{
  "schema": "auturnal.deep-research-review.v1",
  "slug": "causal-inference-ai",
  "title": "从相关到干预：AI 因果推断中的识别、混杂与可迁移性",
  "category": "可信AI与方法论",
  "dek": "用一个可穷举结构模型展示观察差异为何不等于干预效果。",
  "english": "Causal inference requires identification assumptions in addition to predictive accuracy. We distinguish observational conditioning from intervention and work through an explicit binary-confounder model. The calculated association and causal effect differ despite a fixed structural equation. No real treatment effect or policy recommendation is estimated; the example demonstrates assumptions and non-identification risks.",
  "kind": "ai_assisted_research_guide",
  "author": "Codex（AI）",
  "original_language": "zh-Hans",
  "published_date": "2026-10-22",
  "edition": 5,
  "version": 1,
  "previous_version": null,
  "refs": [
    "arxiv-2102.11107",
    "arxiv-1907.02893",
    "arxiv-2012.07421"
  ],
  "sections": [
    [
      "摘要",
      "高预测精度并不能自动回答“改变某个因素会怎样”。本文区分观察条件分布与干预分布，在显式结构方程中推导混杂如何改变关联，并讨论后门调整、表示识别与分布变化的条件。配图来自可穷举的二元混杂模型，变量和概率完全公开，不是实际人群或政策效果研究。",
      [
        "arxiv-2102.11107",
        "arxiv-1907.02893",
        "arxiv-2012.07421"
      ]
    ],
    [
      "1. 预测问题与干预问题使用不同分布",
      "观察P(Y|X=x)描述已发生数据中X取某值的群体；干预P(Y|do(X=x))描述外部设置X后系统的结果。二者可能因共同原因而不同。因果图和结构方程把假设写出来，但图本身不是从高分预测器中自动获得的真理。需要说明变量含义、时间顺序、遗漏因素和可接受干预，尤其不能把统计控制任意变量当作总能消除偏差。",
      [
        "arxiv-2102.11107"
      ]
    ],
    [
      "2. 一个二元混杂模型可精确揭示差异",
      "令U等概率取0或1，P(X=1|U=1)=0.8、P(X=1|U=0)=0.2，结果Y=2X+5U。观察X=1者中U=1的比例为0.8，而X=0者中为0.2，因此观察均值差为2+5×0.6=5。干预X时保持U分布不变，均值差只有2。这里没有抽样误差，差异来自生成机制；现实数据则还需要识别这些条件是否可信。",
      [
        "arxiv-2102.11107"
      ]
    ],
    [
      "3. 后门调整依赖可检验与不可检验条件",
      "若U满足适当的调整条件，干预均值可由按总体P(U)加权的条件均值计算。还需要一致性、相关层内的正值支持以及足够的测量质量。若某类人从未接受某种处理，模型外推不等于数据支持。控制碰撞点或处理后的中介还可能引入新偏差或改变目标效应。报告应先定义估计对象，再说明图与调整集，不能按相关性强弱机械选择变量。",
      [
        "arxiv-2102.11107"
      ]
    ],
    [
      "4. 表示学习的识别有额外结构要求",
      "将高维观察压缩为潜变量，可能有助于因果建模，但仅靠重建良好或独立性正则通常不足以唯一确定因果因素。非线性ICA与因果表示研究讨论了辅助变量、结构限制等识别条件。一个可解释命名的潜维度，不自动成为可干预的真实因素。应明确哪些变量可观测、哪些不可识别，避免把表示旋转后的名称当作因果发现。",
      [
        "arxiv-1907.02893",
        "arxiv-2102.11107"
      ]
    ],
    [
      "5. 分布不变性是一种假设而非万能答案",
      "不同环境中的稳定预测关系可能提供线索，但环境数量、变化机制和隐藏混杂会影响结论。分布变化研究显示，简单的数据增强、预训练和模型选择也会影响跨域表现。因而跨环境准确率高不能独自证明学到了因果结构；应设计能区分竞争解释的干预或反事实检验，并报告哪些变化仍超出训练支持。",
      [
        "arxiv-2012.07421",
        "arxiv-2102.11107"
      ]
    ],
    [
      "6. 可信结论需要敏感性与外部证据",
      "现实研究应记录目标总体、处理、结果、时间窗口、缺失机制和调整依据，使用负对照、敏感性分析或合法随机化证据检验假设。无法识别时可以报告边界或关联，不应强行给出单一因果数值。本文模型只是教学反例：它证明相关与干预可以不同，不证明某个真实应用存在相同比例偏差，也不构成医疗或政策建议。",
      [
        "arxiv-2102.11107",
        "arxiv-2012.07421"
      ]
    ]
  ],
  "equations": [
    {
      "tex": "\\mathbb E[Y\\mid do(X=x)]=\\sum_u\\mathbb E[Y\\mid X=x,U=u]P(U=u)",
      "explanation": "仅在U满足识别与支持等条件时成立；不是对任意控制变量都可用的恒等式。"
    },
    {
      "tex": "Y=2X+5U,\\quad P(U=1)=1/2",
      "explanation": "完全构造的结构模型；X由U影响，X被干预时U分布保持不变。"
    },
    {
      "tex": "\\Delta_{\\mathrm{obs}}=5,\\qquad\\Delta_{\\mathrm{do}}=2",
      "explanation": "由公开概率表精确计算，没有使用真实人群数据或估计真实处理效应。"
    }
  ],
  "table": {
    "headers": [
      "方法或判断",
      "控制变量与测量",
      "失效条件及解释"
    ],
    "rows": [
      [
        "观察预测",
        "报告P(Y|X)与数据来源",
        "预测准确不识别干预效果"
      ],
      [
        "后门调整",
        "给出图、调整集与支持条件",
        "错误控制可增加偏差"
      ],
      [
        "潜在表示",
        "列出识别假设和等价解",
        "可解释命名不等于因果变量"
      ],
      [
        "跨域稳定",
        "设计环境与竞争解释对照",
        "稳定准确率不单独证明因果"
      ]
    ]
  },
  "protocol": [
    "定义目标总体、处理、结果、时间及要估计的效应。",
    "画出因果假设和变量时间顺序，登记隐藏混杂风险。",
    "检查调整集、一致性、正值支持、测量误差和缺失机制。",
    "用可穷举模拟先核对估计器，再在合法数据上分析。",
    "执行敏感性、负对照或外部干预检验，保留竞争解释。",
    "不可识别时明确报告关联或边界，不把预测结果改写成因果结论。"
  ],
  "diagram": [
    "目标干预与变量定义",
    "结构图 / 混杂假设",
    "识别及支持条件",
    "观察分布与调整",
    "敏感性及外部检验",
    "限定总体的因果结论"
  ],
  "analytic": "causal-confounding",
  "limitations": [
    "模型与概率完全构造。",
    "没有估计任何真实干预效果。",
    "识别假设通常不能只由数据证明。",
    "跨域预测不自动证明因果结构。"
  ],
  "keywords": [
    "causal-inference",
    "confounding",
    "identification"
  ],
  "peer_reviewed": false,
  "journal_publication": false,
  "disclosure": {
    "authorship": "AI辅助研究综述；Codex（AI）依据所列一手文献撰写并进行解析推导。",
    "figures": "概念封面与科学图解分开；科学图由公开公式和明示假设计算，原始论文结果均注明归属。",
    "review": "本站没有本篇独立同行评审或原创模型实验记录。"
  },
  "section_kinds": [
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review"
  ],
  "sources": [
    {
      "id": "arxiv-2102.11107",
      "title": "Towards Causal Representation Learning",
      "authors": "Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, Yoshua Bengio",
      "url": "https://arxiv.org/abs/2102.11107",
      "date": "2021-02-22",
      "accessed_at": "2026-10-21",
      "locator": "Abstract and sections on interventions, independent mechanisms and causal representations",
      "note": "原始论文PDF及作者书目信息已核对；因果识别条件不等于相关性预测。"
    },
    {
      "id": "arxiv-1907.02893",
      "title": "Invariant Risk Minimization",
      "authors": ",  ,  ,",
      "url": "https://arxiv.org/abs/1907.02893",
      "date": "2019-07-05",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; 原文方法、理论和实验章节见作者原始论文链接",
      "note": "一手论文；原文研究归属原作者。本文的推导示例与建议实验方案另行标注，不冒充原论文或本站实测。"
    },
    {
      "id": "arxiv-2012.07421",
      "title": "WILDS: A Benchmark of in-the-Wild Distribution Shifts",
      "authors": ",  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  ,  --------------------------------------------------------------------------------",
      "url": "https://arxiv.org/abs/2012.07421",
      "date": "2020-12-14",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; 原文方法、理论和实验章节见作者原始论文链接",
      "note": "一手论文；原文研究归属原作者。本文的推导示例与建议实验方案另行标注，不冒充原论文或本站实测。"
    }
  ],
  "figures": {
    "mechanism": {
      "url": "/guides/assets/v5-mechanism-causal-inference-ai-d38199896b4b5277.png",
      "kind": "conceptual_method_flow",
      "data": [
        "目标干预与变量定义",
        "结构图 / 混杂假设",
        "识别及支持条件",
        "观察分布与调整",
        "敏感性及外部检验",
        "限定总体的因果结论"
      ]
    },
    "analysis": {
      "url": "/guides/assets/v5-analysis-causal-inference-ai-2ea482e19dd71692.png",
      "kind": "analytical_or_constructed_example",
      "data_url": "/guides/assets/v5-data-causal-inference-ai-cdba9205932f8e75.json",
      "assumptions": "P(U=1)=0.5; P(X=1|U=1)=0.8, |U=0=0.2; Y=2X+5U. Exact enumeration, no population data."
    }
  },
  "cover_provenance_url": "/guides/assets/v5-provenance-causal-inference-ai-5771bbb3a6159338.json",
  "source_revision": "ed6feca8fdc75c8de482fd69dedf17cbc746a9d1"
}