We consider the problem of testing a hypothesis about the means of a subset of the components of a multivariate normal distribution with unknown covariance matrix, when the means of a second subset (the covariates) are known. Because of the possible correlation between the two subsets, information provided by the second subset can be useful for inferences about the means of the first subset. In this paper attention is restricted to the class of procedures invariant under the largest group of linear transformations which leaves the problem invariant. The family of tests which are admissible within this class is characterized. This family excludes several well-known tests, thereby proving them to be inadmissible (among all tests), while the admissibility (among invariant tests) of other tests is demonstrated. The powers of the likelihood ratio test LRT, the $D^2_{p+q} - D^2_p$ test, and the overall $T^2$ test are compared numerically; the LRT is deemed preferable on the basis of power and simplicity.

A multivariate version of the Hoerl-Kennard ridge regression rule is introduced. The choice from among a large class of possible generalizations is guided by Bayesian considerations; the result is implicitly in the work of Lindley and Smith although not actually derived there. The proposed rule, in a variety of equivalent forms is discussed and the choice of its ridge matrix considered. As well, adaptive multivariate ridge rules and closely related empirical Bayes procedures are presented, these being for the most part formal extensions of certain univariate rules. Included is the Efron-Morris multivariate version of the James-Stein estimator. By means of an appropriate generalization of a result of Morris (see Thisted) the mean square error of these adaptive and empirical Bayes rules are compared.

We consider the nonparametric pairwise comparisons procedures derived from the Kruskal-Wallis $k$-sample test and from Friedman's test. For large samples the $(k - 1)$-mean significance level is determined, i.e., the probability of concluding incorrectly that some of the first $k - 1$ samples are unequal. We show that in general this probability may be larger than the simultaneous significance level $\alpha$. Even when the $k$th sample is a shift of the other $k - 1$ samples, it may exceed $\alpha$, if the distributions are very skew. Here skewness is defined with Van Zwet's $c$-ordering of distribution functions.

The deficiency of sample quantiles with respect to quasiquantiles is investigated under the assumption that the true density function has bounded derivatives. Then the sample quantile is still an efficient estimator of the true quantile but the relative deficiency of sample quantiles with respect to suitably defined quasiquantiles quickly tends to infinity for increasing sample sizes. If the second derivative of the true density function is bounded, then adaptive estimators will be found which are of a better performance than quasiquantiles. Corresponding results are derived for two-sided confidence intervals which are based on quasiquantiles and adaptive estimators.

In this paper, an asymptotic minimax theory for robust estimation of a one-dimensional parameter is derived, which is an asymptotic counterpart, and generalization to an arbitrary parameter, of Huber's finite sample minimax theory for the location case. A particular variability measure and results from robust asymptotic testing are employed. The results show a relationship of this approach to Hampel's local theory of robustness.

Let $Z_1, Z_2, \cdots$ be i.i.d. standard normal variables. Results are obtained which relate to the tail behavior as $x \rightarrow \infty$ of distributions of the form $F(x) = P\{\sum^\infty_{k=1}\lambda_k\lbrack(Z_k + a_k)^2 - 1\rbrack \leqslant x\}$. For test statistics which have such limiting distributions $F$, asymptotic relative efficiency measures are discussed. One of these is the limiting approximate Bahadur efficiency. Applications are to tests of fit and tests of symmetry.

Consistency and asymptotic normality of least-squares estimators are discussed for the linear autoregressive model with explanatory variables. Few assumptions are made about the error sequence. The case of stochastic explanatory variables is also considered.

Let $\{x_t\}$ be a linear stationary process of the form $x_t + \Sigma_{1\leqslant i<\infty}a_ix_{t-i} = e_t$, where $\{e_t\}$ is a sequence of i.i.d. normal random variables with mean 0 and variance $\sigma^2$. Given observations $x_1, \cdots, x_n$, least squares estimates $\hat{a}(k)$ of $a' = (a_1, a_2, \cdots)$, and $\hat{\sigma}^2_k$ of $\sigma^2$ are obtained if the $k$th order autoregressive model is assumed. By using $\hat{a}(k)$, we can also estimate coefficients of the best predictor based on $k$ successive realizations. An asymptotic lower bound is obtained for the mean squared error of the estimated predictor when $k$ is selected from the data. If $k$ is selected so as to minimize $S_n(k) = (n + 2k)\hat{\sigma}^2_k$, then the bound is attained in the limit. The key assumption is that the order of the autoregression of $\{x_t\}$ is infinite.

Sequences of $m$th order symmetric statistics are examined for convergence in law. Under appropriate conditions, a limiting distribution exists and is equivalent to that of a linear combination of products of Hermite polynomials of independent $N(0, 1)$ random variables. Connections with the work of von Mises, Hoeffding, and Filippova are noted.

Conditional on the value of $\theta, \theta > 0$, let $X(t), t \geqslant 0$, be a homogeneous Poisson process. Sequential estimation procedures of the form $(\sigma, \theta(\sigma))$ are considered. To measure loss due to estimation, a family of functions, indexed by $p$, is used: $L(\theta, \hat{\theta}) = \theta^{-p}(\theta - \hat{\theta})^2$, and the cost of sampling involves cost per arrival and cost per unit time. The notion of "monotone case" for total cost functions of a continuous time process is defined in terms of the characteristic operator of the process at the total cost function. The Bayes sequential procedure is then derived for those cost functions in the monotone case with optimality proven using extensions of Dynkin's identity for the characteristic operator. Finally, the sampling theory properties of these procedures are studied as sampling costs tend to zero.

A new characterization of the Dirichlet distribution using neutrality is given. This settles conjectures of Doksum and Mosimann. The characterization is contrasted with a parallel result for the lognormal distribution. Possible applications to random probabilities and prior distributions in survival analysis are noted.

An ordering of distributions related to a concept of positive dependence is studied and stochastic monotonicity with respect to this ordering is established for a wide class of two-sample test statistics. Asymptotic conservativeness of test sizes under certain departures from independence between samples is discussed. For example, if the observations are paired and the joint density is positive semidefinite then tests such as Kolmogorov-Smirnov, $\chi^2$ and Cramer-von Mises, as well as a large class of linear rank tests, are shown to be asymptotically conservative.

In this paper, the density function for the real and imaginary parts of the quotient of some off-diagonal elements of the three-dimensional complex Wishart matrix is derived. Connection to closed-loop transfer function determination is shown.

Consider the problem of combining two unbiased estimators of a parameter when the estimators are known to be independent normal variables with unknown variances possibly unequal. The two one parameter families of estimators studied in Brown and Cohen, and Khatri and Shah, are accommodated in a single two parameter family studied in this paper and the results in the two papers are unified. For the type of estimators considered by Brown and Cohen, this paper not only offers a generalization but also a significant improvement. This improvement concerns the main result in Theorem 2.1 of Brown and Cohen and has bearing on their entire paper except the last section on interval estimation. Extensions of Brown and Cohen's Theorem 4.1 concerning the point estimation of the common mean of $K$-populations and Theorem 5.1 concerning interval estimation of the common mean of two populations are also presented.

Let $x$ and $y$ be two independent normal variables with mean $\mu$ and variances $\sigma^2_1$ and $\sigma^2_2$ respectively. Also let $S^2_1$ and $S^2_2$ be two independent estimators of $\sigma^2_1$ and $\sigma^2_2$ such that $mS^2_1\sigma^{-2}_1$ and $nS^2_2\sigma^{-2}_2$ are chi-squares with $m$ and $n$ degrees of freedom respectively. The Graybill-Deal estimator of $\mu$ is $\hat{\mu} = (S^{-2}_1x + S^{-2}_2y)/(S^{-2}_1 + S^{-2}_2)$. In this paper an expression for the variance of $\hat{\mu}$ is given. Also bounds for the distribution of $\hat{\mu}$ are studied.

Let $X_1, \cdots, X_n (n \geqslant 2)$ be a random sample on a rv $X$, and $Y_1 < \cdots < Y_n$ be the corresponding order statistics. Define $Z_k = \frac{1}{n-k}\Sigma^n_{i=k+1}(Y_i - Y_k), 1 \leqslant k \leqslant n - 1, W_k = \frac{1}{k-1}\Sigma^{k-1}_{i=1}(Y_k - Y_i),$ $2 \leqslant k \leqslant n$. Using the properties $E(Z_k\mid Y_k = y) = \alpha y + \beta$ and $E(W_k\mid Y_k = y) = \alpha y + \beta$, a.e. $(dF)$, where $\alpha$ and $\beta$ are constants, we obtain characterizations of several distributions which include the exponential, the Pearson (type I) and the Pareto (of the second kind) distributions.

A sequential procedure for estimating the difference of two normal means when the variances are unknown and not necessarily equal is proposed, and an asymptotic expression for "regret" is given. This generalizes the corresponding one sample result of Woodroofe.

Susarla and Van Ryzin exhibited an empirical Bayes estimator of a distribution function $F$ based on randomly right-censored observations. In a later paper they obtained a different estimator which alleviates the weaknesses of their earlier estimator and showed that it is asymptotically optimal with rate of convergence $n^{-1}$. The purpose of this note is to present a slightly different estimator which is simpler and is also asymptotically optimal with the same rate of convergence. Their numerical example is reworked to show that the estimator is a proper distribution function.