The principal components analysis of functional data is often enhanced by the use of smoothing. It is shown that an attractive method of incorporating smoothing is to replace the usual $L^2$-orthonormality constraint on the principal components by orthonormality with respect to an inner product that takes account of the roughness of the functions. The method is easily implemented in practice by making use of appropriate function transforms (Fourier transforms for periodic data) and standard principal components analysis programs. Several alternative possible interpretations of the smoothed principal components as obtained by the method are presented. Some theoretical properties of the method are discussed: the estimates are shown to be consistent under appropriate conditions, and asymptotic expansion techniques are used to investigate their bias and variance properties. These indicate that the form of smoothing proposed is advantageous under mild conditions, indeed milder than those for existing methods of smoothed functional principal components analysis. The choice of smoothing parameter by cross-validation is discussed. The methodology of the paper is illustrated by an application to a biomechanical data set obtained in the study of the behaviour of the human thumb-forefinger system.

The receiver operating characteristic (ROC) curve describes the performance of a diagnostic test used to discriminate between healthy and diseased individuals based on a variable measured on a continuous scale. The data consist of a training set of m responses $X_1, \dots, X_m$ from healthy individuals and n responses $Y_1, \dots, Y_n$ from diseased individuals. The responses are assumed i.i.d. from unknown distributions F and G, respectively. We consider estimation of the ROC curve defined by $1 - G(F^{-1} (1 - t))$ for $0 \leq t \leq 1$ or, equivalently, the ordinal dominance curve (ODC) given by $F(G^{-1} (t))$. First we consider nonparametric estimators based on empirical distribution functions and derive asymptotic properties. Next we consider the so-called semiparametric "binormal" model, in which it is assumed that the distributions F and G are normal after some unknown monotonic transformation of the measurement scale. For this model, we propose a generalized least squares procedure and compare it with the estimation algorithm of Dorfman and Alf, which is based on grouped data. Asymptotic results are obtained; small sample properties are examined via a simulation study. Finally, we describe a minimum distance estimator for the ROC curve, which does not require grouping the data.

Suppose one has a stochastic time-dependent covariate $Z(t)$, and is interested in estimating the hazard relationship $\lambda(t|\overline{Z}(t)) = \omega(Z(t))$, where $\overline{Z}(t)$ denotes the history of $Z(t)$ up to and including time t. In this paper, we consider a model of the form $\exp(s_n(Z(t)))$, where $s_n(Z(t))$ is a spline of finite but arbitrary order, and investigate the behavior of the maximum likelihood estimator of the hazard as the number of knots in the spline function increases with the sample size at some rate $k_n = o(n)$. For twice continuously differentiable $\omega(\cdot)$, we demonstrate that the difference between the estimator $\exp(s_n(\cdot))$ and $\omega(\cdot)$ goes to 0 in probability in sup-norm for any $k_n = n^{\phi}, \phi \epsilon (0, 1)$. In addition, if $\phi > 1/5$, then $\exp(\hat{s}_n(Z(t))) - \omega(Z(t))$, properly normalized, is asymptotically standard normal. A large-sample approximation to the variance is derived in the case where $s_n(\cdot)$ is a linear spline, and exposes some rather interesting behavior.

Two efficient Monte Carlo algorithms are proposed for uniformlygenerating two-way contingency tables with fixed margins. These permit some improvements on recent work of Diaconis, Efron and Gangolli, especially concerning estimates of the total number of such tables.

The Markov chain simulation method has been successfully used in many problems, including some that arise in Bayesian statistics. We give a self-contained proof of the convergence of this method in general state spaces under conditions that are easy to verify.

We apply recent results in Markov chain theory to Hastings and Metropolis algorithms with either independent or symmetric candidate distributions, and provide necessary and sufficient conditions for the algorithms to converge at a geometric rate to a prescribed distribution $\pi$. In the independence case (in $\mathbb{R}^k$) these indicate that geometric convergence essentially occurs if and only if the candidate density is bounded below by a multiple of $\pi$; in the symmetric case (in $\mathbb{R}$ only) we show geometric convergence essentially occurs if and only if $\pi$ has geometric tails. We also evaluate recently developed computable bounds on the rates of convergence in this context: examples show that these theoretical bounds can be inherently extremely conservative, although when the chain is stochastically monotone the bounds may well be effective.

The exact second eigenvalue of the Markov operator of the Gibbs sampler with random sweep strategy for Gaussian densities is calculated. A comparison lemma yields an upper bound on the second eigenvalue for bounded perturbations of Gaussians which is a significant improvement over previous bounds. For two-block Gibbs sampler algorithms with a perturbation of the form $\chi(g_1(x^{(1)}) + g_2(x^{(2)}))$ the derivative of the second eigenvalue of the algorithm is calculated exactly at $\chi = 0$, in terms of expectations of the Hessian matrices of $g_1$ and $g_2$.

Jeffreys' prior, one of the widely used noninformative priors, remains invariant under reparameterization, but does not perform satisfactorily in the presence of nuisance parameters. To overcome this deficiency, recently various noninformative priors have been proposed in the literature.

This article explores the invariance (or lack thereof) of some of these noninformative priors including the reference prior of Berger and Bernardo, the reverse reference prior of J. K. Ghosh and the probability-matching prior of Peers and Stein under reparameterization. Berger and Bernardo's m-group ordered reference prior is shown to remain invariant under a special type of reparameterization. The reverse reference prior of J. K. Ghosh is shown not to remain invariant under reparameterization. However, the probability-matching prior is shown to remain invariant under any reparameterization. Also for spherically symmetric distributions, certain noninformative priors are derived using the principle of group invariance.

Suppose that $X_{\sigma} | \mathbf{\theta} \sim N(\mathbf{\theta}, \sigma^2)$ and that $\sigma \to 0$. For which prior distributions on $\mathbf{\theta}$ is the posterior distribution of $\mathbf{\theta}$ given $X_{\sigma}$ asymptotically $N(X_{\sigma}, \sigma^2)$ when in fact $X_{\sigma} \sim N(\theta_0, \sigma^2)$? It is well known that the stated convergence occurs when $\mathbf{\theta}$ has a prior density that is positive and continuous at $\theta_0$. It turns out that the necessary and sufficient conditions for convergence allow a wider class of prior distributions--the locally uniform and tail-bounded prior distributions. This class includes certain discrete prior distributions that may be used to reproduce minimum description length approaches to estimation and model selection.

We investigate the degree to which posterior expectations are sensitive to prior distributions, using a local method based on functional differentiation. Invariance considerations suggest a family of norms which can be used to measure perturbations to the prior. The sensitivity measure is seen to depend heavily on the choice of norm; asymptotic results suggest which norm will yield the most useful results in practice. We find that to maintain asymptotically sensible behaviour, it is necessary to reduce the richness of the class of prior perturbations as the dimension of the parameter space increases. Jeffreys' prior is characterized as the prior to which inference is least sensitive.

Asymptotically optimal and admissible compound decision rules are obtained in a Hilbert-parameterized Gaussian shift experiment. The component parameter set is restricted to compact. For the squared error loss, every compound Bayes estimator is admissible and every compound estimator Bayes versus full support hyperprior mixture of iid priors on the compound parameter is asymptotically optimal. For the latter class of rules induced by full support hyperpriors, asymptotic optimality and admissibility extend to equi- (in decisions) uniformly continuous and bounded risk functions. Normality of certain mixtures of the standard Gaussian process and qualitative robustness of the component Bayes estimator (results of independent interest used in the paper) are derived.

The empirical Bayes linear loss two-action problem in the continuous one-parameter exponential family is studied. Previous results on this problem construct empirical Bayes tests via kernel density estimates. They also obtain upper bounds for the unconditional regret at some prior distribution. In this paper, we discuss the general question of how difficult the above empirical Bayes problem is, and why empirical Bayes rules based on kernel density estimates are useful. Asymptotic minimax-type lower bounds are obtained for the unconditional regret, and empirical Bayes rules based on kernel density estimates are shown to possess a certain optimal asymptotic minimax property.

This paper proposes a consistent nonparametric empirical Bayes estimator of the prior density for directional data. The methodology is to use Fourier analysis on $S^2$ to adapt Euclidean techniques to this non-Euclidean environment. General consistency results are obtained. In addition, a discussion of efficient numerical computation of Fourier transforms on $S^2$ is given, and their applications to the methods suggested in this paper are sketched.

The restricted maximum likelihood (REML) estimates of dispersion parameters (variance components) in a general (non-normal) mixed model are defined as solutions of the REML equations. In this paper, we show the REML estimates are consistent if the model is asymptotically identifiable and infinitely informative under the (location) invariant class, and are asymptotically normal (A.N.) if in addition the model is asymptotically nondegenerate. The result does not require normality or boundedness of the rank p of design matrix of fixed effects. Moreover, we give a necessary and sufficient condition for asymptotic normality of Gaussian maximum likelihood estimates (MLE) in non-normal cases. As an application, we show for all unconfounded balanced mixed models of the analysis of variance the REML (ANOVA) estimates are consistent; and are also A.N. provided the models are nondegenerate; the MLE are consistent (A.N.) if and only if certain constraints on p are satisfied.

Direct use of the empirical quantile function provides a standard distribution-free approach to constructing confidence intervals and confidence bands for population quantiles. We apply this method to construct confidence intervals and confidence bands for regression quantiles and to construct prediction intervals based on sample regression quantiles. Comparison of the direct method with the studentization and the bootstrap methods are discussed. Simulation results show that the direct method has the advantage of robustness against departure from the normality assumption of the error terms.

We give a stochastic expansion for the empirical distribution function $\hat{F}_n$ of residuals in a p-dimensional linear model. This expansion holds for p increasing with n. It shows that, for high-dimensional linear models, $\hat{F}_n$ strongly depends on the chosen estimator $\hat{\theta}$ of the parameter $\theta$ of the linear model. In particular, if one uses an ML-estimator $\hat{\theta}_{ML}$ which is ML motivated by a wrongly specified error distribution function G, then $\hat{F}_n$ is biased toward G. For p^2 / n \to \infty$, this bias effect is of larger order than the stochastic fluctuations of the empirical process. Hence, the statistical analysis may just reproduce the assumptions imposed.

Under mild misspecifications of model assumptions, maximum likelihood estimates often remain consistent and asymptotically normal. Asymptotic normality will often hold for the signed root of the likelihood ratio statistic and the score statistic as well. However, standard estimates of asymptotic variance are usually inconsistent. This occurs when Bartlett's second identity fails. In the manner of McCullagh and Tibshirani, a variance correction may be used to adjust the profile likelihood so this identity obtains. The resulting likelihood yields the robust versions of the signed root, Wald and score statistic suggested by Kent and Royall.

Assuming model correctness, asymptotic expansions for the first three cumulants of each robust statistic are derived. It is seen that bias and skewness are not severely affected by using a robust statistic. An invariant expression derived for the asymptotic relative efficiency of a robust method allows assessment in numerous examples considered. Even for moderately large sample sizes, losses in efficiency are significant, making the misuse of a robust variance estimate potentially costly. Computer algebra is used in many of the calculations reported in this paper.

This paper derives higher-order terms in the double-saddlepoint expansion of Skovgaard for a unidimensional conditional cumulative distribution function. Expansions for continuous and lattice random variables are derived. Results are applied to the sufficient statistic in logistic regression.

In a recent paper, DiCiccio, Hall and Romano established that Owen's empirical likelihood is Bartlett correctable. This is an intriguing and perhaps surprising result and is the only nonparametric context in which Bartlett correctability is known to hold. An alternative, closely related nonparametric likelihood, referred to here as exponential empirical likelihood, may be constructed using Efron's method of nonparametric tilting. The purpose of this note is to show that exponential empirical likelihood is not Bartlett correctable.

The setting is a stationary, ergodic time series. The challenge is to construct a sequence of functions, each based on only finite segments of the past, which together provide a strongly consistent estimator for the conditional probability of the next observation, given the infinite past. Ornstein gave such a construction for the case that the values are from a finite set, and recently Algoet extended the scheme to time series with coordinates in a Polish space.

The present study relates a different solution to the challenge. The algorithm is simple and its verification is fairly transparent. Some extensions to regression, pattern recognition and on-line forecasting are mentioned.

This paper establishes the asymptotic uniform linearity of M- and R-scores in a family of nonlinear time series and regression models. It also gives an asymptotic expansion of the standardized sequential residual empirical process in these models. These results are, in turn, used to obtain the asymptotic normality of certain classes of M-, R- and minimum distance estimators of the underlying parameters. The classes of estimators considered include analogs of Hodges-Lehmann, Huber and LAD (least absolute deviation) estimators. Some applications to the change point and testing of the goodness-of-fit problems in threshold and amplitude-dependent exponential autoregression models are also given. The paper thus offers a unified functional approach to some aspects of robust inference for a large class of nonlinear time series models.

Consider an ergodic Markov chain on the real line, with parametric models for the conditional mean and variance of the transition distribution. Such a setting is an instance of a quasi-likelihood model. The customary estimator for the parameter is the maximum quasi-likelihood estimator. It is not efficient, but as good as the best estimator that ignores the parametric model for the conditional variance. We construct two efficient estimators. One is a convex combination of solutions of two estimating equations, the other a weighted nonlinear one-step least squares estimator, with weights involving predictors for the third and fourth centered conditional moments of the transition distribution. Additional restrictions on the model can lead to further improvement. We illustrate this with an autoregressive model whose error variance is related to the autoregression parameter.

Let ${X_t, t \epsilon \mathbb{Z}}$ be an observable strictly stationary sequence of random variables and let $X_t = U_t + \varepsilon_t$, where ${U_t}$ is an AR (p) and ${\varepsilon_t}$ is a strictly stationary sequence representing errors of measurement in ${X_t}$, with $E{\varepsilon_1} = 0$. Under some broad assumptions on ${\varepsilon_t}$ we establish the consistency properties as well as the rates of convergence for the standard estimators for the autoregressive parameters computed from a set of modified Yule-Walker equations.

An inequality that relates the probability of false alarm of a change-point detection policy to its average run length to detection is proved. By means of this inequality, a lower bound on the rate of detection, when the change occurs after a long delay, is derived.

A generalization for the theory of comparison of experiments is given to the case of sequential experiments. We investigate only the case of "0" deficiency. Applications are given to the case of exponential experiments.