Semiparametric Mixtures of Generalized Exponential Families

Abstract.  A semiparametric mixture model is characterized by a non-parametric mixing distribution Q (with respect to a parameter θ ) and a structural parameter β common to all components. Much of the literature on mixture models has focused on fixing β and estimating Q . However, this can lead to inconsistent estimation of both Q and the order of the model m . Creating a framework for consistent estimation remains an open problem and is the focus of this article. We formulate a class of generalized exponential family (GEF) models and establish sufficient conditions for the identifiability of finite mixtures formed from a GEF along with sufficient conditions for a nesting structure. Finite identifiability and nesting structure lead to the central result that semiparametric maximum likelihood estimation of Q and β fails. However, consistent estimation is possible if we restrict the class of mixing distributions and employ an information-theoretic approach. This article provides a foundation for inference in semiparametric mixture models, in which GEFs and their structural properties play an instrumental role.     

EXACT P‐VALUES FOR DISCRETE MODELS OBTAINED BY ESTIMATION AND MAXIMIZATION

In constructing exact tests from discrete data, one must deal with the possible dependence of the P‐value on nuisance parameter(s) ψ as well as the discreteness of the sample space. A classical but heavy‐handed approach is to maximize over ψ. We prove what has previously been understood informally, namely that maximization produces the unique and smallest possible P‐value subject to the ordering induced by the underlying test statistic and test validity. On the other hand, allowing for the worst case will be more attractive when the P‐value is less dependent on ψ. We investigate the extent to which estimating ψ under the null reduces this dependence. An approach somewhere between full maximization and estimation is partial maximization, with appropriate penalty, as introduced by Berger & Boos (1994, P values maximized over a confidence set for the nuisance parameter. J. Amer. Statist. Assoc. 89 , 1012–1016). It is argued that estimation followed by maximization is an attractive, but computationally more demanding, alternative to partial maximization. We illustrate the ideas on a range of low‐dimensional but important examples for which the alternative methods can be investigated completely numerically.     

Asymptotic results for model robust regression

Since the mid 1980's many statisticians have studied methods for combining parametric and nonparametric models to improve the quality of fits in a regression problem. Notably Einsporn (1987) proposed the Model Robust Regression 1 estimate (MRRl) in which the parametric function, f, and the nonparametric functiong were combined in a straightforward fashion via the use of a mixing parameter, λ This technique was studied extensively atsmall samples and was shown to be quite effective at modeling various unusual functions. In this paper we have asymptotic results for the MRRl estimate in the case where λ is theoretically optimal, is asymptotically optimal and data driven, and is chosen with the PRESS statistic (Allen, 1971) We demonstrate that the MRRl estimate with λchosen by the PRESS statistic is slightly inferior asymptotically to the other two estimates, but, nevertheless possesses positive asymptotic qualities.     

A note on the moments of stochastic shrinkage parameters in ridge regression

The purpose of this note is to gain insight on the performance of two well known operational Ridge Regression estimators by deriving the moments of their stochastic shrinkage parameters. We also show that, under certain conditions, one of them has bounded moments.     

Optimum Experimental Designs for Choosing Between Competitive and Non Competitive Models of Enzyme Inhibition

In the study of the inhibition of enzyme reactions the Michaelis–Menten model is extended to include two experimental variables and three or more parameters. We combine the three-parameter models for competitive and non competitive inhibition in a four-parameter model and use optimum design theory to find D- and Ds-optimum designs for discriminating between the models. These designs are compared with compound T-optimum designs which provide the most powerful tests for discrimination between models. A single design is found with high discrimination efficiency whichever model is true.     

On confidence limits for the difference of two binomial parameters

In the present article we suggest two new methods for calculating approximate confidence limits for the differences of the two binomial parameters. Different methods for determining the confidence interval are compared.     

A pilot Monte Carlo study of two sequential estimation procedures based on generalized U-statistics

Recently two sequential estimation procedures based on generalized U-statistics have appeared in the statistical literature [Williams and Sen (1973, 1974)]. One of these procedures concerns the multi-sample problem of estimating a vector of parameters when the total sample size is fixed. The other procedure concerns the multi-sample problem of constructing a confidence ellipsoid of bounded maximum width for a vector of parameters. To supplement the asymptotic theory discussed in these earlier papers, a Monte Carlo study investigating the efficiency of these procedures for moderate sample sizes would be useful. This paper describes a preliminary Monte Carlo study utilizing a small number of replications and performed to provide information for the design of a more extensive study.     

Histogram estimation of radon-nikodym derivatives for strong mixing data

Nonparametric inference for point processes is discussed by way of histograms, which provide a nice tool for the analysis of on-line data. The construction of histograms depends on a sequence of partitions, which we take tc be nonenibedded to allow partitions with sets of equal measure. This presents some theoretical problems, which are addressed with an assumption on the decomposition of second order moments. In another direction, we drop the usual independence assumption on the sample, replacing it by a strong mixing assumption. Under this setting, we study the convergence of the histogram in probability, which depends on approximation conditions between the distributions of random pairs and the product of their marginal distributions, and^almost completely, which is based on the decomposition of the second order moments. This last convergence is stated on two versions according to the assumption of Laplace transforms or the Cramer moment conditions. These are somewhat stronger, but enable us to recover the usual condition on the decrease rate of sets on each partition. In the final section we prove that the finite dimensional distributions converge in distribution to a Gaussian centered vector with a specified covariance.