Let $y_i, i = 1, \dots, n$, be independent observations with the density of $y_i$ of the form $h(y_i, f_i) = \exp{y_i f_i - b(f_i) + c(y_i)]$, where b and c are given functions and b is twice continuously differentiable and bounded away from 0. Let $f_i = f(t(i))$, where $t = (t_1, \dots, t_d) \epsilon \mathsf{T}^{(1)} \otimes \dots \otimes \mathsf{T}^{(d)} = \mathsf{T}$, the $\mathsf{T}^{(\alpha)}$ are measurable spaces of rather general form and f is an unknown function on $\mathsf{T}$ with some assumed "smoothness" properties. Given ${y_i, t(i), i = 1, \dots, n}$, it is desired to estimate $f(t)$ for t in some region of interest contained in $\mathsf{T}$. We develop the fitting of smoothing spline ANOVA models to this data of the form $f(t) = C + \sum_{\alpha} f_{\alpha}(t_{\alpha}) + \sum_{\alpha < \beta} f_{\alpha \beta} (t_{\alpha}, t_{\beta}) + \dots$. The components of the decomposition satisfy side conditions which generalize the usual side conditions for parametric ANOVA. The estimate of f is obtained as the minimizer, in an appropriate function space, of $\mathsf{L}(y, f) + \sum_{\alpha} \lambda_{\alpha} J_{\alpha}(f_{\alpha}) + \sum_{\alpha <\beta} \lambda_{\alpha \beta} J_{\alpha \beta}(f_{\alpha \beta}) + \dots$, where $\mathsf{L}(y, f)$ is the negative log likelihood of $y = (y_1, \dots, y_n)'$ given f, the $J_{\alpha}, J_{\alpha \beta}, \dots$ are quadratic penalty functionals and the ANOVA decomposition is terminated in some manner. There are five major parts required to turn this program into a practical data analysis tool: (1) methods for deciding which terms in the ANOVA decomposition to include (model selection), (2) methods for choosing good values of the smoothing parameters $\lambda_{\alpha}, \lambda_{\alpha \beta}, \dots$, (3) methods for making confidence statements concerning the estimate, (4) numerical algorithms for the calculations and, finally, (5) public software. In this paper we carry out this program, relying on earlier work and filling in important gaps. The overall scheme is applied to Bernoulli data from the Wisconsin Epidemiologic Study of Diabetic Retinopathy to model the risk of progression of diabetic retinopathy as a function of glycosylated hemoglobin, duration of diabetes and body mass index. It is believed that the results have wide practical application to the analysis of data from large epidemiological studies.

Additive models are one means of assuaging the curse of dimensionality when nonparametric smoothing methods are used to estimate multivariable regression functions. It is important to have methods for testing the fit of such models, especially in high dimensions where visual assessment of fit becomes difficult. New tests of additivity are proposed in this paper that derive from Fourier series estimators with data-driven smoothing parameters. Other tests related to the classical Tukey test for additivity are also considered. While the new tests are consistent against essentially any "smooth" alternative to additivity, the Tukey-type tests are found to be inconsistent in certain situations. Asymptotic power of both varieties of tests is studied under local alternatives that tend toward additivity at a parametric rate, and small-sample power comparisons are carried out by means of a simulation study.

We analyse methods based on the block bootstrap and leave-out cross-validation, for choosing the bandwidth in nonparametric regression when errors have an almost arbitrarily long range of dependence. A novel analytical device for modelling the dependence structure of errors is introduced. This allows a concise theoretical description of the way in which the range of dependence affects optimal bandwidth choice. It is shown that, provided block length or leave-out number, respectively, are chosen appropriately, both techniques produce first-order optimal bandwidths. Nevertheless, the block bootstrap has far better empirical properties, particularly under long-range dependence.

In the present paper we combine the issues of bandwidth choice and construction of confidence intervals in nonparametric regression. Main emphasis is put on fully data-driven methods. We modify the $\sqrt{n}$-consistent bandwidth selector of Härdle, Hall and Marron such that it is appropriate for heteroscedastic data, and we show how one can optimally choose the bandwidth g of the pilot estimator $\hat{m}_g$. Then we consider classical confidence intervals based on kernel estimators with data-driven bandwidths and compare their coverage accuracy. We propose a method to put undersmoothing with a data-driven bandwidth into practice and show that this procedure outperforms explicit bias correction.

This paper deals with the statistical analysis of matched pairs of shapes of configurations of landmarks in the plane. We provide inference procedures on the complex projective plane for a basic measure of shape change in the plane, on observing that shapes of configurations of $(k + 1)$ landmarks in the plane may be represented as points on $\mathbb{C} P^{k-1}$ and that complex rotations are the only maps on $\mathbb{C} S^{k-1}$ which preserve the usual Hermitian inner product. Specifically, if $u_1, \dots, u_n$ are fixed points on $\mathbb{C} P^{k-1}$ represented as $\mathbb{C} S^{k-1}/U(1)$ and $v_1, \dots, v_n$ are random points on $\mathbb{C} P^{k-1}$ such that the distribution of $v_j$ depends only on $||v_j^* Au_j||^2$ for some unknown complex rotation matrix A, then this paper provides asymptotic inference procedures for A. It is demonstrated that shape changes of a kind not detectable as location shifts by standard Euclidean analysis can be found by this frequency domain method. A numerical example is given.

For a stationary random field Z on $\mathbb{R}^d$, this work studies the asymptotic behavior of predictors of $\int v(x)Z(x)dx$ based on observations on a lattice as the distance between neighbors in the lattice tends to 0. Under a mild condition on the spectral density of Z, an asymptotic expression for the mean-squared error of a predictor of $\int v(x)Z(x)dx$ based on observations on an infinite lattice is derived. For predicting integrals over the unit cube, a simple predictor based just on observations in the unit cube is shown to be asymptotically optimal if v is sufficiently smooth and Z is not too smooth. Modified predictors extend this result to smoother processes.

For predicting $\int_G v(x)Z(x)dx$, where v is a fixed known function and Z is a stationary random field, a good sampling fesign should have a greater density of observations where v is relatively large in absolute value. Designs using this idea when $G = [0, 1]$ have been studied for some time. For G a region in two dimensions, very little is known about the statistical properties of cubature rules based on designs with varying density. This work proposes a class of designs that are locally parallelogram lattices but whose densities can vary. The asymptotic variance of the cubature error for these designs is obtained for a class of isotropic random fields and an asymptotically optimal sequence of cubature rules within this class is found. I conjecture that this sequence of cubature rules is asymptotically optimal with respect to all cubature rules.

Consider an iid sample $Y_1, \dots, Y_n$ of random variables with common distribution function F, whose upper tail belongs to a neighborhood of the upper tail of a generalized Pareto distribution $H_{\beta}, \beta \epsilon \mathbb{R}$. We investigate the testing problem $\beta = \beta_0$ against a sequence $\beta = \beta_n$ of contiguous 0 n alternatives, based on the point processes $N_n$ of the exceedances among $Y_i$ over a sequence of thresholds $t_n$. It turns out that the (random) number of exceedances $\tau (n)$ over $t_n$ is the central sequence for the log-likelihood ratio $d \mathsf{L}_{\beta_n} (N_n)/ d \mathsf{L}_{\beta_0} (N_n)$, yielding its local asymptotic normality (LAN). This result implies in particular that $\tau (n)$ carries asymptotically all the information about the underlying parameter $\beta$, which is contained in $N_n$. We establish sharp bounds for the rate at which $\tau (n)$ becomes asymptotically sufficient, which show, however, that this is quite a poor rate. These results remain true if we add an unknown scale parameter.

We propose a conditional MLE of the index of regular variation when the functional form of a slowly varying function is assumed known in the tail, and we study its asymptotic properties. We prove asymptotic normality of $P_{\theta}^{k_n}$, a probability measure whose density is the product of the joint conditional density of the $k_n$ largest order statistics from $F_{\theta} (x)$ given $Z_{n-k}$, the $$(n-k)$th order statistic, and a density of $Z_{n-k}$ with parameter $\theta$. Based on this result, we show that this conditional MLE is asymptotically normal and asymptotically efficient in many senses whenever $k_n$ is $o(n)$. We also propose an iterative estimator of $\theta$ given only partial knowledge of $L_{\theta}(x)$. This estimator is asymptotically normal, asymptotically unbiased and asymptotically efficient.

Consider a distribution function that belongs to the weak domain of attraction of an extreme value distribution. The extreme value index $\beta$ will be estimated by mixtures of Pickands estimators, where the weights are generated by a probability measure which satisfies a certain integrability condition. We prove a functional limit theorem for a process of Pickands estimators and asymptotic normality of the refined Pickands estimator. For negative $\beta$ the new estimator is asymptotically superior to previously defined estimators. A simulation study also demonstrates the good small-sample performance. In particular, the estimator proves to be robust against an inappropriate choice of the number of upper order statistics used for estimation.

We study properties of the variance function of the least squares estimator for the response surface. For polynomial models, we identify a class of approximate designs for which their variance functions are maximized at the extreme points of the design space. As an application, we examine robustness properties of D-optimal designs and $D_{n-r}$-optimal designs under various polynomial model assumptions. Analytic formulas for the G-efficiencies of these designs are derived, along with their D-efficiencies.

We develop a combinatorial condition necessary for the existence of a saturated asymmetrical orthogonal array of strength 2. This condition limits the choice of integral solutions to the system of equations in the Bose-Bush approach and can thus strengthen considerably the Bose-Bush approach as applied to a symmetrical part of such an array. As a consequence, several nonexistence results follow for saturated and nearly saturated orthogonal arrays of strength 2. One of these leads to a partial settlement of an issue left open in a paper by Wu, Zhang and Wang. Nonexistence of a class of saturated asymmetrical orthogonal arrays of strength 4 is briefly discussed.

For various applications one wants to know the asymptotic behavior of $w(\theta | \overline{X})$, the posterior density of a parameter $\theta$ given the mean $\overline{X}$ of the data rather than the full data set. Here we show that $w(\theta | \overline{X})$ is asymptotically normal in an $L^1$ sense, and we identify the mean of the limiting normal and its asymptotic variance. The main results are first proved assuming that $X_1, \dots, X_n, \dots$ are independent and identical; suitable modifications to obtain results for the nonidentical case are given separately. Our results may be used to construct approximate HPD (highest posterior density) sets for the parameter which is of use in the statistical theory of standardized educational tests. They may also be used to show the covariance between two test items conditioned on the mean is asymptotically nonpositive. This has implications for constructing tests of item independence.

Z.A general (asymptotic) theory of estimation was developed by Ibragimov and Has’minskii under certain conditions on the normalized likelihood ratios. In an earlier work, the present authors studied the limiting behaviour of the posterior distributions under the general setup of Ibragimov and Has’minskii. In particular, they obtained a necessary condition for the convergence of a suitably centered (and normalized) posterior to a constant limit in terms of the limiting likelihood ratio process. In this paper, it is shown that this condition is also sufficient to imply the posterior convergence. Some related results are also presented.

We investigate diagnostics for quantifying the effect of small changes to the prior distribution over a k-dimensional parameter space. We show that several previously suggested diagnostics, such as the norm of the Fréchet derivative, diverge at rate $n^{k/2}$ if the base prior is an interior point in the class of priors, under the density ratio topology. Diagnostics based on $\phi$-divergences exhibit similar asymptotic behavior. We show that better asymptotic behavior can be obtained by suitably restricting the classes of priors. We also extend the diagnostics to see how various marginals of the prior affect various marginals of the posterior.

This essay considers decision-theoretic foundations for robust Bayesian statistics. We modify the approach of Ramsey, de Finetti, Savage and Anscombe and Aumann in giving axioms for a theory of robust preferences. We establish that preferences which satisfy axioms for robust preferences can be represented by a set of expected utilities. In the presence of two axioms relating to state-independent utility, robust preferences are represented by a set of probability/utility pairs, where the utilities are almost state-independent (in a sense which we make precise). Our goal is to focus on preference alone and to extract whatever probability and/or utility information is contained in the preference relation when that is merely a partial order. This is in contrast with the usual approach to Bayesian robustness that begins with a class of "priors" or "likelihoods," and a single loss function, in order to derive preferences from these probability/utility assumptions.

This paper proposes a fully sequential procedure for constructing a fixed-width confidence band for an unknown density on a finite interval and shows the procedure has the desired coverage probability asymptotically as the width of the band approaches zero. The procedure is based on a result of Bickel and Rosenblatt. Its implementation in the sequential setting cannot be obtained using Anscombe's theorem, because the normalized maximal deviations between the kernel estimate and the true density are not uniformly continuous in probability (u.c.i.p.). Instead, we obtain a slightly weaker version of the u.c.i.p. property and a correspondingly stronger convergence property of the stopping rule. These together yield the desired results.

We address the empirical bandwidth choice problem in cases where the range of dependence may be virtually arbitrarily long. Assuming that the observed data derive from an unknown function of a Gaussian process, it is argued that, unlike more traditional contexts of statistical inference, in density estimation there is no clear role for the classical distinction between short- and long-range dependence. Indeed, the "boundaries" that separate different modes of behaviour for optimal bandwidths and mean squared errors are determined more by kernel order than by traditional notions of strength of dependence, for example, by whether or not the sum of the covariances converges. We provide surprising evidence that, even for some strongly dependent data sequences, the asymptotically optimal bandwidth for independent data is a good choice. A plug-in empirical bandwidth selector based on this observation is suggested. We determine the properties of this choice for a wide range of different strengths of dependence. Properties of cross-validation are also addressed.

We show that, for $L^p$ convergence of the mode of a nonparametric density estimator to the mode of an unknown probability density, finiteness of the pth moment of the underlying distribution is both necessary and sufficient. The basic requirement of existence of finite variance has been overlooked by statisticians, who have earlier considered mean square convergence of nonparametric mode estimators; they have focussed on mean squared error of the asymptotic distribution, rather than on asymptotic mean squared error. The effect of bandwidth choice on the rate of $L^p$ convergence is analysed, and smoothed bootstrap methods are used to develop an empirical approximation to the $L^p$ measure of error. The resulting bootstrap estimator of $L^p$ error may be minimised with respect to the bandwidth of the nonparametric density estimator, and in this way an empirical rule may be developed for selecting the bandwidth for mode estimation. Particular attention is devoted to the problem of selecting the appropriate amount of smoothing in the bootstrap algorithm.

Various smoothing methods for quantile density estimation are unified into a generalized kernel smoothing. Based on a stochastic upper bound of the derivatives sequence for a sequence of smoothed Brownian bridges, uniform in-probability consistency of generalized kernel quantile density estimators on any closed subinterval of the open unit interval is derived.