Genetic epidemiology is almost unique among the sciences in that computation of a likelihood function is the accepted approach to statistical inference. In the context of genetic linkage analysis, in which genes are mapped by analysing the dependence in inheritance of different traits, the use of likelihood dates back to the early work of Fisher and Haldane, and has seldom been seriously challenged. After introducing the underlying genetic concepts, this paper reviews the history of the statistics of linkage analysis, from 1913 to 1980, and its dependence on the development of likelihood inference.

With the sudden increase in genetic marker data deriving from new DNA technology, the potential for mapping the genes contributing to complex genetic traits is markedly increased, but the difficulties of likelihood analysis are also multiplied. With increasing complexity of models and the desire to make maximum use of available data on individuals not closely related, the likelihood approach to human linkage analysis faces new computational and methodological challenges. New methods are meeting some of these challenges; likelihood and linkage seem as closely interwoven as ever.

In this article we study function estimation via wavelet shrinkage for data with long-range dependence. We propose a fractional Gaussian noise model to approximate nonparametric regression with long-range dependence and establish asymptotics for minimax risks. Because of long-range dependence, the minimax risk and the minimax linear risk converge to 0 at rates that differ from those for data with independence or short-range dependence. Wavelet estimates with best selection of resolution level-dependent threshold achieve minimax rates over a wide range of spaces. Cross-validation for dependent data is proposed to select the optimal threshold. The wavelet estimates significantly outperform linear estimates. The key to proving the asymptotic results is a wavelet-vaguelette decomposition which decorrelates fractional Gaussian noise. Such wavelet-vaguelette decomposition is also very useful in fractal signal processing.

Z.Consider the problem of estimating $\int \Phi(f)$, where $\Phi$ is a smooth function and f is a density with given order of regularity s. Special attention is paid to the case $\Phi(t) = t^3$. It has been shown that for low values of s the $n^{-1/2}$ rate of convergence is not achievable uniformly over the class of objects of regularity s. In fact, a lower bound for this rate is $n^{-4s/(1+4s)}$ for $0 < s \leq 1/4$. As for the upper bound, using a Taylor expansion, it can be seen that it is enough to provide an estimate for the case $\Phi(x) = x^3$. That is the aim of this paper. Our method makes intensive use of special algebraic and wavelet properties of the Haar basis.

Density estimation is a commonly used test case for nonparametric estimation methods. We explore the asymptotic properties of estimators based on thresholding of empirical wavelet coefficients. Minimax rates of convergence are studied over a large range of Besov function classes $B_{\sigma pq}$ and for a range of global $L'_p$ error measures, $1 \leq p' < \infty$. A single wavelet threshold estimator is asymptotically minimax within logarithmic terms simultaneously over a range of spaces and error measures. In particular, when $p' > p$, some form of nonlinearity is essential, since the minimax linear estimators are suboptimal by polynomial powers of n. A second approach, using an approximation of a Gaussian white-noise model in a Mallows metric, is used to attain exactly optimal rates of convergence for quadratic error $(p' = 2)$.

The maximum likelihood estimator (MLE) for the proportional hazards model with "case 1" interval censored data is studied. It is shown that the MLE for the regression parameter is asymptotically normal with $\sqrt{n}$ convergence rate and achieves the information bound, even though the MLE for the baseline cumulative hazard function only converges at $n^{1/3}$ rate. Estimation of the asymptotic variance matrix for the MLE of the regression parameter is also considered. To prove our main results, we also establish a general theorem showing that the MLE of the finite-dimensional parameter in a class of semiparametric models is asymptotically efficient even though the MLE of the infinite-dimensional parameter converges at a rate slower than $\sqrt{n}$. The results are illustrated by applying them to a data set from a tumorigenicity study.

In survival analysis with censored data, we consider three closely related survival function estimators: the Kaplan-Meier, Nelson and moment estimators. We derive the Edgeworth expansions for these three estimators with Studentization. Edgeworth expansions for the corresponding bootstrap statistics are also given. It is found that the bootstrap approximation is better than the normal approximation for the Studentized Kaplan-Meier and Nelson estimators, but not so for the Studentized moment estimator. With these results, we construct bootstrap-based confidence intervals with better coverage probabilities. We also include some simulations which show strong agreement with our theoretical findings.

The NPMLE in the bivariate censoring model is not consistent for continuous data. The problem is caused by the singly censored observations. In this paper we prove that if we observe the censoring times or if the censoring times are discrete, then a NPMLE based on a slightly reduced data set, in particular, we interval censor the singly censored observations, is asymptotically efficient for this reduced data and moreover if we let the width of the interval converge to zero slowly enough, then the NPMLE is also asymptotically efficient for the original data. We are able to determine a lower bound for the rate at which the bandwidth should converge to zero. Simulation results show that the estimator for small bandwidths has a very goodperformance. The efficiency proof uses a general identity which holds for NPMLE of a linear parameter in convex models. If we neither observe the censoring times nor the censoring times are discrete, then we conjecture that our estimator based on simulated censoring times is also asymptotically efficient.

The purpose of this paper is to derive confidence bands for quantile functions using a nonparametric likelihood ratio approach. The method is easy to implement and has several appealing properties. It applies to right-censored and left-truncated data, and it does not involve density estimation or even require the existence of a density. Previous approaches (e.g., bootstrap) have imposed smoothness conditions on the density. The performance of the proposed method is investigated in a Monte Carlo study, and an application to real data is given.

Hall and Wellner proposed a natural extension of the Kolmogorov-Smirnov simultaneous confidence band for survival curve using the Kaplan-Meier estimator. They and Gill conjectured that the confidence band holds for all t up to the last observed failure time. A counterexample is given herein, showing that this may not always be true.

A model of interval censorship of a failure time T is considered when there is only one inspection time Y. The observable data are n independent copies of the pair $(Y, \delta)$, where $\delta = [T \leq Y]$. We construct a class of self-consistent estimators of the survival function of T defined implicitly through two equations and show their strong consistency under certain conditions. The properties of the onparametric maximum likelihood estimator are also investigated.

We consider the problem of estimating a functional of a density of the type $\int \phi (f, \cdot)$. Starting from efficient estimators of linear and quadratic functionals of f and using a Taylor expansion of $\phi$, we build estimators that achieve the $n^{-1/2}$ rate whenever f is smooth enough. Moreover, we show that these estimators are efficient. Concerning the estimation of quadratic functionals (more precisely, of integrated squared density) Bickel and Ritov have already built efficient estimators. We propose here an alternative construction based on projections, which seems more natural.

Bickel and Ritov suggested an optimal estimator for the integral of the square of the kth derivative of a density when the unknown density belongs to a Lipschitz class of a given order $\beta$. In this context optimality means that the estimate is asymptotically efficient, that is, it has the best constant and rate of risk convergence, whenever $\beta > 2k + 1/4$, and it is rate optimal otherwise. The suggested optimal estimator crucially depends on the value of $\beta$ which is obviously unknown. Bickel and Ritov conjectured that the method of cross validation leads to a corresponding adaptive estimator which has the same optimal statistical properties as the optimal estimator based on prior knowledge of $\beta$.

We show for probability densities supported over a finite interval that when $\beta > 2k + 1/4$ adaptation is not necessary for the construction of an asymptotically efficient estimator. On the other hand, it is not possible to construct an adaptive estimator which has the same rate of convergence as the optimal nonadaptive estimator as soon as $k < \beta \leq 2k + 1/4$.

We present general sufficient conditions for the almost sure $L_1$-consistency of histogram density estimates based on data-dependent partitions. Analogous conditions guarantee the almost-sure risk consistency of histogram classification schemes based on data-dependent partitions. Multivariate data are considered throughout.

In each case, the desired consistency requires shrinking cells, subexponential growth of a combinatorial complexity measure and sublinear growth of the number of cells. It is not required that the cells of every partition be rectangles with sides parallel to the coordinate axis or that each cell contain a minimum number of points. No assumptions are made concerning the common distribution of the training vectors.

We apply the results to establish the consistency of several known partitioning estimates, including the $k_n$-spacing density estimate, classifiers based on statistically equivalent blocks and classifiers based on multivariate clustering schemes.

In the multivariate calibration problem using a multivariate linear model, some conservative confidence regions are constructed. The regions are nonempty and invariant under nonsingular transformations. Situations where the explanatory variable occurs nonlinearly in the model are also considered. Computational aspects of the confidence region and its practical implementation are discussed. The results are illustrated using two examples. The examples show that our confidence regions are much more satisfactory compared to those based on some of the existing procedures. Furthermore, simulation results for the examples reveal that the coverage probability of the conservative confidence regions are very close to the assumed confidence level.

We discuss maximum likelihood methods for fitting a broad class of multivariate categorical response data models. In particular, we derive the large-sample distributions for maximum likelihood estimators of parameters of product-multinomial generalized log-linear models. The large-sample behavior of other relevant likelihood-based statistics such as goodness-of-fit statistics and adjusted residuals is also described. The asymptotic results are derived within the framework of the constraint specification, rather than the more common freedom specification, of the model. We also outline an improved fitting algorithm for computing parameter maximum likelihood estimates and other relevant statistics. The broad class of multivariate categorical response data models, which are referred to as generalized log-linear models, can imply structure on several response configuration distributions (e.g., joint and marginal distributions). These models, which include as special cases log-linear, logit and cumulative-logit models, enjoy a wide breadth of application including longitudinal, rater-agreement and crossover data analyses.

Under appropriate conditions the probability of a convex symmetric set decreases as the spread or scatter of the distribution increases. This paper studies the conditions when the random vector has a symmetric unimodal distribution.

We characterize the Wishart distributions on a symmetric cone C. If $C = (0, +\infty)$, this has been done by Lukacs in 1955. If C is the cone of positive definite symmetric matrices, this has been done by Olkin and Rubin in 1962. We both shorten and extend the Olkin-Rubin proof (sometimes obscure) by using three modern ideas: (i) try to avoid artificial coordinates in differential geometry; (ii) the variance function of a natural exponential family F characterizes F; (iii) symmetric matrices are a particular example of a Euclidean simple Jordan algebra.

Some properties of the ordinary two-parameter Cauchy family, the circular or wrapped Cauchy family, and their connection via Möbius transformation are discussed. A key simplification is achieved by taking the parameter $\theta = \mu + i \sigma$ to be a point in the complex plane rather than the real plane. Maximum likelihood estimation is studied in some detail. It is shown that the density of any equivariant estimator is harmonic on the upper half-plane. In consequence, the maximum likelihood estimator is unbiased for $n \geq 3$, and every harmonic or analytic function of the maximum likelihood estimator is unbiased if its expectation is finite. The joint density of the maximum likelihood estimator is obtained in exact closed form for samples of size $n \leq 4$, and in approximate form for $n \geq 5$. Various marginal distributions, including that of Student's pivotal ratio, are also obtained. Most results obtained in the context of the real Cauchy family also apply to the wrapped Cauchy family by Möbius transformation.

For a broad class of error distributions that includes the spherically symmetric ones, we give a short proof that the usual estimator of the mean in a d-dimensional shift model is inadmissible under quadratic loss when $d \geq 3$. Our proof involves representing the error distribution as that of a stopped Brownian motion and using elementary stochastic analysis to obtain a generalization of an integration by parts lemma due to Stein in the Gaussian case.

We give a geometric proof that the estimates of a regression model derived by using partial least squares shrink the ordinary least squares estimates. The proof is based on a sequential construction algorithm of partial least squares. A discussion of the nature of shrinkage is included.

A two-way multivariate normal model is proposed and attention is focused on estimation of the mean values when the common variance of the observations is unknown. A class of empirical Bayes estimators is proposed and mean-squared errors are given. A lower bound on the mean-squared error is found and related to risk asymptotics. A James-Stein-type estimator is derived and compared with its competitor--a modal estimator that is obtained from a hierarchical prior for the unknown parameters.

Tests of hypotheses about finite-dimensional parameters in a semiparametric model are studied from Pitman's moving alternative (or local) approach using Le Cam's local asymptotic normality concept. For the case of a real parameter being tested, asymptotically uniformly most powerful (AUMP) tests are characterized for one-sided hypotheses, and AUMP unbiased tests for two-sided ones. An asymptotic invariance principle is introduced for multidimensional hypotheses, and AUMP invariant tests are characterized. These provide optimality for Wald, Rao (score), Neyman-Rao (effective score) and likelihood ratio tests in parametric models, and for Neyman-Rao tests in semiparametric models when constructions are feasible. Inversions lead to asymptotically uniformly most accurate confidence sets. Examples include one-, two- and k-sample problems, a linear regression model with unknown error distribution and a proportional hazards regression model with arbitrary baseline hazards. Results are presented in a format that facilitates application in strictly parametric models.

We consider maximum likelihood estimation in several examples of semiparametric mixture models, including the exponential frailty model and the errors-in-variables model. The observations consist of a sample of size n from the mixture density $\int p_{\theta}(x|z) d \eta(z)$. The mixing distribution is completely unknown. We show that the first component $\hat{\theta}_n$ of the joint maximum likelihood estimator , $(\hat{\theta}_n \hat{\eta}_n)$ is asymptotically normal and asymptotically efficient in the semiparametric sense.

In this paper we show that the well-known asymptotic efficiency bounds for full mixture models remain valid if individual sequences of nuisance parameters are considered. This is made precise both for some classes of random (i.i.d.) and nonrandom nuisance parameters. For the random case it is shown that superefficiency of the kind given by an example of Pfanzagl can happen only with low probability. The nonrandom case deals with permutation-invariant estimators under one-dimensional nuisance parameters. It is shown that the efficiency bounds remain valid for individual nonrandom arrays of nuisance parameters whose empirical process, if it is centered around its limit and standardized, satisfies a compactness condition. The compactness condition is satisfied in the random case with high probability. The results make use of basic LAN theory. Regularity conditions are stated in terms of $L^2$-differentiability.

Some conditions which are usually found in the literature on sufficiency and invariance are considered, with counterexamples given to clarify the relationship between these conditions.

In this paper results and counterexamples are given to study the relationship between some conditions which appear in the literature on sufficiency, invariance and conditional independence.