In Sections 1-3, the classical theory of the comparison of two experiments is reviewed with particular reference to the comparison of two location experiments. It is shown that the requirement of domination of one experiment by another for all decision problems is too strong to provide a reasonable basis for comparison. For one-parameter problems with monotone likelihood ratio, it is therefore proposed to restrict the comparison to decision problems that are monotone in the sense of Karlin and Rubin (1956). Application of this weaker definition to the location problem is shown to give satisfactory results. A scale-free comparison of this type leads to a new tail-ordering of distributions, and this is explored in Section 6.

In this paper we introduce $E$-ancillarity and complete $E$-sufficiency, natural extensions of the definitions of ancillarity and complete sufficiency to a space of estimating or inference functions. These are functions of both the data and the parameter. We begin either with a space of all such functions or with a subset defined to exploit special features of a model; for example, we allow restrictions to inference functions that are linear in the observations or linear in the parameter. Subsequently, a reduction analogous to complete sufficiency is carried out, and within the complete $E$-sufficient space of inference functions, one is chosen with properties that we deem desirable.

The minimum distance (MD) functional defined by a distance $\mu$ is automatically robust over contamination neighborhoods defined by $\mu$. In fact, when compared to other Fisher-consistent functionals, the MD functional was no worse than twice the minimum sensitivity to $\mu$-contamination, and at least half the best possible breakdown point. In invariant settings, the MD functional has the best attainable breakdown point against $\mu$-contamination among equivariant functionals. If $\mu$ is Hilbertian (e.g., the Hellinger distance), the MD functional has the smallest sensitivity to $\mu$-contamination among Fisher-consistent functionals. The robustness of the MD functional is inherited by MD estimates, both estimates based on "weak" distances and estimates based on "strong" distances, when the empirical distribution is appropriately smoothed. These facts are general and apply not just in simple location models, but also in multivariate location-scatter and in semiparametric settings. Of course, this robustness is formal because $\mu$-contamination neighborhoods may not be large enough to contain realistic departures from the model. For the metrics we are interested in, robustness against $\mu$-contamination is stronger than robustness against gross errors contamination; and for "weak" metrics (e.g., $\mu = \text{Cramer-von Mises, Kolmogorov})$, robustness over $\mu$-neighborhoods implies robustness over Prohorov neighborhoods.

Minimum distance estimates are studied at the $N(\theta, 1)$ model. Estimates based on a non-Hilbertian distance $\mu (\mu = \text{Kolmogorov-Smirnov, Levy, Kuiper, variation and Prohorov})$ can exhibit very large variances, or even outright inconsistency, at distributions arbitrarily close to the model in terms of $\mu$-distance. For Hilbertian distances $(\mu = \text{Cramer-von Mises, Hellinger})$ this problem does not seem to occur. Geometric motivation for these results is provided.

Some standard robustness concepts, developed for linear data, are found wanting in the case of directional data. We introduce a standardized bias robustness allowing for uniform robustness considerations of statistics for bounded parameter spaces. Specifically, we verify the nonrobustness in this sense of the directional mean, the directional dispersion and the maximum likelihood estimator of the concentration parameter of the von Mises-Fisher distributions.

Let $\hat{H}_n(\alpha) (0 < \alpha < 1)$ denote the length of the shortest $\alpha$-fraction of the ordered sample $X_{1:n}, X_{2:n}, \cdots, X_{n:n}$, i.e., $\hat{H}_n(\alpha) = \min\{X_{k + j:n} - X_{k:n}: 1 \leq k \leq k + j \leq n; (j + 1)/n \geq \alpha\}.$ Such quantities arise in the context of robust scale estimation. Using the concept of compact derivatives of statistical functionals, the asymptotic behaviour of $\hat{H}_n(\alpha)$ as $n \rightarrow \infty$ is investigated.

A mode of a probability density $f(t)$ is a value $\theta$ that maximizes $f$. The problem of estimating the location of the mode is considered here. Estimates of the mode are considered via kernel density estimates. Previous results on this problem have several serious drawbacks. Conditions on the underlying density $f$ are imposed globally (rather than locally in a neighborhood of $\theta$). Moreover, fixed bandwidth sequences are considered, resulting in an estimate of the location of the mode that is not scale-equivariant. In addition, an optimal choice of bandwidth depends on the underlying density, and so cannot be realized by a fixed bandwidth sequence. Here, fixed and random bandwidths are considered, while relatively weak assumptions are imposed on the underlying density. Asymptotic minimax risk lower bounds are obtained for estimators of the mode and kernel density estimates of the mode are shown to possess a certain optimal local asymptotic minimax risk property. Bootstrapping the sampling distribution of the estimates is also discussed.

Maximum likelihood estimation for the covariance $R$ of a zero-mean Gaussian process is considered, with no assumptions on the covariance or the "time" parameter set $T$. It is shown that the likelihood function is a.s. unbounded in general, and a sieve estimator $\hat{R}$ is constructed. The distribution of $\hat{R}$, considered as a process on $T \times T$, can be described exactly if a certain technical assumption is satisfied concerning the bivariate series expansion of $R$. It is then shown that $\hat{R}(s, t)$ is asymptotically unbiased and consistent (weakly and in mean square) at each $(s, t) \in T \times T$, and that $\hat{R}$ is strongly consistent (globally) in an appropriate norm.

The nonparametric maximum likelihood estimators of the distribution functions of observations in the truncation model are represented as iid mean processes, with a remainder term of order $o(n^{-1/2})$ a.s.

Let $X$ be a positive random variable with Lebesgue density $f_\theta(x)$, where $\theta$ is the scale parameter, and let $Y$ be a positive random variable independent of $X$. We consider two models of truncation: the LHS model, where the data consist only of those observations of $X$ for which $X > Y$; and the RHS model, where the data consist of those observations of $X$ for which $X \leq Y$. Consider the problem of estimating $\theta^s, s \neq 0$, under a normalized squared error loss function. It is shown that under appropriate assumptions, if $f_1(\cdot)$ varies regularly at 0 (or $+ \infty$), then the minimax value in the RHS (LHS) model is equal to 1 for arbitrarily large sample size. This implies the existence of trivial minimax and admissible estimators, which do not depend on the sample at all, in contrast with the scale model without truncation.

Cox's regression model has been successfully used for censored survival data. It can be adapted to model a counting process having a periodic underlying intensity. In survival analysis, the asymptotic properties, as studied by Andersen and Gill, correspond to a large number of processes running parallel over the same time interval. Here a single point process is observed over a large number of successive periods. Cox's model can easily be adapted to this situation and conditions are given which ensure the estimators have the classical large sample properties. Proofs use both martingale techniques and theorems for convergence of empirical probability measures. Finally, an example concerning the feeding pattern of domestic rabbits is included.

The Tikhonov method of regularization (MOR) estimator provides a general method for estimation of a nonparametric regression parameter in an abstract linear model with discrete noisy data. An asymptotic analysis is given in which the discrete estimation problem is approximated by a continuous one. Rates of convergence are calculated in a family of norms natural to the problem. The general theory is applied to the estimation of functions from noisy evaluations of the function and one of its derivatives.

For a regression model $y_i = \theta(x_i) + \varepsilon_i$, the unknown function $\theta$ is estimated by least squares on a subspace $\Lambda_m = \operatorname{span}\{\psi_1, \psi, \cdots, \psi_m\}$, where the basis functions $\psi_i$ are predetermined and $m$ is varied. Assuming that the design is suitably approximated by an asymptotic design measure, a general method is presented for approximating the bias and variance in a scale of Hilbertian norms natural to the problem. The general theory is illustrated with two examples: truncated Fourier series regression and polynomial regression. For these examples, we give rates of convergence of derivative estimates in (weighted) $L_2$ norms and establish consistency in supremum norm.

The problem of testing the correctness of a nonlinear response function against unspecified general alternatives is considered. The proposed test statistic is a modification of a nonlinear analogue to the well-known linear regression lack-of-fit test and can be used with or without replication. Asymptotically valid critical points can be obtained from a central $F$-distribution. Also, when the null model is the orthogonal projection of the true model, the test statistic is asymptotically comparable to a random variable with a noncentral $F$-distribution.

In monotone regression procedures one utilizes only the monotonicity of the regression function. In nonparametric regression one utilizes only the assumed smoothness. The analytic and asymptotic properties of the estimator are superior in the latter case; however, monotonicity is not guaranteed. We study a hybrid procedure that produces monotone estimators with properties similar to those of nonparametric regression estimators.

We consider an experiment which consists of $k$ treatment groups and a control group. Let the sample means $\bar{Y}_0, \bar{Y}_1, \cdots, \bar{Y}_k$ be independent normal variates with expected values $\mu_0, \mu_1, \cdots, \mu_k$ and with variances $\sigma^2/n_0, \sigma^2/n_1, \cdots, \sigma^2/n_k$. Let $w_0, w_1, \cdots, w_k$ be positive weights and let $\mu^\ast_0, \mu^\ast_1, \cdots, \mu^\ast_k$ be the weighted least squares estimators subject to the constraints $\mu_0 \leq \mu_i, i = 1, \cdots, k$. We establish that for large $k, E(\mu^\ast_0 - \mu_0)^2 > E(\bar{Y}_0 - \mu_0)^2$ when $w_i = n_i, i = 0, 1, \cdots, k$. Under suitable conditions, we show that $E(\mu^\ast_i - \mu_i)^2 < E(\bar{Y}_i - \mu_i)^2, i = 0, 1, \cdots, k$.

Asymptotic properties of estimators for the confirmatory factor analysis model are discussed. The model is identified by restrictions on the elements of the factor loading matrix; the number of restrictions may exceed that required for identification. It is shown that a particular centering of the maximum likelihood estimator derived under assumed normality of observations yields an asymptotic normal distribution that is common to a wide class of distributions of the factor vectors and error vectors. In particular, the asymptotic covariance matrix of the factor loading estimator derived under the normal assumption is shown to be valid for the factor vectors containing a fixed part and a random part with any distribution having finite second moments and for the error vectors consisting of independent components with any distributions having finite second moments. Thus the asymptotic standard errors of the factor loading estimators computed by standard computer packages are valid for virtually any type of nonnormal factor analysis. The results are extended to certain structural equation models.

For independent $d$-variate random samples $X_1, \cdots, X_{n_1}$ i.i.d. $f(x), Y_1, \cdots, Y_{n_2}$ i.i.d. $g(x)$, where the densities $f$ and $g$ are assumed to be continuous a.e., consider the number $T$ of all $k$ nearest neighbor comparisons in which observations and their neighbors belong to the same sample. We show that, if $f = g$ a.e., the limiting (normal) distribution of $T$, as $\min(n_1, n_2) \rightarrow \infty, n_1/(n_1 + n_2) \rightarrow \tau, 0 < \tau < 1$, does not depend on $f$. An omnibus procedure for testing the hypothesis $H_0: f = g$ a.e. is obtained by rejecting $H_0$ for large values of $T$. The result applies to a general distance (generated by a norm on $\mathbb{R}^d$) for determining nearest neighbors, and it generalizes to the multisample situation.

Stationary vector ARMA processes $x(t), t = 0, \pm 1, \pm 2, \cdots$, of $n$ components are considered that are of full rank, and the situation where there are linear functions of the future $x(t), t > 0$, and the past $x(t), t \leq 0$ (more properly the present and the past) that have unit correlation. It is shown that the number of linearly independent such pairs (i.e., the number of unit canonical correlations between future and past) is the number of zeros of the determinant of the transfer functions, from innovations to outputs, that lie on the unit circle, counting these with their multiplicities.

We consider estimation of a regression model with long-memory stationary errors. First, we estimate the regression parameters by the least-squares estimator (LSE) and, next, those describing the correlation structure of the error terms by using the residuals obtained from the LSE. Certain regularity conditions introduced to develop the asymptotic theory no longer hold in this model. In this situation we derive asymptotic properties of the preceding estimation procedure.

To estimate the parameters of a stationary process, Whittle (1953) introduced an approximation to the Gaussian likelihood function. Although the Whittle estimate is asymptotically efficient, the small sample behavior may be poor if the spectrum of the process contains peaks. We introduce a mathematical model that covers such small sample effects. We prove that the exact maximum likelihood estimate is still optimal in this model, whereas the Whittle estimate and the conditional likelihood estimate are not. Furthermore, we introduce tapered Whittle estimates and prove that these estimates have the same optimality properties as exact maximum likelihood estimates.

Properties of a probability density estimator having the rational form of an ARMA spectrum are investigated. Under various conditions on the underlying density's Fourier coefficients, the ARMA estimator is shown to have asymptotically smaller mean integrated squared error (MISE) than the best tapered Fourier series estimator. The most interesting cases are those in which the Fourier coefficients $\phi_j$ are asymptotic to $Kj^{-p}$ as $j \rightarrow \infty$, where $\rho > 1/2$. For example, when $\rho = 2$ the asymptotic MISE of a certain ARMA estimator is only about 63% of that for the optimum series estimator. For a density $f$ with support in $\lbrack 0, \pi \rbrack$, the condition $\rho = 2$ occurs whenever $f'(0 +) \neq 0, f'(\pi -) = 0$ and $f"$ is square integrable.

A simple class of sequential tests is proposed for testing the one-sided composite hypotheses $H_0: \theta \leq \theta_0$ versus $H_1: \theta \geq \theta_1$ for the natural parameter $\theta$ of an exponential family of distributions under the 0-1 loss and cost $c$ per observation. Setting $\theta_1 = \theta_0$ in these tests also leads to simple sequential tests for the hypotheses $H: \theta < \theta_0$ versus $K: \theta > \theta_0$ without assuming an indifference zone. Our analytic and numerical results show that these tests have nearly optimal frequentist properties and also provide approximate Bayes solutions with respect to a large class of priors. In addition, our method gives a unified approach to the testing problems of $H$ versus $K$ and also of $H_0$ versus $H_1$ and unifies the different asymptotic theories of Chernoff and Schwarz for these two problems.

Two stochastic processes, or "arms," that yield dichotomous responses are available for use in a two-stage decision problem. During the first stage, arms are chosen sequentially; the resulting observations are discounted by a fixed value $\beta$. A single arm must be used in the second stage, in which observations are not discounted. The decision to end the first stage is based on the data obtained. Optimal strategies are considered in the presence of the random discount sequence that arises in this setting. This extends the work of Berry and Fristedt (1979).

Practical Bayesian statistics with realistic models usually gives posterior distributions that are analytically intractable, and inferences must be made via numerical integration. In many cases, the integrands can be transformed into periodic functions on the unit $d$-dimensional cube, for which quasirandom sequences are known to give efficient numerical integration rules. This paper reviews some relevant theory, defines new criteria for identifying suitable quasirandom sequences and suggests some extensions to the basic integration rules. Various quasirandom methods are then compared on the sort of integrals that arise in Bayesian inference and are shown to be much more efficient than Monte Carlo methods.

Constrained majorization orderings and entropy functions are used to study the class of probability matrices associated with paired comparisons. Majorization orderings are also defined to handle the cases of order effects and/or ties. Results are obtained for maximal and minimal probability matrices with respect to the majorization ordering; these are related to transitivity conditions. The Bradley-Terry and Thurstone-Mosteller models are shown to be maximum entropy models. New models based on maximum entropy are obtained for the cases of order effects and ties; these models are compared with the Davidson and Beaver, the Rao and Kupper and the Davidson models. Applications to professional baseball and hockey are given.