先看這個問題為什麼值得解Why this problem is worth solving
情境先行Start with the situation
資料顯示帶雨傘的人更常遇到下雨。預測模型可以利用這個關聯，但強迫每個人帶傘不會增加降雨。 Data show that people carrying umbrellas more often encounter rain. A predictive model can use this association, but forcing everyone to carry an umbrella will not cause more rain.
若把 P(Y|X) 當成 P(Y|do(X))，高準確度模型也可能導出方向錯誤的政策。 If P(Y|X) is treated as P(Y|do(X)), even a highly accurate model can recommend a policy in the wrong direction.
先備知識Prerequisite
- 理解條件機率與相關性 Understand conditional probability and correlation
讀完你能做到After this path
- 先定義要估計的量（estimand），再選模型與資料 Write the estimand before choosing a model and data
- 區分預測不確定性與因果識別 Distinguish predictive uncertainty from causal identification
- 用敏感度分析呈現無法直接檢驗的假設 Expose untestable assumptions with sensitivity analysis
閱讀框架Concept frame
- 核心問題Question
- 模型能準確預測結果，為什麼仍可能答錯『採取某項行動後會怎樣』？ Why can a model predict an outcome accurately yet answer incorrectly what happens after an action?
- 對照軸Comparison axis
- 未觀測混淆可能有多強 Plausible strength of unobserved confounding
- 判讀方式Evidence to inspect
- 並排閱讀三種相對情境，再回到論文檢查消融實驗、反事實測試與適用範圍。Read three relative scenarios side by side, then inspect the paper's ablations, counterfactual tests, and scope statements.
回答後查看解析Answer, then view the explanation
你的預測Make a prediction
在觀測資料中，處置與結果都受未觀測的病情嚴重度影響時，增加樣本量最直接改善什麼？ When treatment and outcome are both affected by unobserved severity, what does adding samples most directly improve?
看現象，不急著講故事Observe before narrating
三個診斷訊號Three diagnostic signals
這些是待檢查的診斷假說。下一步用條件比較、消融實驗或反事實測試排除其他解釋。These are diagnostic hypotheses to test. Use condition comparisons, ablations, or counterfactual tests to rule out alternatives.
一次看完三種情境Read three scenarios side by side
三種情境對照Three-scenario comparison
未觀測混淆增強到什麼程度時，你的決策會翻轉？ At what confounding strength would your decision reverse?
對照軸：Comparison axis: 未觀測混淆可能有多強 Plausible strength of unobserved confounding
可聚焦檢查資料重疊、估計量變異與校準。 Focus on overlap, estimator variance, and calibration.
報告敏感度曲線，而不是只給一個效果值。 Report sensitivity curves rather than one effect estimate.
效果的符號或大小可能不可識別，結論應降級。 The sign or magnitude may be unidentified, so downgrade the conclusion.
從直覺到機制From intuition to mechanism
把機制講清楚Explain the mechanism
機率模型描述我們看見什麼與不知道什麼；因果模型還要說明世界在介入後如何改變。 Probabilistic models describe what we observe and do not know; causal models additionally specify how the world changes under intervention.
分析從資料、要估計的量與結構假設出發，得到預測分布或帶有不確定性的效果估計；這些結果只能在已明示的條件下用於預測、解釋或決策。 Input: data, an estimand, and structural assumptions. Output: a predictive distribution or effect estimate with uncertainty. Objective: support prediction, explanation, or decisions under explicit conditions.
- 01 先定義研究族群、介入方式、結果，以及要估計的量。 Define the population, intervention, outcome, and estimand first.
- 02 用因果圖、研究設計或模型說明識別假設。 State identification assumptions through a graph, design, or model.
- 03 估計之後檢查資料重疊、校準、敏感度與結果能否移用。 After estimation, inspect overlap, calibration, sensitivity, and transport.
用更複雜的預測模型，掩蓋估計目標不清或因果識別不成立。 Using a more complex predictor to obscure an unclear estimand or failed identification.
四步閱讀：要估計什麼、靠哪些假設識別、如何估計，以及違反假設時結果有多敏感。 Read in four steps: the estimand, identifying assumptions, estimation, and sensitivity to violations.
不用回看頁面的自我檢查Check your recall without looking back
三個回想題Three recall questions
先口頭回答，再展開解析。Answer aloud before opening each explanation.
01 估計目標（estimand）是什麼？ What is an estimand?
研究要估計的明確數量，應說清楚研究族群、處置或條件、結果與比較方式。 The precise quantity of interest, including population, treatment or condition, outcome, and contrast.
02 因果識別與統計估計有何不同？ How do identification and estimation differ?
因果識別問觀測分布加上假設後，能否確定研究目標；統計估計則問如何用有限資料逼近它。 Identification asks whether the observed distribution plus assumptions determines the target; estimation asks how to approximate it from finite data.
03 為何後驗分布很窄仍可能不可靠？ Why can a narrow posterior still be unreliable?
它通常只反映既定模型內的不確定性，不包含模型設定錯誤或因果識別失敗。 It often reflects uncertainty within the chosen model, not misspecification or identification failure.
把工具帶回你的研究Take the tool back to your research
最後一個任務One final task
選一篇機率或因果論文，只從書目資料與摘要寫出估計目標、依賴的假設、不確定性，以及尚未進行的敏感度檢查；無法判斷的項目保留為未知。 Choose a probabilistic or causal paper and, from metadata or abstract only, write its estimand, assumptions, uncertainty, and unreported sensitivity check, keeping indeterminate fields unknown.