Nonparametric Sequential Tests And Change Point Detection Problems

The explosive development of information science and technology puts in new problems involving statistical data analysis. These problems result from higher re quirements concerning the reliability of statistical decisions, the accuracy of math ematical models and the quality of control in complex systems. A new aspect of statistical analysis has emerged, closely connected with one of the basic questions of cynergetics: how to "compress" large volumes of experimental data in order to extract the most valuable information from data observed. De tection of large "homogeneous" segments of data enables one to identify "hidden" regularities in an object's behavior, to create mathematical models for each seg ment of homogeneity, to choose an appropriate control, etc. Statistical methods dealing with the detection of changes in the characteristics of random processes can be of great use in all these problems. These methods have accompanied the rapid growth in data beginning from the middle of our century. According to a tradition of more than thirty years, we call this sphere of statistical analysis the "theory of change-point detection. " During the last fifteen years, we have witnessed many exciting developments in the theory of change-point detection. New promising directions of research have emerged, and traditional trends have flourished anew. Despite this, most of the results are widely scattered in the literature and few monographs exist. A real need has arisen for up-to-date books which present an account of important current research trends, one of which is the theory of non parametric change--point detection.

Sequential Analysis: Hypothesis Testing and Changepoint Detection systematically develops the theory of sequential hypothesis testing and quickest changepoint detection. It also describes important applications in which theoretical results can be used efficiently. The book reviews recent accomplishments in hypothesis testing and changepoint detecti

This book covers the development of methods for detection and estimation of changes in complex systems. These systems are generally described by nonstationary stochastic models, which comprise both static and dynamic regimes, linear and nonlinear dynamics, and constant and time-variant structures of such systems. It covers both retrospective and sequential problems, particularly theoretical methods of optimal detection. Such methods are constructed and their characteristics are analyzed both theoretically and experimentally. Suitable for researchers working in change-point analysis and stochastic modelling, the book includes theoretical details combined with computer simulations and practical applications. Its rigorous approach will be appreciated by those looking to delve into the details of the methods, as well as those looking to apply them.

Recently there has been a keen interest in the statistical analysis of change point detec tion and estimation. Mainly, it is because change point problems can be encountered in many disciplines such as economics, finance, medicine, psychology, geology, litera ture, etc. , and even in our daily lives. From the statistical point of view, a change point is a place or time point such that the observations follow one distribution up to that point and follow another distribution after that point. Multiple change points problem can also be defined similarly. So the change point(s) problem is two fold: one is to de cide if there is any change (often viewed as a hypothesis testing problem), another is to locate the change point when there is a change present (often viewed as an estimation problem). The earliest change point study can be traced back to the 1950s. During the fol lowing period of some forty years, numerous articles have been published in various journals and proceedings. Many of them cover the topic of single change point in the means of a sequence of independently normally distributed random variables. Another popularly covered topic is a change point in regression models such as linear regres sion and autoregression. The methods used are mainly likelihood ratio, nonparametric, and Bayesian. Few authors also considered the change point problem in other model settings such as the gamma and exponential.

A study of sequential nonparametric methods emphasizing the unified Martingale approach to the theory, with a detailed explanation of major applications including problems arising in clinical trials, life-testing experimentation, survival analysis, classical sequential analysis and other areas of applied statistics and biostatistics.

In our research, we exploit the relationship between properties of U-statistics and Energy statistics (V-statistics) to come up with non-parametric tests in change-point analysis. \cite{lee} provided a wide discussion on asymptotic behaviour, and connection between U-statistics and V-statistics when large samples. Many other researchers such as \cite{sen1974}, \cite{serfling1980} and \cite{neuhaus1977} studied connections between U-statistic and V-statistics and their asymptotic properties. We first propose a non-parametric test to detect change in the distribution based on MIC using energy statistics. The proposed energy-statistic based MIC is used for model selection between null and alternative hypothesis models. We achieve this by adopting the idea of the works of \cite{chen} and \cite{pan} and apply energy distance statistic. To test the performance of our proposed test, we assess the finite sample properties and compare efficiency and powers of different methods with that of our method. We then discuss applications of our proposed test in two different real-life examples for detecting change in mean and variance, respectively. Since the underlying distribution is unknown, we use bootstrap approximations for the p-values as proposed by \cite{hangfen2009} in detecting unknown change points in means and variances. In the second part of my dissertation, we propose a non-parametric sequential test based on energy statistics \cite{rizzo2013} to detect changes in distribution for independent random variables. In their study, \cite{Oscar} considered backward-looking windows each of length $L$ across the pooled data, and then retrospectively investigate if there is evidence for a change point between the times $\text{max}\{t-L,1\}$ and $t$, for any given time $t$. We adopt this idea to come up with a test statistic similar in structure based on energy statistics. We compare the performance of this method in terms of false-alarm rates and powers to existing sequential methods such as sequential KS, Generalized Likelihood Ratio test and others for detecting change in distribution. We apply our proposed method and others to the problem of detecting radiological anomalies.

The analysis of numerical sequential data, such as time series, is a frequent practice in both academic and industrial settings. Offline change detection segments the data retrospectively and is useful for uncovering events and systematic behaviors in data analysis tasks. It is applied in a variety of fields including finance, genomics and energy consumption. Furthermore, in the potential presence of change points, utilizing change detection prior to data modeling can help prevent building inappropriate models under the assumption of data homogeneity, and consequently supports improved prediction and statistical inference. In this thesis, we propose three methods that study the offline change point detection problem from different aspects and application domains. The first method is a nonparametric procedure that can provide computational speedups to simultaneously detect multiple change points. The second method models the relationship between the different channels of multivariate observations to detect change points and anomalies. The third method focuses on the specific biomedical domain of cell culture monitoring to detect the transition from cell growth to confluence. All proposed methods are evaluated through simulations and real-world data applications.

A technically precise yet clear presentation of modern sequential methodologies having immediate applications to practical problems in the real world, Applied Sequential Methodologies communicates invaluable techniques for data mining, agricultural science, genetics, computer simulation, finance, clinical trials, sonar signal detection, randomization, multiple comparisons, psychology, tracking, surveillance, and numerous additional areas of application. Includes more than 500 references, 165 figures and tables, and over 25 pages of subject and author indexes. Applied Sequential Methodologies brings the crucial nature of sequential approaches up to speed with recent theoretical gains, demonstrating their utility for solving real-life problems associated with Change-point detection in multichannel and distributed systems Best component selection for multivariate distributions Multistate processes Approximations for moving sums of discrete random variables Interim and terminal analyses of clinical trials Adaptive designs for longitudinal clinical trials Slope estimation in measurement-error models Tests for randomization and target tracking Appropriate count of simulation runs Stock price models Orders of genes Size and power control in multiple comparisons Authored by 33 leading scientists, this volume will greatly benefit sequential analysts, data analysts, applied statisticians, biometricians, clinical trialists, and upper-level undergraduate and graduate students in these disciplines.