold (1973) it was showing how to convert results about known problems to results about their product.

Let $$ be a continuous mean regression function defined on $$, the unit cube in $$-dimensional Euclidean space. Let $$ be a distribution function with support in $$, and let $$ denote the indefinite integral of $$ with respect to $$. This paper provides consistency results, including rates of convergence, for a certain estimator of $$ in the case that the $$th estimate is based on observations at points $$ of $$. The estimator is the $$-dimensional analogue of that considered by Brunk (1970).