A data depth can be used to measure the “depth” or “outlyingness” of a given multivariate sample with respect to its underlying distribution. This leads to a natural center-outward ordering of the sample points. Based on this ordering, quantitative and graphical methods are introduced for analyzing multivariate distributional characteristics such as location, scale, bias, skewness and kurtosis, as well as for comparing inference methods. All graphs are one-dimensional curves in the plane and can be easily visualized and interpreted. A “sunburst plot” is presented as a bivariate generalization of the box-plot. DD-(depth versus depth) plots are proposed and examined as graphical inference tools. Some new diagnostic tools for checking multivariate normality are introduced. One of them monitors the exact rate of growth of the maximum deviation from the mean, while the others examine the ratio of the overall dispersion to the dispersion of a certain central region. The affine invariance property of a data depth also leads to appropriate invariance properties for the proposed statistics and methods.

We study a simple “horizon model” for the problem of recovering an image from noisy data; in this model the image has an edge with $\alpha$-Hölder regularity. Adopting the viewpoint of computational harmonic analysis, we develop an overcomplete collection of atoms called wedgelets, dyadically organized indicator functions with a variety of locations, scales and orientations. The wedgelet representation provides nearly optimal representations of objects in the horizon model, as measured by minimax description length. We show how to rapidly compute a wedgelet approximation to noisy data by finding a special edgelet-decorated recursive partition which minimizes a complexity-penalized sum of squares. This estimate, using sufficient subpixel resolution, achieves nearly the minimax mean-squared error in the horizon model. In fact, the method is adaptive in the sense that it achieves nearly the minimax risk for any value of the unknown degree of regularity of the horizon, $1 \leq \alpha \leq 2$. Wedgelet analysis and denoising may be used successfully outside the horizon model. We study images modelled as indicators of star-shaped sets with smooth boundaries and show that complexity-penalized wedgelet partitioning achieves nearly the minimax risk in that setting also.

We study wavelet function estimation via the approach of block thresholding and ideal adaptation with oracle. Oracle inequalities are derived and serve as guides for the selection of smoothing parameters. Based on an oracle inequality and motivated by the data compression and localization properties of wavelets, an adaptive wavelet estimator for nonparametric regression is proposed and the optimality of the procedure is investigated. We show that the estimator achieves simultaneously three objectives: adaptivity, spatial adaptivity and computational efficiency. Specifically, it is proved that the estimator attains the exact optimal rates of convergence over a range of Besov classes and the estimator achieves adaptive local minimax rate for estimating functions at a point. The estimator is easy to implement, at the computational cost of $O(n)$. Simulation shows that the estimator has excellent numerical performance relative to more traditional wavelet estimators.

This paper is motivated by the problem of detecting local changes or differences in shape between two samples of objects via the nonlinear deformations required to map each object to an atlas standard. Local shape changes are then detected by high values of the random field of Hotelling’s $T^2$ statistics for detecting a change in mean of the vector deformations at each point in the object. To control the null probability of detecting a local shape change, we use the recent result of Adler that the probability that a random field crosses a high threshold is very accurately approximated by the expected Euler characteristic (EC) of the excursion set of the random field above the threshold. We give an exact expression for the expected EC of a Hotelling’s $T^2$ field, and we study the behavior of the field near local extrema. This extends previous results for Gaussian random fields by Adler and $\chi^2$, $t$ and $F$ fields by Worsley and Cao. For illustration, these results are applied to the detection of differences in brain shape between a sample of 29 males and 23 females.

We consider score tests of the null hypothesis $H_0: \theta = 1/2$ against the alternative hypothesis $H_1: 0 \leq \theta < 1/2$, based upon counts multinomially distributed with parameters $n$ and $\rho(\theta,\pi)_{1 \times m} = \pi_{1\times m}T(\theta)_{m \times m}$, where $T(\theta)$ is a transition matrix with $T(0)=I$ , the identity matrix, and $T(1/2)=(1,\dots,1)^T (\alpha_1,\dots,\alpha_m)$. This type of testing problem arises in human genetics when testing the null hypothesis of no linkage between a marker and a disease susceptibility gene, using identity by descent data from families with affected members. In important cases in this genetic context, the score test is independent of the nuisance parameter $\pi$ and based on a widely used test statistic in linkage analysis. The proof of this result involves embedding the states of the multinomial distribution into a continuous-time Markov chain with infinitesimal generator $Q$. The second largest eigenvalue of $Q$ and its multiplicity are key in determining the form of the score statistic. We relate $Q$ to the adjacency matrix of a quotient graph in order to derive its eigenvalues and eigenvectors.

We consider hierarchical mixtures-of-experts (HME) models where exponential family regression models with generalized linear mean functions of the form $\psi(\alpha + \mathbf{x}^T \mathbf{\beta})$ are mixed. Here $\psi(\cdot)$ is the inverse link function. Suppose the true response $y$ follows an exponential family regression model with mean function belonging to a class of smooth functions of the form $\psi(h(\mathbf{x}))$ where $h(\cdot)\in W_{2; K_0}^{\infty}$ (a Sobolev class over $[0, 1]^s$). It is shown that the HME probability density functions can approximate the true density, at a rate of $O(m^{-2/s})$ in Hellinger distance and at a rate of $O(m^{-4/s})$ in Kullback–Leibler divergence, where $m$ is the number of experts, and $s$ is the dimension of the predictor $x$. We also provide conditions under which the mean-square error of the estimated mean response obtained from the maximum likelihood method converges to zero, as the sample size and the number of experts both increase.

In this paper we study the problem of testing the functional form of a given regression model. A consistent test is proposed which is based on the difference of the least squares variance estimator in the assumed regression model and a nonparametric variance estimator. The corresponding test statistic can be shown to be asymptotically normal under the null hypothesis and under fixed alternatives with different rates of convergence corresponding to both cases. This provides a simple asymptotic test, where the asymptotic results can also be used for the calculation of the type II error of the procedure at any particular point of the alternative and for the construction of tests for precise hypotheses. Finally, the finite sample performance of the new test is investigated in a detailed simulation study, which also contains a comparison with the commonly used tests.

We are concerned here with establishing the consistency and asymptotic normality for the maximum likelihood estimator of a “merit vector” $(u_0,\dots,u_t)$, representing the merits of $t +1$ teams (players, treatments, objects), under the Bradley–Terry model, as $t \to \infty$. This situation contrasts with the well-known Neyman–Scott problem under which the number of parameters grows with $t$ (the amount of sampling), and for which the maximum likelihood estimator fails even to attain consistency. A key feature of our proof is the use of an effective approximation to the inverse of the Fisher information matrix. Specifically, under the Bradley–Terry model, when teams $i$ and $j$ with respective merits $u_i$ and $u_j$ play each other, the probability that team $i$ prevails is assumed to be $u_i/(u_i + u_j)$. Suppose each pair of teams play each other exactly $n$ times for some fixed $n$. The objective is to estimate the merits, $u_i$’s, based on the outcomes of the $nt(t +1)/2$ games. Clearly, the model depends on the $u_i$’s only through their ratios. Under some condition on the growth rate of the largest ratio $u_i/u_j (0 \leq i, j \leq t)$ as $t \to \infty$, the maximum likelihood estimator of $(u_1/u_0,\dots,u_t/u_0)$ is shown to be consistent and asymptotically normal. Some simulation results are provided.

A central limit theorem for generalized trimmed $k$-means is obtained in a very general framework that covers the multivariate setting, general penalty functions and general $k \geq 1$. Several applications, including the location estimator case $(k = 1)$ for elliptical distributions and the construction of multivariate (not necessarily connected) tolerance zones, are also given.

Given a statistical functional T and a parametric family of distributions, a bias reduced functional $\tilde{T}$ is defined by setting the expected value of the statistic equal to the observed value. Under certain regularity conditions this new statistic, called the target estimator, will have smaller bias and mean square error than the original estimator. The theoretical aspects are analyzed by using higher order von Mises expansions. Several examples are given, including $M$-estimates of location and scale. The procedure is applied to an autoregressive model, the errors-in-variables model and the logistic regression model. A comparison with the jackknife and the bootstrap estimators is also included.

The connection between large and small deviation results for the signed square root statistic $R$ is studied, both for likelihoods and for likelihood-like criterion functions. We show that if $p - 1$ Barlett identities are satisfied to first order, but the $p$th identity is violated to this order, then $\operator{cum}_q(R) = O(n^{-q/2})$ for $3 \leq q < p$, whereas $\operator{cum}_p(R) = \kappa_p n^{-(p-2)/2} + O(n^{-p/2})$. We also show that the large deviation behavior of $R$ is determined by the values of $p$ and $\kappa_p$. The latter result is also valid for more general statistics. Affine (additive and/or multiplicative) correction to $R$ and $R^2$ are special cases corresponding to $p = 3$ and 4. The cumulant behavior of $R$ gives a way of characterizing the extent to which $R$-statistics derived from criterion functions other than log likelihoods can be expected to behave like ones derived from true log likelihoods, by looking at th number of Bartlett identities that are satisfied. Empirical and nonparametric survival analysis type likelihoods are analyzed from this perspective via the device of “dual criterion functions.”