GUIDO IMBENS, DONALD RUBIN, Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. New York: Cambridge University Press.
GUIDO IMBENS, DONALD RUBIN, Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. New York: Cambridge University Press.
复制标题
DOI：
10.1111/biom.12615
复制
发表时间：
2016-12
期刊：
影响因子：
1.9
通讯作者：
B. Shepherd;Ryan T. Jarrett;L. Fu
中科院分区：
文献类型：
--
作者：
B. Shepherd;Ryan T. Jarrett;L. Fu
这是一些恐惧，我们提供我们的统计，社会和生物医学科学的因果推理：介绍。因果推理的领域是广泛的，并有一些强烈的个性。在谷歌上搜索一下，就能看出人们对这本书的反应是多么的不同，从毫不掩饰的赞扬到嘲笑。我（BES）同意复习这本书，因为我一直希望开发一门关于因果推理的课程，并觉得这本书可以成为我课程的关键组成部分。我从我们的生物统计学系招募了两名研究生，一个是第一年，一个是第二年，在过去的春季学期和我一起阅读。这次审查代表了我们的共同努力，但我（BES）肯定偏向他们的想法。这本书是由该领域的两位著名研究人员撰写的。唐·鲁宾在70年代和80年代发表的有影响力的论文为当前因果推理领域的大部分研究奠定了基础。鲁宾博士对现代统计学的巨大影响--不仅仅是在因果推理方面--是众所周知的。另一位作者Guido Imbens也不是一个懒鬼，他在经济学和统计学期刊上发表了一系列令人印象深刻的重要论文。他们合在一起，完全有资格为因果推理写一篇介绍性的文章。这本书写得很好。它的目的是为广大观众，在大多数方面，它成功地提出了材料的方式，是一个合理的熟悉数学和统计学的读者访问。有些章节有相当数量的数学内容，但作者始终保持着叙述性的，易于阅读的写作风格。文本非常有条理。每一章都有一个介绍，提供了一个清晰而有趣的概述。大多数章节然后提出了一个数据的例子，用于框架的材料，并证明了方法的应用。整个过程中使用了丰富的、刺激的和可理解的数据示例，这是一个很大的优势，尽管限制是没有提供数据集和分析代码。此外，有时示例数据集不是理想的拟合。例如，在第23章中，作者在一项社区随机干预的研究中引入了更复杂的平均因果效应;这个概念在个体随机干预的情况下可能会更简单。作者预示未来的内容几乎是一个错误，反复说在部分/章节/第X节，我们将讨论Y。过于清晰与过于简洁是更可取的，虽然这本书有时有点缓慢和乏味，但很容易遵循。整本书都有错别字（可能每章有一两处，而在令人惊讶的容易出错的结论一章中，错别字更多），但没有一处是特别严重的，而且几乎在所有情况下，真正的含义都很容易辨别。第1-3章提供了一个直观和清晰的介绍“鲁宾因果模型”，其中关于潜在结果的推断被视为缺失数据问题。由于作者只关注单个时间点的二元治疗，因此每个人只有两种潜在结果，其中一种从未观察到。基本假设和哲学，分配机制，并提供了一个简短的历史的潜在结果的因果推理方法。其中有些观点并没有被一致接受。例如，“没有操纵就没有因果关系”这一格言仍在争论之中。而且，阅读唐·鲁宾的因果推理史有点像阅读托马斯·杰斐逊的美国建国史：当然有趣，毫无疑问是由一位创始人写的，但与亚历山大汉密尔顿的写作方式不同。
It is with some trepidation that we offer our review of Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. The field of causal inference is broad and known to have some strong personalities. A quick Google search shows how varied the responses have been to this book, ranging from unmitigated praise to derision. I (BES) agreed to review the book because I have been hoping to develop a course on causal inference and felt that this book could be a key component of my course. I recruited two graduate students from our biostatistics department, one in their first year and one in their second year, to read it with me this past spring semester. This review represents our combined efforts, but I (BES) certainly biased their thinking. The book is written by two prominent researchers in the area. Don Rubin’s influential papers in the 70s and 80s provided the foundation for much of the current field of causal inference. Dr. Rubin’s enormous impact on modern statistics– not just in causal inference–is well recognized. The other author, Guido Imbens, is no slacker either with a very impressive bibliography of important papers in both economics and statistics journals. Together, they are more than qualified to write an introductory text to causal inference. The book is well written. It is intended for a broad audience and, in most respects, it succeeds in presenting material in a manner that is accessible to readers with a reasonable familiarity with mathematics and statistics. Some sections have a fair amount of mathematical content, but the authors maintain a narrative, easy-to-read writing style throughout. The text is very structured. Every chapter begins with an introduction, which provides a clear and interesting overview. Most chapters then present a data example that is used to frame the material and to demonstrate the application of the methods. The rich use of diverse, stimulating, and understandable data examples throughout is a big plus, although a limitation is that datasets and analysis code are not provided. Also, occasionally the example datasets are not ideal fits. For example, in chapter 23, the authors introduce complier average causal effects in a study with a community randomized intervention; the concept may have been introduced more simply with an individually randomized intervention. The authors foreshadow future content almost to a fault, repeatedly saying in part/chapter/section X we will discuss Y . It is preferable to be overly clear versus overly terse, and although the book is a little slow and tedious at times, it is easy to follow. There are typos throughout (probably one or two per chapter, and more than that in the surprisingly error-prone Conclusion chapter), but none are particularly serious, and in almost all cases the true meaning is easily discerned. Chapters 1–3 provide an intuitive and well-articulated introduction to the “Rubin Causal Model,” in which inference regarding potential outcomes is treated as a missing data problem. As the authors focus solely on binary treatments at a single time point, there are only two potential outcomes for each individual, one of which is never observed. Basic assumptions and philosophies, assignment mechanisms, and a brief history of the potential outcomes approach to causal inference are provided. Some of these ideas are not uniformly accepted. For example, the dictum “no causation without manipulation” is still debated. And, reading a history of causal inference by Don Rubin kind of feels like reading a history of the establishment of the United States by Thomas Jefferson: certainly interesting and undoubtedly written by a founder, but different from how Alexander Hamilton would write it. With that said, these chapters contain some of the