Examples are presented of statistical techniques for the analysis of random process data and of their uses in the substantive fields of seismology and neurophysiology. The problems addressed include frequency estimation for decaying cosinusoids, signal estimation, association measurement, causal connection assessment, estimation of speed and direction and structural modeling. The techniques employed include complex demodulation, nonlinear regression, probit analysis, deconvolution, maximum likelihood, singular value decomposition, Fourier analysis and averaging.

Best linear unbiased predictors of a random field can be obtained if the covariance function of the random field is specified correctly. Consider a random field defined on a bounded region $R$. We wish to predict the random field $z(\cdot)$ at a point $x$ in $R$ based on observations $z(x_1), z(x_2), \ldots, z(x_N)$ in $R$, where $\{x_i\}^\infty_{i = 1}$ has $x$ as a limit point but does not contain $x$. Suppose the covariance function is misspecified, but has an equivalent (mutually absolutely continuous) corresponding Gaussian measure to the true covariance function. Then the predictor of $z(x)$ based on $z(x_1), \ldots, z(x_N)$ will be asymptotically efficient as $N$ tends to infinity.

A case-cohort design was recently proposed [Prentice (1986)] as a means of reducing cost in large epidemiologic cohort studies. A "pseudolikelihood" procedure was described for relative risk regression parameter estimation. This procedure involves covariate data only on subjects who develop disease and on a random subset of the entire cohort. In contrast, the usual partial likelihood estimation procedure requires covariate histories on the entire cohort. Accordingly, a case-cohort design may affect cost saving, particularly with large cohorts and infrequent disease occurrence. Asymptotic distribution theory for such pseudolikelihood estimators, along with that for corresponding cumulative failure rate estimators, are presented here. Certain asymptotic efficiency expressions relative to full-cohort estimators are developed and tabulated in situations of relevance to the design of large-scale disease prevention trials. The theoretical developments make use of martingale convergence results and finite population convergence results.

This paper is concerned with parametric regression models of the form $Y_{ij} = f(t_{ij}, \theta_i) + \text{error}, i = 1, \ldots, n, j = 1, \ldots, T_i$, where the continuous function $f$ may depend nonlinearly on the known regressors $t_{ij}$ and the unknown parameter vectors $\theta_i$. The assumption of an a priori known $f$ is dropped and replaced by the requirement that qualitative information about the structure of the model is available or can be generated by a preliminary exploratory data analysis. This framework--allowing both $f$ and the individual parameter vectors to be unknown--necessitates a detailed discussion of identifiability of model and parameters. A method is then proposed for the simultaneous estimation of $f$ and $\theta_i$ by making use of the prior information. An iterative algorithm simplifying computation of the estimates is presented, and for $\min\{n, T_1, \ldots, T_n\} \rightarrow \infty$ conditions for strong uniform consistency of the resulting estimators of $f$ and strong consistency of the estimators of $\theta_i$ are established. Some examples illustrating the method are included.

Cox and Koh (1986) considered the model $y_i = f(x(i)) + \varepsilon_i, \varepsilon_i$ i.i.d. $N(0, \sigma^2)$, with the (parametric) null hypothesis $f(x), x \in \lbrack 0, 1 \rbrack$, a polynomial of degree $m - 1$ or less, versus the alternative $f$ is "smooth," based on the Bayesian model for $f$ which leads to polynomial smoothing spline estimates for $f$. They showed that there was no uniformly most powerful test and found the locally most powerful (LMP) test. We extend their result to the generalized smoothing spline models of Wahba (1985) and to the partial spline models proposed and studied by Engle, Granger, Rice and Weiss (1986), Shiller (1984), Green, Jennison and Seheult (1985), Wahba (1984), Heckman (1986) and others. We also show that the test statistic has an intimate relationship with the behavior of the generalized cross validation (GCV) function at $\lambda = \infty$. If the GCV function has a minimum at $\lambda = \infty$, then GCV has chosen the (parametric) model corresponding to the null hypothesis; we show that if the LMP test statistic is no larger than a certain multiple of the residual sum of squares after (parametric) regression, then the GCV function will have a (possibly local) minimum at $\lambda = \infty$.

Let $\{X_i, Y_i\}^n_{i=1} \subset \mathbb{R}^d \times \mathbb{R}$ be independent identically distributed random variables. If the conditional distribution $F(y \mid x)$ can be parametrized by $F(y \mid x) = F_0((y - m(x))/\sigma(x))$ with a fixed and known distribution $F_0$, the regression curve $m(x)$ and scale curve $\sigma(x)$ could be estimated by some parametric method. More generally, we assume that $F$ is unknown and consider nonparametric simultaneous $M$-type estimates of the unknown functions $m(x)$ and $\sigma(x)$, using kernel estimators for the conditional distribution function $F(y \mid x)$. We show pointwise consistency and asymptotic normality of these estimates. The rate of convergence is optimal in the sense of Stone (1980). The asymptotic bias term of this robust estimate turns out to be the same as for the linear Nadaraya-Watson kernel estimate.

Consider the regression model $Y_i = X'_i\beta + g(t_i) + e_i$ for $i = 1, \cdots, n$. Here $g$ is an unknown Holder continuous function of known order $p$ in $R, \beta$ is a $k \times 1$ parameter vector to be estimated and $e_i$ is an unobserved disturbance. Such a model is often encountered in situations in which there is little real knowledge about the nature of $g$. A piecewise polynomial $g_n$ is proposed to approximate $g$. The least-squares estimator $\hat\beta$ is obtained based on the model $Y_i = X'_i\beta + g_n(t_i) + e_i$. It is shown that $\hat\beta$ can achieve the usual parametric rates $n^{-1/2}$ with the smallest possible asymptotic variance for the case that $X$ and $T$ are correlated.

Estimation of the parameters of the nonlinear functional model with known error covariance matrix is discussed. Asymptotic properties of the maximum likelihood estimator for the implicit functional model are presented. The approximate bias in the maximum likelihood estimator due to the nonlinearity of the relationship is given and a bias-adjusted estimator is suggested. Numerical and theoretical results support the superiority of the bias-adjusted estimator relative to the maximum likelihood estimator.

The selection of the order, i.e., number of vanishing moments, of the kernel in a kernel density estimator is considered from two points of view. First, theoretical properties are investigated by a mean integrated squared error analysis of the problem. Second, and perhaps more importantly, cross validation is proposed as a practical method of choice, and theoretical backing for this is provided through an asymptotic optimality result.

The multivariate calibration problem is considered, in which a sample of $n$ observations on vectors $\xi_{(i)}$ (of "true values") and $Y_{(i)}$ (of less accurate but more easily obtained values) are to be used to estimate the unknown $\xi$ corresponding to a future $Y$. It is assumed that $Y = BX + \varepsilon$, where $\varepsilon$ is multivariate normal and $X = h(\xi)$ for known $h$. Current methods for obtaining a confidence region $C$ for $\xi$, which consist of computing a region $R$ for $X$ and then taking $C = h^{-1}(R)$, have the disadvantage that although the region $R$ might be nicely behaved, the region $C$ need not be. An alternative method is proposed which gives a well-behaved region (corresponding to the uniformly most accurate translation-invariant region when $h$ is linear, $B$ is known and the covariance matrix of $\varepsilon$ is a known multiple of the identity). An application is given to the estimation of gestational age using ultrasound fetal bone measurements.

Consider a sequence of independent random variables $\{X_i: 1 \leq i \leq n\}$ having cdf $F$ for $i \leq \theta n$ and cdf $G$ otherwise. A class of strongly consistent estimators for the change-point $\theta \in (0, 1)$ is proposed. The estimators require no knowledge of the functional forms or parametric families of $F$ and $G$. Furthermore, $F$ and $G$ need not differ in their means (or other measure of location). The only requirement is that $F$ and $G$ differ on a set of positive probability. The proof of consistency provides rates of convergence and bounds on the error probability for the estimators. The estimators are applied to two well-known data sets, in both cases yielding results in close agreement with previous parametric analyses. A simulation study is conducted, showing that the estimators perform well even when $F$ and $G$ share the same mean, variance and skewness.

A sequence of independent random variables $X_1, X_2, \cdots, X_N$ is said to have a change point if $X_1, X_2, \cdots, X_n$ have a common distribution $F$ and $X_{n+1}, \cdots, X_N$ have a common distribution $G, G \neq F$. Consider the problem of testing the null hypothesis of no change against the alternative of a change $G < F$ at an unknown change point $n$. Two classes of statistics based upon two-sample linear rank statistics (max- and sum-type) are compared in terms of their Bahadur efficiency. It is shown that for every sequence of sum-type statistics a sequence of max-type statistics can be constructed with at least the same Bahadur slope at all possible alternatives. Special attention is paid to alternatives close to the null hypothesis.

The LBI (locally best invariant) test is suggested under normality for the constancy of regression coefficients against the alternative hypothesis that one component of the coefficients follows a random walk process. We discuss the limiting null behavior of the test statistic without assuming normality under two situations, where the initial value of the random walk process is known or unknown. The limiting distribution is that of a quadratic functional of Brownian motion and the characteristic function is obtained from the Fredholm determinant associated with a certain integral equation. The limiting distribution is then computed by numerical inversion of the characteristic function.

The problem of detecting a change in drift of Brownian motion is considered in the Bayesian framework with the time of change having a (prior) exponential distribution. To the well known problem of finding an optimal stopping rule for "declaring a change," we add the option of continuously controlling the sampling rates--resulting in controlling the variance coefficient of the process. The combined problem of finding an optimal rate function (dynamic sampling) together with an optimal stopping rule is solved and explicit expressions for the quantities of interest are derived. The dynamic sampling procedure is shown to be significantly superior to constant rate sampling. The comparison is most favorable when the expected time until change tends to infinity, where the relative efficiency between the two procedures tends to infinity.

A general model for a two-armed bandit with delayed responses is introduced and solved with dynamic programming. One arm has geometric lifetime with parameter $\theta$, which has prior distribution $\mu$. The other arm has known lifetime with mean $\kappa$. The response delays completely change the character of the optimal strategies from the no delay case; in particular, the bandit is no longer a stopping problem. The delays also introduce an extra parameter $p$ into the state space. In clinical trial applications, this parameter represents the number of patients previously treated with the unknown arm who are still living. The value function is introduced and investigated as $p, \mu$ and $\kappa$ vary. Under a regularity condition on the discount sequence, there exists a manifold in the state space such that both arms are optimal on the manifold, arm $x$ is optimal on one side and arm $y$ on the other. Properties of the manifold are described.

Simultaneous confidence regions and simultaneous prediction regions are considered using the expected coverage measure criterion of Naiman. Invariance methods are used to construct minimax procedures in multivariate linear models.

This paper characterizes admissible linear estimators of multiple Poisson parameters under entropy loss. Estimators dominating some of the standard estimators are given. Further, hierarchical Bayes estimators are generated and conditions under which they dominate the standard estimators are also given. Monte Carlo simulations are undertaken to indicate the extent of the risk dominance.

We investigate the problem of testing multinormality against alternatives invariant with respect to the affine group of transformations $G$ and against left-bounded alternatives defined by Szkutnik. The last problem remains invariant under a suitably chosen subgroup $G^\ast$ of $G$. Using Wijsman's theorem we find general forms of the most powerful $G$- and $G^\ast$-invariant tests for multinormality which opens the way to an extension of the one-dimensional results of Uthoff to the bivariate case. We find explicit forms of tests against bivariate exponential and bivariate uniform alternatives. A Monte Carlo approximation of the power of these tests is given. This provides us with upper bounds for the power of all invariant tests for binormality against the alternatives considered. The maximum property of the tests obtained is also studied.

It is proved in this paper that covariance hypotheses which are linear in both the covariance and the inverse covariance are products of models each of which consists of either (i) independent identically distributed random vectors which have a covariance with a real, complex or quaternion structure or (ii) independent identically distributed random vectors with a parametrization of the covariance which is given by means of the Clifford algebra. The models (i) are well known. For models (ii) we have found, under the assumption that the distribution is normal, the exact distributions of the maximum likelihood estimates and the likelihood ratio test statistics.

Suppose $s$ is a random variate that follows a statistical model with parameter $\omega$, and let $s_1, s_2, \cdots, s_n, \cdots$ be independent and identically distributed observations of $s$. The model is reproductive in $s$ and $\omega$ if for any $n$ the mean $\bar{s} = (s_1 + \cdots + s_n)/n$ follows the same model as $s$ but with parameter $n\omega$ instead of $\omega$. Suitable combinations of reproductive models yield reproductive models for higher-dimensional variates. This combination technique is discussed and illustrated by examples. It is possible, in particular, to construct reproductive combinations of gamma, inverse-Gaussian and Gaussian distributions, determined by a regression structure, which may conveniently be described in graph-theoretic terms. The graph-theoretical interpretation makes it feasible to draw conclusions about conditional independencies in the models concerned, by means of a very general result for Markovian-type probability laws on graphs due to Kiiveri, Speed and Carlin (1984). Most of the models discussed are exponential, of a form, which in conjunction with the reproductivity, implies various useful distributional properties, derivable from the general theory of reproductive exponential models.

Let $X_1, \cdots, X_n$ be iid observations coming from an exponential family. The problem of interest is this: Given a finite number of models $m_j$ (smoothly curved manifolds in $\mathbb{R}^k$), choose the best model to fit the observations, with some penalty for choosing models with dimensions which are too large. A result of Schwarz is made more specific and is extended to the case where the models are curved manifolds. If $S(Y, n, j)$ is--up to a constant $C(n)$ independent of the model--the log of the posterior probability of the $j$th model, where the sample mean $Y_n = (1/n)\sum^n_{i=1}X_i$ has been replaced by $Y$, Schwarz suggested an asymptotic expansion of $S(Y, n, j)$ whose leading terms are $\gamma(Y, n, j) = n \sup_{\psi \in m_j \cap \Theta}(Y\psi - b(\psi)) - \frac{1}{2}k_j \log n$, in the case where the models are affine subspaces of $\mathbb{R}^k$. We establish a similar asymptotic expansion, including the next term, with uniform bounds for $Y$ in a compact neighborhood of $\nabla b(\theta)$, where $\theta$ is the true value of the parameter. We suggest a criterion for the choice of the best model that consists of maximizing the three leading terms in the expansion $S(Y, n, j)$. We show that the criterion gives the correct model with probabilities $P^n_\theta \rightarrow 1$ as $n \rightarrow + \infty$.

Consider a sample of size $n$ from a regular exponential family in $p_n$ dimensions. Let $\hat\theta_n$ denote the maximum likelihood estimator, and consider the case where $p_n$ tends to infinity with $n$ and where $\{\theta_n\}$ is a sequence of parameter values in $R^{p_n}$. Moment conditions are provided under which $\|\hat\theta_n - \theta_n\| = O_p(\sqrt{p_n/n})$ and $\|\hat\theta_n - \theta_n - \overline{X}_n\| = O_p (p_n/n)$, where $\overline{X}_n$ is the sample mean. The latter result provides normal approximation results when $p^2_n/n \rightarrow 0$. It is shown by example that even for a single coordinate of $(\hat\theta_n - \theta_n), p^2_n/n \rightarrow 0$ may be needed for normal approximation. However, if $p^{3/2}_n/n \rightarrow 0$, the likelihood ratio test statistic $\Lambda$ for a simple hypothesis has a chi-square approximation in the sense that $(-2 \log \Lambda - p_n)/\sqrt{2p_n} \rightarrow_D \mathscr{N}(0, 1)$.

An autoregressive process $y_n = \beta_1y_{n-1} + \cdots + \beta_py_{n-p} + \varepsilon_n$ is said to be unstable if the characteristic polynomial $\phi(z) = 1 - \beta_1z - \cdots - \beta_pz^p$ has all roots on or outside the unit circle. The limiting distribution of the least squares estimate of $(\beta_1, \cdots, \beta_p)$ is derived and characterized as a functional of stochastic integrals under a $2 + \delta$ moment assumption on $\varepsilon_n$. Up to the present, distributional results were available only with substantial restrictions on the possible roots which did not suggest the form of the distribution for the general case. To establish the limiting distribution, a result concerning the weak convergence of a sequence of random variables to a stochastic integral, which is of independent interest, is also developed.

The problem of testing a given ARMA model (in which the density of the generating white noise is unspecified) against other ARMA models is considered. A distribution-free asymptotically most powerful test, based on a generalized linear serial rank statistic, is provided against contiguous ARMA alternatives with specified coefficients. In the case when the ARMA model in the alternative has unspecified coefficients, the asymptotic sufficiency (in the sense of Le Cam) of a finite-dimensional vector of rank statistics is established. This asymptotic sufficiency is used to derive an asymptotically maximin most powerful test, based on a generalized quadratic serial rank statistic. The asymptotically maximin optimal test statistic can be interpreted as a rank-based, weighted version of the classical Box-Pierce portmanteau statistic, to which it reduces, in some particular problems, asymptotically and under Gaussian assumptions.

The asymptotic distribution of quadratic forms in uniform order statistics is studied under contiguous alternatives. Using minimal conditions on the sequence of forms, the limiting distribution is shown to be the convolution of a sum of weighted noncentral chi-squares and a normal variate. The results give approximate distribution theory even when no limit exists. As an example, high-order spacings statistics are shown to have trivial asymptotic power unless the order of the spacings grows linearly with the sample size. The results are derived from a modification of an invariance principle for quadratic forms due to Rotar, which we prove by martingale central limit methods.

Though jackknifing serves well the dual purpose of bias reduction and variance estimation, the pseudovariables it generates may not generally preserve robustness for general statistical functionals. These pseudovariables are incorporated in a differentiable functional detour of jackknifing, and along with its rationality, the related asymptotic theory is studied systematically. A two-step jackknifing is considered for the variance estimation. Second-order asymptotic distributional representations for the classical jackknifed estimators are also considered.

In this paper we consider a partial ordering that is "between" the stochastic ordering defined by Lehmann (1955) and an ordering associated with the monotone likelihood ratio property. A tail ordering deduced from it is applied to the comparison of the asymptotic efficiencies of rank tests in the two-sample problem. In particular, we show that the asymptotic relative efficiency of two rank tests preserve this tail ordering if one score function is "more convex" than the other.