A hot new statistical technology like the pure prediction algorithms divides the discipline into insiders and outsiders. The insiders, caught up in the excitement of new lands to conquer, are usually in no mood to think critically about methodological weaknesses. Outsiders, everyone else, may lack enough information and insight to mount an effective critique, especially if the new methods are intricate, sprawling, and somewhat mysterious. Eventually though, there is enough mixing between the two groups to allow the critical process to begin. Statistics has a strong tradition of internal criticism—witness the Bayesian/frequentist debates—with discussion papers like this one serving as useful outlets. The discussants here span the insider/outsider spectrum, including some of the most prominent developers of modern prediction methods. Predictably, the insiders are the happiest with the fit between traditional and algorithmic methods: they “go hand in glove” say Professors Friedman, Hastie, and Tibshirani, (henceforth FHT), speaking with authority as the authors of seminal texts on statistical learning, as well as key technologies such as the lasso. “Data Science Process: One Culture” is the title of Professor Yu and Dr Barter’s commentary, a reverse echo of Leo Breiman’s famous “two cultures” paper, which, contrariwise, claimed a stark dichotomy between traditional methods and the “prediction culture.” Again, Yu/Barter speak as experienced and accomplished inside developers of modern prediction methodology. The outsiders are less happy with the current state of connection between traditional and algorithmic regression ideas. (The term “outsider” here is only relative to algorithmic prediction; all of the discussants have been influential developers of modern data-analytic techniques such as false discovery rates, genomic analysis, and the bootstrap.) This paper was written from an outsider’s point of view. Rereading it now, a year later, made me aware of some outsider limitations: there is in fact more connective tissue between the two cultures than I credited, both FHT and Yu/Barker giving convincing examples. That being said, an extra year of experience has not changed my belief in a disconnect between traditional statistical regression methods and the pure prediction algorithms. Section 8 of the paper, featuring Table 5, makes the case directly in terms of six criteria. Five of the six emerged more or less unscathed from the discussion. Criteria 2, long-time scientific truth versus possibly short-term prediction accuracy, was doubted by FHT and received vigorous push-back from Yu/Barter: ... but in our experience the “truth” that traditional regression methods supposedly represent is rarely justified or validated...

中文翻译：

---

反驳

像纯预测算法这样的热门新统计技术将学科分为内部人员和外部人员。内部人员沉浸在征服新土地的兴奋中，通常没有心情批判性地思考方法论的弱点。局外人，其他所有人，可能缺乏足够的信息和洞察力来进行有效的批评，尤其是在新方法复杂、庞大且有些神秘的情况下。但最终，两组之间有足够的混合，可以开始关键过程。统计学有着强大的内部批评传统——见证贝叶斯/频率论者的辩论——像这样的讨论文件可以作为有用的出口。这里的讨论者涵盖内部/外部范围，包括现代预测方法的一些最杰出的开发者。可以预见，业内人士对传统方法和算法方法之间的契合感到最满意：他们“携手共进”，弗里德曼教授、哈斯蒂教授和蒂布希拉尼教授（以下简称 FHT）说，作为统计学习开创性文本的作者，他们发表了权威演讲作为关键技术，如套索。“数据科学过程：一种文化”是于教授和 Barter 博士评论的标题，与 Leo Breiman 著名的“两种文化”论文相反，后者声称传统方法和“预测文化”之间存在明显的二分法。再说一次，Yu/Barter 在现代预测方法的开发人员中表现出经验丰富和成就。局外人对传统回归思想和算法回归思想之间的当前联系状态不太满意。（这里的“局外人”一词仅与算法预测相关；所有讨论者都是现代数据分析技术（例如错误发现率、基因组分析和引导程序）的有影响力的开发者。）本文是从局外人的角度撰写的看法。现在重读，一年后，让我意识到一些局外人的局限性：实际上，两种文化之间的结缔组织比我想象的要多，FHT 和 Yu/Barker 都给出了令人信服的例子。话虽如此，额外一年的经验并没有改变我对传统统计回归方法与纯预测算法之间脱节的信念。论文的第 8 节，以表 5 为特色，直接根据六个标准来说明情况。六人中有五人在讨论中或多或少毫发无损。标准 2,