加入新推出的
Discord 社区,展开实时讨论,获得同行支持,并直接与 Meridian 团队互动!
必要假設
透過集合功能整理內容
你可以依據偏好儲存及分類內容。
一般來說,迴歸中沒有潛在結果的概念,因為迴歸模型會估算回應變數的條件預期值。不過,在有條件可互換性和一致性這兩項重要假設下:
$$
E \Biggl(
\overset \sim Y_{g,t}^{
\left(\left\{
x_{g,t,m}^{(\ast)}
\right\}\right)
} \Big| \bigl\{z_{g,t,c}\bigr\}
\Biggr) = E \Biggl(
\overset \sim Y_{g,t} \Big|
\bigl\{z_{g,t,c}\bigr\}, \big\{x_{g,t,m}^{(\ast)}\bigr\} \Biggr)
$$
主要假設
條件式可替換性:
\( \overset \sim Y_{g,t}^{(\{ x_{g,t,m}^{(\ast)} \})} \)與任何對照情境\(\bigl\{ x_{g,t,m}^{(\ast)} \bigr\}\)的隨機變數\(\bigl\{ X_{g,t,m}^{(\ast)} \bigr\}\) 無關。因此,潛在結果集合與廣告主過去的媒體執行決策無關。
一致性:
\( \overset \sim Y_{g,t} = \overset \sim Y_{g,t}^{
(\{ x_{g,t,m}^{(\ast)} \})
} \) when \(\bigl\{ X_{g,t,m}^{(\ast)} \bigr\} =
\bigl\{ x_{g,t,m}^{(\ast)} \bigr\}\)。因此,觀察到的 KPI 實現值,是對照假設情境的潛在結果,與廣告客戶的歷來媒體執行作業相符。
根據這些假設,您會得到先前所述的結果:
$$
E \Biggl( \overset \sim Y_{g,t}^{
\left(\left\{ x_{g,t,m}^{\ast} \right\}\right)
} \Big| \bigl\{ z_{g,t,m} \bigr\} \Biggr)
\overset{\text{exchangeability}}{=} E \Biggl( \overset \sim Y_{g,t}^{
\left(\left\{ x_{g,t,m}^{\ast} \right\}\right)
} \Big| \bigl\{ z_{g,t,c} \bigr\},\ \bigl\{ x_{g,t,m}^{(\ast)} \bigr\} \Biggr)
\overset{\text{consistency}}{=} E \Biggl( \overset \sim Y_{g,t}\ \Big|
\bigl\{ z_{g,t,c} \bigr\},\ \bigl\{ x_{g,t,m}^{(\ast)} \bigr\}
\Biggr)
$$
一致性假設相當直覺,除非對照組定義不佳,或在資料中未正確呈現,否則這項假設會持續有效。詳情請參閱 Hernan MA, Robins JM, (2020) Causal Inference: What If。
條件可替換性假設則不太直觀。如果所有混淆變數都已評估並納入控制陣列 \(\{z_{g,t,c}\}\),則這項假設就成立。混淆變數是指對觀察到的治療方法 \(\{x_{g,t,m}\}\) 和結果\(\{\overset \sim y_{g,t}\}\)都具有因果效應的任何變數。對實驗處理的因果效應,可能會影響廣告主的整體預算水準、各管道的分配方式、各地區的分配方式,或各時段的分配方式。在實際操作中,很難知道是否已測量所有混淆變數,因為這純粹是一種假設,而且沒有統計測試可從資料中判斷這一點。不過,如果您假設因果圖符合「後門條件」 (Pearl, J., 2009)。詳情請參閱「因果圖」。
除非另有註明,否則本頁面中的內容是採用創用 CC 姓名標示 4.0 授權,程式碼範例則為阿帕契 2.0 授權。詳情請參閱《Google Developers 網站政策》。Java 是 Oracle 和/或其關聯企業的註冊商標。
上次更新時間:2024-11-14 (世界標準時間)。
[null,null,["上次更新時間:2024-11-14 (世界標準時間)。"],[[["\u003cp\u003eRegression models can be used to estimate potential outcomes under the assumptions of conditional exchangeability and consistency.\u003c/p\u003e\n"],["\u003cp\u003eConditional exchangeability implies that potential outcomes are independent of historical media execution decisions, given confounding variables.\u003c/p\u003e\n"],["\u003cp\u003eConsistency means the observed outcome matches the potential outcome for the actual historical media execution.\u003c/p\u003e\n"],["\u003cp\u003eConfounding variables, which affect both treatment and outcome, must be measured and included for conditional exchangeability to hold.\u003c/p\u003e\n"],["\u003cp\u003eWhile there's no statistical test to guarantee conditional exchangeability, causal graphs and the backdoor criterion can help assess it.\u003c/p\u003e\n"]]],["Regression models typically lack potential outcomes, but under conditional exchangeability and consistency, we can derive a relevant result. Conditional exchangeability means potential outcomes are independent of historical media execution. Consistency dictates that observed outcomes match potential outcomes when treatment equals historical media execution. The key result is derived by first exchanging outcomes with potential outcomes, then aligning them with observed values under these assumptions. Conditional exchangeability relies on all confounders (variables affecting both treatment and outcome) being measured and can be assessed with causal graph analysis.\n"],null,["# Required assumptions\n\nGenerally speaking, there is no concept of potential outcomes in regression\nbecause regression models estimate conditional expectations of a response\nvariable. However, under the key assumptions of *conditional exchangeability*\nand *consistency*: \n$$ E \\\\Biggl( \\\\overset \\\\sim Y_{g,t}\\^{ \\\\left(\\\\left\\\\{ x_{g,t,i}\\^{(\\\\ast)} \\\\right\\\\}\\\\right) } \\\\Big\\| \\\\bigl\\\\{z_{g,t,i}\\\\bigr\\\\} \\\\Biggr) = E \\\\Biggl( \\\\overset \\\\sim Y_{g,t} \\\\Big\\| \\\\bigl\\\\{z_{g,t,i}\\\\bigr\\\\}, \\\\big\\\\{x_{g,t,i}\\^{(\\\\ast)}\\\\bigr\\\\} \\\\Biggr) $$\n\n**Key assumptions**\n\n- Conditional exchangeability:\n\n \\\\( \\\\overset \\\\sim Y_{g,t}\\^{(\\\\{ x_{g,t,i}\\^{(\\\\ast)} \\\\})} \\\\)\n is independent of the random variables\n \\\\(\\\\bigl\\\\{ X_{g,t,i}\\^{(\\\\ast)} \\\\bigr\\\\}\\\\) for any counterfactual scenario\n \\\\(\\\\bigl\\\\{ x_{g,t,i}\\^{(\\\\ast)} \\\\bigr\\\\}\\\\). So, the set of potential outcomes\n is conditionally independent of the advertiser's historical media execution\n decision.\n- Consistency:\n\n \\\\( \\\\overset \\\\sim Y_{g,t} = \\\\overset \\\\sim Y_{g,t}\\^{\n (\\\\{ x_{g,t,i}\\^{(\\\\ast)} \\\\})\n } \\\\) when \\\\(\\\\bigl\\\\{ X_{g,t,i}\\^{(\\\\ast)} \\\\bigr\\\\} =\n \\\\bigl\\\\{ x_{g,t,i}\\^{(\\\\ast)} \\\\bigr\\\\}\\\\). So, the observed KPI realization of\n the potential outcome for the counterfactual scenario matching the\n advertiser's historical media execution.\n\nUnder these assumptions, you have the previously stated result: \n$$ E \\\\Biggl( \\\\overset \\\\sim Y_{g,t}\\^{ \\\\left(\\\\left\\\\{ x_{g,t,i}\\^{\\\\ast} \\\\right\\\\}\\\\right) } \\\\Big\\| \\\\bigl\\\\{ z_{g,t,i} \\\\bigr\\\\} \\\\Biggr) \\\\overset{\\\\text{exchangeability}}{=} E \\\\Biggl( \\\\overset \\\\sim Y_{g,t}\\^{ \\\\left(\\\\left\\\\{ x_{g,t,i}\\^{\\\\ast} \\\\right\\\\}\\\\right) } \\\\Big\\| \\\\bigl\\\\{ z_{g,t,i} \\\\bigr\\\\},\\\\ \\\\bigl\\\\{ x_{g,t,i}\\^{(\\\\ast)} \\\\bigr\\\\} \\\\Biggr) \\\\overset{\\\\text{consistency}}{=} E \\\\Biggl( \\\\overset \\\\sim Y_{g,t}\\\\ \\\\Big\\| \\\\bigl\\\\{ z_{g,t,i} \\\\bigr\\\\},\\\\ \\\\bigl\\\\{ x_{g,t,i}\\^{(\\\\ast)} \\\\bigr\\\\} \\\\Biggr) $$\n\nThe consistency assumption is fairly intuitive, and holds unless the\ncounterfactual is poorly defined or is not accurately represented in the data.\nFor more information, see [Hernan MA, Robins JM, (2020) Causal Inference: What\nIf](https://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/).\n\nThe conditional exchangeability assumption is a bit less intuitive. This\nassumption holds if all confounding variables are measured and included in the\ncontrol array \\\\(\\\\{z_{g,t,i}\\\\}\\\\). *Confounding variables* are anything that has\na causal effect on both the observed treatment \\\\(\\\\{x_{g,t,i}\\\\}\\\\) and outcome\n\\\\(\\\\{\\\\overset \\\\sim y_{g,t}\\\\}\\\\). A causal effect on treatment can mean an effect\nof the advertiser's overall budget level, the allocation across channels, the\nallocation across geos, or the allocation across time periods. In practice, it\nis difficult to know whether all of the confounding variables are measured\nbecause it is purely an assumption, and there is no statistical test to\ndetermine this from your data. However, it can be helpful to know that the\nconditional exchangeability assumption holds if you assume a causal graph that\nmeets a condition known as the *backdoor criterion* (Pearl, J., 2009). For more\ninformation, see [Causal graph](/meridian/docs/basics/causal-graph)."]]