Missing data in longitudinal studies: Strategies for Bayesian modeling and sensitivity analysis M. J. Daniels and J. W. Hogan, Chapman & Hall/CRC, Boca Raton, 2008. No. of pages: 303. Price: $83.95. ISBN: 978-1-58488-609-9
Missing data in longitudinal studies: Strategies for Bayesian modeling and sensitivity analysis M. J. Daniels and J. W. Hogan, Chapman & Hall/CRC, Boca Raton, 2008. No. of pages: 303. Price: $83.95. ISBN: 978-1-58488-609-9
复制标题
纵向研究中的缺失数据：贝叶斯建模和敏感性分析策略 M. J. Daniels 和 J. W. Hogan, Chapman
DOI：
10.1002/sim.4162
复制
发表时间：
2011
影响因子：
2
通讯作者：
Bartlett J
中科院分区：
文献类型：
--
作者：
Bartlett J
数据缺失是流行病学和临床研究中的一个问题，但在人类受试者的纵向研究中尤为普遍，参与者通常无法返回一次或多次计划的随访。大量的文献和方法收集已经发展成为处理这种缺失数据的原则性方法。虽然有一些现有的优秀的文本描述这一领域，丹尼尔斯和霍根的是第一个明确关注缺失数据的纵向研究的背景下。这本书由三个部分组成，第一个介绍建模的纵向研究中，有没有缺失的数据。第一章介绍了六个激励性的例子数据集，这些数据集在整个专著中用于说明所描述的概念和方法。除了纵向队列研究外，所有病例研究均为临床试验。在第2章中，概述了纵向数据的各种建模方法，包括连续和离散结果。考虑到整本书都可以而且已经专门讨论这些主题，因此报道必须简洁。第3章也是如此，它简要介绍了无缺失数据时的贝叶斯推理。第4章通过将模型拟合到第1章中介绍的一些数据集的完整案例来说明前两章的材料。本书的第二部分从第5章开始，介绍了完整数据和观察数据的概念，缺失数据机制，以及根据鲁宾分类法的分类（完全随机缺失（MCAR），随机缺失（MAR）和非随机缺失（MNAR））。特别关注的是由于辍学的单调缺失MAR的影响。介绍了MNAR模型的主要类型（选择模型、模式混合模型和共享参数模型）。第6章考虑了可验证性下的模型估计和推断。需要强调的是，虽然在数据完整的情况下，协方差结构的正确规范对于平均参数的一致估计可能不是必要的，但在MAR机制下存在缺失数据的情况下，情况并非如此。这导致考虑协方差结构的适当规格的高度重要性。第7章说明了推理下的可验证性假设在一些激励的例子数据集。第8章更详细地讨论了不可识别（MNAR）缺失。Daniels和Hogan主张模型可以用“敏感性参数”来参数化，敏感性参数是未识别的参数，它表征了给定观测数据的缺失数据的条件分布。结果表明，模式混合模型更容易借给自己这样的因子分解，在选择模型。本章包括一个很好的说明性例子，说明为什么参数选择模型可能是完全可识别的，但这严重依赖于某些参数假设，这些假设无法根据观察到的数据进行经验验证。第9章提倡一种策略，用于执行分析，以评估结论对MAR假设的敏感性，无论是在模式混合和选择建模框架内。本书最后使用三个激励性的例子进行了说明性的敏感性分析。在这样做的过程中，它表明，即使在最简单的纵向设置中，执行敏感性分析也是一个具有挑战性的目标。
Missing data is a problem throughout epidemiology and clinical studies, but is particularly prevalent in longitudinal studies of human subjects, in which participants often fail to return to one or more planned follow-up visits. A vast literature and collection of methodology have developed towards principled approaches to handling such missing data. While there are a number of existing excellent texts describing this field, Daniels and Hogan’s is the first to explicitly focus on missing data in the context of longitudinal studies.The book consists of three parts, with the first introducing modelling of longitudinal studies in which there are no missing data. The first chapter introduces six motivating example data sets that are used throughout the monograph to illustrate the concepts and methods that are described. The case studies are all clinical trials, with the exception of a longitudinal cohort study. In Chapter 2, an overview is given of various modelling approaches for longitudinal data, covering both continuous and discrete outcomes. Given that whole books can and have been devoted to these topics by themselves, the coverage is necessarily concise. The same is true of Chapter 3, which gives a whistle-stop tour of Bayesian inference when there are no missing data. Chapter 4 illustrates the material of the preceding two chapters by fitting models to the complete cases in some of the data sets introduced in Chapter 1. The second part of the book begins in Chapter 5 by introducing the concepts of full and observed data, the missing data mechanism, and its classification according to Rubin’s taxonomy (missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR)). Particular focus is given to the implications of MAR for monotone missingness due to dropout. The main types of MNAR models (selection, pattern-mixture, and shared-parameter) are also introduced. Chapter 6 considers model estimation and inference under ignorability. It is emphasized that while with complete data, correct specification of covariance structures may not be necessary for consistent estimation of mean parameters, the same is not true in the presence of missing data under an MAR mechanism. This leads to a heightened importance to considering the appropriate specification of covariance structures. Chapter 7 illustrates inference under the ignorability assumption in a number of the motivating example data sets. Chapter 8 discusses nonignorable (MNAR) missingness in greater detail. Daniels and Hogan advocate models that can be parametrized in terms of ‘sensitivity parameters’—nonidentified parameters that characterize the conditional distribution of missing data given observed. It is shown that pattern-mixture models lend themselves more readily to such a factorization, in contrast to selection models. The chapter includes an excellent illustrative example of why parametric selection models may be fully identifiable, but that this relies critically on certain parametric assumptions that cannot be empirically verified based on the observed data. Chapter 9 advocates a strategy for performing analyses to assess the sensitivity of conclusions to the MAR assumption, both within the pattern-mixture and selection modelling frameworks. The book concludes with illustrative sensitivity analyses using three of the motivating examples. In doing so, it illustrates that performing sensitivity analyses is a challenging aim in even the simplest of longitudinal settings.