Time-varying clustering of multivariate longitudinal observations

Abstract

We propose a statistical method for clustering multivariate longitudinal data into homogeneous groups. This method relies on a time-varying extension of the classical K-means algorithm, where a multivariate vector autoregressive model is additionally assumed for modeling the evolution of clusters' centroids over time. Model inference is based on a least-squares method and on a coordinate descent algorithm. To illustrate our work, we consider a longitudinal dataset on human development. Three variables are modeled, namely life expectancy, education and gross domestic product.     

Tukey-Type Distributions in the Context of Financial Data

Abstract

In one-parameter (θ) families, we were not aware of explicit hypothesis testing scenarios where maximal invariant statistics failed to distinguish the models. We start with a concrete example (Sec. 2.2) to highlight such a hypothesis testing problem involving markedly different models. In this problem, because of the absence of a nontrivial uniformly most powerful invariant (UMPI) test, we briefly suggest two approaches to test the hypothesis. The first resolution (Sec. 3.1) is frequentist in nature. It utilizes a weight function on the parameter space and compares “average” distributions obtained under the null and alternative models in the sense of Wald (1947 Wald , A. ( 1947 ). Sequential Analysis . New York : Wiley . [Google Scholar] 1950 Wald , A. ( 1950 ). Statistical Decision Functions . New York : Wiley . [Google Scholar]). In contrast, a fully Bayesian resolution (Sec. 3.2) is also included. The note ends with a series of other interesting examples involving one-parameter families where maximal invariant statistics fail to distinguish the hypothesized models. The examples include easy-to-construct families of probability models involving only a single location or scale parameter θ.     

Hypothesis Testing Problems in an Unbalanced Longitudinal Ophthalmology Study

ABSTRACT

The analysis of clustered data in a longitudinal ophthalmology study is complicated by correlations between repeatedly measured visual outcomes of paired eyes in a participant and missing observations due to the loss of follow-up. In the present article we consider hypothesis testing problems in an ophthalmology study, where eligible eyes are randomized to two treatments (when two eyes of a participant are eligible, the paired eyes are assigned to different treatments), and vision function outcomes are repeatedly measured over time. A large sample-based nonparametric test statistic and a nonparametric Bootstrap test analog are proposed for testing an interaction effect of two factors and testing an effect of a eye-specific factor within a level of the other person-specific factor on visual function outcomes. Both test statistics allow for missing observations, correlations between repeatedly measured outcomes on individual eyes, and correlations between repeatedly measured outcomes on both eyes of each participant. A simulation study shows that these proposed test statistics maintain nominal significance levels approximately and comparable powers to each other, as well as higher powers than the naive test statistic ignoring correlations between repeated bilateral measurements of both eyes in the same person. For illustration, we apply the proposed test statistics to the changes of visual field defect score in the Advanced Glaucoma Intervention Study.     

Goodness-of-Fit Tests for Multivariate Distributions

We propose three new statistics, Z _p , C _p , and R _p for testing a p-variate (p ≥ 2) normal distribution and compare them with the prominent test statistics. We show that C _p is overall most powerful and is effective against skew, long-tailed as well as short-tailed symmetric alternatives. We show that Z _p and R _p are most powerful against skew and long-tailed alternatives, respectively. The Z _p and R _p statistics can also be used for testing an assumed p-variate nonnormal distribution.     

Estimation of Survivorship Function Based on Doubly Censored Data with Application to HIV/AIDS

Left censoring concept has been defined in different ways in statistical applications. Turnbull (1974 Turnbull , B. W. ( 1974 ). Nonparametric estimation of a survivorship function with doubly censored data . J. Amer. Statist. Assoc. 69 : 169 – 173 .[Taylor & Francis Online], [Web of Science ®] , [Google Scholar]) defines it in a particular way. Whereas in recent literature, especially in epidemiological studies, it has been defined in another way. This difference between the two approaches is the main reason that despite simplicity, Turnbull method cannot be applicable in all cases of doubly censored data. In this article we present a modified Turnbull method for analysis of doubly censored data adequate with recent definition. Comparison has been done with other statistical methods, including imputation estimator, full likelihood-based and conditional likelihood-based approach using Iranian HIV data.     

Strong Gaussian Approximations of Product-Limit and Quantile Processes for Strong Mixing and Censored Data

In this article, we consider the product-limit quantile estimator of an unknown quantile function under a censored dependent model. This is a parallel problem to the estimation of the unknown distribution function by the product-limit estimator under the same model. Simultaneous strong Gaussian approximations of the product-limit process and product-limit quantile process are constructed with rate O[(log n)^?λ] for some λ > 0. The strong Gaussian approximation of the product-limit process is then applied to derive the laws of the iterated logarithm for product-limit process.     

Estimation Bias in Complete-Case Analysis in Crossover Studies with Missing Data

Crossover designs are used often in clinical trials. It is not uncommon that subjects discontinue before completing all treatment periods in a crossover study. Despite availability of statistical methodologies utilizing all available data and software for obtaining valid inferences under the assumption of missing at random (MAR), naïve approaches, such as the complete case (CC) analysis, which is only valid with a strong assumption of missing completely at random are still widely used in practice. In this article, we obtain the analytical form of the estimation bias of treatment effects with CC for linear mixed models. We use simulation studies to examine the inflation of Type I error and efficiency loss in the inferences with CC under MAR. Invalidity and inefficiency of two other commonly used approaches for defining analyzed data in the presence of missing data, including data from at least two periods in three period crossover and available cases for a specific comparison of interest, are also demonstrated through simulation studies.     

On asymptotic non-normality of null distributions of mrpp statistics

Severe departures from normality occur frequently for null distributions of statistics associated with applications of mulLi-response permutation procedures (MRPP) for either small or large finite populations. This paper describes the commonly encountered situation associated with asymptotic non-normality for null distributions of MRPP statistics which does not depend on the underlying multivariate distribution. In addition, this paper establishes the existence of a non-degenerate underlying distribution for which the null distributions of MRPP statistics are asymptotically non-normal for essentially all size structure configurations. It is known that MRPP statistics are symmetric versions of a broader class of statistics, most of which are asymmetric. Because of the non-normality associated with null distributions of MRPP statistics, this paper includes necessary results for inferences based on the exact first three moments of anv statistic in this broader class (analogous to existing results for MRPP statistics).     

Univaiuate anu multivariate categorical data analysis for block designs

Analysis for univariate and multivariate categorical data in block designs is given and illustrated through examples. The univariate analysis compares the treatments on the basis of their pooled frequency distributions (pooled over blocks). The test statistic used is called Q after Cochran (1950). The large sample null distribution of Q is a chi-square. Analysis of p-variate categorical data (kth variable having ck classes, K=1,...,p) can be done by treating it as a univariate categorical problem with [d] classes. Very often [d] is large in relation to the size of the experiment. This makes the expected frequencies for some of the cells very small, making the univariate method inapplicable. In these circumstances it may be reasonable to compare the treatments on the basis of marginal distributions up to the mth dimension, 1[d] , which is given in this paper. This method is also illustrated for missing observations