Masafumi Akahira – författare

This monograph is a collection of results recently obtained by the authors. Most of these have been published, while others are awaitlng publication. Our investigation has two main purposes. Firstly, we discuss higher order asymptotic efficiency of estimators in regular situa­ tions. In these situations it is known that the maximum likelihood estimator (MLE) is asymptotically efficient in some (not always specified) sense. However, there exists here a whole class of asymptotically efficient estimators which are thus asymptotically equivalent to the MLE. It is required to make finer distinctions among the estimators, by considering higher order terms in the expansions of their asymptotic distributions. Secondly, we discuss asymptotically efficient estimators in non­ regular situations. These are situations where the MLE or other estimators are not asymptotically normally distributed, or where l 2 their order of convergence (or consistency) is not n / , as in the regular cases. It is necessary to redefine the concept of asympto­ tic efficiency, together with the concept of the maximum order of consistency. Under the new definition as asymptotically efficient estimator may not always exist. We have not attempted to tell the whole story in a systematic way. The field of asymptotic theory in statistical estimation is relatively uncultivated. So, we have tried to focus attention on such aspects of our recent results which throw light on the area.

In order to obtain many of the classical results in the theory of statistical estimation, it is usual to impose regularity conditions on the distributions under consideration. In small sample and large sample theories of estimation there are well established sets of regularity conditions, and it is worth while to examine what may follow if any one of these regularity conditions fail to hold. "Non-regular estimation" literally means the theory of statistical estimation when some or other of the regularity conditions fail to hold. In this monograph, the authors present a systematic study of the meaning and implications of regularity conditions, and show how the relaxation of such conditions can often lead to surprising conclusions. Their emphasis is on considering small sample results and to show how pathological examples may be considered in this broader framework.

In order to obtain many of the classical results in the theory of statistical estimation, it is usual to impose regularity conditions on the distributions under consideration. In small sample and large sample theories of estimation there are well established sets of regularity conditions, and it is worth while to examine what may follow if any one of these regularity conditions fail to hold. "Non-regular estimation" literally means the theory of statistical estimation when some or other of the regularity conditions fail to hold. In this monograph, the authors present a systematic study of the meaning and implications of regularity conditions, and show how the relaxation of such conditions can often lead to surprising conclusions. Their emphasis is on considering small sample results and to show how pathological examples may be considered in this broader framework.

This monograph is a collection of results recently obtained by the authors. Most of these have been published, while others are awaitlng publication. Our investigation has two main purposes. Firstly, we discuss higher order asymptotic efficiency of estimators in regular situa­ tions. In these situations it is known that the maximum likelihood estimator (MLE) is asymptotically efficient in some (not always specified) sense. However, there exists here a whole class of asymptotically efficient estimators which are thus asymptotically equivalent to the MLE. It is required to make finer distinctions among the estimators, by considering higher order terms in the expansions of their asymptotic distributions. Secondly, we discuss asymptotically efficient estimators in non­ regular situations. These are situations where the MLE or other estimators are not asymptotically normally distributed, or where l 2 their order of convergence (or consistency) is not n / , as in the regular cases. It is necessary to redefine the concept of asympto­ tic efficiency, together with the concept of the maximum order of consistency. Under the new definition as asymptotically efficient estimator may not always exist. We have not attempted to tell the whole story in a systematic way. The field of asymptotic theory in statistical estimation is relatively uncultivated. So, we have tried to focus attention on such aspects of our recent results which throw light on the area.

This book presents new findings on nonregular statistical estimation. Unlike other books on this topic, its major emphasis is on helping readers understand the meaning and implications of both regularity and irregularity through a certain family of distributions. In particular, it focuses on a truncated exponential family of distributions with a natural parameter and truncation parameter as a typical nonregular family. This focus includes the (truncated) Pareto distribution, which is widely used in various fields such as finance, physics, hydrology, geology, astronomy, and other disciplines. The family is essential in that it links both regular and nonregular distributions, as it becomes a regular exponential family if the truncation parameter is known. The emphasis is on presenting new results on the maximum likelihood estimation of a natural parameter or truncation parameter if one of them is a nuisance parameter. In order to obtain more information on the truncation, the Bayesian approach is also considered. Further, the application to some useful truncated distributions is discussed. The illustrated clarification of the nonregular structure provides researchers and practitioners with a solid basis for further research and applications.

This book is a reconstruction of Masafumi Akahira’s works on statistical estimation and its related fields, especially higher order asymptotics and non-regular cases with consideration based on information amounts. There have been few books on higher order asymptotics and non-regular cases, but the book helps the reader to understand the meaning and implications of the hierarchical structure of higher order asymptotics and gives some insights into the structure of non-regular estimation. After the results on the second and third order asymptotic efficiency in the volume entitled “Joint Statistical Papers of Akahira and Takeuchi” are summarized, the positive resolution of the conjecture “third order efficiency implies fourth order efficiency” of J. K. Ghosh is described, from which the fourth order asymptotic efficiency of the bias-adjusted maximum likelihood estimator and the bias-adjusted generalized Bayes estimator is shown. In non-regular situations, the (maximum) order of consistency and the (second order) asymptotic sufficiency are discussed including the view of loss of information, and in regular cases, the asymptotic deficiency of asymptotically efficient estimators is stated. For a truncated family of distributions, the influence of a nuisance parameter on the estimation of the interest parameter is investigated through maximum likelihood estimators. Also discussed is the higher order sequential estimation with the Bhattacharyya type bound, including the presence of a nuisance parameter. In interval estimation, a systematic method of the construction of a confidence interval for the difference between means is discussed including the Behrens–Fisher type problem, and also ordinary, Bayesian, likelihood ratio and combined Bayesian–frequentist type confidence intervals for a positive parameter are provided and compared. Further, the higher order approximations to percentage points of the non-central t-distribution and the distribution of a non-central t-statistic without the normality assumptions are given. Finally, the large deviation efficiency and large deviation approximations are discussed up to the higher order. Masafumi Akahira is Professor Emeritus at the University of Tsukuba. He has served as Vice President of the University of Tsukuba and President of the Japan Statistical Society.

This book is a reconstruction of Masafumi Akahira’s works on statistical estimation and its related fields, especially higher order asymptotics and non-regular cases with consideration based on information amounts. There have been few books on higher order asymptotics and non-regular cases, but the book helps the reader to understand the meaning and implications of the hierarchical structure of higher order asymptotics and gives some insights into the structure of non-regular estimation. After the results on the second and third order asymptotic efficiency in the volume entitled “Joint Statistical Papers of Akahira and Takeuchi” are summarized, the positive resolution of the conjecture “third order efficiency implies fourth order efficiency” of J. K. Ghosh is described, from which the fourth order asymptotic efficiency of the bias-adjusted maximum likelihood estimator and the bias-adjusted generalized Bayes estimator is shown. In non-regular situations, the (maximum) order of consistency and the (second order) asymptotic sufficiency are discussed including the view of loss of information, and in regular cases, the asymptotic deficiency of asymptotically efficient estimators is stated. For a truncated family of distributions, the influence of a nuisance parameter on the estimation of the interest parameter is investigated through maximum likelihood estimators. Also discussed is the higher order sequential estimation with the Bhattacharyya type bound, including the presence of a nuisance parameter. In interval estimation, a systematic method of the construction of a confidence interval for the difference between means is discussed including the Behrens–Fisher type problem, and also ordinary, Bayesian, likelihood ratio and combined Bayesian–frequentist type confidence intervals for a positive parameter are provided and compared. Further, the higher order approximations to percentage points of the non-central t-distribution and the distribution of a non-central t-statistic without the normality assumptions are given. Finally, the large deviation efficiency and large deviation approximations are discussed up to the higher order. Masafumi Akahira is Professor Emeritus at the University of Tsukuba. He has served as Vice President of the University of Tsukuba and President of the Japan Statistical Society.