If there are many independent, identically distributed observations governed by a smooth, finite-dimensional statistical model, the Bayes estimate and the maximum likelihood estimate will be close. Furthermore, the posterior distribution of the parameter vector around the posterior mean will be close to the distribution of the maximum likelihood estimate around truth. Thus, Bayesian confidence sets have good frequentist coverage properties, and conversely. However, even for the simplest infinite-dimensional models, such results do not hold. The object here is to give some examples.

In the regression model $\mathbf{Y} = \eta + \mathbf{BX} + \mathbf{Z}$ with $\mathbf{Z}$ unobserved, $\mathscr{E}\mathbf{Z} = \mathbf{0}$ and $\mathscr{E}\mathbf{ZZ}' = \mathbf{\Sigma}_{ZZ}$, the least squares estimator of $\mathbf{B}$ is $\hat{\mathbf{B}} = \mathbf{S}_{YX}\mathbf{S}_{XX}^{-1}$. If the rank of $\mathbf{B}$ is known to be $k$ less than the dimensions of $\mathbf{Y}$ and $\mathbf{X}$, the reduced rank regression estimator of $\mathbf{B}$, say $\mathbf{B}_k$, depends on the first $k$ canonical variates of $\mathbf{Y}$ and $\mathbf{X}$. The asymptotic distribution of $\hat{\mathbf{B}}_k$ is obtained and compared with the asymptotic distribution of $\hat{\mathbf{B}}$. The advantage of $\hat{\mathbf{B}}_k$ is characterized.

Strong consistency for maximum quasi-likelihood estimators of regression parameters in generalized linear regression models is studied. Results parallel to the elegant work of Lai, Robbins and Wei and Lai and Wei on least squares estimation under both fixed and adaptive designs are obtained. Let $y_1,\dots, y_n$ and $x_1,\dots, x_n$ be the observed responses and their corresponding design points ($p \times 1$ vectors), respectively. For fixed designs, it is shown that if the minimum eigenvalue of $\Sigma x_i x^\prime_i$ goes to infinity, then the maximum quasi-likelihood estimator for the regression parameter vector is strongly consistent. For adaptive designs, it is shown that a sufficient condition for strong consistency to hold is that the ratio of the minimum eigenvalue of $\Sigma x_i \x^\prime_i$ to the logarithm of the maximum eigenvalues goes to infinity. Use of the results for the adaptive design case in quantal response experiments is also discussed.

The breakdown point behavior of $M$-estimators in linear models with fixed designs, arising from planned experiments or qualitative factors, is characterized. Particularly, this behavior at fixed designs is quite different from that at designs which can be corrupted by outliers, the situation prevailing in the literature. For fixed designs, the breakdown points of robust $M$-estimators (those with bounded derivative of the score function), depend on the design and the variation exponent (index) of the score function. This general result implies that the highest breakdown point within all regression equivariant estimators can be attained also by certain $M$-estimators: those with slowly varying score function, like the Cauchy or slash maximum likelihood estimator. The $M$-estimators with variation exponent greater than 0, like the $L_1$ or Huber estimator, exhibit a consider-ably worse breakdown point behavior.

In this paper, we address the problem of testing hypotheses using the likelihood ratio test statistic in nonidentifiable models, with application to model selection in situations where the parametrization for the larger model leads to nonidentifiability in the smaller model. We give two major applications: the case where the number of populations has to be tested in a mixture and the case of stationary ARMA$(p, q)$ processes where the order $(p, q)$ has to be tested. We give the asymptotic distribution for the likelihood ratio test statistic when testing the order of the model. In the case of order selection for ARMAs, the asymptotic distribution is invariant with respect to the parameters generating the process. A locally conic parametrization is a key tool in deriving the limiting distributions; it allows one to discover the deep similarity between the two problems.

A novel approach for constructing goodness-of-fit techniques in arbitrary finite dimensions is presented. Testing problems are considered as well as the construction of diagnostic plots. The approach is based on some new notions of mass concentration, and in fact, our basic testing problems are formulated as problems of ‘‘goodness-of-concentration.’’ It is this connection to concentration of measure that makes the approach conceptually simple. The presented test statistics are continuous functionals of certain processes which behave like the standard one-dimensional uniform empirical process. Hence, the test statistics behave like classical test statistics for goodness-of-fit. In particular, for simple hypotheses they are asymptotically distribution free with well-known asymptotic distribution. The simple technical idea behind the approach may be called a generalized quantile transformation, where the role of one-dimensional quantiles in classical situations is taken over by so-called minimum volume sets.

We consider the Wasserstein distance between a sample distribution and the set of normal distributions as a measure of nonnormality. By considering the standardized version of this distance we obtain a version of Shapiro–Wilk’s test of normality. The asymptotic behavior of the statistic is studied using approximations of the quantile process by Brownian bridges. This method differs from the “ad hoc” method of de Wet and Venter and permits a similar analysis for testing other location scale families.

Fractional factorial (FF) designs are commonly used in industrial ex-periments to identify factors affecting a process. When it is expensive or difficult to change the levels of some of the factors, fractional factorial split-plot (FFSP) designs represent a practical design option. Though FFSP de-sign matrices correspond to FF design matrices, the randomization struc-ture of the FFSP design is different. In this paper, we discuss the impact of randomization restrictions on the choice of FFSP designs and develop theoretical results. Some of these results are very closely related to those available for FF designs while others are more specific to FFSP designs and are more useful in practice. We pay particular attention to the min-imum aberration criterion (MA) and emphasize the differences between FFSP and FF designs.

A projective geometric characterization is given of the existence of any regular main effect $s^{n-k}$ design in $s^{\gamma}$ blocks. It leads to a constructive method for finding a maximal blocking scheme for any given fractional factorial design. A useful sufficient condition for admissible block designs is given in terms of the minimum aberration property of a certain unblocked design.

In this paper $D$-optimal designs for the weighted polynomial regression model of degree $p$ with efficiency function $(1 + x^2)^{-n}$ are presented. Interest in these designs stems from the fact that they are equivalent to locally $D$-optimal designs for inverse quadratic polynomial models. For the unrestricted design space $\mathbb{R}$ and $p < n$, the $D$-optimal designs put equal masses on $p + 1$ points which coincide with the zeros of an ultraspherical polynomial, while for $p = n$ they are equivalent to $D$-optimal designs for certain trigonometric regression models and exhibit all the curious and interesting features of those designs. For the restricted design space $[1, 1]$ sufficient, but not necessary, conditions for the $D$-optimal designs to be based on $p + 1$ points are developed. In this case the problem of constructing ($p + 1$)-point $D$-optimal designs is equivalent to an eigenvalue problem and the designs can be found numerically. For $n = 1$ and 2, the problem is solved analytically and, specifically, the $D$-optimal designs put equal masses at the points $\pm 1$ and at the $p - 1$ zeros of a sum of $n + 1$ ultraspherical polynomials. A conjecture which extends these analytical results to cases with $n$ an integer greater than 2 is given and is examined empirically.

Usually, when testing the null hypothesis that a distribution has one mode against the alternative that it has two, the null hypothesis is interpreted as entailing that the density of the sampling distribution has a unique point of zero slope, which is a local maximum. In this paper we argue that a more appropriate null hypothesis is that the density has two points of zero slope, of which one is a local maximum and the other is a shoulder. We show that when a test for a mode-with-shoulder is properly calibrated, so that it has asymptotically correct level, it is generally conservative when applied to the case of a mode without a shoulder. We suggest methods for calibrating both the bandwidth and dip-excess mass tests in the setting of a mode with a shoulder. We also provide evidence in support of the converse: a test calibrated for a single mode without a shoulder tends to be anticonservative when applied to a mode with a shoulder. The calibration method involves resampling from a ‘‘template’’ density with exactly one mode and one shoulder. It exploits the following asymptotic factorization property for both the sample and resample forms of the test statistic: all dependence of these quantities on the sampling distribution cancels asymptotically from their ratio. In contrast to other approaches, the method has very good adaptivity properties.

Groeneboom introduced a jump process that can be used (among other things) to study the asymptotic properties of the Grenander estimator of a monotone density. In this paper we derive the asymptotic normality of a suitably rescaled version of the $L_1$ error of the Grenander estimator, using properties of this jump process.

The nonparametric empirical likelihood approach is used to obtain simultaneous confidence tubes for multiple quantile plots based on $k$ independent (possibly right-censored) samples. These tubes are asymptotically distribution free, except when both $k \geq 3$ and censoring is present. Pointwise versions of the confidence tubes, however, are asymptotically distribution free in all cases. The various confidence tubes are valid under minimal conditions. The proposed methods are applied in three real data examples.

We consider the problem of estimating a mixture proportion using data from two different distributions as well as from a mixture of them. Under the model assumption that the log-likelihood ratio of the two densities is linear in the observations, we develop an empirical likelihood ratio based statistic for constructing confidence intervals for the mixture proportion. Under some regularity conditions, it is shown that this statistic converges to a chi-squared random variable. Simulation results indicate that the performance of this statistic is satisfactory. As a by-product, we give estimators for the two distribution functions. Connections with case-control studies and discrimination analysis are pointed out.

Locally asymptotically optimal tests based on autoregression rank scores are constructed for testing linear constraints on the structural parameters of AR processes. Such tests are asymptotically distribution free and do not require the estimation of nuisance parameters. They constitute robust, flexible and quite powerful alternatives to existing methods such as the classical correlogram-based parametric tests, the Gaussian Lagrange multiplier tests, the optimal non-Gaussian and ranked residual tests described by Kreiss, as well as to the aligned rank tests of Hallin and Puri. Optimality requires a nontrivial extension of existing asymptotic representation results to the case of unbounded score functions (such as the Gaussian quantile function). The problem of testing AR$(p - 1)$ against AR$(p)$ dependence is considered as an illustration. Asymptotic local powers and asymptotic relative efficiencies are explicitly computed. In the special case of van der Waerden scores, the asymptotic relative efficiency with respect to optimal correlogram-based procedures is uniformly larger than one.

This paper discusses the properties of an estimator of the memory parameter of a stationary long-memory time-series originally proposed by Robinson. As opposed to ‘‘narrow-band’’ estimators of the memory parameter (such as the Geweke and Porter-Hudak or the Gaussian semiparametric estimators) which use only the periodogram ordinates belonging to an interval which degenerates to zero as the sample size $n$ increases, this estimator builds a model of the spectral density of the process over all the frequency range, hence the name, “broadband.” This is achieved by estimating the ‘‘short-memory’’ component of the spectral density, $f*(x) = |1 - e^{ix}|^{2d}f(x)$, where $d \epsilon (-1/2, 1/2)$ is the memory parameter and $f(x)$ is the spectral density, by means of a truncated Fourier series estimator of log $f*$. Assuming Gaussianity and additional conditions on the regularity of $f*$ which seem mild, we obtain expressions for the asymptotic bias and variance of the long-memory parameter estimator as a function of the truncation order. Under additional assumptions, we show that this estimator is consistent and asymptotically normal. If the true spectral density is sufficiently smooth outside the origin, this broadband estimator outperforms existing semiparametric estimators, attaining an asymptotic mean-square error $O(\log(n)/n)$ .