Series ID: EDKS-ADV-VOC-ZH-0042 · Advanced English Vocabulary (Chinese Edition) · Lesson No.042 · C1 → C2 · 简体中文辅助
Association asks what moves together. Causation asks what would change if the cause were changed while the relevant comparison were held conceptually fixed.
association 问什么一起变化；causation 问如果改变原因，在合适的比较下，结果会发生什么变化。
Advanced causal English includes cause, effect, causal effect, mechanism, pathway, confounder, mediator, moderator, interaction, intervention, treatment, exposure, outcome, counterfactual, adjustment and identification. These are not decorative synonyms for “X is related to Y.” Each term places the claim at a different point in the causal argument.
Harvard’s freely available Causal Inference: What If frames causal questions in terms of interventions and counterfactual outcomes, while OpenStax’s research-methods materials emphasise that ordinary correlation is insufficient for cause-and-effect claims and that confounding creates alternative explanations. This lesson uses that boundary as a language discipline rather than as a statistics manual.
A causal verb is a methodological commitment.
每一个因果动词，都是一种方法论承诺。
Part I — Build the causal map | 第一部分：建立“因果语言地图”
1. Cause names a producing influence | cause
To say X causes Y is stronger than saying X is associated with Y. It implies that changing X, under the causal contrast being considered, changes Y.
2. Effect names the outcome difference attributable to a cause | effect
A causal effect is not merely an observed difference. It compares what would happen under different interventions or exposure states under a causal framework.
3. Causal effect is a technical phrase | causal effect
Use causal effect when the design/analysis aims to estimate an intervention contrast. Do not upgrade a regression coefficient to a causal effect without identification assumptions.
4. Treatment is not always medicine | treatment
In causal inference, treatment can be any intervention/exposure being compared: a policy, teaching method, price, message or medical treatment.
5. Exposure is broader than treatment | exposure
Exposure often describes what participants experience or possess, whether or not researchers assign it. Observational studies commonly use exposure/outcome language.
6. Outcome is what is affected/measured | outcome
The outcome is the response of interest. “Outcome” does not imply causation by itself; it simply identifies the dependent/result variable in the causal question.
7. Intervention makes causal meaning explicit | intervention
An intervention changes a treatment/exposure according to a defined rule. Asking what would happen if we intervened helps distinguish causal from merely associational questions.
8. Counterfactual means an alternative outcome under another condition | counterfactual
A counterfactual outcome is what would have happened under a different treatment/exposure state. Causal effects compare such potential outcomes conceptually, even though one person cannot usually be observed simultaneously under both states.
9. Potential outcome is formal counterfactual vocabulary | potential outcome
Potential-outcome notation formalises outcomes under specified interventions. General readers may prefer “what would have happened if…”.
10. Mechanism asks how the cause produces the effect | mechanism
A mechanism is the process through which X brings about Y. Evidence of a causal effect does not automatically identify the mechanism.
11. Pathway is a route linking cause and outcome | pathway
A causal pathway can contain intermediate variables. “Pathway” can be biological, behavioural, economic or institutional depending domain.
12. Direct effect excludes specified mediated routes | direct effect
In formal mediation analysis, “direct effect” has technical definitions depending framework. It is not simply “the obvious effect.” Use carefully.
13. Indirect effect works through a mediator | indirect effect
An indirect/mediated effect describes a causal pathway through an intermediate variable under formal assumptions. Observing that the mediator correlates with both variables is not sufficient.
14. Mediator lies on a causal pathway | mediator
A mediator is caused by the exposure and in turn affects the outcome in the causal model. Calling a variable “mediator” already commits to causal structure.
15. Moderator changes the size/direction of a relationship/effect | moderator
A moderator identifies conditions under which an effect/association differs. In causal language, treatment effects may vary across levels of another variable.
16. Interaction is not identical to moderation in every framework | interaction
Harvard’s causal-inference text treats interaction as effect differences under joint interventions and notes dependence on effect scale. Statistical interaction, moderation and causal interaction overlap but are not universal synonyms.
17. Effect modification is domain-specific heterogeneity language | effect modification
Epidemiology often uses effect modification when causal effects differ across levels of another factor. It is not confounding.
18. Confounder distorts causal comparison | confounder
A confounder is a variable that creates mixing of causal effects because treatment groups differ in ways relevant to outcome. Formal causal criteria depend on the causal graph/design, not just correlations.
19. Confounding is not ordinary confusion | confounding
Confounding is a technical causal-inference problem. Saying “the results are confounded by age” means age distorts the exposure–outcome comparison under the design, not that age makes the report confusing.
20. Common cause is a useful intuitive confounder picture | common cause
If Z causes both X and Y, an X–Y association can appear even without direct causal effect. But not every variable associated with both is automatically a confounder after conditioning/causal structure are considered.
21. Collider is a common effect of two variables | collider
Conditioning on a common effect can create a non-causal association between its causes. Collider bias is why “adjust for everything” is not a safe causal strategy.
22. Selection bias can behave like collider conditioning | selection
Entering a study, hospital, platform or analysis sample can depend on variables that create distorted relationships. Selection is a causal-structure problem, not merely small sample size.
23. Adjustment is not automatic confounding control | adjustment
Including covariates in regression can reduce confounding, introduce bias or change the estimand depending on which variables are adjusted for. “Adjusted for” is a method description, not a causal guarantee.
24. Covariate is a neutral modelling term | covariate
A covariate is a variable included in analysis. It may be a confounder, predictor, mediator, precision variable or something else. Do not call every covariate a confounder.
25. Control variable is ambiguous causal language | control
“We controlled for age” usually means age entered an adjustment procedure. It does not mean age’s influence has been perfectly removed.
26. Random assignment targets confounding balance | randomisation
Randomised treatment assignment helps make treatment groups comparable in expectation and supports causal inference under the experiment’s conditions. It does not guarantee perfect balance in every finite sample.
27. Random sampling and random assignment are different | sampling vs assignment
Random sampling addresses population representation; random assignment addresses treatment comparability/causal effect. One does not substitute for the other.
28. Control group provides a comparison condition | control group
A control/comparison group gives a baseline for outcome differences. The validity of causal interpretation depends on assignment, adherence, measurement and other design features.
29. Placebo controls expectations in some trials | placebo
Placebos can help isolate treatment components and expectations where appropriate. The word is specific to certain experimental contexts.
30. Blinding reduces some measurement/behaviour biases | blinding
Blinding/masking can reduce bias from knowing assignment. It is not itself proof that a causal effect exists.
31. Observational causal inference is possible but assumption-heavy | observational
Modern causal inference can estimate causal effects from observational data when assumptions, design emulation and identification strategies are credible. “Observational” therefore does not mean “causal claims impossible,” but it requires more explicit justification.
32. Natural experiment is not simply “data from nature” | natural experiment
A natural/quasi-experimental setting exploits external assignment or variation approximating an experiment under specific assumptions. The label should not be used casually.
33. Instrumental variable is a specialised identification strategy | instrument
An instrument affects treatment/exposure and supports causal identification only under strong assumptions about pathways and independence. It is not just any predictor correlated with treatment.
34. Difference-in-differences is a causal design under assumptions | DiD
Difference-in-differences compares changes across treated and comparison groups, relying on assumptions such as parallel trends. The method name does not guarantee validity.
35. Regression discontinuity uses threshold assignment | RD
Regression discontinuity can estimate local causal effects near a cutoff under continuity/no-manipulation assumptions. It is not ordinary regression.
36. Target trial emulation clarifies observational causal questions | target trial
Harvard causal-inference work popularises emulating a hypothetical randomised trial using observational data: eligibility, treatment strategies, follow-up, outcome and analysis are specified explicitly.
37. Identification asks whether assumptions link data to causal quantity | identification
A causal effect can be conceptually defined but not identifiable from available data without assumptions. Identification is not the same as statistical estimation.
38. Estimation comes after identification | estimation
Once the causal quantity is identified under assumptions, statistical methods estimate it from data. Sophisticated estimation cannot rescue a non-identified causal question.
39. Assumption is part of the causal claim | assumptions
No unmeasured confounding, consistency, positivity, correct model structure and other assumptions can be central. Strong causal language should acknowledge what the result depends on.
40. Positivity means relevant treatment options exist across covariate strata | positivity
Formal positivity requires nonzero probability of treatment alternatives in relevant strata. It is specialist recognition vocabulary.
41. Consistency links observed and counterfactual outcomes | consistency
In causal inference, consistency is a technical assumption connecting the observed outcome under the treatment actually received with the corresponding potential outcome. It is not ordinary “results are consistent.”
42. Exchangeability formalises comparability | exchangeability
Exchangeability expresses that treatment groups are comparable with respect to potential outcomes under a causal framework. Randomisation helps support it; observational analyses attempt to justify conditional versions.
43. Internal validity concerns causal credibility inside the study | internal validity
A study can be internally valid for its participants/settings yet poorly generalise elsewhere. Causal truth in one context and transportability are separate questions.
44. External validity concerns generalisation/transport | external validity
External validity asks whether results apply beyond the study population or conditions. A causal effect can be valid locally but differ elsewhere.
45. Part I checkpoint: a causal claim needs a contrast, design and assumptions | 第一部分检查点
Before saying causes, affects, leads to, reduces, increases, identify the intervention/contrast, outcome, comparison group, design, confounding strategy and assumptions. The verb commits you to that architecture.
Part II — Calibrate causal verbs, mechanisms and pathways | 第二部分：校准因果动词、机制与路径
46. Associated with is non-causal by default | associated with
“X was associated with Y” reports a relationship. It does not state what would happen if X were changed.
47. Predicts is predictive, not automatically causal | predicts
“X predicts Y” can mean X helps forecast Y in a model. It does not mean intervening on X changes Y.
48. Contributes to is causal but often partial | contributes to
Contributes to suggests X plays a causal role among multiple factors. It is weaker than “fully causes” but stronger than association.
49. Affects implies causal influence | affects
“X affects Y” states causal influence. Do not use it merely because a regression coefficient differs from zero.
50. Influences often carries causal suggestion | influences
Influences can be slightly softer than causes but still suggests directional effect. Method should justify it.
51. Leads to is strongly directional | leads to
“X leads to Y” normally implies a causal pathway. It should not replace “is followed by” when only temporal ordering is known.
52. Results in is a strong causal construction | results in
Results in says Y occurs as a consequence of X. This is inappropriate for mere association.
53. Causes is the clearest causal verb | causes
Use when the design and assumptions justify a causal conclusion. Strong wording is valuable when earned; false caution can also be unhelpful.
54. Drives suggests an important mechanism or force | drives
“X drives Y” often implies X is a major causal force. It is stronger and more mechanistic than “associated with.”
55. Triggers means initiates an event/process | triggers
Trigger implies X initiates Y, often rapidly or conditionally. It does not necessarily mean X alone is sufficient.
56. Produces implies generation | produces
“The reaction produces heat” is mechanistically causal. In social-science prose, produces can sound too deterministic if evidence is probabilistic.
57. Enables makes an outcome possible | enables
Enables describes a facilitating cause/condition rather than a sufficient cause. It can be appropriate for infrastructure, tools and institutional processes.
58. Facilitates makes an outcome easier/more likely | facilitates
Facilitation implies a causal contribution without claiming inevitability. It is common in education and organisational writing.
59. Prevents is a causal blocking verb | prevents
To say X prevents Y means intervention on X reduces occurrence of Y under the causal contrast. Observing lower Y among people with X is not enough by itself.
60. Reduces/increases can be causal or descriptive | reduces/increases
“X reduced Y” usually sounds causal. “Y decreased among those with X” is descriptive. Sentence grammar can silently upgrade evidence.
61. Explains can mean statistical or causal explanation | explains
“X explains 30% of variance” is statistical. “X explains why Y occurred” is causal/mechanistic. Keep the sense explicit.
62. Accounts for is similarly polysemous | accounts for
Accounts for can mean constitutes a share, statistically explains variation, or causally explains. Avoid ambiguity.
63. Mechanism evidence differs from effect evidence | mechanism vs effect
A randomised trial can establish an effect while leaving the mechanism uncertain. Conversely, a plausible mechanism does not prove the net effect occurs in real populations.
64. Mechanistic plausibility supports but does not replace design | plausibility
A biologically or logically plausible mechanism can strengthen interpretation, but observational association remains vulnerable to confounding unless design assumptions are addressed.
65. Mediator is downstream of exposure | mediator
If X changes M and M changes Y, M may mediate part of X’s effect on Y. This is a causal pathway claim, not merely “M sits statistically between X and Y.”
66. Confounder is not on the causal pathway | confounder vs mediator
Confounders precede/create treatment-outcome mixing; mediators carry part of the treatment effect. Adjusting for a mediator can remove part of the effect you want to estimate.
67. Moderator is not mediator | moderator vs mediator
A moderator changes effect magnitude across contexts; a mediator transmits effect through a pathway. One answers “for whom/when?”; the other “through what route?”
68. Interaction depends on effect scale | scale dependence
An interaction may exist on risk-difference scale but not on ratio scale. Saying “there is no interaction” without naming scale can be incomplete.
69. Heterogeneous treatment effect means effects differ | heterogeneity
Average treatment effect can hide stronger benefit in one group and weaker/no benefit in another. The average causal effect is not every individual’s effect.
70. Necessary cause is required but may not be sufficient | necessary
A necessary condition must be present for the outcome under the defined mechanism, but may not produce the outcome alone.
71. Sufficient cause can produce outcome under conditions | sufficient
A sufficient causal set can produce the outcome, though individual components may not be sufficient alone. Formal causal theory can define this more precisely.
72. Risk factor is not automatically a cause | risk factor
A risk factor can be predictive/associated, causal, or a marker depending evidence and usage. Avoid assuming the label means modifiable cause.
73. Protective factor is similarly ambiguous | protective factor
“Protective” can sound causal. In observational research, specify whether evidence is associational or causal.
74. Marker is often deliberately non-causal | marker
A marker indicates status/risk without claiming intervention on it changes outcome. This is useful when causal role is unknown.
75. Determinant can be causal but broad | determinant
Public health and social science use determinant for factors shaping outcomes. It often implies causal influence but may bundle complex pathways.
76. Driver is strong systems language | driver
Driver suggests an important force producing change. Use only when mechanism/evidence supports it.
77. Root cause is a diagnostic claim | root cause
A root cause is an underlying causal source in engineering/operations. Finding a correlated symptom does not identify the root cause.
78. Proximate and distal causes describe causal distance | proximate/distal
Proximate causes operate nearer the outcome; distal causes act through longer chains. “Closer” does not mean more important.
79. Upstream/downstream are pathway metaphors | upstream/downstream
Policy and health writing use upstream factors for earlier structural causes and downstream factors for later consequences/interventions. The metaphor encodes causal ordering.
80. Necessary temporal order does not prove causation | temporality
Cause must precede effect in many frameworks, but earlier occurrence alone does not establish cause. A rooster crows before sunrise without causing it.
Part III — Mandarin-to-English causal control | 第三部分：中文母语学习者的因果词汇转换
81. 导致 = cause / lead to / result in | 导致
All three are causal. If the source only shows association, translate more cautiously: was associated with, not led to.
82. 造成 = cause / produce / result in | 造成
Often strong and consequence-focused. “造成损失” → caused/resulted in losses. Do not weaken or strengthen causality accidentally.
83. 引起 = cause / trigger / give rise to | 引起
Trigger suits initiation; cause broad production; give rise to formal. Choose from mechanism and register.
84. 影响 = affect / influence / be associated with | 影响
Chinese 影响 is often used more loosely than English affect. If the evidence is observational, is associated with may be safer than a causal translation.
85. 促成 = contribute to / facilitate | 促成
Use when X is one contributing condition among several. It avoids claiming X alone is sufficient.
86. 驱动 = drive | 驱动
Drive is strong mechanistic language. In business reports it is often overused. Verify whether the evidence supports a driver claim.
87. 机制 = mechanism | 机制
A mechanism explains how causal influence is transmitted. “可能机制” → possible/plausible mechanism if not established.
88. 路径 = pathway | 路径
“作用路径” → causal pathway/mechanism. A statistical sequence is not automatically causal.
89. 混杂因素 = confounder/confounding factor | 混杂
Use the technical term when the variable distorts the causal comparison. Do not translate as “confusing factor.”
90. 中介变量 = mediator | 中介
A mediator lies on the hypothesised causal pathway. Calling it mediator requires more than statistical association with exposure/outcome.
91. 调节变量 = moderator / effect modifier | 调节
Moderator is common in psychology/social science; effect modifier common in epidemiology. Both concern differences in relationships/effects across conditions, but frameworks differ.
92. 交互作用 = interaction | 交互作用
Translate as interaction and specify scale/model if technical interpretation matters.
93. 干预 = intervention | 干预
In causal/research contexts, intervention is a deliberate action/change. In everyday education, support/strategy may sound more natural depending use.
94. 反事实 = counterfactual | 反事实
Formal causal inference uses counterfactual outcomes. General explanation: “what would have happened if the treatment had been different.”
95. 因果效应 = causal effect | 因果效应
Do not translate any observed difference as causal effect. The term presupposes a causal estimand/design.
96. 控制了 = adjusted for / controlled for | 控制
“控制年龄后” → after adjusting/controlling for age. This describes modelling, not proof that all age-related confounding is removed.
97. 独立影响 can overclaim | 独立影响
Chinese research prose often says “independent effect.” If the study is associational, independent association after adjustment may be safer than independent causal effect.
98. 因素 = factor is causally ambiguous | 因素
Factor can mean predictor, correlate, determinant or cause. Specify role instead of relying on the vague noun.
99. Part III checkpoint: Chinese causal verbs may be stronger than the design | 第三部分检查点
Translate according to evidence, not rhetorical habit. If a Chinese summary says 导致 but the study only reports correlation, accurate English may need to weaken the causal verb rather than reproduce the overclaim.
Part IV — Causal failure laboratory, FENCE and mastery | 第四部分：因果失误实验室、FENCE 与掌握系统
100. Adjusted association is not automatically causal | 调整后相关 ≠ 因果
A model adjusts for age and sex, but unmeasured confounding remains possible. “Independent predictor” should not be rewritten as “independent cause.”
101. Overadjusting for a mediator can remove part of the effect | 过度调整
If treatment affects mediator M which affects outcome, adjusting for M changes the estimand. “More adjusted” is not automatically “more causal.”
102. Adjusting for a collider can create bias | collider bias
Conditioning on a common effect of exposure and outcome causes can open a non-causal path. “Control every available variable” is unsafe.
103. Randomisation supports causality but chance imbalance remains | finite samples
Random assignment removes systematic treatment selection in expectation; a finite experiment can still show chance imbalances and statistical uncertainty.
104. Random sampling without random assignment does not establish effect | sampling trap
A representative observational sample can generalise an association well while still leaving confounding. External representativeness and causal identification are different.
105. Temporal order alone is insufficient | before ≠ because
X happened before Y. That rules out some reverse-causation stories but leaves common causes and coincidence. “Before” is not “because.”
106. Dose-response pattern supports but does not prove mechanism | dose response
Higher exposure aligns with stronger outcome. This can support a causal interpretation but can also arise from confounding or selection.
107. Plausible mechanism plus association can still be wrong | mechanism fallacy
Humans can tell convincing causal stories about coincidental patterns. Mechanistic plausibility is evidence, not a substitute for design.
108. A causal effect can exist without known mechanism | effect before mechanism
A well-designed experiment may show an intervention changes outcomes before scientists fully understand the pathway. Mechanism uncertainty does not erase the effect.
109. Average effect can hide harm in a subgroup | heterogeneity
A treatment helps most people but harms one subgroup. “The treatment is beneficial” may be true on average but incomplete for decisions.
110. No average effect can hide opposite subgroup effects | cancellation
Strong benefit in one group and strong harm in another can average to zero. “No effect” may conceal heterogeneity.
111. Mediator–outcome association can be confounded | mediation trap
Even if treatment affects M and M correlates with Y, the M→Y causal pathway can be confounded. Formal mediation requires more assumptions than a sequence of significant coefficients.
112. Moderator is not “variable with significant interaction term” only | moderation
Statistical interaction depends on scale/model specification; causal moderation requires an interpretable effect-heterogeneity question.
113. Confounder is not “anything correlated with both variables” | confounder rule
A mediator can correlate with exposure and outcome but should not be treated as a baseline confounder for total-effect estimation. Causal ordering matters.
114. Proxy adjustment may leave residual confounding | proxy
Adjusting for an imperfect proxy of socioeconomic status may reduce but not eliminate confounding. Measurement quality enters causal inference.
115. Measurement error can create causal bias | measurement
Misclassified exposure, outcome or confounders can distort estimated effects. Better statistical sophistication cannot fully rescue poor measurement.
116. Attrition can create selection bias | dropout
If follow-up depends on treatment and outcome risk, complete-case analysis can lose the original comparability of groups.
117. Noncompliance complicates treatment assignment | adherence
Assignment to treatment and treatment actually received can differ. Intention-to-treat and per-protocol effects answer different causal questions.
118. Intention-to-treat is an assignment effect | ITT
ITT compares groups as randomised, preserving randomisation. It estimates effect of assignment/strategy, not necessarily effect of actually receiving treatment.
119. Per-protocol effect needs stronger adjustment | per protocol
Comparing adherers can reintroduce confounding because adherence is not random. Formal methods attempt to adjust for this selection.
120. Surrogate outcome can mislead mechanism | surrogate
An intervention may improve a surrogate marker without improving the final outcome. Mechanistic relevance does not guarantee patient/system benefit.
121. Build the causal FENCE | 因果 FENCE
| Fence | Question | Job |
|---|---|---|
| F0 Question | What intervention/contrast? | define causal target |
| F1 Outcome | What changes? | define Y |
| F2 Design | Randomised, observational, quasi-experimental? | claim strength |
| F3 Confounding | What alternative common causes? | backdoor control |
| F4 Selection | Who enters/remains? | selection bias |
| F5 Pathway | Confounder, mediator, collider? | causal roles |
| F6 Heterogeneity | For whom/when? | moderation/interaction |
| F7 Assumptions | What must be true? | identification |
| F8 Verb | Associated, contributes, causes? | language calibration |
122. Practice A — verb ladder | 练习 A
- associated with
- predicts
- contributes to
- affects
- leads to
- causes
- drives
For each verb, write the minimum evidence/design you would want before using it in a formal report.
123. Practice B — role classification | 练习 B
Given X → M → Y with Z → X and Z → Y, classify X, M, Y and Z as exposure, mediator, outcome and confounder. Then explain why adjusting for M changes the total-effect question.
124. Practice C — confounder or collider? | 练习 C
Draw two arrows into Z (X → Z ← Y). Explain why conditioning on Z can create association between X and Y. Then contrast with Z → X and Z → Y.
125. Practice D — Mandarin repair | 练习 D
- 该研究发现相关，但不能证明导致。
- 调整年龄后，关联仍然存在。
- 这个变量可能是中介，而不是混杂因素。
- 治疗效果在不同年龄组中不同。
- 一个可能机制是……但机制尚未得到直接验证。
126. A 20-minute causal-language session | 20 分钟训练
| Time | Task |
|---|---|
| 0–4 | Write the intervention question. |
| 4–8 | Draw exposure/outcome/confounders/pathways. |
| 8–12 | Identify design and assumptions. |
| 12–16 | Choose causal verb strength. |
| 16–20 | Write one cautious and one fully justified version. |
127. Seven-day mastery route | 七天路线
| Day | Focus |
|---|---|
| 1 | cause/effect/intervention |
| 2 | confounder/collider/selection |
| 3 | mediator/pathway/mechanism |
| 4 | moderator/interaction/heterogeneity |
| 5 | design/identification/assumptions |
| 6 | Mandarin causal verbs |
| 7 | full FENCE audit |
128. First weak-link diagnostic | 弱点诊断
| Symptom | Weakness | Repair |
|---|---|---|
| I say causes after regression. | design blindness | verb ladder |
| I adjust for everything. | causal-role blindness | confounder/mediator/collider drills |
| I call mediator a moderator. | path vs heterogeneity | through-what vs for-whom test |
| I think random sample = causal. | sampling/assignment | design contrast |
| I use mechanism as proof. | mechanism/effect confusion | two-evidence-track rule |
| I translate 影响 as affect automatically. | bilingual overclaim | association fallback |
129. Reading harvest | 阅读提取
Collect phrases such as causal effect of, under the assumption that, after adjustment for, potential mechanism, mediated through, effect differed by, consistent with a causal interpretation. These qualifiers encode the evidential contract.
130. Writing activation | 写作激活
Use: causal question → design → effect estimate → assumptions → mechanism status → limits. This keeps the result separate from the causal story.
131. Paraphrase test | 改述测试
Original: “The adjusted association remained statistically significant.” Unsafe: “X had an independent causal effect.” Safe: “The association persisted after adjustment for the measured covariates included in the model.”
132. Canonical ownership boundary | 本课所有权边界
Lesson 042 owns causal claim language: cause/effect, intervention, counterfactual, confounder/collider/selection, mechanism/pathway, mediator/moderator/interaction, identification and causal verb strength. Lesson 041 owns association before causation.
133. Recommended reference floor | 推荐参考资源
- Hernán & Robins — Causal Inference: What If — counterfactuals, interventions, interaction and modern causal inference.
- Harvard CAUSALab — causal-inference research and training.
- OpenStax — Experimental Design — random assignment and confounding.
- OpenStax — Correlation and Causation.
134. Final rules | 最后的规则
A causal verb is a methodological commitment.
每一个因果动词，都是一种方法论承诺。
Do not adjust for a variable until you know what causal role you think it plays.
