MinutePhysics Uses Causal Networks to Prove Statistical Causation

Correlation CAN Imply Causation! | Statistics Misconceptions — minutephysics

minutephysics

Evaluating multiple correlations across interconnected variables allows researchers to systematically eliminate incompatible cause-and-effect pathways and prove direct causality.

AI-assisted summary of Correlation CAN Imply Causation! | Statistics Misconceptions — original source. Check the original for full context.

Source published:

Key ideas

  • The common statistical adage that correlation cannot imply causation is an overcorrection; analyzing networks of multiple correlations can systematically prove or eliminate causal pathways. (0:01)
  • Using a three-variable system (height, cat ownership, and island habitat), researchers can prune 19 theoretically possible causal models down to just two valid configurations through conditional observations. (0:01)
  • Classical causal models break down in quantum mechanics, where entanglement experiments produce statistical correlations that mathematically rule out every classical cause-and-effect explanation. (2:25)

Detailed notes

## Background Created by Henry Reich and published on the YouTube channel minutephysics on August 17, 2017, this video tackles the statistical misconception that correlation can never establish causality. According to the page description, it demonstrates how causal networks allow researchers to infer causation from correlation, while touching on Reichenbach's principle, common causes, and quantum entanglement.

## TL;DR While an isolated correlation between two variables cannot reveal the direction of cause and effect, the widespread belief that statistics can never prove causality is mistaken. By mapping multiple interconnected variables into causal networks and testing for conditional correlations—demonstrated through a thought experiment involving height, cat ownership, and island environments—researchers can systematically rule out dozens of competing models until only plausible causal chains remain. The primary exception to this classical framework occurs in quantum mechanics, where entanglement experiments produce statistical correlations that mathematically defy all classical causal models.

## Atmosphere & style The presentation is brisk, energetic, and intellectually playful, characteristic of MinutePhysics' visual, thought-experiment-driven educational approach. It creates an engaging eureka moment by turning a dogmatic statistical rule on its head through a whimsical island scenario, before pivoting into the weird, counterintuitive realities of quantum physics.

## Bulletin - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) The common statistical adage that correlation cannot imply causation is an overcorrection; analyzing networks of multiple correlations can systematically prove or eliminate causal pathways. - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) Using a three-variable system (height, cat ownership, and island habitat), researchers can prune 19 theoretically possible causal models down to just two valid configurations through conditional observations. - [02:25](https://www.youtube.com/watch?v=HUti6vGctQM&t=145s) Classical causal models break down in quantum mechanics, where entanglement experiments produce statistical correlations that mathematically rule out every classical cause-and-effect explanation.

## Key takeaways - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) Assuming an isolated correlation between two variables proves direct causation is a fundamental statistical error. - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) A raw correlation between height and cat ownership does not reveal whether height causes cat ownership, cats cause height, or a third factor causes both. - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) The strict belief that statistics can never demonstrate causality is an incorrect overcorrection to the standard mantra. - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) Causal networks allow researchers to evaluate the presence and absence of correlations across multiple interconnected variables to deduce causality. - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) In a three-variable scenario without context, 19 distinct causal models can theoretically explain an observed correlation (or 20 if pure chance is included). - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) Restricting mobility so that people remain on their birth island eliminates any causal model where height determines island habitat. - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) Observing conditional independence—finding zero correlation between height and cat ownership within each island—eliminates any direct causal link between the two. - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) Pruning hypotheses leaves only two viable models: a common cause model (island affects both) or a causal chain (cats affect the island, which affects height). - [02:25](https://www.youtube.com/watch?v=HUti6vGctQM&t=145s) Systematic evaluation of causal models demonstrates that statistical correlation can indeed imply causation. - [02:25](https://www.youtube.com/watch?v=HUti6vGctQM&t=145s) Quantum entanglement represents an exception to classical causal discovery, producing correlations incompatible with any classical causal network.

## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) Beyond the Mantra: How Causal Networks Infer Causality

A standard pitfall in statistical reasoning is assuming that an observed correlation proves a direct cause. For instance, if data shows that taller people are more likely to own cats, a superficial reading might conclude that being tall causes someone to get a cat. In reality, a raw correlation between height and feline ownership cannot reveal the direction of causation: it could just as well be that owning a cat somehow triggers human growth, or that an unseen third factor dictates both. A whimsical scenario illustrates the point: suppose people and cats are distributed across two distinct islands—one a lush, nutrient-rich paradise providing abundant food for humans to grow tall and care for pets, and the other a barren wasteland that stunts physical growth and suppresses pet ownership. In such cases, the correlation between height and cats is genuine, but neither directly causes the other.

```mermaid flowchart TD accTitle: Narrowing Causal Hypotheses accDescr: Flowchart illustrating how 19 potential causal models between height, cats, and islands are reduced to two valid causal paths.

subgraph Initial["Initial Hypotheses"] A["19 Possible Causal Models<br/>(20 if correlation is accidental)"] end

subgraph Filters["Elimination via Additional Correlations"] B["Observation 1: People stay on birth island<br/>(Height cannot cause Island)"] C["Observation 2: No correlation within each individual island<br/>(Height and Cats do not directly influence each other)"] end

subgraph Outcomes["Surviving Causal Models"] D["Option 1: Common Cause<br/>Island environment causes both Cat Ownership and Height"] E["Option 2: Causal Chain<br/>Cat Ownership causes Island conditions, which cause Height"] end

A --> B B --> C C --> D C --> E ```

However, aggressively repeating the rule that correlation does not mean causation creates an opposing error: the belief that statistics can never establish causality at all. As Henry Reich emphasizes, this overcorrection is mistaken: “But this oft-repeated mantra leads to another common misconception – the idea that you can’t infer any causality from statistics. You can!” While a single isolated correlation cannot untangle cause and effect on its own, causality can be systematically deduced by evaluating multiple correlations across interconnected variables using causal networks.

When analyzing the three variables—height, cat ownership, and island habitat—without any background context, there are 19 distinct possible causal configurations capable of explaining the correlation (or 20 if one allows that the correlation occurred by pure random chance). Introducing just two additional statistical observations drastically prunes these possibilities: - **Observation 1 (Mobility restriction):** People remain on the island where they were born, meaning adult height cannot determine which island a person inhabits; this eliminates every model where height influences island status. - **Observation 2 (Conditional independence):** Examining each island individually reveals no correlation whatsoever between an individual's height and whether they own a cat; this eliminates all configurations where height and cats exert a direct influence on one another.

Once those hypotheses are eliminated, exactly two plausible models survive: either the island environment is a common cause that independently influences both height and cat ownership, or cat ownership influences the state of the island, which in turn alters the height of subsequent generations (for example, if a massive population of cats cultivated the island into an abundant paradise). Using nothing more than correlations across a causal network, 19 theoretical relationships are narrowed down to two.

[02:25](https://www.youtube.com/watch?v=HUti6vGctQM&t=145s) The Quantum Exception and Causal Models

The island puzzle illustrates a broader mathematical reality: for any collection of variables, analyzing the presence or absence of correlations among them allows researchers to rule out incompatible cause-and-effect paths until only viable models remain. In this structural way, statistical correlation genuinely can imply causation. There is, however, a bizarre breakdown of this framework at the microscopic level: certain entanglement experiments in quantum mechanics yield statistical correlations that mathematically rule out *every single* possible classical cause-and-effect relationship. Reich captures the refined takeaway in an updated adage: “Correlation doesn’t necessarily imply causation, but it can if you use it to evaluate causal models. …Except in quantum mechanics.”

Following this statistical walkthrough, Reich highlights the sponsor Brilliant.org, a platform centered on active learning in STEM. He argues that passive consumption is insufficient for mastering analytical subjects, noting that “the only way to truly learn and understand much of math and physics is to actively work through the material yourself – videos only get you so far.” Brilliant structures guided puzzles, quizzes, and problems across skill levels, offering courses in subjects such as Probability, Classical Mechanics, Gravitational Physics, and the Physics of the Everyday. Reich concludes by pointing viewers to a companion footnote video exploring related edge cases that were omitted from the main explanation, including statistical feedback loops and correlations generated purely by coincidence.

## Notable quotes > [02:25](https://www.youtube.com/watch?v=HUti6vGctQM&t=145s) Correlation doesn’t necessarily imply causation, but it can if you use it to evaluate causal models. — Henry Reich

## Mentioned Henry Reich, MinutePhysics, Brilliant.org

## Video description Have fun improving your math & physics skills! Head to https://brilliant.org/minutephysics/

Footnote video: https://www.youtube.com/watch?v=iMbcMMe0D_Y

This video is about how causal models (which use causal networks) allow us to infer causation from correlation, proving the common refrain not entirely accurate: statistics CAN be used to prove causality! Including: Reichenbach's principle, common causes, feedback, entanglement, EPR paradox, and so on.

REFERENCES: Causal Discovery Algorithm in Quantum Mechanics Paper: https://arxiv.org/pdf/1208.4119.pdf Causal Models overview (Quantum and Classical): https://arxiv.org/pdf/1609.09487.pdf

Support MinutePhysics on Patreon! http://www.patreon.com/minutephysics Link to Patreon Supporters: http://www.minutephysics.com/supporters/

MinutePhysics is on twitter - @minutephysics And facebook - http://facebook.com/minutephysics And Google+ (does anyone use this any more?) - http://bit.ly/qzEwc6

Minute Physics provides an energetic and entertaining view of old and new problems in physics -- all in a minute!

Created by Henry Reich

[Source](https://www.youtube.com/watch?v=HUti6vGctQM)

中文简报

MinutePhysics利用因果网络证明统计因果关系 评估相互关联变量间的多个相关性，使研究人员能够系统性排除不相容的因果路径并证明直接因果关系。 尽管两个变量之间孤立的相关性无法揭示因果方向，但“统计学永远无法证明因果关系”这一普遍观点其实是错误的。通过将多个相互关联的变量映射到因果网络中并检验条件相关性——正如涉及身高、养猫以及岛屿环境的思想实验所展示的那样——研究人员可以系统地排除数十种竞争模型，直到只剩下合理的因果链。这种经典框架的主要例外出现在量子力学中，在量子力学里，纠缠实验产生的统计相关性在数学上违背了所有经典因果模型。 “相关不代表因果”这句常见统计学谚语属于矫枉过正；分析多个相关性的网络可以系统地证明或排除因果路径。 (0:01) 利用一个三变量系统（身高、养猫和岛屿栖息地），研究人员可以通过条件观察，将理论上可能的19种因果模型精简至仅剩两种有效构型。 (0:01) 经典因果模型在量子力学中失效，在量子力学中，纠缠实验产生的统计相关性在数学上排除了所有经典的因果解释。 (2:25)

Reading summary

Henry Reich of minutephysics argues that while an isolated correlation cannot establish a direct cause, evaluating networks of correlations across multiple variables allows researchers to systematically eliminate incompatible hypotheses and infer causal relationships. Pruning Causal Hypotheses Through Variable Networks (0:01) A common statistical mistake is assuming that an observed correlation—such as taller people being more likely to own cats—proves direct causation. While an unseen third variable, like living on a lush island versus a barren one, could influence both traits, Reich notes that it is equally mistaken to believe statistics can never determine cause and effect. By mapping variables into causal networks and testing for conditional independence, researchers can systematically rule out competing explanations. In the three-variable scenario of height, cat ownership, and island environment, there are initially 19 potential causal models explaining the correlation (or 20 if accounting for pure coincidence). Observing that people do not leave their birth island rules out any configuration where adult height determines island residency. Furthermore, finding zero correlation between height and cats within individual islands eliminates any direct influence between the two. These statistical observations narrow the 19 initial configurations down to just two viable paths: the island acts as a common cause for both, or cat ownership alters island conditions, which then affects human height. The Mathematical Framework and its Quantum Limit (2:25) The broader principle demonstrated by causal networks is that tracking the presence or absence of correlations across interconnected variables enables researchers to eliminate mathematically incompatible causal structures. Reich summarizes this refined principle in an updated maxim: correlation does not necessarily imply causation on its own, but it can do so when used systematically to evaluate causal models. However, Reich highlights a notable boundary to this framework in quantum mechanics. In quantum entanglement experiments, particles exhibit statistical correlations that mathematically rule out every possible classical cause-and-effect configuration, representing an exception where classical causal modeling breaks down completely. While causal networks allow researchers to deduce cause and effect in classical statistics, edge cases such as statistical feedback loops, accidental correlations, and the unique paradoxes of quantum entanglement present ongoing limits to standard causal inference.

精炼阅读摘要

MinutePhysics创作者Henry Reich指出，常言“相关不等于因果”导致了一种过度修正的误区，即误以为统计学无法推断任何因果关系；实际上，通过在因果网络中评估多个变量之间的关联结构，研究人员能够系统性排除不兼容的假设，进而推导出因果路径。 统计教条的常见误区与因果网络推理 (0:01) 在统计推理中，常见的陷阱是将表面观察到的相关性直接等同于因果关系。例如，数据若显示身材较高的人更有可能养猫，浅层的解读可能会草率断定身高促成了养猫行为。然而，单凭两个变量间的原始相关性无法揭示因果方向：它既可能是养猫促进了人体发育，也可能是未被观察到的第三变量同时决定了这两者。Reich举例设想一种情境：人们与猫分布在两座完全不同的岛屿上，一座是资源富饶、营养充足且适宜生长与饲养宠物的乐土，另一座则是抑制发育且不利于宠物的荒凉废土。在这种设定下，身高与养猫的关联真实存在，但彼此之间并无直接因果作用。 然而，反复强调“相关不代表因果”也催生了另一种极端看法，即认为统计学根本无法建立因果推断。Reich强调这种过度修正是错误的：孤立的单一相关性固然无法理清因果机制，但若将多个相互关联的变量纳入因果网络中进行综合评估，因果关系完全可以通过统计学手段被系统性地推导出来。 利用额外关联筛选假说的岛屿模型 (0:01) 以身高、养猫所有权以及岛屿栖息地三个变量为例，在缺乏先验背景的情况下，能够解释现有相关性的潜在因果模型共有19种；若算入因纯粹随机巧合导致相关的可能性，则多达20种。通过引入两项额外的统计观察结果，研究人员可以大幅削减这些候选假设： 其一，人员流动限制观察表明，所有人自出生起就停留在原生岛屿上，这意味着成年后的身高不可能决定个体所栖息的岛屿，由此即可彻底排除所有“身高影响岛屿状态”的模型。其二，条件独立性观察显示，在单座岛屿内部孤立考察时，个人的身高与是否养猫之间毫无关联，这进而排除了身高与养猫存在直接相互影响的全部假设。历经这两轮排除后，原本的19种假设仅剩下两种成立可能：要么是岛屿环境充当了共同原因，分别独立决定了身高与养猫行为；要么是养猫影响了岛屿生态，进而改变了后代的身高状态。仅凭因果网络中的统计关联，潜在假说便被有效收敛至两种可能。 模型普适性边界与量子力学特例 (2:25) 这种因果推断方法展示了更广泛的数学逻辑：对于任意变量集合，分析其间相关性的存在与否，允许研究者逐步剔除不相容的因果链条，直至剩下唯一或极少数可行的因果模型。因此，在结构性网络的支撑下，统计相关性确实能够推导并暗示因果关系。基于此，Reich提出了修正后的格言：“相关性未必意味着因果关系，但若用于评估因果模型，它确实可以推导因果——量子力学除外。” 然而，这一统计框架在微观物理尺度上遭遇了颠覆。量子力学中的特定纠缠实验展现出的统计关联，从数学上排除了任何符合经典物理规则的因果模型。此外，针对现实应用中更复杂的边界情况，统计反馈循环以及纯粹由巧合构成的虚假关联，仍是因果网络分析中需要审慎处理的特殊课题。 Reich的核心论点在于，单一统计关联虽不足以证明因果，但多变量因果网络能够通过系统性排除不相容假说来建立宏观世界的因果推断；然而，量子纠缠现象打破了所有经典因果模型，构成了经典统计因果推断的理论极限。

Detailed notes — zh

## Background 该视频由科普频道 minutephysics（由 Henry Reich 创作）于 2017 年 8 月 17 日发布在 YouTube 平台。视频简介指出，本期内容旨在探讨如何利用因果模型（因果网络）从相关性中推导因果关系，打破“相关不等于因果”的刻板印象，证明统计学确实可以用来证明因果关系，并触及了赖兴巴赫共同原因原理、量子纠缠与 EPR 佯谬等物理与统计学课题。 ## TL;DR 视频首先指出了“相关不等于因果”这一经典警示的合理性，但随后纠正了人们由此产生的误区——即认为统计数据完全无法证明因果。作者通过身高、养猫与岛屿环境的趣味模型展示，仅凭单对相关性固然无法定论，但只要结合多重相关性检验与因果网络，便能系统性地排除不合理的假设，将 19 种潜在因果模型大幅压缩至仅剩 2 种。最后，视频指出量子力学构成了经典统计因果推断的例外，粒子间的诡异纠缠能够同时排除所有经典因果结构，从而提出了一个更严谨的新格言：相关性虽不必然意味着因果，但借助因果模型即可实现推断，除非是在量子力学之中。 ## Atmosphere & style 整部短片保持了 minutephysics 标志性的快节奏与手绘涂鸦风格，叙事明快利落、幽默诙谐。Henry Reich 的讲解语调生动风趣，将枯燥的统计学概念和反直觉的因果图谱拆解得浅显易懂，让观众在轻快烧脑的推演中迅速体会到逻辑反转的乐趣。 ## Bulletin - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 人们常说的“相关性不意味着因果性”导致了一个普遍误解，即认为统计学无法证明因果关系；但实际上，借助因果模型，统计学完全可以用来推导因果。 - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 以身高、养猫和岛屿环境三变量为例，在没有任何限制时数学上存在 19 种因果模型（若计入巧合则为 20 种），但仅通过两项变量相关性约束即可将可能性筛减至 2 种。 - [02:39](https://www.youtube.com/watch?v=HUti6vGctQM&t=159s) 因果网络方法在经典物理与统计中行之有效，但在量子力学领域失效，因为量子纠缠所展现出的相关性会同时排除所有局域经典因果解释。 ## Key takeaways - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) 统计学中最常见的误区是直接认为“相关性意味着因果性”，单凭两件事物相关无法确定其因果流向或背后是否存在共同原因。 - [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) 在缺乏额外信息时，高个子养猫比例更高的现象存在多种解释，例如养猫让人变高、长得高让人想养猫，或者第三个环境变量（如生活在物产丰饶的岛屿）同时促成了这两者。 - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) “相关性不意味着因果性”这句警示本身是正确的，但它让许多人矫枉过正，误以为完全无法从统计数据中推导因果关系。 - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 当两件事物存在相关性时，背后通常存在特定原因；即使单一相关性无法直接揭示原因，多组相关性与因果网络（Causal Networks）的结合却能推导出因果。 - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 在三变量案例中，在初始状态下存在 19 种不同因果结构可以解释相关性（若算上“纯属巧合”则共有 20 种）。 - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 通过引入居民不离岛的事实，可以排除所有“身高影响岛屿环境”的因果模型。 - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 通过发现“在单座岛屿内部身高与养猫无关”（条件独立性），可以进一步排除所有“身高与养猫直接相互影响”的模型，最终将 19 种假设大幅压缩至仅剩 2 种候选模型。 - [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 通过系统性地检验变量间的相关性与不相关性（条件独立），统计学能够消去不符合事实的因果假设，进而直接推导并证明现实中的因果机制。 - [02:39](https://www.youtube.com/watch?v=HUti6vGctQM&t=159s) 经典统计与因果网络在量子力学中遭遇颠覆，微观粒子的强相关性会排除一切经典因果解释，构成了该方法的例外。 - [02:39](https://www.youtube.com/watch?v=HUti6vGctQM&t=159s) 面对量子实验的冲击，更严谨的格言应修正为：“相关性不一定意味着因果性，但若用于评估因果模型则可以推导因果——除非是在量子力学中”。 ## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=HUti6vGctQM&t=1s) 相关性与因果性的常见误区 统计学中最常见的误区之一，就是理所当然地认为“相关性意味着因果性”（correlation implies causation）。例如，如果数据显示个子高的人养猫的比例更高，有人可能会草率地推断：长得高会让一个人更有可能去养猫。然而，仅凭身高与养猫之间存在相关性，根本无法判断因果关系的流向——事实完全有可能是反过来的，比如养猫导致人长得更高；或者，真正的起因完全是另一回事，比如两组人和猫分别生活在两个互不相通的岛屿上，其中一座岛是物产丰饶的天堂，有充足的食物让人长高并供养宠物猫，而另一座岛则是一片荒芜的废土，既限制了人的身高，也限制了养猫的可能。 这类例子的核心目的在于说明：仅仅注意到两件事物之间存在关联，并不意味着其中一者就是另一者的起因。正如 Henry Reich 所言：“Hence the common refrain: correlation doesn’t imply causation. And it’s true – it doesn’t!” 这句常挂在人们嘴边的警示确实没有错，相关性本身确实不等于因果性。 ### [00:44](https://www.youtube.com/watch?v=HUti6vGctQM&t=44s) 如何通过因果网络从统计中推导因果 然而，这句被反复念诵的格言也催生了另一个广泛存在的误区——人们误以为完全无法从统计数据中推导因果关系。事实并非如此，统计学确实可以用来证明因果关系。当两件事物存在相关性时，背后往往存在某种原因，即便单单一对相关性无法直接揭晓答案。有时我们可以借助额外的信息来推断因果（比如知道某一事件在时间上先于另一事件发生），但即便仅靠相关性本身，只要我们掌握不止一组相关性，并结合所谓的“因果网络”（causal networks / causal models），同样能够直接推断因果。 在猫、身高与岛屿的例子中，如果除“养猫与身高相关”之外一无所知，在数学上总共有多达 19 种不同的因果关系结构可以解释该现象（如果算上“纯属巧合”的情况，则共有 20 种）。然而，假设我们掌握了另外两项关键事实： 1. **出生地固定**：假定出生在某座岛上的人一辈子都留在那里，因此身高绝不可能反过来决定一个人住在哪个岛上，这便直接排除了所有“身高影响岛屿”的因果猜想。 2. **条件独立性**：假定在任意一座岛屿内部单独观察时，身高与养猫之间其实没有任何相关性。这意味着我们可以排除所有“身高与养猫直接相互影响”的选项。 排除了绝大部分可能性后，原本复杂的 19 种因果关系被一举收窄到了仅剩的 2 种可能：要么岛屿环境是导致身高和养猫的共同起因（即丰饶岛屿对人与猫皆有益）；要么是养猫影响了岛屿，进而由岛屿影响了身高（例如猫群的繁盛把岛屿改造成了天堂，从而促进了后代猫主人的身高增长）。仅仅通过组合使用多组变量之间的相关与不相关，就能将因果假设从 19 个压缩到 2 个。正如 Henry Reich 总结道：“So, starting with 19 possible causal relationships, we used correlations to narrow things down to just 2 options – not bad!” ```mermaid flowchart TD accTitle: 变量关系消除法 accDescr: 展示如何利用已知条件将19种因果网络排除至仅剩2种可能 Start["初始状态<br/>已知'身高与养猫相关'<br/>（共 19 种可能模型）"] --> Step1["规则 1：居民不离岛<br/>排除'身高 -> 岛屿'模型"] Step1 --> Step2["规则 2：岛内无直接相关<br/>排除'身高 <-> 养猫'直接影响模型"] Step2 --> Result1["候选模型 A（共同原因）：<br/>岛屿 -> 身高 且 岛屿 -> 养猫"] Step2 --> Result2["候选模型 B（因果链条）：<br/>养猫 -> 岛屿 -> 身高"] ``` - **初始模型空间**：在未知额外条件前，三个变量之间可能构建出 19 种因果关系模型（若计入完全无因果的巧合则为 20 种）。 - **因果筛选逻辑**：利用变量之间的多种相关性或“缺乏相关性”（即条件独立），能够系统性地消去不符合事实的因果链条。 - **核心结论**：针对现实中的任何一组事物，只要借助多重相关性检验因果网络，相关性确实能够推导并证明因果关系。 ### [02:39](https://www.youtube.com/watch?v=HUti6vGctQM&t=159s) 量子力学带来的颠覆与新格言 虽然基于因果模型的统计推断在经典世界中无往不利，但在微观领域却遭遇了颠覆性挑战。在量子力学的一系列实验中，粒子之间展现出的特定相关性极其诡异，竟然会同时排除掉所有可能的因果关系解释。面对经典统计与量子物理的这种冲突，Henry Reich 幽默地提议为这句传世格言换上一个更严谨、更现代的新版本：“‘Correlation doesn’t necessarily imply causation, but it can if you use it to evaluate causal models. …Except in quantum mechanics.’” - **量子力学的例外**：量子纠缠与贝尔实验（以及 EPR 佯谬相关研究）所呈现的强相关性，排除了任何局域经典因果网络的描述。 - **赞助与补充延伸**：视频后半部分介绍了本期赞助商 Brilliant.org，强调被动观看视频的作用有限，唯有通过动手解决趣味谜题与练习才能真正掌握物理与数学技能。最后，作者提到还有一部单独的“脚注视频”（footnotes video），专门补充因时长原因剪掉的进阶内容，例如因果反馈回路（feedback loops）以及纯粹由于随机巧合而产生的假性相关。 ## Notable quotes > [02:39](https://www.youtube.com/watch?v=HUti6vGctQM&t=159s) Correlation doesn’t necessarily imply causation, but it can if you use it to evaluate causal models. …Except in quantum mechanics. — Henry Reich ## Mentioned Henry Reich, MinutePhysics, Brilliant.org, EPR 佯谬 (EPR Paradox), 赖兴巴赫原则 (Reichenbach's Principle), Patreon ## Video description Have fun improving your math & physics skills! Head to https://brilliant.org/minutephysics/ Footnote video: https://www.youtube.com/watch?v=iMbcMMe0D_Y This video is about how causal models (which use causal networks) allow us to infer causation from correlation, proving the common refrain not entirely accurate: statistics CAN be used to prove causality! Including: Reichenbach's principle, common causes, feedback, entanglement, EPR paradox, and so on. REFERENCES: Causal Discovery Algorithm in Quantum Mechanics Paper: https://arxiv.org/pdf/1208.4119.pdf Causal Models overview (Quantum and Classical): https://arxiv.org/pdf/1609.09487.pdf Support MinutePhysics on Patreon! http://www.patreon.com/minutephysics Link to Patreon Supporters: http://www.minutephysics.com/supporters/ MinutePhysics is on twitter - @minutephysics And facebook - http://facebook.com/minutephysics And Google+ (does anyone use this any more?) - http://bit.ly/qzEwc6 Minute Physics provides an energetic and entertaining view of old and new problems in physics -- all in a minute! Created by Henry Reich [Source](https://www.youtube.com/watch?v=HUti6vGctQM)

Source quotes

[2:39] Henry Reich: Correlation doesn’t necessarily imply causation, but it can if you use it to evaluate causal models. …Except in quantum mechanics.

Source quotes

[2:25] Henry Reich: Correlation doesn’t necessarily imply causation, but it can if you use it to evaluate causal models.

Original transcript

[0:01] A common misconception in statistics is to think that correlation implies causation – like, [0:06] if more tall people have cats, you might think that means being tall makes people more likely [0:09] to get a cat. [0:11] However, simply knowing a correlation between height and cat ownership can’t tell us which [0:14] way the causality goes – it may instead be that having a cat causes people to grow [0:18] taller – or perhaps the real cause is something else altogether, like that the people and [0:23] cats live on two separate islands, one a lush paradise with enough food for growing tall [0:27] and feeding pet cats, and the other a wasteland that limits both height and cat ownership. [0:32] The point of examples like this is that noticing a correlation between two things doesn’t [0:36] imply that one of those things causes the other. [0:38] Hence the common refrain: correlation doesn’t imply causation. [0:42] And it’s true – it doesn’t! [0:44] But this oft-repeated mantra leads to another common misconception – the idea that you [0:48] can’t infer any causality from statistics. [0:51] You can! [0:52] I mean, it’s quite reasonable to think that, if two things are correlated, there’s likely [0:55] some reason, , even if a single correlation can’t tell you. [0:58] Sometimes you can infer the causality from additional information – like knowing that [1:02] one thing happened before the other – but you can also infer causality directly from [1:06] correlations – you just need more than one, together with something called causal [1:09] networks. [1:10] Like, in our cat-height-island example, we know that cat ownership and height are correlated, [1:15] but we don’t know what the cause of that correlation is. [1:17] If we don’t know anything else, then there are 19 – yes 19! [1:21] – different causal relationships that could explain the situation. [1:24] 20 if you think the correlation is just an accident. [1:29] However, perhaps we know two other things: first, suppose people born on a particular [1:34] island stay there, so their height doesn’t influence what island they live on, and we [1:38] can rule out the relationships where height influences island. [1:41] Second, suppose that on either island, taken by itself, there isn’t any correlation between [1:45] height and cat ownership; then we can rule out all the options where height and cats [1:49] influence each other directly . This leaves us with just two options: either the islands [1:54] are the causal explanation for both height and cat ownership (maybe, as before, one island [1:58] is a lush, healthy paradise for both people and cats), or else cat ownership is the causal [2:03] explanation for the islands which are the causal explanation for height, (like, maybe [2:07] an abundance of cats turned the island into a paradise, thereby influencing the height [2:11] of future cat owners). [2:12] So, starting with 19 possible causal relationships, we used correlations to narrow things down [2:17] to just 2 options – not bad! [2:25] Of course, this is just a simple example, but for any group of things, you can use the [2:29] various correlations between them (or lack of correlations) to eliminate some of the [2:33] possible cause-and-effect relationships. [2:36] And that’s how correlations CAN imply causation. [2:39] There is one problem, though… [2:41] some experiments in quantum mechanics have correlations that rule out ALL possible cause [2:46] and effect relationships. [2:47] We’ll have to save the details for a later video, but until then, may I suggest a new [2:50] version of the famous refrain? [2:52] “Correlation doesn’t necessarily imply causation, but it can if you use it to evaluate [2:55] causal models. [2:56] …Except in quantum mechanics.” [2:59] I’ve got a little more about statistics and causality after this, but first I’m [3:03] excited to introduce the very relevant sponsor for this video: Brilliant.org. [3:08] Brilliant is a problem solving website designed to help you practice and learn math and science [3:11] via guided problems, puzzles and quizzes: I know that sounds kind of nerdy, but the [3:16] truth is that the only way to truly learn and understand much of math and physics is [3:20] to actively work through the material yourself – videos only get you so far. [3:24] And Brilliant does a brilliant job of making that easy, sneakily enticing you into doing [3:29] math and physics problems by means of intriguing questions structured for all ability and knowledge [3:34] levels. [3:35] I say this from experience, because if you haven’t done a problem for a few days, Brilliant [3:38] will send you an attention-grabbing puzzle , and I’ve been sucked in by quite a few [3:42] of them. [3:43] If you want to try out Brilliant (which I recommend), heading to brilliant.org/minutephysics [3:47] will let them know you came from here, and you can check out their courses on Probability, [3:52] the Physics of the Everyday, Classical Mechanics, Gravitational Physics and so on. [3:56] Hey, glad you’re still here – in case you’re interested, there’s a footnotes [4:00] video covering a few things that got cut out of this one, like feedback loops and correlations [4:04] that arise just by chance. [4:06] The link’s on screen and in the video description.