{"id":30349,"date":"2021-11-25T22:07:46","date_gmt":"2021-11-25T14:07:46","guid":{"rendered":"https:\/\/swarma.org\/?p=30349"},"modified":"2021-11-25T22:07:46","modified_gmt":"2021-11-25T14:07:46","slug":"%e5%9b%a0%e6%9e%9c%e6%8e%a8%e6%96%ad%ef%bc%9a%e5%8e%9f%e7%90%86%e4%b8%8e%e6%96%b9%e6%b3%95","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=30349","title":{"rendered":"\u56e0\u679c\u63a8\u65ad\uff1a\u539f\u7406\u4e0e\u65b9\u6cd5"},"content":{"rendered":"
\n
\n
\n
\n

<\/p>\n<\/section>\n

\n
\n
\n
\n
\n
\n

<\/section>\n<\/section>\n
\n
\n

\u5bfc\u8bed<\/strong><\/p>\n<\/section>\n<\/section>\n

\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n

<\/strong>\u7edf\u8ba1\u56e0\u679c\u63a8\u65ad\u8fd1\u671f\u5728\u5de5\u4e1a\u754c\u5e94\u7528\u8d8a\u6765\u8d8a\u591a\uff0c\u5f88\u591a\u65b0\u4eba\u5728\u8fd9\u4e2a\u5feb\u9910\u5f0f\u3001\u788e\u7247\u5316\u4fe1\u606f\u7e41\u591a\u4e14\u6742\u7684\u65f6\u4ee3\uff0c\u5b66\u4e60\u65b0\u77e5\u8bc6\u6e20\u9053\u591a\uff0c\u4f46\u96be\u5ea6\u66f4\u5927\u3002<\/span>\u60c5\u62a5\u641c\u96c6\u3001\u6574\u7406\u5e76\u5185\u5316\u662f\u5f53\u4eca\u5fc5\u8981\u7684\u65b9\u6cd5\u3002<\/span><\/strong><\/p>\n


<\/p>\n

\u521a\u5f00\u59cb\u63a5\u89e6\u56e0\u679c\u63a8\u65ad\uff08\u5176\u4ed6\u6bd4\u5982\u91cf\u5b50\u673a\u5668\u5b66\u4e60\uff09\uff0c\u542c\u5230\u5404\u79cd\u540d\u8bcd\u3001\u65b9\u6cd5\uff0c\u67e5\u9605\u5305\u62ec\u8bf7\u6559\u8eab\u8fb9\u7684\u4eba\uff0c\u4e00\u6bb5\u65f6\u95f4\u53d1\u73b0\u5f88\u591a\u6982\u5ff5\u8fd8\u662f\u4e0d\u591f\u7406\u89e3\uff0c\u7cfb\u7edf\u6027\u5f88\u5f31\uff0c\u592a\u8fc7\u4e8e\u788e\u7247\u5316\uff0c\u7f3a\u5c11\u65b9\u6cd5\u8bba\u7684\u8054\u7cfb\uff0c\u751a\u81f3\u4e0d\u77e5\u9053\u4e00\u4e9b\u65b9\u6cd5\u7684\u4f18\u5316\u70b9\u3002\u8fd9\u6837\u5bf9\u81ea\u5df1\u7684\u7406\u89e3\u975e\u5e38\u4e0d\u5229\uff0c\u751a\u81f3\u963b\u788d\u81ea\u5df1\u601d\u8003\u548c\u521b\u65b0\u3002<\/strong><\/p>\n


<\/strong><\/p>\n

\u6545\u6b64\u7bc7\uff0c\u9488\u5bf9\u65b0\u624b\uff0c\u4ece\u6574\u4f53\u6846\u67b6\u7684\u89d2\u5ea6\u4ecb\u7ecd\u56e0\u679c\u63a8\u65ad\u7684\u539f\u7406\u548c\u4e00\u4e9b\u5e38\u89c1\u65b9\u6cd5\u8bba\uff0c\u5e76\u7b80\u5355\u56de\u987e\u5de5\u4e1a\u754c\u4e00\u4e9b\u65b9\u6cd5\u6240\u5904\u4f4d\u7f6e\u3002\u672c\u6587\u76ee\u6807\u7528\u6237\u4e3a\u5bf9\u6982\u7387\u8bba\u3001\u6982\u7387\u56fe\u6a21\u578b\u3001\u56e0\u679c\u63a8\u65ad\u611f\u5174\u8da3\u7684\u65b0\u8001\u624b\u670b\u53cb \":)\"<\/strong><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/strong><\/section>\n<\/section>\n
\n

\u9b4f\u731b<\/span> | \u4f5c\u8005<\/strong><\/span><\/p>\n

\u77e5\u4e4e | <\/strong><\/span>\u6765\u6e90<\/strong><\/span><\/p>\n<\/section>\n


<\/span><\/p>\n


<\/p>\n

\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u672c\u6587\u5927\u7eb2<\/strong> (\u8fd9\u91cc\u5047\u8bbe\u4e00\u4e9b\u57fa\u672c\u77e5\u8bc6\u548c\u7b26\u53f7\u5927\u5bb6\u90fd\u4e86\u4e86\u89e3)<\/span><\/p>\n

    \n
  • \n
    Causal Inference\uff1a\u76ee\u6807\u3001\u5047\u8bbe\u3001\u6846\u67b6 \uff08\u590d\u4e60\u56de\u987e\u5e76\u603b\u7ed3\u6846\u67b6\uff09<\/span><\/section>\n<\/li>\n
  • \n
    Causal Model\uff1a\u57fa\u7840\u6982\u7387\u56fe\u3001\u56e0\u679c\u56fe\u77e5\u8bc6<\/span><\/section>\n<\/li>\n
  • \n
    Identification\uff1abackdoor\u3001frontdoor criterion\u3001 do-calculus<\/span><\/section>\n<\/li>\n
  • \n
    Evaluation<\/span><\/section>\n<\/li>\n
      \n
    • \n
      Conditional Outcome Modeling (COM\u3001S-Learner)<\/span><\/section>\n<\/li>\n
    • \n
      Grouped COM (GCOM\u3001T-learner)<\/span><\/section>\n<\/li>\n
    • \n
      X-learner<\/span><\/section>\n<\/li>\n
    • \n
      Propensity score and IPW<\/span><\/section>\n<\/li>\n
    • \n
      others<\/span><\/section>\n<\/li>\n
        \n
      • \n
        Doubly Robust Methods<\/span><\/section>\n<\/li>\n
      • \n
        Matching<\/span><\/section>\n<\/li>\n
      • \n
        Double Machine Learning (DML)<\/span><\/section>\n<\/li>\n
      • \n
        Causal Trees & Forests<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n
      • \n
        \u5de5\u4e1a\u754c\u56e0\u679c\u63a8\u65ad\u5e94\u7528\uff1auplift\u3001debias\u3001\u53ef\u89e3\u91ca\u6027<\/span><\/section>\n<\/li>\n
      • \n
        \u603b\u7ed3<\/span><\/section>\n<\/li>\n<\/ul>\n

        <\/section>\n


        <\/strong><\/span><\/h2>\n

        \n
        \n
        \n
        \n
        \n

        <\/section>\n

        <\/section>\n
        \n
        <\/svg><\/section>\n<\/section>\n<\/section>\n
        \n

        1. <\/strong><\/strong>Causal Inference\uff1a<\/strong>\u76ee\u6807\u3001\u5047\u8bbe\u3001\u6846\u67b6 <\/strong><\/p>\n<\/section>\n

        \n

        <\/section>\n

        <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


        <\/p>\n

        \u7814\u7a76\u76ee\u6807\u53d8\u91cfT\u5bf9\u7ed3\u679c\u53d8\u91cfY\u7684\u4f5c\u7528 (effect of T on Y)<\/span><\/section>\n

        <\/span><\/section>\n
        \u4e4b\u524d\u4ecb\u7ecd\u8fc7\u57fa\u4e8ePotential Outcomes\u57fa\u4e8e{Y(1), Y(0)}\u8868\u8fbe\u8ba1\u7b97causal effect\uff0c\u5e76\u8bf4\u660e\u4e86\u7531\u4e8eFundamental Problem<\/span> (Counterfactual)<\/span> \u5bfc\u81f4\u4e2a\u4f53\u6548\u5e94ITE\u65e0\u6cd5\u8bc6\u522b\uff0c\u6240\u4ee5\u8ba1\u7b97\u603b\u4f53\u5e73\u5747\u6c34\u5e73ATE\u3002\u5e76\u8bf4\u660e\u4e86\u4e3a\u4ec0\u4e48\u76f8\u5173\u6027<\/span>\u56e0\u679c\u6027\uff0c\u4e14\u4ecb\u7ecd\u5f53\u6ee1\u8db3\u54ea\u4e9b\u6761\u4ef6\u65f6\uff0c\u53ef\u4ee5\u76f8\u5173\u6027\u21d2\u56e0\u679c\u6027\uff0c\u8fdb\u800c\u53ef\u4ee5\u57fa\u4e8e\u6570\u636e\u8ba1\u7b97\u5f97\u5230ATE\u3002\u6211\u4eec\u5728\u6b64\u5148\u505a\u4e00\u4e2a\u603b\u7ed3\uff01<\/span><\/section>\n

        <\/span><\/section>\n
        <\/section>\n

        <\/section>\n
        \u5f62\u8c61\u5316\u4f8b\u5b50\u5982\u4e0b<\/span>
        <\/section>\n

        <\/p>\n

        graphical example<\/span><\/p>\n


        <\/span><\/section>\n
        \u7b80\u5355\u63cf\u8ff0\u4e00\u4e0b\uff0c\u6211\u4eec\u76ee\u6807\u662f\u8ba1\u7b97average treatment effect <\/span>(ATE)<\/span>\uff0c\u7814\u7a76T\u5bf9Y\u7684\u56e0\u679c\u6548\u5e94\u3002Y(t)\u662f potential outcome\u7684\u7b80\u5199\uff0c\u8fd9\u662f\u4e00\u4e2acausal quantity\uff0c\u65e0\u6cd5\u76f4\u63a5\u8ba1\u7b97<\/span>\uff08\u6211\u4eec\u83b7\u53d6\u7684\u6570\u636e\u90fd\u662f\u8ba1\u7b97statistical quantities\u7684\uff09<\/span>\u3002\u4e4b\u524d\u4e5f\u4ecb\u7ecd\u8fc7\u968f\u673a\u8bd5\u9a8cRCTs\u5bf9\u4e8e\u56e0\u679c\u63a8\u65ad\u662f\u795e\u5723\u7684\uff01\u8fd9\u4e5f\u5f88\u597d\u7406\u89e3\uff0c\u4f46\u662f\u4e00\u822c\u5f88\u591a\u539f\u56e0<\/span>\uff08\u6bd4\u5982\u4f26\u7406\u539f\u56e0\u3001\u53ef\u884c\u6027\u3001\u8d44\u6e90\u5229\u7528\u7b49\uff09<\/span>\u5bfc\u81f4\u83b7\u53d6\u6570\u636e\u5e76\u975e\u662f\u968f\u673a\u7684 (w.r.t T)\u3002\u5728\u8fd9\u60c5\u51b5\u4e0b\uff0c\u5982\u679cX\u5b8c\u5168\u63cf\u8ff0\u4e86\u6570\u636e\u5206\u914d\u7684\u5173\u7cfb<\/span> (sufficient adjustment sets)<\/span>\uff0c\u90a3\u6211\u4eec\u8fd8\u662f\u53ef\u4ee5\u4e00\u81f4\u4e14\u65e0\u504f\u4f30\u8ba1ATE<\/span> (consistent & unbiased)<\/span>\uff0c\u8fd9\u91cc\u6d89\u53ca\u4e00\u4e9b\u7814\u7a76\u65f6\u7684\u5047\u8bbe<\/span>\uff08\u76ee\u7684\u662f: Identifiability)<\/span><\/section>\n

        <\/span><\/section>\n
          \n
        1. \n
          Unconfoundedness (conditional ignorability \/ conditional exchangeability)<\/span><\/section>\n<\/li>\n
            \n
          1. \n
            <\/sub>\u8bf4\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u200d<\/span>\u660e\u7ed9\u5b9aX\u65f6\uff0cpotential outcome Y(t)\u4e0eT\u65e0\u5173\uff0c\u8fd9\u6b65\u5bf9\u4e8eATE\u975e\u5e38\u5173\u952e\uff0c\u662f\u4e0a\u9762\u516c\u793a\u7b2c\u4e8c\u884c\u7b49\u53f7\u6210\u7acb\u7684\u6761\u4ef6\uff01<\/span><\/sub> <\/span><\/section>\n<\/li>\n
          2. \n
            conditional exchangeability \u4e3a <\/span><\/span><\/sub> \uff0c\u8bf4\u660e\u4e86Y(1)\u4e0d\u8bba\u5728treatment\u7ec4\u8fd8\u662fcontrol\u7ec4\u671f\u671b\u662f\u4e0d\u53d8\u7684\uff0c\u4e5f\u8bf4\u660e\u4e86Y(t)\u4e0eT\u65e0\u5173<\/span><\/section>\n<\/li>\n
          3. \n

            \u5e76\u4e14\u8fd9\u4e2a\u6761\u4ef6\u662f\u65e0\u6cd5\u6d4b\u8bd5\u7684\uff0c\u60f3\u60f3\u4e3a\u4ec0\u4e48\uff1f<\/span>[\u56e0\u4e3a\u4e16\u754c\u4e4b\u5927]<\/span><\/p>\n<\/li>\n<\/ol>\n

          4. \n
            Positivity<\/span><\/section>\n<\/li>\n
              \n
            1. \n
              <\/span><\/sub>\u770b\u8d77\u6765\u4f3c\u4e4e\u6ca1\u4ec0\u4e48\uff0c\u4f46\u4e5f\u662f\u975e\u5e38\u5173\u952e\u7684\u4e00\u4e2a\u5047\u8bbe\uff01intuition\u4e3a\u5982\u679c\u8fb9\u754c\u8fbe\u6210\uff0c\u5219\u67d0\u4e2ax\u4e0b\uff0c\u8981\u4e48\u5168\u662ftreatment\u8981\u4e48\u5168\u662fcontrol\uff0c\u5219\u65e0\u6cd5\u8ba1\u7b97\u771f\u5b9eATE\uff0c\u6b64\u65f6causal effect\u662fill-defined<\/span>\uff08\u4e14\u5728\u90e8\u5206\u6982\u7387\u63a8\u5230\u4e2d\u5bfc\u81f4\u96640\uff09<\/span><\/section>\n<\/li>\n
            2. \n
              \u6839\u636e\u8d1d\u53f6\u65af\u516c\u5f0f\uff0c\u8fd9\u6761\u5047\u8bbe\u4e5f\u53ebOverlap between P(X|T=1) & P(X|T=0)<\/span><\/section>\n<\/li>\n
            3. \n

              \u5982\u679c\u67d0\u4e9b\u53d8\u91cf\u8fdd\u80cc\u6b64\u5047\u8bbe\uff0c\u53ebpositivity violation\uff0c\u67d0\u4e9b\u60c5\u51b5\u4e0b\u6211\u4eec\u53ef\u4ee5extrapolate\uff0c\u53ef\u4ee5\u60f3\u60f3\u54ea\u4e9b\u60c5\u51b5\u4e0b\uff0c\u53ef\u4ee5\u5916\u63a8\u5462\uff1f<\/span>\uff08\u6bd4\u5982\u95e8\u5f53\u6237\u5bf9\u65f6\u5019\uff09<\/span><\/p>\n<\/li>\n<\/ol>\n

            4. \n
              No interference<\/span><\/section>\n<\/li>\n
                \n
              1. \n
                <\/span><\/sub>\u6b64\u5047\u8bbe\u4e00\u822c\u9ed8\u8ba4\u6210\u7acb\uff0c\u662f\u8bf4\u6211\u7684\u7ed3\u679c\u4e0e\u4ed6\u4eba\u65e0\u5173<\/span><\/section>\n<\/li>\n<\/ol>\n
              2. \n

                Consistency<\/span><\/p>\n<\/li>\n

                  \n
                1. \n
                  <\/span><\/sub>\u6b64\u5047\u8bbe\u4e00\u822c\u9ed8\u8ba4\u6210\u7acb\uff0c\u662f\u8bf4T\u5bf9\u7ed3\u679c\u662f\u4e00\u81f4\u7684\u3002\u7b80\u5355\u4e3e\u4e2a\u4f8b\u5b50\uff0cT\u4e3a\u662f\u5426\u517b\u732b\uff0cY\u4e3a\u662f\u5426\u5f00\u5fc3\uff0c\u4e0d\u5e0c\u671b\u517b\u4e86\u4e00\u53ea\u52a0\u83f2T=1, \u7ed3\u679cY=1\uff1b\u517b\u4e86\u4e00\u53ea\u7f8e\u77edT=1, \u7ed3\u679cY=0\u3002\u8bf4\u660eT\u5b9a\u4e49\u4e0d\u597d\uff0c\u8fd9\u5e38\u5728\u8bbe\u8ba1\u5b9e\u9a8c\u65f6\u5bb9\u6613\u51fa\u9519\u3002\u4e5f\u53ebNo multiple version of Treatment<\/span><\/section>\n<\/li>\n<\/ol>\n<\/ol>\n

                  <\/section>\n
                  \u8fd9\u91cc\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cUnconfoundedness \u548c Positivity\u770b\u8d77\u6765\u662f\u76f8\u4e92\u5236\u7ea6\u7684tradeoff\u7684\u5047\u8bbe\uff0c\u5728\u6587\u7ae0\u6700\u540e\u6211\u4eec\u518d\u6765\u8ba8\u8bba\u3002<\/span>\uff08\u5e0c\u671b\u80fd\u770b\u5230\u6700\u540e\uff5e\uff09<\/span><\/section>\n

                  <\/section>\n
                  \u4ece\u516c\u5f0f\u770b\u5230\uff0c\u6211\u4eec\u5f00\u59cb\u4f30\u8ba1\u7684ATE\u4e3acausal quantities\uff0c\u57fa\u4e8e\u4e00\u7cfb\u5217\u5047\u8bbe\u6700\u540e\u5f97\u5230\u4e00\u4e2astatistical quantities\uff0c\u8fdb\u800c\u5c31\u53ef\u5229\u7528data\u8fdb\u884c\u6570\u503c\u8ba1\u7b97\uff01\u7b80\u5355\u56de\u987e\u5b8c\u4e4b\u524d\u7684\u4e1c\u897f\uff0c\u6211\u4eec\u8fd9\u91cc\u603b\u7ed3\u4e00\u4e0bCausal Inference\u6574\u4f53\u6846\u67b6<\/span><\/section>\n

                  <\/span><\/section>\n

                  \u4ecb\u7ecd\u4e00\u4e9b\u7edf\u8ba1\u8bdd\u672f<\/span><\/p>\n

                    \n
                  • \n

                    Estimand: \u6211\u4eec\u611f\u5174\u8da3\u7684\u91cf\uff0c\u6bd4\u5982\u671f\u671b\uff0cATE<\/span><\/p>\n<\/li>\n

                  • \n

                    Estimate: \u6211\u4eec\u5229\u7528\u6570\u636e\u4f30\u8ba1\u7684\u503c<\/span><\/p>\n<\/li>\n

                  • \n
                    Estimation\uff1aEstimand\u5230Estimate\u8fc7\u7a0b<\/span><\/section>\n<\/li>\n<\/ul>\n

                    <\/section>\n
                    \n
                    \n
                    \n

                    <\/section>\n

                    <\/section>\n<\/section>\n
                    \n
                    \n
                    Causal Inference\u6574\u4f53\u6846\u67b6\uff1a\u5927\u81f4\u5206\u4e3a2\u6b65
                    <\/span><\/section>\n
                      \n
                    1. \n
                      Identification\uff1a\u6211\u4eec\u611f\u5174\u8da3\u7684\u662fCausal Estimand (\u5982ATE)\uff0c\u4f46\u662f\u6ca1\u6cd5\u76f4\u63a5\u8ba1\u7b97\uff08\u56e0\u4e3a\u6211\u4eec\u53ea\u80fd\u8ba1\u7b97\u7edf\u8ba1\u91cf\uff09\uff0c\u9700\u8981\u8f6c\u5316\u4e3astatistical Estimand (\u7edf\u8ba1\u91cf)\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u53ebIdentification<\/span><\/section>\n<\/li>\n
                    2. \n
                      Estimation\uff1a\u6839\u636estatistical Estimand\uff0c\u5229\u7528data\u8fdb\u884c\u4f30\u8ba1\u7684\u6570\u503c\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u53ebEstimation<\/span><\/section>\n<\/li>\n<\/ol>\n
                      \n


                      <\/p>\n

                      <\/p>\n

                      The Identification-Estimation Flowchart<\/span>
                      <\/p>\n


                      <\/span><\/p>\n<\/figure>\n

                      \u63a5\u4e0b\u6765\u57fa\u4e8e\u8fd9\u4e2aCausal Inference\u7684\u6846\u67b6\u6d41\u7a0b\uff0c\u7cfb\u7edf\u4ecb\u7ecd\u4e00\u4e0b\u539f\u7406\u4e0e\u65b9\u6cd5\uff0c\u5e2e\u4f60\u5bf9\u6574\u4f53\u4e0e\u7ec6\u8282\u90fd\u6709\u4e00\u5b9a\u4e86\u89e3\u3002<\/span><\/p>\n<\/section>\n<\/section>\n

                      \n

                      <\/section>\n

                      <\/section>\n<\/section>\n<\/section>\n<\/section>\n


                      <\/p>\n


                      <\/p>\n

                      \n
                      \n
                      \n
                      \n
                      \n

                      <\/section>\n

                      <\/section>\n
                      \n
                      <\/svg><\/section>\n<\/section>\n<\/section>\n
                      \n

                      2. <\/strong><\/strong>Causal Model\uff1a\u57fa\u7840\u6982\u7387\u56fe<\/strong><\/p>\n<\/section>\n

                      \n

                      <\/section>\n

                      <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


                      <\/p>\n

                      \u672c\u8282\u76ee\u7684\uff1a\u4ecb\u7ecd\u6574\u4f53\u6846\u67b6\u975e\u5e38\u91cd\u89812\u6b65\u5185\u5bb9\u4e4b\u524d\uff0c\u5148\u4ecb\u7ecd\u4e00\u4e9bCausal Model\u5305\u62ecgraph\u5185\u5bb9\uff0c\u56e0\u4e3a\u6839\u636e\u4e4b\u524d\u4ecb\u7ecd\u53ef\u4ee5\u770b\u51fa\uff0c\u6211\u4eec\u8ba1\u7b97ATE\u65f6\u4f7f\u7528\u4e86\u5f88\u591a\u5047\u8bbe\uff0c\u8fd9\u4e9b\u5047\u8bbe\u662f\u57fa\u4e8eCausal Model\u800c\u6765\uff0c\u6240\u4ee5\u4e3a\u4e86\u66f4\u597d\u8bb2\u89e3Identification\uff0c\u5fc5\u987b\u5148\u77e5\u9053\u8fd9\u4e9b\u5047\u8bbe\u6765\u7531\uff01<\/span>\uff08\u5f53\u7136\u4e5f\u53ef\u4ee5\u8df3\u8fc7\uff0c\u5f71\u54cd\u4e0d\u5927\uff0c\u5173\u6ce8\u4e00\u4e0b\u57fa\u7840\u7ed3\u6784\u548cd-seperation\u5373\u53ef\uff09<\/span><\/section>\n

                      <\/span><\/section>\n
                      graph model\u5185\u5bb9\u975e\u5e38\u591a\uff0c\u8fd9\u91cc\u53ea\u4ecb\u7ecd\u9488\u5bf9\u56e0\u679c\u63a8\u65ad\u76f8\u5173\u5185\u5bb9\uff0c\u5305\u62ec\u5982\u4f55\u8868\u8fbeintervention\uff0c\u53caintervention\u5206\u5e03\u7b49\uff0cCausal model\u4e00\u822c\u4e3agraph model (directed graph)\uff0c\u5148\u975e\u5e38\u7b80\u5355\u4ecb\u7ecd\u4e00\u4e9b\u76f8\u5173\u4e14\u91cd\u8981\u7684graph model\uff0c\u5177\u4f53\u7ec6\u8282\u4ee5\u540e\u518d\u8865\u5145\u3002<\/span><\/section>\n

                      <\/span><\/section>\n
                      \u7814\u7a76\u56e0\u679c\uff0c\u901a\u5e38\u6211\u4eec\u671f\u671b\u7684\u56e0\u679c\u5173\u7cfb\u662f\u975e\u5bf9\u79f0\u7684\uff0c\u4e00\u822c\u57fa\u4e8e\u4f20\u7edfdirected graph (Bayesian Network)\u5efa\u6a21\uff0cBayesian graph\u63d0\u4f9b\u4e86\u5f88\u591a\u53ef\u7528\u5de5\u5177\uff0c\u6bd4\u5982\u7b80\u5316\u6982\u7387\u5efa\u6a21\u7684bayesian network factorization\u7b49\uff0c\u6574\u4e2a\u6a21\u578b\u601d\u8def\u5c31\u662f\uff0c\u57fa\u4e8e\u6982\u7387\u56fe\u57fa\u7840\u4e0a\u518d\u589e\u52a0\u4e00\u4e9bcausal\u5047\u8bbe\uff0c\u53ef\u4ee5\u8fdb\u884ccausal\u7684\u64cd\u4f5c\u3002<\/span><\/section>\n

                      <\/span><\/section>\n
                        \n
                      • \n

                        Bayesian Network Factorization (Markov Assumption)\uff1a<\/sub><\/span><\/p>\n<\/li>\n

                          \n
                        • \n

                          \u4e3b\u8981\u5bf9\u56fe\u8d4b\u4e88\u4e86independent\u5173\u7cfb(parent-children)\uff0c\u8fdb\u800c\u7b80\u5316\u6982\u7387\u5206\u5e03<\/span><\/p>\n<\/li>\n<\/ul>\n

                        • \n
                          Minimality\/Causal-Edges Assumption\uff1a\u8fde\u63a5\u8282\u70b9\u5177\u6709\u76f8\u5173\u6027<\/span><\/section>\n<\/li>\n<\/ul>\n

                          <\/section>\n
                          \u4e4b\u524d\u4ecb\u7ecd\u4e2d\u4e5f\u63d0\u8fc7\uff0c\u76f8\u5173\u6027<\/span>\u56e0\u679c\u6027<\/span>\uff0c\u5c31\u662f\u56e0\u4e3a\u5728\u7814\u7a76\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u4e0d\u540c\u7684\u5047\u8bbe\uff0c\u5bf9graph<\/span> (DAG) <\/span>\u8d4b\u4e88\u4e86\u6982\u7387\u6027<\/span> (independent & dependent) <\/span>\u548c\u56e0\u679c\u6027<\/span> (dependent)<\/span>\uff0c\u901a\u5e38\u7684\u89c2\u6d4b\u6570\u636e\u4e2d\u65e2\u5305\u542bcausal association\u4e5f\u5305\u542b non-causal association\uff01<\/span><\/section>\n

                          <\/span><\/section>\n
                          <\/figure>\n

                          <\/section>\n
                          \u6240\u4ee5\uff0c\u7ed9\u5b9a\u56fe\u6211\u4eec\u6700\u5173\u5fc3\u5c31\u662f\u90a3\u4e9bassociation\u662f\u5982\u4f55flow\u7684\uff0c\u8fd9\u6837\u6211\u4eec\u5c31\u53ef\u4ee5\u5206\u79bb\u51fa\u6765\u6211\u4eec\u5173\u7cfb\u7684causal association\uff01DAG\u56fe\u4e2d\u6709\u4e09\u79cd\u57fa\u672c\u7ed3\u6784\uff0c\u5e2e\u52a9\u6211\u4eec\u5206\u6790\uff1a<\/span><\/section>\n

                          <\/span><\/section>\n
                          \u9996\u5148\u4ecb\u7ecd\u5e38\u7528\u672f\u8bed\uff0c\u5728graph\u4e2d\uff0c\u963b\u65ad\u4e86association\u79f0\u4e3ablock the path\uff0c\u5426\u5219unblock the path\u3002\u5728\u6982\u7387\u4e2d\u6211\u4eecblock\u4e3b\u8981\u624b\u6bb5\u5c31\u662fconditioning<\/span>\uff08\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38general\u7684\u8bdd\u9898\uff0cconditioning strategy\u8fde\u63a5\u5e76\u603b\u7ed3\u4e86matching\u548cregression\u65b9\u6cd5\uff0c\u4ee5\u540e\u518d\u5199\uff09<\/span>\uff0c\u4e0b\u9762\u4ecb\u7ecd\u57fa\u672c\u7ed3\u6784<\/span><\/section>\n

                          <\/span><\/section>\n
                            \n
                          • \n
                            Chain\uff1ax1<\/sub><\/span>\u2192x2<\/sub><\/span>\u2192x3<\/sub><\/span> <\/span>; Fork\uff1ax1<\/sub><\/span>\u2190x2<\/sub><\/span>\u2192x3<\/sub><\/span><\/span><\/section>\n<\/li>\n
                              \n
                            • \n
                              \u6ca1\u6709condition on x2<\/sub><\/span>\u65f6\uff0cassociation\u4ecex1<\/sub><\/span>\u6d41\u5411x3<\/sub><\/span>\uff08flows from x1<\/sub><\/span> to x3<\/sub><\/span>)<\/span><\/section>\n<\/li>\n
                            • \n
                              condition on x2<\/sub><\/span>, block the path<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n
                              \n

                              <\/section>\n
                              <\/section>\n
                              blocked path\uff08\u5de6\uff1achain \u53f3\uff1afork\uff09<\/span>
                              <\/section>\n

                              <\/span><\/section>\n<\/figure>\n
                                \n
                              • \n
                                immorality<\/span> (collider)<\/span>\uff1a <\/span>x<\/span>1<\/sub><\/span>\u2192<\/span>x<\/span>2<\/sub><\/span>\u2190<\/span>x<\/span>3<\/sub><\/span><\/section>\n<\/li>\n
                                  \n
                                • \n
                                  \u8fd9\u4e2a\u7ed3\u6784\u6b63\u597d\u4e0e\u4e0a\u9762\u76f8\u53cd\uff0c\u6ca1\u6709condition on x2<\/sub><\/span>\u65f6\uff0cblock the path<\/span><\/section>\n<\/li>\n
                                • \n
                                  condition on x2<\/sub><\/span> (or descendent of x2<\/sub><\/span>)\uff0cunblock the path<\/span><\/section>\n<\/li>\n
                                • \n
                                  \u8fd9\u4e2a\u76f4\u89c2\u4e0a\u53ef\u4ee5\u7406\u89e3\u4e3a\uff1ax2<\/sub><\/span>\u4e3a\u9884\u6d4b\u76ee\u6807\u6bd4\u5982\u662f\u5426\u4eba\u7c7b\u9ad8\u8d28\u91cf\uff0cx1<\/sub><\/span>\u548cx3<\/sub><\/span>\u662f2\u4e2a\u7279\u5f81\uff0c\u6bd4\u5982\u91d1\u94b1\u548c\u5e74\u9f84\uff1b\u91d1\u94b1\u548c\u5e74\u9f84\u672c\u6765\u662f\u6ca1\u6709\u76f8\u5173\u6027\uff0c\u4f46\u662f\u5982\u679c\u6211\u9650\u5236\u4e86x2<\/sub><\/span>=\u662f\u4eba\u7c7b\u9ad8\u8d28\u91cf\uff0c\u90a3\u53ef\u80fd\u5fc5\u987b\u94b1\u591a\u5e74\u8f7b\u624d\u884c\uff0c\u5bfc\u81f4x1<\/sub><\/span>\u4e0ex3<\/sub><\/span>\u5728x2<\/sub><\/span>=\u662f\u4eba\u7c7b\u9ad8\u8d28\u91cf\u65f6\u662f\u76f8\u5173\u7684<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n
                                  \n

                                  <\/section>\n
                                  <\/section>\n
                                  blocked path for collider<\/span>
                                  <\/section>\n

                                  <\/span><\/section>\n<\/figure>\n
                                  \u57fa\u4e8e\u4e09\u79cd\u57fa\u672c\u7ed3\u6784\u548c\u5404\u81ea\u5982\u4f55block the path\uff0c\u5728\u4efb\u610fgraph\u4e2d\u6211\u4eec\u5c31\u53ef\u4ee5\u77e5\u9053association flows & blocked.<\/span><\/section>\n

                                  <\/span><\/section>\n
                                  blocked path rules<\/span><\/section>\n

                                  <\/span><\/section>\n
                                    \n
                                  1. \n
                                    block chain\u548cfork\u7ed3\u6784\u7684\u4e2d\u95f4\u8282\u70b9 <\/span><\/sub><\/span><\/section>\n<\/li>\n
                                  2. \n
                                    unblock collider\u53ca\u5176\u5b50\u5b59<\/span><\/section>\n<\/li>\n<\/ol>\n
                                    <\/span><\/section>\n

                                    <\/section>\n
                                    \u57fa\u4e8epath block\u89c4\u5219\uff0c\u518d\u4ecb\u7ecd\u4e00\u4e2a\u7ecf\u5e38\u89c1\u5230\u7684\u6982\u5ff5\uff0cd-seperation<\/span> (X and Y are d-seperated by Z)<\/span>\uff0c\u662f\u8bf4condition on Z\uff0cblock X\u5230Y\u7684\u5168\u90e8path<\/span><\/section>\n

                                    <\/span><\/section>\n
                                    <\/p>\n
                                    T and Y are d-seperated by W&amp;amp;amp;M<\/span><\/section>\n

                                    <\/section>\n<\/figure>\n


                                    <\/span><\/p>\n

                                    \n
                                    \n
                                    \n
                                    \n
                                    \n

                                    <\/section>\n

                                    <\/section>\n
                                    \n
                                    <\/svg><\/section>\n<\/section>\n<\/section>\n
                                    \n

                                    3. <\/strong><\/strong>Causal Model\uff1a\u56e0\u679c\u56fe<\/strong><\/p>\n<\/section>\n

                                    \n

                                    <\/section>\n

                                    <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


                                    <\/p>\n

                                    \u4e4b\u524d\u4ecb\u7ecd\u7684\u8d1d\u53f6\u65af\u56fe\u66f4\u591a\u6d89\u53castatistics\uff0c\u57fa\u4e8e\u6b64\u5f15\u5165\u66f4\u591a\u56e0\u679c\u64cd\u4f5c\u53ca\u5047\u8bbe\uff0c\u6700\u7ec8\u5f97\u4ee5\u8ba1\u7b97\u56e0\u679c\u6548\u5e94\u3002\u4e4b\u524d\u4ecb\u7ecd\u8fc7\u56e0\u679c\u6a21\u578b\u76f8\u5bf9\u4e8e\u4f20\u7edf\u7edf\u8ba1\u6a21\u578b\uff0c\u6700\u5927\u7684\u533a\u522b\u5728\u4e8e\u4e00\u4e2a\u662f\u7814\u7a76\u5e72\u9884<\/span> (intervention)<\/span>\uff0c\u8fd8\u6709\u5c31\u662f\u7814\u7a76\u53cd\u4e8b\u5b9e<\/span> (conterfactual)<\/span>\u3002<\/span><\/section>\n

                                    <\/span><\/section>\n
                                    \u9996\u5148\u4ecb\u7ecd\u76f8\u5bf9\u4e8eintervening\u7684\u8868\u8fbedo-operator\uff0c\u8fd9\u662f\u4e00\u4e2acausal op\uff0c\u5bf9\u5e94\u4e00\u4e2acausal distribution<\/span> (interventional distribution)<\/span>\uff0c\u56e0\u4e3a\u662f\u5f15\u5165\u6982\u7387\u6a21\u578b\u6982\u5ff5\uff0c\u6211\u4eec\u53ef\u4ee5\u5bf9\u6bd4\u5206\u6790\u3002<\/span><\/section>\n

                                    <\/span><\/section>\n
                                    \u6838\u5fc3\u533a\u522b\u5728\u4e8e\u7edf\u8ba1\u5206\u6790\u662f\u57fa\u4e8e\u89c2\u6d4b\u6570\u636e\uff0c\u53ef\u4ee5conditioning\u64cd\u4f5c\u5bf9\u6570\u636e\u5212\u5206\u7ea6\u675f<\/span> (sub-population)<\/span>\uff0c\u5c5e\u4e8e\u88ab\u52a8\u5206\u6790\uff1b\u800c\u56e0\u679c\u63a8\u65ad\u5e0c\u671b\u4e3b\u52a8\u83b7\u53d6\u6570\u636e\uff0c\u901a\u8fc7intervening\u64cd\u4f5c\u5e72\u9884\u6570\u636e\u5206\u914d<\/span> (whole-population)<\/span>\uff0c\u5c5e\u4e8e\u4e3b\u52a8\u5206\u6790\u3002<\/span><\/section>\n

                                    <\/span><\/section>\n
                                      \n
                                    • \n
                                      Interventional distribution: <\/sub><\/span><\/section>\n<\/li>\n
                                        \n
                                      • \n
                                        \u7528do(T=t)\u6765\u8868\u8fbe\u6211\u4eec\u5f3a\u5236\u7528\u6237\u8fdb\u884ctreatment\u5206\u914d(T = t)<\/span><\/section>\n<\/li>\n
                                      • \n
                                        causal quantities<\/span> (\u65e0\u6cd5\u76f4\u63a5\u8ba1\u7b97\uff0c\u9700\u8981\u8bc6\u522b(Identification)\u8f6c\u5316\u4e3a\u7edf\u8ba1\u91cf\u624d\u80fd\u8ba1\u7b97)<\/span><\/section>\n<\/li>\n
                                      • \n
                                        \u53ef\u4ee5\u901a\u8fc7experimental<\/span> (\u968f\u673a\u8bd5\u9a8c)<\/span>\u83b7\u53d6<\/span> (\u76f8\u5bf9\u4e8estatistical distribution P(X,Y,T), \u53ef\u76f4\u63a5\u4ecedata\u4e2d\u83b7\u5f97)<\/span><\/section>\n<\/li>\n
                                      • \n
                                        ATE\u5219\u4e3a<\/sub><\/span>\u4e00\u9636\u77e9\u7684\u5dee<\/sub><\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n
                                        <\/span><\/section>\n

                                        <\/section>\n
                                        \u57fa\u4e8e\u4ee5\u4e0a\u4ecb\u7ecd\uff0c\u4f60\u53ef\u4ee5\u5bf9\u6bd4<\/sub>\u548c<\/sub>\u4e86\u4e48\uff1f\u4eceIdentification\u89d2\u5ea6\u6765\u8bf4\uff0c\u6211\u4eec\u5e0c\u671b\u8f6c\u5316causal estimand<\/span> (\u542b\u6709do op)<\/span>\u4e3astatistical estimand<\/span> (\u4e0d\u542bdo op)<\/span> \uff01<\/span><\/section>\n

                                        <\/span><\/section>\n
                                        <\/section>\n

                                        <\/span><\/section>\n
                                        Causal Model\u5c31\u662f\u57fa\u4e8egraph\u83b7\u53d6\u4fe1\u606f<\/span> (Assumption)<\/span>\uff0c\u8fdb\u800c\u4ecegraph\u4e2d\u5206\u79bb\u51facausation\uff0c\u4e3e\u4e2a<\/span> (Assumption) <\/span>\u4f8b\u5b50\u5982\u4e0b:<\/span><\/section>\n

                                        <\/span><\/section>\n
                                          \n
                                        • \n

                                          no confounder\uff1a<\/sub><\/span> \uff0c\u8fd9\u79cd\u60c5\u51b5 \u76f8\u5173\u6027\u2192\u56e0\u679c\u6027<\/span><\/p>\n<\/li>\n

                                        • \n
                                          confounder X\uff1a <\/span><\/sub><\/section>\n<\/li>\n<\/ul>\n

                                          <\/section>\n
                                          \u63a5\u4e0b\u6765\u4ecb\u7ecd\u4e00\u4e2a\u91cd\u8981\u4e14\u76f4\u89c2\u7684\u5047\u8bbe\uff0cModularity Assumption\uff0c\u5b83\u4e3a\u6211\u4eecformalize interventional distribution\u63d0\u4f9b\u4e86\u65b9\u6cd5\u3002\u8fd9\u4e2a\u5047\u8bbe\u4e5f\u5f88\u76f4\u89c2\uff0c\u6211\u4eecintervening on <\/span>Xi<\/sub> \uff0c\u5219graph\u91cc\u9762\u9664\u4e86<\/sub><\/span>\u6539\u53d8\uff0c\u5176\u4ed6factor\u4fdd\u6301\u4e0d\u53d8\u3002<\/span><\/section>\n

                                          <\/span><\/section>\n
                                          Modularity Assumption\uff1aIntervening on sets S, factor of variables i
                                          1. <\/sub>\u4e0d\u53d8<\/span>
                                          2. <\/sub><\/span><\/section>\n

                                          <\/section>\n
                                          \u6839\u636eBaysian Network Factorization\u548cModularity Assumption\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5199\u51fainterventional distribution\u7684\u5206\u5e03\uff0c\u5f97\u5230truncated factorization<\/span>\uff08\u628a\u5e72\u9884\u53d8\u91cf\u56e0\u5b50\u66ff\u63620\u62161\uff09<\/span>\u3002 <\/sub>. graph\u76f8\u5e94\u6539\u53d8\u4e3amanipulated graph\uff0c\u5220\u6389\u6307\u5411\u5e72\u9884\u53d8\u91cf\u7684\u8fb9\uff0c\u5176\u4ed6\u4fdd\u6301\u4e0d\u53d8\uff01<\/span><\/section>\n

                                          <\/span><\/section>\n
                                          <\/p>\n
                                          manipulated graph<\/span><\/section>\n

                                          <\/span><\/section>\n<\/figure>\n
                                          \u6700\u540e\u4ee5\u4e00\u4e2a\u4f8b\u5b50\u770b\u5982\u4f55\u8fd0\u7528\u8fd9\u4e9b\u77e5\u8bc6\uff0c\u76ee\u6807\u4e3aIdentify<\/sub><\/span><\/section>\n

                                          <\/sub><\/span><\/section>\n
                                          \n
                                          <\/section>\n
                                          causal graph to manipulated graph
                                          <\/span><\/section>\n

                                          <\/span><\/section>\n<\/figure>\n
                                            \n
                                          1. \n
                                            Bayesian Network Factorization: <\/span><\/sub><\/section>\n<\/li>\n
                                          2. \n
                                            Truncated Factorization: <\/span><\/sub><\/section>\n<\/li>\n
                                          3. \n
                                            Marginalize: <\/sub><\/span><\/section>\n<\/li>\n<\/ol>\n

                                            <\/section>\n


                                            <\/span><\/h2>\n

                                            \n
                                            \n
                                            \n
                                            \n
                                            \n

                                            <\/section>\n

                                            <\/section>\n
                                            \n
                                            <\/svg><\/section>\n<\/section>\n<\/section>\n
                                            \n

                                            4. <\/strong><\/strong>Identification\uff1a<\/strong>backdoor\u3001<\/strong><\/p>\n

                                            frontdoor criterion\u3001 do-calculus<\/strong><\/p>\n<\/section>\n

                                            \n

                                            <\/section>\n

                                            <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


                                            <\/p>\n

                                            \u63a5\u4e0b\u6765\u4ecb\u7ecd\u56e0\u679c\u63a8\u65ad\u6838\u5fc3\u6d41\u7a0b\u4e4b\u4e00\uff0cIdentification\uff0c\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u5e76\u4e0d\u662f\u6240\u6709\u7684\u56e0\u679c\u5173\u7cfb\u90fd\u662fidentifiable\uff0c\u5176\u4e2d\u6709\u5f88\u591a\u5b9e\u73b0\u65b9\u6cd5\uff0c\u8fd9\u91cc\u4ecb\u7ecd\u5e38\u7528\u4e14\u76f4\u89c2\u7684\u5165\u95e8\u65b9\u6cd5\u3002<\/span><\/section>\n

                                            <\/span><\/section>\n
                                            \u57fa\u4e8e\u521a\u624d\u7684\u4ecb\u7ecd\uff0c\u518d\u56de\u5fc6\u4e00\u4e0b\u4e3a\u4ec0\u4e48\u968f\u673a\u8bd5\u9a8c<\/span> (RCTs) <\/span>\u5bf9\u4e8e\u56e0\u679c\u63a8\u65ad\u5982\u6b64\u795e\u5723\uff1f<\/span><\/section>\n

                                            <\/span><\/section>\n
                                            \u7531\u6b64\u5f15\u51faIdentification\u7684\u6838\u5fc3\u601d\u8def\u4e4b\u524d\u63d0\u5230\u8fc7\uff0c\u5927\u90e8\u5206\u57fa\u4e8e\u89c2\u6d4b\u6570\u636e\u7684\u7814\u7a76\uff0c\u56e0\u679c\u56fe\u7c7b\u4f3c\u4e0a\u9762\u56fe\u5de6\uff0c\u901a\u5e38\u5b58\u5728confounder X<\/span> (T\u5e76\u975e\u968f\u673a)<\/span>\uff0c\u5bfc\u81f4\u76f8\u5173\u6027<\/span>\u56e0\u679c\u6027<\/span> \uff0c \u5373<\/sub>\u3002\u5f53\u6211\u4eec\u8fdb\u884c\u968f\u673a\u5b9e\u9a8c<\/span> (intervening on T) <\/span>\u65f6\uff0c\u5207\u65ad\u4e86T\u7684in-coming edge\uff0c\u6d88\u9664\u4e86confounder <\/span>(block non-causal path)<\/span>\uff0c\u5219\u53ef\u4ee5\u76f4\u63a5\u8fdb\u884c\u56e0\u679c\u8ba1\u7b97\uff01\u6240\u4ee5\u6211\u4eec\u5e0c\u671b\u57fa\u4e8e\u89c2\u6d4b\u6570\u636e\uff0c\u6a21\u62df\u8fd9\u4e2a\u64cd\u4f5c\u57fa\u4e8e\u7edf\u8ba1\u7684conditioning that block the non-causal path\uff0c\u5b9e\u73b0Identification\uff01\u8fd9\u4e9bnon-causal path\u4e3abackdoor path\u3002<\/span><\/section>\n

                                            <\/span><\/section>\n
                                            \u503c\u5f97\u591a\u8bf4\u7684\u662f\uff0c\u5728\u4e00\u4e9b\u7ecf\u5178\u9886\u57df\u6bd4\u5982\u793e\u4f1a\u7ecf\u6d4e\u533b\u5b66\u7b49\uff0c\u5e38\u5e38\u63d0\u5230balance the data\uff0c\u5728conditioning strategy\u6216\u8005matching method\u7b49\uff0c\u6211\u4eec\u8bf4Identification\u901a\u8fc7balanced data\u5b9e\u73b0\uff0c\u8fd9\u91cc\u548c\u5176\u4ed6\u7bc7\u5e45\u5199\u7684comparable group\u5bf9\u5e94\uff0c\u4e5f\u662fRCTs\u81ea\u5e26\u5c5e\u6027\u4e4b\u4e00\uff01\u5373<\/sub><\/span><\/section>\n

                                            <\/section>\n
                                            backdoor adjustment<\/span><\/strong><\/section>\n

                                            <\/span><\/section>\n
                                            Backdoor Path\uff1a\u6211\u4eec\u7814\u7a76T on Y\u7684\u56e0\u679c\u6548\u5e94\uff0cbackdoor path\u4e3a\u90a3\u4e9bT\u5230Y\u4e14\u6307\u5411T\u7684path\uff0c\u8fd9\u4e9bpath\u4e3anon-causal association\uff0c\u5e0c\u671bblock\u5206\u79bb\u51facausation\uff0c\u7ed3\u5408\u4e4b\u524d\u4ecb\u7ecd\u7684graph\u4e2dblock path\u65b9\u6cd5\uff0c\u5f97\u5230\u7b2c\u4e00\u4e2aidentification\u65b9\u6cd5 backdoor adjustment.<\/span><\/section>\n

                                            <\/span><\/section>\n
                                            backdoor criterion: variable sets W satisfy 1. block all backdoor path 2. not contains descents of T. then W is called sufficient adjustment sets.
                                            backdoor adjustment: if W satisfy backdoor criterion, then ATE is identified
                                            <\/span><\/section>\n
                                            <\/section>\n
                                              \n
                                            • \n
                                              \u8bf4\u660e<\/span>\uff08\u76f4\u89c2\u8bc1\u660e\uff09<\/span><\/section>\n<\/li>\n
                                                \n
                                              • \n
                                                \u7b2c\u4e8c\u4e2a\u7b49\u53f7\u4e2d<\/sub><\/span>\u7684do(t)\u662f\u56e0\u4e3aW\u4e3asufficient adjustment set\uff0c\u5207\u65ad\u4e86\u6240\u6709backdoor path\uff0cT->Y\u4ec5\u7559\u4e0bcausation<\/span><\/section>\n<\/li>\n
                                              • \n
                                                \u7b2c\u4e09\u4e2a\u7b49\u53f7\u6d88\u9664<\/sub><\/span>\u7684do(t)\u662f\u56e0\u4e3a 1. do(t)\u5207\u65ad\u4e86T\u7684parents\uff0c\u6ca1\u6709in-edge association\u6d41\u5411T->W; 2. \u5982\u679c\u5b58\u5728association\u5219\u4e3aT \u7684 out-edge association\uff0c\u5219\u4f1a\u4e0eY\u5f62\u6210\u4e00\u4e2acollider\uff0cassociation \u88ab collider \u5207\u65ad\u3002\u6240\u4ee5T\u4e0eW\u72ec\u7acb\uff01\u4e0b\u9762\u4e3e\u4e2a\u4f8b\u5b50\uff0c<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n
                                                <\/p>\n
                                                blocking backdoor path<\/span><\/section>\n

                                                <\/span><\/section>\n<\/figure>\n
                                                  \n
                                                    \n
                                                  • \n
                                                    \u4e0a\u56fe\u67092\u6761backdoor path\uff1a<\/sub> <\/span> ; <\/span><\/section>\n<\/li>\n
                                                  • \n
                                                    conditioning sets <\/span><\/sub><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n

                                                    <\/section>\n
                                                    \u56de\u987e\u4e00\u4e0b\u4e0a\u9762\u4ecb\u7ecd\u7684d-seperation\uff0c\u6211\u4eec\u53ef\u4ee5\u7528d-seperation\u6765\u8868\u8fbebackdoor path <\/span><\/sub>\uff0c<\/sub><\/span>\u8868\u8fbemanipulated graph\uff0c\u5220\u6389T outgoing edge\u3002<\/span><\/section>\n

                                                    <\/span><\/section>\n
                                                    \u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cbackdoor criterion\u4e2d\u7b2c\u4e8c\u6761\uff0c\u4e0d\u8981\u963b\u65adT\u7684\u5b50\u5b59\uff0c\u8fd9\u4e2a\u76f4\u89c2\u6765\u8bf4\u5c31\u662f\u522b\u628aT\u7684\u6548\u5e94\u63d0\u524d\u963b\u65ad\u4e86\uff0c\u8fd8\u6709\u4e00\u79cd\u662f\u57fa\u4e8eSCMs\u6765\u8bf4\u660e\uff0cSCMs\u672c\u8d28\u4e5f\u5e94\u8be5\u5728Causal Model\u4ecb\u7ecd\uff0c\u66f4\u591a\u4f5c\u7528\u662f\u5728\u540e\u9762\u4ecb\u7ecdcounterfactuals\u65f6\u548c\u5b58\u5728unobserved variables\u4f7f\u7528\uff0c\u4e3a\u4e86\u7bc7\u5e45\u5728\u8fd9\u5148\u4e0d\u4ecb\u7ecd\u4e86\u3002<\/span><\/section>\n

                                                    <\/span><\/section>\n
                                                    frontdoor adjustment<\/span><\/strong><\/section>\n

                                                    <\/span><\/section>\n
                                                    \u524d\u9762\u63d0\u8fc7unconfoundedness\u662funtestable\u7684\uff0c\u4e5f\u5373\u5b58\u5728unobserved variable\u4e3aconfounder\uff0c\u6211\u4eec\u6ca1\u6cd5conditioning\uff0c\u8fd9\u4e2a\u65f6\u5019backdoor adjustment\u5c31\u6ca1\u6cd5\u4f7f\u7528\u4e86\uff0c\u5982\u4e0b\u6240\u793a<\/span><\/section>\n
                                                    <\/section>\n
                                                    unobserved backdoor path<\/span>
                                                    <\/section>\n
                                                    \n

                                                    <\/span><\/section>\n<\/figure>\n
                                                    \u8fd9\u91cc\u6838\u5fc3\u70b9\u662f\uff0cM\u627f\u8f7d\u4e86\u5168\u90e8T -> Y\u7684causation\uff0c\u6211\u4eec\u901a\u8fc7focus on M\u6765\u5206\u79bbcausation\uff0c\u5b9e\u73b0frontdoor adjustment\uff0c\u4e3b\u8981\u75313\u6b65\u7ec4\u6210\uff0c\u603b\u7ed3\u5982\u4e0b\u3002<\/span><\/section>\n

                                                    <\/span><\/section>\n
                                                    frontdoor criterion: variable sets M satisfy 1. all causal path from T on Y through M 2. no unblocked backdoor path from T to M 3. T block all backdoor path from M to Y
                                                    frontdoor adjustment: <\/span><\/section>\n
                                                    <\/section>\n

                                                    <\/section>\n
                                                      \n
                                                    • \n
                                                      \u8bf4\u660e<\/span>\uff08\u76f4\u89c2\u8bc1\u660e\uff09<\/span><\/section>\n<\/li>\n
                                                    • \n
                                                      step 1\uff0cT->M\uff0c\u6ca1\u6709backdoor path<\/span><\/section>\n<\/li>\n
                                                    • \n
                                                      step 2\uff0cM->Y\uff0c\u5b58\u5728backdoor path M\u2190T\u2190<\/span>W\u2192<\/span>Y<\/span>\uff0c\u53ef\u4ee5\u88abT block<\/span><\/section>\n<\/li>\n<\/ul>\n

                                                      <\/span><\/section>\n
                                                      Pearl’s do-calculus<\/span><\/strong><\/section>\n

                                                      <\/span><\/section>\n
                                                      \u4e4b\u524d\u4ecb\u7ecdbackdoor\u3001frontdoor adjustment\u90fd\u662f\u5f88\u76f4\u89c2\u76f4\u63a5\u53ef\u4ee5\u4ececausal graph\u4e2d\u5224\u65adidentifiable\u3002\u540c\u65f6\u4e5f\u5b58\u5728\u4ee5\u4e0a\u4e24\u79cd\u65b9\u6cd5\u4e0d\u80fdIdentity\u7684\u60c5\u51b5\u3002\u8fd9\u91cc\u6211\u4eec\u4ecb\u7ecd\u4e00\u4e2acomplete \u65b9\u6cd5\uff0c\u53ea\u8981\u662fidentifiable\u7684\u90fd\u53ef\u4ee5\u901a\u8fc7do-calculus\u8bc6\u522b\uff01\u4e0b\u9762\u8fdb\u884c\u4ecb\u7ecd\u8bf4\u660e\uff1a<\/span><\/section>\n
                                                        \n
                                                      • \n
                                                        Rule 1: <\/span><\/sub><\/section>\n<\/li>\n
                                                          \n
                                                        • \n
                                                          \u63cf\u8ff0\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0cZ\u4e0eY\u72ec\u7acb\uff0c\u5219\u53ef\u4ee5\u6d88\u9664condition on z<\/span><\/section>\n<\/li>\n
                                                        • \n
                                                          \u4fbf\u4e8e\u76f4\u89c2\u4e0a\u7406\u89e3\uff0c\u6211\u4eec\u5148\u628ado(t) \u5220\u6389\uff0c\u5f97\u5230<\/sub><\/span>\uff0c\u53ef\u4ee5\u53d1\u73b0\u8fd9\u4e2a\u5c31\u662fd-seperation\uff0c\u6240\u4ee5\u6b64\u4e3ad-seperation\u5728intenventional distribution\u63a8\u5e7f<\/span><\/section>\n<\/li>\n<\/ul>\n
                                                        • \n
                                                          Rule 2: <\/span><\/sub><\/section>\n<\/li>\n
                                                            \n
                                                          • \n
                                                            \u63cf\u8ff0\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u53ea\u6709causation\uff0c\u5219\u53ef\u4ee5\u6d88\u9664do(z)<\/span><\/section>\n<\/li>\n
                                                          • \n
                                                            \u540c\u6837\u4fbf\u4e8e\u76f4\u89c2\u4e0a\u7406\u89e3\uff0c\u6211\u4eec\u628ado(t) \u5220\u6389\uff0c\u5f97\u5230<\/span><\/sub>\uff0c\u53ef\u4ee5\u53d1\u73b0\u8fd9\u4e2a\u5c31\u662fbackdoor adjustment\uff0c\u6240\u4ee5\u6b64\u4e3abackdoor adjustment\u5728intenventional distribution\u63a8\u5e7f<\/span><\/section>\n<\/li>\n<\/ul>\n
                                                          • \n
                                                            Rule 3: <\/sub><\/span><\/section>\n<\/li>\n
                                                              \n
                                                            • \n
                                                              Z(W)\u4e3avariable sets Z\u4e2d\u6392\u9664W\u7956\u5148\u7684\u90a3\u4e9b<\/span><\/section>\n<\/li>\n
                                                            • \n
                                                              \u63cf\u8ff0\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0cdo(Z)\u4e0d\u5f71\u54cdY<\/span><\/section>\n<\/li>\n
                                                            • \n
                                                              \u56e0\u4e3acondition on W\uff0c\u5b58\u5728Z\u4e3acollider\uff0c\u5bb9\u6613block collider\u7684\u5b50\u5b59\u5e26\u6765\u989d\u5916association\uff0c\u6240\u4ee5\u8fd9\u91cc\u4ec5\u5207\u6389Z(W)\u7684incoming-edge<\/span><\/section>\n<\/li>\n<\/ul>\n<\/ul>\n

                                                              <\/section>\n
                                                              \u6700\u540e\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cdo-calculus\u662fIdentification\u7684\u5145\u8981\u6761\u4ef6<\/span> (\u53ef\u4ee5\u8bc6\u522b\u7684\u60c5\u51b5\u4e0b)<\/span>\uff0c\u4f46\u5b83\u4e0d\u662f\u57fa\u4e8eGraph\u7684\uff0c\u6709\u4e9b\u65b9\u6cd5\u53ef\u4ee5\u76f4\u63a5\u4ececausal graph\u5224\u65adIdentification\u3002\u6bd4\u5982\u4f5c\u4e3a\u5145\u5206\u6761\u4ef6\u7684unconfounded children criterion\uff1b\u5145\u5206\u5fc5\u8981\u6761\u4ef6hedge criterion\uff0c\u611f\u5174\u8da3\u53ef\u4ee5\u81ea\u884c\u9605\u8bfb\u3002<\/span><\/section>\n

                                                              <\/section>\n


                                                              <\/h2>\n

                                                              \n
                                                              \n
                                                              \n
                                                              \n
                                                              \n

                                                              <\/section>\n

                                                              <\/section>\n
                                                              \n
                                                              <\/svg><\/section>\n<\/section>\n<\/section>\n
                                                              \n

                                                              5. <\/strong><\/strong>Evaluation<\/strong><\/p>\n<\/section>\n

                                                              \n

                                                              <\/section>\n

                                                              <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


                                                              <\/p>\n

                                                              \u524d\u9762\u4ecb\u7ecd\u5b8cCausal Model\u548cIdentification\uff0c\u5e2e\u52a9\u6211\u4eec\u628acausal estimand\u8f6c\u5316\u4e3a\u53ef\u4ee5\u4f30\u8ba1\u7684statistical estimand\uff0c\u63a5\u4e0b\u6765\u6211\u4eec\u5c31\u662f\u8ba1\u7b97\u7684\u90e8\u5206\uff0c\u5982\u4f55\u7528data\u6765\u8ba1\u7b97\u5177\u4f53\u7684casual effect\u7684\u6570\u503c\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u5c31\u662fEstimation\u3002\u9996\u5148\u4ecb\u7ecd\u4e00\u4e9b\u5e38\u7528\u7684\u4f30\u8ba1\u91cf\uff0c\u6211\u4eec\u4e4b\u524d\u4e00\u76f4\u4ee5ATE\u4e3a\u4f8b\uff0c\u5b9e\u9645\u89c2\u6d4b\u6570\u636e\u4e2d\u6211\u4eec\u7ecf\u5e38\u4f30\u8ba1conditional ATE<\/span> (CATE)<\/span>\uff0c\u6709\u4e9b\u7814\u7a76\u53ef\u80fd\u53ea\u5173\u6ce8treatment\u6216\u8005control\u7ec4\u7684ATE\uff0c\u5206\u522b\u53ebATT\u548cATC<\/span>\uff08\u6709\u4e9b\u60c5\u51b5\u4e0bATE\u65e0\u6cd5\u4f30\u8ba1\uff09<\/span>\uff0c\u6211\u4eec\u4ee5\u540e\u6709\u673a\u4f1a\u518d\u7ec6\u8bb2\uff0c\u8fd9\u91cc\u5148\u4ecb\u7ecd\u5e38\u7528\u7684CATE<\/span><\/sub><\/section>\n

                                                              <\/span><\/section>\n
                                                              \u672c\u5c0f\u8282\u9ed8\u8ba4\u5047\u5b9aunconfoundedness & positivity\uff0c\u5373Identification\u6210\u7acb\uff0c\u4e14CATE\u4e2dX\u548ccovariates W\u90fd\u662fsufficient adjustment sets. (X\u4e0d\u9700\u8981all observed)\uff0c\u63a5\u4e0b\u6765\u4ecb\u7ecd\u5e38\u7528\u65b9\u6cd5\u3002<\/span><\/section>\n

                                                              <\/span><\/section>\n
                                                              Conditional Outcome Modeling \uff08COM\uff09<\/span><\/strong><\/section>\n

                                                              <\/span><\/strong><\/section>\n
                                                              \u57fa\u4e8e\u4e4b\u524d\u8ba8\u8bba\u6211\u4eec\u5f97\u5230ATE\u7684\u8bc6\u522b\u4e3a<\/sub><\/span>\u53f3\u8fb9\u5f0f\u5b50\u53ea\u5305\u542bstatistical quantities\uff0c\u662fstatistical estimand\u3002estimation\u9700\u8981\u6211\u4eec\u7528data\u4f30\u8ba1estimate\u3002\u6700\u76f4\u63a5\u7684\u529e\u6cd5\u5c31\u662f\u6211\u4eec\u53ef\u4ee5\u5148\u4f7f\u7528statistical model\u6216\u8005ML model\u4f30\u8ba1conditional expectation\uff0c\u4e4b\u540e\u8fdb\u884c\u6837\u672c\u4f30\u8ba1\u3002<\/span><\/section>\n
                                                                \n
                                                              • \n
                                                                ATE<\/span><\/section>\n<\/li>\n
                                                              • \n
                                                                step1: model conditional expectation: <\/span><\/sub><\/section>\n<\/li>\n
                                                              • \n
                                                                step2: average over sample: <\/span><\/sub><\/span><\/sub><\/section>\n<\/li>\n
                                                              • \n
                                                                CATE<\/span><\/section>\n<\/li>\n
                                                              • \n
                                                                step1<\/span>: <\/span><\/sub><\/section>\n<\/li>\n
                                                              • \n
                                                                step2: average over sample: <\/span><\/sub><\/section>\n<\/li>\n<\/ul>\n

                                                                <\/span><\/section>\n
                                                                COM\u6709\u5f88\u591a\u5176\u4ed6\u540d\u5b57\uff0c\u6bd4\u5982S-learner\u7b49\uff0cS\u4ee3\u8868singe model\uff0c\u8868\u8fbe\u5bf9\u4e8etreatment\u548ccontrol\u7ec4\u4f7f\u7528\u7684\u662f\u540c\u4e00\u4e2amodel\u3002<\/span><\/section>\n

                                                                <\/span><\/section>\n
                                                                \u5b58\u5728\u95ee\u9898\uff1a\u5f53\u6570\u636e\u7eac\u5ea6\u53d8\u5927\u65f6\uff0c\u5bb9\u6613\u4ea7\u751f\u4f30\u8ba1\u504f\u5dee<\/span> (estimate biased towards 0\uff01\uff09<\/span>\uff0c\u8fd9\u4e2a\u5f88\u76f4\u89c2\uff0c\u5047\u8bbe\u6570\u636e\u6709100\u7ef4\uff0cT\u53ea\u662f\u5176\u4e2d1\u7ef4\uff0c\u6a21\u578b\u8bad\u7ec3\u65f6\u5bb9\u6613\u5ffd\u7565\u8fd9\u4e2a\u7ef4\u5ea6\uff0c\u5bfc\u81f4\u8ba1\u7b97\u51fa\u6765ATE -> 0\uff0c\u8fdb\u800c\u5f15\u51faGrouped COM\u3002<\/span><\/section>\n

                                                                <\/span><\/section>\n
                                                                Grouped COM (GCOM)<\/span><\/strong><\/section>\n

                                                                <\/span><\/section>\n
                                                                GCOM\u5f15\u5165\u4e3a\u4e86\u89e3\u51b3COM\u4e2destimate biased towards to 0\u7684\u95ee\u9898\uff0c\u4e3a\u4e86\u52a0\u5f3a\u6a21\u578b\u5bf9T\u7684\u6743\u91cd\uff0c\u6211\u4eec\u5728\u4f30\u8ba1<\/span>\u03bc\u65f6<\/span>\uff0c\u5206\u522b\u5bf9treatment\u7ec4\u548ccontrol\u7ec4\u4f7f\u7528\u4e0d\u540c\u7684\u6a21\u578b\uff0c\u5176\u4ed6\u6d41\u7a0b\u4e0eCOM\u4e00\u6837\uff0c\u4e5f\u5e38\u53ebT-Learner\u3002<\/span><\/section>\n

                                                                <\/span><\/section>\n
                                                                  \n
                                                                • \n
                                                                  ATE<\/span><\/section>\n<\/li>\n
                                                                • \n
                                                                  step1<\/span><\/section>\n<\/li>\n
                                                                    \n
                                                                  • \n
                                                                    model <\/span><\/sub><\/section>\n<\/li>\n
                                                                  • \n
                                                                    model <\/span><\/sub><\/section>\n<\/li>\n<\/ul>\n
                                                                  • \n
                                                                    step2: average over sample: <\/span><\/sub><\/section>\n<\/li>\n<\/ul>\n
                                                                    <\/span><\/section>\n

                                                                    <\/section>\n
                                                                    \u5b58\u5728\u95ee\u9898\uff1a\u6bcf\u4e2aestimator\u53ea\u4f7f\u7528\u90e8\u5206\u6570\u636e\uff0c\u5c24\u5176\u5f53\u6837\u672c\u4e0d\u8db3\u6216\u8005treatment\u3001control\u6837\u672c\u91cf\u5dee\u522b\u8f83\u5927\u65f6\uff0c\u6a21\u578bvariance\u8f83\u5927<\/span>\uff08\u5bf9\u6570\u636e\u5229\u7528\u6548\u7387\u4f4e\uff09<\/span>\uff0c\u6211\u4eec\u9700\u8981\u63d0\u5347\u6570\u636e\u5229\u7528\u7387\uff01\u8fdb\u800c\u5f15\u51fa\u63a5\u4e0b\u67652\u4e2amodel TARNet\u548cX-Learner\u3002<\/span><\/section>\n

                                                                    <\/section>\n
                                                                    TARNet<\/span><\/strong><\/section>\n

                                                                    <\/span><\/section>\n
                                                                    \u501f<\/span>\u9274\u795e\u7ecf\u7f51\u7edcshare-bottom\u7ed3\u6784\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u5982\u4e0b\u7ed3\u6784\uff0c\u589e\u5f3a\u6570\u636e\u5229\u7528\u3002<\/span><\/section>\n
                                                                    \n

                                                                    <\/span><\/p>\n

                                                                    TARNet<\/span><\/section>\n

                                                                    <\/span><\/section>\n<\/figure>\n
                                                                    \u53ef\u4ee5\u53d1\u73b0\uff0c\u5176\u5b9ebottom\u7ed3\u6784\u53c2\u6570\u4f7f\u7528\u4e86\u5168\u90e8\u6570\u636e\uff0c\u4f46\u662fhead\u7ed3\u6784\u8fd8\u662f\u4ec5\u4f7f\u7528\u90e8\u5206\u6570\u636e\uff0c\u6570\u636e\u5229\u7528\u7387\u5e76\u4e0d\u662f\u6700\u9ad8\u7684\u3002\u8fd9\u79cd\u7ed3\u6784\u5728\u795e\u7ecf\u7f51\u7edc\u7ecf\u5e38\u7528\u5230\u4e14\u95ee\u9898\u4e0d\u5927\uff0c\u6211\u731c\u4e3b\u8981\u539f\u56e0\u662f\u795e\u7ecf\u7f51\u7edc\u4e00\u822c\u6570\u636e\u91cf\u90fd\u592a\u5927\uff0c\u8fd9\u79cd\u95ee\u9898\u53ef\u4ee5\u5ffd\u7565\u3002<\/span><\/section>\n

                                                                    <\/span><\/section>\n
                                                                    X-Learner<\/span><\/strong><\/section>\n

                                                                    <\/span><\/section>\n
                                                                    \u5145\u5206\u5229\u7528\u6570\u636e\u4f30\u8ba1\u6bcf\u4e2agroup\u7684estimator\uff0c\u5bf9\u4e8e\u6570\u636e\u503e\u659c\u5f88\u4e25\u91cd\u7684\u4f30\u8ba1\u6709\u5f88\u597d\u7684\u5f25\u8865\u4f5c\u7528\u3002X-learner\u4f30\u8ba1\u6b65\u9aa4\u5982\u4e0b\uff1a<\/span><\/section>\n

                                                                    <\/span><\/section>\n
                                                                      \n
                                                                    • \n
                                                                      step1 :Estimate <\/sub><\/span> \uff0c\u8fd9\u4e2a\u548cT-Learner\u4e00\u6837\uff0c\u5206\u522b\u7528treatment\u548ccontrol\u6570\u636e\u4f30\u8ba1\u6761\u4ef6\u671f\u671b<\/span><\/section>\n<\/li>\n
                                                                    • \n
                                                                      step2: Impute ITEs<\/span><\/section>\n<\/li>\n
                                                                        \n
                                                                      • \n
                                                                        2a. \u9488\u5bf9treatment\u7ec4\uff0c<\/sub><\/span> \uff1b\u9488\u5bf9control\u7ec4\uff0c <\/sub><\/span><\/section>\n<\/li>\n
                                                                      • \n
                                                                        2b. \u5206\u522b\u5b66\u4e602\u4e2amodel <\/span><\/sub>\u548c<\/span><\/sub>\uff0c\u5176\u4e2d <\/span><\/sub><\/span><\/span>\u4f7f\u7528treatment\u7ec4\u6570\u636e\u9884\u6d4b <\/sub><\/span>\uff0c<\/sub><\/span><\/span> \u4f7f\u7528control\u7ec4\u6570\u636e\u9884\u6d4b<\/sub><\/span><\/section>\n<\/li>\n<\/ul>\n
                                                                      • \n
                                                                        step3: \u5229\u7528weight function<\/sub><\/span>\u7ec4\u54082\u4e2amodel<\/span> <\/span><\/sub><\/span>\u548c<\/sub><\/span><\/span>\u8fdb\u884c\u9884\u6d4b\uff1a<\/sub><\/span><\/section>\n<\/li>\n<\/ul>\n

                                                                        <\/span><\/section>\n
                                                                        \u89c2\u5bdf\u53ef\u4ee5\u53d1\u73b0\uff0cX-Learner\u5728step1\u4e2d\u8bad\u7ec3 <\/span> \u65f6\uff0c\u4e0eT-Learner\u4e00\u6837\u4ec5\u4f7f\u7528treatment\u7ec4\u6570\u636e\u4f46\u662f\u5728step2\u8bad\u7ec3<\/sub><\/span>\u65f6\u662f\u57fa\u4e8e<\/span><\/sub>\u8bad\u7ec3\u7684\uff0c\u5176\u4e2d<\/span><\/section>\n
                                                                        \u4f7f\u7528control\u7ec4\u7684potential outcome <\/span>\uff0c\u6240\u4ee5\u5728\u8bad\u7ec3<\/sub><\/span><\/span>\u65f6\u4f7f\u7528\u4e86\u5168\u90e8\u7684\u6570\u636e\uff0c\u540c\u6837\u8bad\u7ec3 <\/span><\/sub><\/span><\/span>\u4e5f\u662f\u4e00\u6837\u3002<\/span>
                                                                        <\/section>\n

                                                                        <\/span><\/section>\n
                                                                        \u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cpaper\u8bf4\u7528g(x) propensity score\u6548\u679c\u4e0d\u9519\uff1b\u6216\u8005\u5176\u4ed6\u4e5f\u884c\uff0c\u4e00\u822c\u6570\u636e\u91cf\u591a\u7684\u7ec4\u6743\u91cd\u7ed9\u5927\u4e00\u4e9b\uff1b\u6216\u8005\u989d\u5916\u5b66\u4e60g(x)\uff0c\u7528\u4ee5\u6700\u5c0f\u5316\u6a21\u578b\u65b9\u5dee\u7b49\u3002<\/span><\/section>\n

                                                                        <\/span><\/section>\n
                                                                        Propensity score and IPW<\/span><\/strong><\/section>\n

                                                                        <\/span><\/section>\n
                                                                        propensity score based\u7684\u65b9\u6cd5\u6709\u5f88\u6df1\u7684\u80cc\u666f\u548c\u601d\u60f3\uff0c\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u4ece\u5f62\u5f0f\u4e0a\u89e3\u51b3\u4e86\u4e00\u4e9b\u4e0d\u53ef\u884c\u7684\u60c5\u51b5\uff0c\u8fd9\u91cc\u6211\u4eec\u7b80\u5355\u4ecb\u7ecd\u4e9b\u57fa\u7840\uff0c\u4ee5\u540e\u6df1\u5165\u8ba8\u8bba\u3002<\/span><\/section>\n

                                                                        <\/span><\/section>\n
                                                                        propensity score\u5b9a\u4e49\u4e3a <\/span> \uff0c\u8861\u91cf\u4e86\u5728W=w\u4e0b\u5206\u914dT=1\u7684\u6982\u7387\uff0c\u76f4\u89c9\u4e0a\u5982\u679cW\u662fsufficient adjustment sets\u4e14e(W)\u4f30\u8ba1\u6a21\u578b\u6b63\u786e\u7684\uff0c\u90a3e(W)\u5305\u542b\u4e86\u5168\u90e8treatment assignment\u7684\u4fe1\u606f\uff0c\u90a3\u6211\u4eec\u4ec5\u901a\u8fc7e(W)\u5c31\u53ef\u4ee5consistent & unbiased\u4f30\u8ba1ATE\u4e86\uff01<\/span><\/section>\n

                                                                        <\/section>\n
                                                                        Propensity Score Theorem
                                                                        \u5982\u679c <\/span>\uff0c\u5219 <\/span><\/section>\n
                                                                        <\/section>\n

                                                                        <\/section>\n
                                                                        \u8fd9\u91cc\u6709\u4e2a\u5f88\u660e\u663e\u7684\u597d\u5904\u662fe(W)\u4ec5\u662f\u4e00\u7ef4\u7684\uff01\u8bb0\u5f97\u5f00\u5934\u6211\u4eec\u4f7f\u7528potential outcome model\u8bc6\u522bATE\u63d0\u5230\u7684\u4e24\u4e2a\u6838\u5fc3\u5047\u8bbeunconfoundedness \u548c positivity\u662f\u4e00\u5bf9tradeoff\uff0c\u56e0\u4e3a\u6211\u4eec\u8003\u8651\u66f4\u591a\u53d8\u91cf\u52a0\u5165sufficient adjustment sets\uff0c\u5219\u6211\u4eec\u6709\u66f4\u591a\u7684\u673a\u4f1ablock all backdoor path\uff0c\u5b9e\u73b0unconfoundedness\uff0c\u4f46\u662f\u8003\u8651\u7684\u7ef4\u5ea6\u8d8a\u9ad8\uff0c\u6211\u4eec\u5b9e\u73b0positivity\u96be\u5ea6\u5c31\u8d8a\u5927<\/span>\uff08\u7ef4\u5ea6\u707e\u96be\uff09<\/span>\uff0c\u6837\u672c\u7a00\u758f\u5ea6\u5448\u6307\u6570\u7ea7\u589e\u52a0\uff01<\/span><\/section>\n

                                                                        <\/span><\/section>\n
                                                                        \u65e2\u7136\u7ef4\u5ea6\u707e\u96be\u5b58\u5728\u5bfc\u81f4Identification\u7684\u5047\u8bbe\u5b58\u5728tradeoff\uff0c\u90a3\u662f\u5426\u57fa\u4e8epropensity score\u5c31\u53ef\u4ee5\u5b8c\u7f8e\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u4e86\uff1f\u7b54\u6848\u80af\u5b9a\u662f\u5426\u5b9a\u7684\uff01\u524d\u9762\u4e5f\u63d0\u5230e(W)\u6307\u5b9a\u6b63\u786e\u7684\u6a21\u578b\u4e5f\u662f\u5f88\u5173\u952e\u7684\uff0c\u8fd9\u91cc\u53ea\u662f\u628a\u9ad8\u7ef4\u95ee\u9898\u8f6c\u5316\u6210\u4e86e(W)\u9884\u4f30\u51c6\u786e\u7684\u95ee\u9898\uff0c\u540c\u6837\u662f\u68d8\u624b\u7684\u95ee\u9898\uff01<\/span><\/section>\n

                                                                        <\/span><\/section>\n
                                                                        \u8fd9\u91cc\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u65f6\u5e38\u4f1a\u6709\u4eba\u628a\u95ee\u9898\u5f62\u5f0f\u8f6c\u5316\u800c\u8ba4\u4e3a\u95ee\u9898\u5f97\u5230\u89e3\u51b3\uff0c\u8fd9\u662f\u6b20\u7f3a\u601d\u8003\u7684\uff01\u6bd4\u5982\u7528ML\u9884\u6d4b\u91d1\u878d\u5e02\u573a\u96be\u5ea6\u5927\uff0c\u800c\u4f7f\u7528RL\u8fdb\u884c\u9884\u6d4b\uff0c\u6216\u8005\u7b80\u5355\u66f4\u6362\u4e0d\u540c\u7684\u6a21\u578b\u7ed3\u6784\u7b49\u7b49\uff0cNo Free Lunch\u662f\u6c38\u6052\u6210\u7acb\u7684\uff01<\/span><\/section>\n

                                                                        <\/span><\/section>\n
                                                                        \u63a5\u4e0b\u6765\u4ecb\u7ecdInverse Probability Weighting (IPW)\u5982\u4f55\u4f30\u8ba1ATE\uff0cas always\uff0c\u4e3a\u4e86\u76f4\u89c2\u4ecb\u7ecd\u5176\u601d\u60f3\uff0c\u6211\u4eec\u5148\u4ecb\u7ecdPseudo-population\u3002<\/span><\/section>\n

                                                                        <\/span><\/section>\n
                                                                        \u901a\u5e38\u60c5\u51b5\u4e0b\uff0c\u89c2\u6d4b\u6570\u636e\u4e2d\u5b58\u5728backdoor\uff0c\u5982\u4e0b\u56fe\u5de6\u6240\u793a\uff0c\u65e0\u6cd5\u76f4\u63a5\u8ba1\u7b97ATE\uff0c\u56e0\u4e3a <\/span> \u3002\u5728\u4ecb\u7ecdbackdoor\u65f6\uff0c\u6211\u4eec\u5f97\u5230\u5f53W\u6ee1\u8db3backdoor criterion\u65f6\uff0cATE\u53ef\u4ee5\u8bc6\u522b (\u516c\u5f0f\u89d2\u5ea6\uff0c\u56fe\u5f62\u89d2\u5ea6\u53c2\u8003\u524d\u8ff0\uff09<\/span><\/section>\n