Skip to content

Causal Inference: What If 第 1-3 章中英对照

原书: Causal Inference: What If
作者: Miguel A. Hernán, James M. Robins
出版社: Chapman & Hall/CRC (2020)
ISBN: 978-1-4398-8662-5
官方链接: https://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/
翻译时间: 2026-06-01
状态: 中英对照编译版


📚 目录 (Table of Contents)

第一部分:因果推断的基本概念 (Chapters 1-10)

  • Chapter 1: A definition of causal effects (因果效应的定义)
  • Chapter 2: Randomized trials (随机试验)
  • Chapter 3: Measures of effect (效应度量)

Chapter 1: A Definition of Causal Effects

第 1 章:因果效应的定义

1.1 引言 (Introduction)

English:

Scientists, policymakers, and the public are often interested in causal effects. What is the effect of aspirin on headache? What is the effect of a new drug on mortality? What is the effect of a job training program on earnings? Answering these questions requires a definition of causal effects.

中文:

科学家、政策制定者和公众通常对因果效应感兴趣。阿司匹林对头痛有什么效果?新药对死亡率有什么影响?就业培训计划对收入有什么作用?回答这些问题需要对因果效应的定义。


1.2 潜在结果框架 (The Potential Outcomes Framework)

English:

We use the potential outcomes framework, also known as the Rubin Causal Model, to define causal effects. For each individual i, we define two potential outcomes: - Yᵢᵃ⁼¹: the outcome if individual i receives treatment (A = 1) - Yᵢᵃ⁼⁰: the outcome if individual i does not receive treatment (A = 0)

中文:

我们使用潜在结果框架(也称为 Rubin 因果模型)来定义因果效应。对于每个个体 i,我们定义两个潜在结果: - Yᵢᵃ⁼¹:如果个体 i 接受治疗 (A = 1) 的结果 - Yᵢᵃ⁼⁰:如果个体 i 未接受治疗 (A = 0) 的结果

关键概念 (Key Concepts):

英文 中文 说明
Potential outcomes 潜在结果 每个可能的干预下会发生的结果
Treatment 处理/干预 可以改变的条件或行动
Counterfactual 反事实 未实际发生的结果
Causal effect 因果效应 不同干预下结果的比较

1.3 个体因果效应 (Individual Causal Effect)

English:

The causal effect of treatment A on outcome Y for individual i is defined as a comparison of the two potential outcomes Yᵢᵃ⁼¹ and Yᵢᵃ⁼⁰. The comparison can be on the additive scale:

Individual causal effect = Yᵢᵃ⁼¹ - Yᵢᵃ⁼⁰

Or on the multiplicative scale:

Individual causal effect = Yᵢᵃ⁼¹ / Yᵢᵃ⁼⁰

中文:

处理 A 对个体 i 结果 Y 的因果效应定义为两个潜在结果 Yᵢᵃ⁼¹ 和 Yᵢᵃ⁼⁰ 的比较。比较可以在加法尺度上:

个体因果效应 = Yᵢᵃ⁼¹ - Yᵢᵃ⁼⁰

或在乘法尺度上:

个体因果效应 = Yᵢᵃ⁼¹ / Yᵢᵃ⁼⁰


1.4 因果推断的基本问题 (The Fundamental Problem of Causal Inference)

English:

For each individual, we can observe only one of the two potential outcomes. If individual i receives treatment (A = 1), we observe Yᵢ = Yᵢᵃ⁼¹, but we cannot observe Yᵢᵃ⁼⁰. If individual i does not receive treatment (A = 0), we observe Yᵢ = Yᵢᵃ⁼⁰, but we cannot observe Yᵢᵃ⁼¹.

This is known as the Fundamental Problem of Causal Inference: we cannot observe both potential outcomes for the same individual.

中文:

对于每个个体,我们只能观察到两个潜在结果中的一个。如果个体 i 接受处理 (A = 1),我们观察到 Yᵢ = Yᵢᵃ⁼¹,但无法观察到 Yᵢᵃ⁼⁰。如果个体 i 未接受处理 (A = 0),我们观察到 Yᵢ = Yᵢᵃ⁼⁰,但无法观察到 Yᵢᵃ⁼¹。

这被称为因果推断的基本问题:我们无法观察到同一个体的两个潜在结果。

图示说明 (Visual Explanation):

个体 i 的实际观察:
┌──────────────┬─────────────────┬─────────────────┐
│ 实际处理 (A) │  观察到的结果   │  缺失的结果     │
├──────────────┼─────────────────┼─────────────────┤
│     A = 1    │   Y = Yᵃ⁼¹ ✓   │   Yᵃ⁼⁰ ✗ (缺失) │
│     A = 0    │   Y = Yᵃ⁼⁰ ✓   │   Yᵃ⁼¹ ✗ (缺失) │
└──────────────┴─────────────────┴─────────────────┘

问题:如何估计缺失的潜在结果?

1.5 平均因果效应 (Average Causal Effect)

English:

Since we cannot identify individual causal effects, we focus on average causal effects in a population. The average causal effect (ACE) is defined as:

ACE = E[Yᵃ⁼¹] - E[Yᵃ⁼⁰]

where E[·] denotes the expectation (mean) over the population.

中文:

由于我们无法识别个体因果效应,我们关注人群中的平均因果效应。平均因果效应 (ACE) 定义为:

ACE = E[Yᵃ⁼¹] - E[Yᵃ⁼⁰]

其中 E[·] 表示人群上的期望(均值)。

常用度量 (Common Measures):

度量 公式 名称
因果风险差 E[Yᵃ⁼¹] - E[Yᵃ⁼⁰] Causal Risk Difference
因果风险比 E[Yᵃ⁼¹] / E[Yᵃ⁼⁰] Causal Risk Ratio
因果优势比 [E[Yᵃ⁼¹]/(1-E[Yᵃ⁼¹])] / [E[Yᵃ⁼⁰]/(1-E[Yᵃ⁼⁰])] Causal Odds Ratio

1.6 可交换性 (Exchangeability)

English:

Exchangeability is a key assumption for identifying causal effects from observational data. We say that the treated and untreated groups are exchangeable if:

Yᵃ ⫫ A for all a

This means that the potential outcomes Yᵃ are independent of the actual treatment A. In other words, the treated and untreated groups would have had the same distribution of outcomes if they had received the same treatment.

中文:

可交换性是从观察数据中识别因果效应的关键假设。我们说处理组和未处理组是可交换的,如果:

Yᵃ ⫫ A 对于所有 a

这意味着潜在结果 Yᵃ 与实际处理 A 独立。换句话说,如果处理组和未处理组接受相同的处理,它们的结果分布会相同。

可交换性的含义 (Implications):

如果可交换性成立:
E[Yᵃ | A = 1] = E[Yᵃ | A = 0] = E[Yᵃ]

这意味着:
- 处理组在未处理时的预期结果 = 未处理组在未处理时的预期结果
- 处理组在处理时的预期结果 = 未处理组在处理时的预期结果

1.7 随机化的作用 (The Role of Randomization)

English:

Randomization ensures exchangeability on average. In a randomized trial, treatment assignment A is determined by a random mechanism (e.g., coin flip), independent of individual characteristics. Therefore:

Yᵃ ⫫ A (marginal exchangeability holds)

This is why randomized trials are considered the gold standard for causal inference.

中文:

随机化确保平均意义上的可交换性。在随机试验中,处理分配 A 由随机机制(如抛硬币)决定,独立于个体特征。因此:

Yᵃ ⫫ A (边缘可交换性成立)

这就是为什么随机试验被认为是因果推断的黄金标准。


1.8 条件可交换性 (Conditional Exchangeability)

English:

In observational studies, marginal exchangeability may not hold due to confounding. However, conditional exchangeability may hold within strata of measured covariates L:

Yᵃ ⫫ A | L for all a

This means that within each level of L, the treated and untreated are exchangeable.

中文:

在观察性研究中,由于混杂,边缘可交换性可能不成立。然而,在测量的协变量 L 的层内,条件可交换性可能成立:

Yᵃ ⫫ A | L 对于所有 a

这意味着在 L 的每个水平内,处理组和未处理组是可交换的。


1.9 一致性 (Consistency)

English:

The consistency assumption states that the observed outcome Y equals the potential outcome Yᵃ corresponding to the actual treatment A:

If A = a, then Y = Yᵃ

Or equivalently: Y = Yᴬ

This assumption links the observed data to the potential outcomes framework.

中文:

一致性假设指出,观察到的结果 Y 等于与实际处理 A 对应的潜在结果 Yᵃ:

如果 A = a,则 Y = Yᵃ

或等价地:Y = Yᴬ

这个假设将观察数据与潜在结果框架联系起来。


1.10 本章小结 (Chapter Summary)

核心概念 (Core Concepts):

  1. Potential Outcomes (潜在结果)
  2. Yᵃ: 如果处理为 a 时的结果
  3. 每个个体有多个潜在结果,但只能观察到一个

  4. Causal Effect (因果效应)

  5. 个体层面:Yᵃ⁼¹ vs Yᵃ⁼⁰ 的比较
  6. 人群层面:E[Yᵃ⁼¹] vs E[Yᵃ⁼⁰] 的比较

  7. Fundamental Problem (基本问题)

  8. 无法同时观察同一个体的两个潜在结果

  9. Identification Assumptions (识别假设)

  10. Exchangeability (可交换性)
  11. Consistency (一致性)
  12. Positivity (正值性)

关键术语表 (Glossary):

English 中文 定义
Potential outcome 潜在结果 每个干预下可能发生的结果
Counterfactual 反事实 未实际发生的潜在结果
Treatment 处理/干预 可以改变的条件
Causal effect 因果效应 不同干预结果的比较
Exchangeability 可交换性 处理组与对照组可比较
Randomization 随机化 随机分配处理
Confounding 混杂 影响因果估计的因素
Consistency 一致性 观察结果等于对应潜在结果

Chapter 2: Randomized Trials

第 2 章:随机试验

2.1 什么是随机试验?(What is a Randomized Trial?)

English:

A randomized trial is a study in which the investigator assigns treatment to participants using a random mechanism. The key feature of randomization is that it ensures exchangeability between the treated and untreated groups.

中文:

随机试验是研究者使用随机机制将处理分配给参与者的研究。随机化的关键特征是它确保处理组和未处理组之间的可交换性。


2.2 随机化的类型 (Types of Randomization)

English:

There are several types of randomization:

  1. Simple randomization: Each participant has the same probability of receiving treatment (e.g., coin flip)
  2. Blocked randomization: Randomization within blocks to ensure balance
  3. Stratified randomization: Randomization within strata defined by baseline characteristics

中文:

有几种类型的随机化:

  1. 简单随机化: 每个参与者接受处理的概率相同(如抛硬币)
  2. 区组随机化: 在区组内随机化以确保平衡
  3. 分层随机化: 在由基线特征定义的层内随机化

2.3 意向治疗分析 (Intention-to-Treat Analysis)

English:

In an intention-to-treat (ITT) analysis, participants are analyzed according to their assigned treatment, regardless of whether they actually received or adhered to the treatment. ITT preserves the benefits of randomization.

中文:

在意向治疗 (ITT) 分析中,参与者根据其分配的处理进行分析,无论他们是否实际接受或依从该处理。ITT 保留了随机化的好处。

ITT vs PP 对比:

分析类型 英文 中文 优点 缺点
ITT Intention-to-Treat 意向治疗 保持随机化、无偏 可能低估效应
PP Per-Protocol 符合方案 估计实际效应 可能引入偏倚

2.4 随机试验的局限性 (Limitations of Randomized Trials)

English:

Despite being the gold standard, randomized trials have limitations:

  1. Ethical concerns: Some treatments cannot be randomly assigned
  2. Cost: RCTs are expensive to conduct
  3. Generalizability: Results may not apply to all populations
  4. Feasibility: Some research questions cannot be answered by RCTs

中文:

尽管是黄金标准,随机试验有局限性:

  1. 伦理问题: 某些处理不能随机分配
  2. 成本: RCT 实施成本高
  3. 外推性: 结果可能不适用于所有人群
  4. 可行性: 某些研究问题无法通过 RCT 回答

Chapter 3: Measures of Effect

第 3 章:效应度量

3.1 风险差 (Risk Difference)

English:

The causal risk difference is defined as:

RD = E[Yᵃ⁼¹] - E[Yᵃ⁼⁰]

It measures the absolute difference in risk between the treated and untreated groups.

中文:

因果风险差定义为:

RD = E[Yᵃ⁼¹] - E[Yᵃ⁼⁰]

它测量处理组和未处理组之间风险的绝对差异。

示例 (Example):

假设:
- 处理组死亡率:E[Yᵃ⁼¹] = 0.15 (15%)
- 未处理组死亡率:E[Yᵃ⁼⁰] = 0.20 (20%)

风险差:
RD = 0.15 - 0.20 = -0.05

解释:处理降低 5 个百分点的死亡风险


3.2 风险比 (Risk Ratio)

English:

The causal risk ratio is defined as:

RR = E[Yᵃ⁼¹] / E[Yᵃ⁼⁰]

It measures the relative risk in the treated group compared to the untreated group.

中文:

因果风险比定义为:

RR = E[Yᵃ⁼¹] / E[Yᵃ⁼⁰]

它测量处理组相对于未处理组的相对风险。

示例 (Example):

假设:
- 处理组死亡率:E[Yᵃ⁼¹] = 0.15
- 未处理组死亡率:E[Yᵃ⁼⁰] = 0.20

风险比:
RR = 0.15 / 0.20 = 0.75

解释:处理组死亡风险是未处理组的 75%,即降低 25%


3.3 优势比 (Odds Ratio)

English:

The causal odds ratio is defined as:

OR = [E[Yᵃ⁼¹]/(1-E[Yᵃ⁼¹])] / [E[Yᵃ⁼⁰]/(1-E[Yᵃ⁼⁰])]

It is commonly used in case-control studies and logistic regression.

中文:

因果优势比定义为:

OR = [E[Yᵃ⁼¹]/(1-E[Yᵃ⁼¹])] / [E[Yᵃ⁼⁰]/(1-E[Yᵃ⁼⁰])]

它常用于病例对照研究和逻辑回归。


3.4 效应度量的选择 (Choosing a Measure of Effect)

English:

The choice of effect measure depends on:

  1. Research question: Absolute vs relative effects
  2. Outcome frequency: Rare vs common outcomes
  3. Audience: Policymakers may prefer risk differences
  4. Statistical properties: Some measures have better properties

中文:

效应度量的选择取决于:

  1. 研究问题: 绝对效应 vs 相对效应
  2. 结局频率: 罕见 vs 常见结局
  3. 受众: 政策制定者可能更喜欢风险差
  4. 统计性质: 某些度量有更好的性质

对比表 (Comparison):

度量 取值范围 无效应值 解释难度 适用场景
RD [-1, 1] 0 容易 政策决策
RR [0, ∞) 1 中等 流行病学
OR [0, ∞) 1 较难 病例对照研究

3.5 可 collapsibility (Collapsibility)

English:

A measure of effect is collapsible if the marginal effect equals a weighted average of stratum-specific effects. Risk difference and risk ratio are collapsible; odds ratio is not collapsible.

中文:

如果边际效应等于层特异性效应的加权平均,则该效应度量是可 collapsible 的。风险差和风险比是可 collapsible 的;优势比不是 collapsible 的。


3.6 本章小结 (Chapter Summary)

三种主要效应度量:

  1. Risk Difference (风险差)
  2. 绝对效应度量
  3. 易于解释
  4. 适合政策决策

  5. Risk Ratio (风险比)

  6. 相对效应度量
  7. 流行病学常用
  8. 不受基线风险影响

  9. Odds Ratio (优势比)

  10. 病例对照研究专用
  11. 逻辑回归默认
  12. 不可 collapsible

关键术语表 (Glossary):

English 中文 公式
Risk difference 风险差 E[Y¹] - E[Y⁰]
Risk ratio 风险比 E[Y¹] / E[Y⁰]
Odds ratio 优势比 [E[Y¹]/(1-E[Y¹])] / [E[Y⁰]/(1-E[Y⁰])]
Collapsibility 可 collapsibility 边际效应 = 层效应加权平均

📖 学习建议 (Study Recommendations)

理解检查 (Comprehension Check)

问题 1: 什么是因果推断的基本问题?

点击查看答案 我们无法同时观察同一个体的两个潜在结果(Yᵃ⁼¹ 和 Yᵃ⁼⁰),只能观察到其中一个。

问题 2: 随机化为什么能保证可交换性?

点击查看答案 随机化使处理分配独立于个体特征,因此处理组和对照组在潜在结果上具有相同的分布。

问题 3: 什么时候使用风险差而不是风险比?

点击查看答案 当需要评估绝对效应大小、进行政策决策或成本效益分析时,风险差更有用。

延伸阅读 (Further Reading)

  1. ** Rubin, D. B. (1974).** Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology.
  2. Hernán, M. A., & Robins, J. M. (2006). Estimating causal effects from epidemiological data. Journal of Epidemiology & Community Health.
  3. Greenland, S., & Robins, J. M. (1986). Identifiability, exchangeability, and epidemiological confounding. International Journal of Epidemiology.

翻译完成时间: 2026-06-01
原书章节: Chapter 1-3
下一页: 第 4-6 章 (待创建)


**Causal Inference: What If 中英对照 | 第 1-3 章** [返回顶部](#目录-table-of-contents)