《Causal Inference》经典读后感有感

《Causal Inference》是一本由Hernán MA / Robins JM著作，Boca Raton: Chapman & Hall/CRC出版的2020图书，本书定价：311，页数：，特精心收集的读后感，希望对大家能有帮助。

《Causal Inference》读后感(一)：本书缺点

虽然本书是一本优秀的教材，但是依然存在一些不足之处。我冒昧指出其中几点。

第一，中介效应分析（mediation analysis）是因果推断的重要内容之一，然而本书基本没有讨论这一内容。我猜想，主要原因可能是本书作者的同事Tyler VanderWeele教授，一位对中介效应分析论著颇丰的学者，在本书面世之前已经单独出版了一本中介效应分析的专著，因而为了避免重复，本书没有纳入中介效应分析。

第二，本书作者都是哈佛大学的学者，因而本书的主体内容大多来自于哈佛学派的论述。虽然能感受到作者试图兼顾其他学校学者的成果，然而其对哈佛学派的侧重依然非常明显。同时，哈佛学派的某些论述并不一定被所有学者认同接受，而本书却也大篇幅地介绍讨论。所以，读者在阅读本书时最好配合阅读其他学者的论文，学会甄别本书内容。

第三，本书假设读者具有一定的统计基础，因而未对统计知识做过多介绍。本书给出的统计内容不具备知识梯度上的连续性，读者在阅读时若遇见过于复杂的统计论证，可以暂时选择跳过，不要被吓到失去信心。

第四，本书原文语言有时过于重复琐碎。

《Causal Inference》读后感(二)：因果推断，反事实思维

随着业界的不断应用，”因果推断“在广告、推荐等领域应用越来越多，在风控领域还甚少看到相关应用，之前对这块也知之甚少，所以学习了解一下相关知识，几年看过另一本经典入门书籍《The book of why》，主要是以causal diagram的方式来介绍因果推断，从相关关系的observational analysis,到引入因果推断的do analysis和counterfactual analysis，建立了基于causal diagram的分析框架。

本书侧重从分析、实验和计量的角度来对因果推断的概念和方法进行介绍，对因果推断的由来和基础概念介绍的比较清楚，有了这些基础的sense再去学习应用具体的方法会更能掌握其原理。全书分为3个部分，Causal inference without models, Causal inference with models, Causal inference from complex logitudinal data，重点在第一部分，并且侧重于bias的介绍，virance较少涉及。

从本质上来讲，如果进行完全随机实验或者conditional随机实验后对结果按照不同样本比例进行处理（如standardization或inverse probability weighthing等）则能得到causal effect的估计，导致causal effect不等于associational effect的原因往往是因为存在混淆变量（confounding variables），即存在同时对treatment和outcome有影响的变量使得基于观测数据评估treatment对outcome的causal effect时存在bias，所以能准确地识别出混淆变量非常关键，这除了统计方法外往往需要结合一定的专家经验，standardization或inverse probability weighthing等方法是为了准确评估总体人群中的causal effect，也可以基于confounding variables对人群分层后评估每个segment的causal effect，其实和我们在风控领域建模前做的一系列segment analysis类似。除了confounding以外值得关注的问题还有effect modification, selection bias, measurement bias, random variability，其中effect modification, selection bias问题和confounding会有些相似性。

然而现实世界和现实问题往往特征维度很高，若高维特征组合分类则类别过多不太可行，这时可以应用倾向性评分等方法，倾向性评分先是基于confounding variables L建立回归模型预测treatment A的probability作为倾向性得分，这个倾向性得分作为在评估treatment A对outcome Y的因果效应时的权重，或者可以建立样本带权重的模型，即将可分类的权重变为了连续型预测值作为权重。

工具变量instrument variable没看太明白，第三部分纵向数据logitudinal data主要讲时序treatment的causal inference，由于不同时间点不同treatment的组合会有很多，且时序路径上会存在影响，所以问题会更复杂，由于目前还未遇到实际的应用问题，所以这块也没细看，等遇到实际问题后再仔细研究。总体来讲，这本基础性的书籍能帮助我们更好地掌握很多基本概念，涉及到具体的常用的统计方法和机器学习算法与因果推断的结合应用还需要去学习对应的具体的模型。

《Causal Inference》读后感(三)：摘录&总结

Introduction

走马观花的看了一下，了解了下基本原理。第二部分的代码也还没顾得上仔细研读。这就算第一个完成了吧，等下一轮迭代吧

【Definition of Causal Inference】

1. We compare (usually only mentally) the outcome when an action is taken with the outcome when the action is withheld. 2. Individual Causal Effects:

3. the treatment has a causal effect on an individual’s outcome if $Y^{a=1} neq Y^{a=0}$ for the individual ($Y^{a=1}, Y^{a=0}$ are refered as potential outcome or counterfactual outcome). 4. Average causal effects in a population: indicidual causal effects cann't be measured 5. Measure of causal effects: 1.causal risk difference

6. risk ratio 7. odds ratio 8. Association is not causation 9. To analysis causal effects, there are two methods:

10. Randomied experiments 11. Observational studies

【Randomized Experimens】

1. Randomization

2. Conditional randomization 3. Marginally randomization 4. Approaches to estimate average causal effects:

5. Standardiztion 6. Inverse probability weighting

【Observational Studies】

1. Sometimes randomized experiments are not possible, we need to get conclusion from the observation 2. An observational study can be conceptualized as a conditionally randomized experiment under the following three conditions:

3. Consistency 4. Exchangeability 5. Positivity

【Effect Modification】

1. Instead of average causal effects in the whole population, many causal questions are about the subset of the population. 2. Effect modifier:

3. We say that M is the modifier of effect A on Y when the average causal effect of A on Y vaies across the levels of M. 4. A stratified(分层) analysis is the natural way to identiy effect modifier. 5. In the absence of marginal randomization, achieving the goals requires ajustments to the variable L to ensure conditional exchangeablity of the treated and untreated. Forms of ajustments:

6. Statification 7. Matching

2. 8. Construct a subset of the population in which the varibles L have the same distribution in both treated and untreated. 9. Four methods to estimate average causal effect

10. Both marginal and conditional effects:

11. IP weighting and standardization 12. Conditional effects in a subset of population

13. Statification and matching

【Interaction】

Many problems are about the interaction of several treatments. This chapter provides definition of interaction both in the counterfactual framework and the sufficient-component framework.

1. We refer to interventions on two or more treatments as joint interventions. 2. Identification of interaction:

3. The three key identifying conditions were exchangeability, positivity, and consistency. Because interaction is concerned with the joint effect of two (or more) treatments  and , identifying interaction requires exchangeability, positivity, and consistency for both treatments. 4. suffient-component framework:

5. The graphical representation of sufficient-component causes helps visualize a key consequence of effect modification: 6. Two frameworks:

7. The sufficient-component-cause framework and the counterfactual (potential outcomes) framework address different questions. The counterfactual approach addresses the question “what happens？” Thesufficient-component-cause approach addresses the question “how does it happen？”

8. The sufficient component cause model considers sets of actions, events, or states of nature which together inevitably bring about the outcome under consideration. 9. the potential outcomes framework addresses the question, “What would have occurred if a particular factor were intervened upon and thus set to a different level than it in fact was？”

【Graphical Representation of Causal Effects】

1. Causal diagrams:

2. The modern theory of diagrams for causal inference arose within the disciplines of computer science and artificial intelligence.

【Bias】

There are two qualitatively different reasons why causal inferences may be wrong:

1. systematic bias

2. selection bias both of which may arise in observational studies and in randomized experiments 3. measurement bias–both of which may arise in observational studies and in randomized experiments 4. unmeasuredconfounding–which is not expected in randomized experiments. 5.

random variability.

6.

If those factors affect the risk of developing the outcome (e.g., another person’s looking up), then the effects of those factors become entangled with the effect of treatment. We then say that there is confounding, which is just a form of lack of exchangeability between the treated and the untreated.

7. Confounding is often viewed as the main shortcoming of observational studies.

Unlike confounding, this type of bias is not due to the presence of common causes of treatment and outcome, and can arise in both randomized experiments and observational studies. Like confounding, selection bias is just a form of lack of exchangeability between the treated and the untreated

【Why Models？】

1. Estimators

2. Nonparametric estimators 3. Parametric estimators

【Three Methods to Estimate Average Causal Effects】

1. Three methods to estimate the average causal effects (these three methods are often collectively referred to as g-methods because they are designed for application to generalized treatment contrasts involving treatments that vary over time.):

2. how to use IP weighting to estimate this effect from observational data. 3. how to use standardization to estimate the average causal effect of smoking cessation on body weight gain. 4. g-estimation.

【Instrumental Variable Estimation】

1. The causal inference methods described so far in this book rely on a key untestable assumption: all variables needed to adjust for confounding and selection bias have been identified and correctly measured.It turns out that there exist other methods that can validly estimate causal effects under an alternative set of assumptions that do not require measuring all adjustment factors. Instrumental variable estimation is one of those methods.

【Casual Survival Analysis】

1. we have been concerned with causal questions about the treatment effects on outcomes occurring at a particular time point.Many causal questions, however, are concerned with treatment effects on the time until the occurrence of an event of interest. 2. The term “survival analysis”, or the equivalent term “failure time analysis”, is applied to any analyses about time to an event

【Time-varying Treatments】

1. So far this book has dealt with fixed treatments which do not vary over time. However, many causal questions involve treatments that vary over time. 2. Sequential exchangeability is a key condition to identify the causal effects of time-varying treatments. 3. When treatment-confounder feedback exists, using traditional adjustment methods may introduce bias in the effect estimates. 4. G-methods can be ajusted for time-varying treatments