Skip to main content

Chapter_21

Aaron L9/17/25About 9 minCausal inferencegpt-4.1response_md

Chapter 21: G-Methods for Time-Varying Treatments（时间变化处理的g方法）

This chapter provides a comprehensive overview of three main g-methods—the g-formula, IP weighting, and g-estimation—and their doubly/multiply robust extensions for estimating causal effects in the presence of time-varying treatments with treatment-confounder feedback. The chapter also addresses the issue of censoring as a time-varying process and generalizes the notion of identification through the "big g-formula". Below, I summarize all technical, fine, and key points, and provide main formulas in both English and Chinese with standard LaTeX math formatting.

1. The Challenge

  • Time-varying treatments complicate confounding adjustment due to treatment-confounder feedback (i.e., confounders affected by or affecting subsequent treatments).
  • Traditional adjustment methods (e.g., standard regression) fail under feedback: even with no causal effect, they can produce a spurious effect estimate.

2. G-Methods Overview

Main Idea

  • G-methods are designed to yield causally correct estimates under treatment-confounder feedback by satisfying three identifiability conditions:
    1. Sequential exchangeability (sequential ignorability)
    2. Positivity
    3. Consistency

3. The G-Formula for Time-Varying Treatments

A. For Two Time Points

Suppose we observe baseline and follow-up treatments (A0,A1)(A_0, A_1), time-varying confounders (L0,L1)(L_0, L_1), and outcome YY.

G-Formula (Static Strategy):
[
E[Y_{a_0, a_1}] = \sum_{l_1} E[Y|A_0=a_0, A_1=a_1, L_1=l_1] ; f(l_1|a_0)
]

  • f(l1∣a0)f(l_1|a_0) is the distribution of L1L_1 given A0=a0A_0 = a_0 in the study population.

中文：
[
E[Y_{a_0, a_1}] = \sum_{l_1} E[Y|A_0=a_0, A_1=a_1, L_1=l_1] \cdot P(L_1=l_1|A_0=a_0)
]

B. For KK Time Points and General Settings

[
E[Y_{\bar{a}}] = \sum_{\bar{l}} E[Y|\bar{A} = \bar{a}, \bar{L} = \bar{l}] \prod_{k=0}^K f(l_k | \bar{a}{k-1}, \bar{l})
]

  • Here, Aˉ=(A0,...,AK)\bar{A} = (A_0,...,A_K), Lˉ=(L0,...,LK)\bar{L} = (L_0,...,L_K).

中文：
[
E[Y_{\bar{a}}] = \sum_{\bar{l}} E[Y|\bar{A}=\bar{a}, \bar{L}=\bar{l}] \prod_{k=0}^K P(L_k | \bar{A}{k-1}, \bar{L})
]

Key Points about the G-Formula

  • The formula standardizes the outcome to the confounder distribution under the intervention.
  • The set of confounders included ("history") is what is needed to achieve conditional exchangeability for AkA_k (Fine Point 21.1).
  • The positivity condition: There must be observed data supporting all relevant treatment/covariate histories.
  • In high-dimensional data, parameter estimation requires plug-in (possibly parametric) models (parametric g-formula).
  • The formula generalizes to both deterministic (fixed path of treatment) and random (probabilistic) strategies.

G-Formula Density (Technical Point 21.1)

For outcome YY and covariates LL:
[
f(y|\bar{a}K, \bar{l}K) \prod^K f(l_k | \bar{a}, \bar{l}_{k-1})
]

  • Marginalizing over LL yields the density for YY under treatment strategy aˉ\bar{a}.

中文：
处理结果YY与协变量LL的g公式密度为：
[
f(y|\bar{a}K, \bar{l}K) \prod^K f(l_k | \bar{a}, \bar{l}_{k-1})
]

G-Formula for Random Strategies

[
E[Y] = \sum_{\bar{a}, \bar{l}} E[Y|\bar{A} = \bar{a}, \bar{L} = \bar{l}]
\prod_{k=0}^K f(l_k | \bar{a}{k-1}, \bar{l}) \prod_{k=0}^K f_{\text{int}}(a_k | \bar{a}_{k-1}, \bar{l}_k)
]

中文：
对于随机策略，g公式为：
[
E[Y] = \sum_{\bar{a}, \bar{l}} E[Y|\bar{A}=\bar{a}, \bar{L}=\bar{l}] \prod_{k=0}^K P(L_k|\bar{A}{k-1}, \bar{L}) \prod_{k=0}^K P_{\text{int}}(A_k|\bar{A}_{k-1},\bar{L}_k)
]

C. Technical/Fine Points

  • Inclusion of Variables is Crucial: Omitting relevant confounders breaks the identification (removes causal interpretation).
  • Component-wise Causal Meaning: The overall g-formula may be causal, but its pieces may lack causal interpretation (unless in sequentially randomized data).
  • "History" Might Not Match Time: "History" for adjustment is the set of variables required for exchangeability, not necessarily all earlier-in-time variables.

4. IP Weighting for Time-Varying Treatments

A. Definition

For treatment history Aˉ=(A0,…,AK)\bar{A} = (A_0,\ldots,A_K) and confounder history Lˉ=(L0,...,LK)\bar{L} = (L_0,...,L_K), the Inverse Probability (IP) Weight at time kk:

Non-stabilized weight:
[
W_{\bar{A}} = \prod_{k=0}^K \frac{1}{f(A_k|\bar{A}_{k-1}, \bar{L}_k)}
]

Stabilized weight:
[
SW_{\bar{A}} = \prod_{k=0}^K \frac{f(A_k|\bar{A}{k-1})}{f(A_k|\bar{A}, \bar{L}_k)}
]

中文：

非稳态IP权重：
[
W_{\bar{A}} = \prod_{k=0}^K \frac{1}{P(A_k|\bar{A}_{k-1}, \bar{L}_k)}
]

稳态（稳定化）IP权重：
[
SW_{\bar{A}} = \prod_{k=0}^K \frac{P(A_k|\bar{A}{k-1})}{P(A_k|\bar{A}, \bar{L}_k)}
]

B. Implementation

  • Fit models for f(Ak∣Aˉk−1,Lˉk)f(A_k|\bar{A}_{k-1},\bar{L}_k) (often logistic regression).
  • Calculate weights for each individual and time point.
  • Weighted estimation: Use these weights to create a pseudo-population in which treatment is randomized.

Key Outcome: Under correct assumptions, IP weighting "removes" confounding by simulating a situation where treatment/exposure is "randomized" with respect to confounders.

C. Estimation of Causal Effects

Main approach:

  • Estimate average outcome conditional on being in a treatment strategy via weighted average:
    [
    E^{ps}[Y|A=\bar{a}] = \frac{E[SW_{\bar{A}} Y I(A=\bar{a})]}{E[SW_{\bar{A}} I(A=\bar{a})]}
    ]

中文：
[
E^{ps}[Y|A=\bar{a}] = \frac{E[SW_{\bar{A}} Y I(A=\bar{a})]}{E[SW_{\bar{A}} I(A=\bar{a})]}
]

For the causal effect:
[
E^{ps}[Y|A=\bar{a}] - E^{ps}[Y|A=\bar{a}']
]

D. Technical Points

  • In dynamic treatments, stabilized weights no longer work and nonstabilized weights are required (Technical Point 21.2).
  • For dynamic strategies: only the denominator of the weight is used; numerator isn’t well-defined.
  • Model misspecification in either weights or g-formula can produce bias; comparing both approaches is a robustness check.
  • Marginal structural mean models (e.g., E[Yaˉ]=β0+β1cum⁡(aˉ)E[Y_{\bar{a}}]=\beta_0+\beta_1 \operatorname{cum}(\bar{a})) are used to reduce model complexity and to estimate overall causal effect.

E. Censoring as a Time-Varying Treatment

When dealing with right-censoring (loss to follow-up), treat censoring as time-varying "treatment" variables C1,...,CK+1C_1, ..., C_{K+1} (0 = not censored, 1 = censored).

Update g-formula and weights:

[
W_{\bar{C}} = \prod_{k=1}^{K+1} \frac{1}{P(C_k=0|C_{k-1}=0, \bar{A}{k-1},\bar{L})}
]

[
\begin{align*}
\text{G-formula with censoring: } & \
E[Y_{\bar{a}, \bar{c}=0}] &= \sum_{\bar{l}} E[Y | \bar{C} = 0, \bar{A}=\bar{a}, \bar{L} = \bar{l}] \prod_{k=0}^K f(l_k | c_k = 0, \bar{a}{k-1}, \bar{l})
\end{align*}
]

中文：

[
W_{\bar{C}} = \prod_{k=1}^{K+1} \frac{1}{P(C_k = 0 | C_{k-1}=0, \bar{A}{k-1}, \bar{L})}
]

带删失的g公式：
[
E[Y_{\bar{a}, \bar{c}=0}] = \sum_{\bar{l}} E[Y | \bar{C} = 0, \bar{A} = \bar{a}, \bar{L} = \bar{l}] \prod_{k=0}^K P(L_k | c_k=0, \bar{A}{k-1}, \bar{L})
]

Censoring IP weights, when used, create a pseudo-population with no censoring.

5. Doubly Robust and Multiply Robust Estimation

A. Doubly Robust (DR) Estimator

For Time-Fixed Treatment

Given

  • Treatment model: f^(A∣L)\hat{f}(A|L)
  • Outcome model: E^[Y∣A,L]\hat{E}[Y|A, L]

Estimator for E[Ya]E[Y^a]:
[
\hat{E}[Y^a] = \frac{1}{n} \sum_{i=1}^n \hat{E}[Y|A=a, L_i] + \frac{I(A_i = a)}{\hat{f}(a | L_i)} (Y_i - \hat{E}[Y|A=a, L_i])
]

中文：
[
\hat{E}[Y^a] = \frac{1}{n} \sum_{i=1}^n \left[ \hat{E}[Y|A=a, L_i] + \frac{I(A_i = a)}{\hat{f}(a | L_i)} (Y_i - \hat{E}[Y|A=a, L_i]) \right]
]

Key property: Consistent if either the treatment model or the outcome model is correct.

For Time-Varying Treatments (Bang & Robins 2005, TMLE):

  • Sequentially fit treatment and outcome models at each time.
  • Construct weights and sequential outcome fits.
  • The estimator is doubly robust (consistent if at least one set of models is correct), and, under certain algorithms/combinations, even multiply robust (consistent in a broader range of scenarios, e.g., if correct for some subset of times).

See Technical Point 21.4/21.5/21.6 for the "K+2"-robust and "2K+1"-robust estimators.

Multiply Robust Estimation (Technical Points 21.4–21.6):

  • Estimators exist (e.g., TMLE) that are consistent if at least any one of several model blocks (either for outcome or treatment at any time) are correct.
  • In practice, fits involve iterated regression and reweighting.

6. G-Estimation for Time-Varying Treatments

A. Structural Nested Mean Models (SNMM)

  • For each time kk, define "blip" functions γk(ak−1,lk,β)\gamma_k(a_{k-1}, l_k, \beta) for causal effect of AkA_k at level ak=1a_k=1.
  • The observed data is related to counterfactual outcomes via:

[
Y_{a_0,0} = Y_{0,0} + \psi_0 a_0
]
[
Y_{a_0,a_1} = Y_{a_0,0} + \psi_{11} a_1 + \psi_{12} a_1 L_{a_0,1} + \psi_{13} a_1 a_0 + \psi_{14} a_1 a_0 L_{a_0,1}
]

Estimation Strategy (G-Estimation):

  • Use observed data to construct candidate "blip-adjusted" outcomes Hk(ψ†)H_k (\psi^\dagger).
  • Solve estimating equations (via regression): the g-estimate ψ^\hat{\psi} is the value for which Hk(ψ†)H_k(\psi^\dagger) yields no dependence of AkA_k on Hk(ψ†)H_k(\psi^\dagger), conditional on prior history.

For multiple time points:

  • General form:
    [
    E[Y_{a_{k-1}, a_k, 0_{k+1}} - Y_{a_{k-1}, 0_k} | L_{a_{k-1},k} = l_k, A_{k-1} = a_{k-1}, A_k = a_k ] = a_k \gamma_k(a_{k-1}, l_k, \beta)
    ]

Key property: G-Estimation avoids certain null paradoxes of the g-formula; more efficient than IP weighting when model specified correctly, but more sensitive to model misspecification.

Relationship to Marginal Structural Models (Technical Point 21.7):

  • SNMM becomes a marginal structural mean model (MSM) if effect modification by past covariate history is absent.
  • G-estimation and IP weighting are dual under MSM, but g-estimation more efficient if the SNMM is correct.

B. Simulation for Dynamic Regimes (Technical Point 21.9)

  • After estimating blips, can simulate outcomes under dynamic strategies by
    1. Simulating histores of confounders under the strategy,
    2. Applying the estimated blip effects.

7. The G-Null Paradox (Technical Point 21.3)

  • Parametric g-formula can yield spurious effects under the null if models for each component are misspecified, despite no true treatment effect.
  • IPW and g-estimation methods do not suffer from this paradox under the sharp null (no treatment effect for any subject).

8. The Big G-Formula

Key Concept

  • If all variables (measured and unmeasured) that are parents of treatment are included, the g-formula based on (A,L,U)(A, L, U) will identify the counterfactual mean for any regime.
  • In practice, unmeasured confounders (UU) can't be included, so question reduces to: can the g-formula be written as a function of observed data only? (i.e., is the effect identifiable).
  • The front-door formula is an example: under certain DAG conditions, the effect is identifiable even when unmeasured confounding exists, via mediators.

Formal Expression (see Technical Point 21.11):
[
E[Y^a] = \sum_m P(M=m|A=a) \sum_{a'} E[Y|M=m, A=a'] P(A=a')
]

中文：
[
E[Y^a] = \sum_m P(M=m|A=a) \sum_{a'} E[Y|M=m, A=a'] P(A=a')
]

  • General results by Tian & Pearl, Shpitser & Pearl: If identification is possible, there exists such a formula.

9. Formal Definition of General SNMM (Technical Point 21.13)

  • The SNMM can be defined relative to any strategy gg.

  • The blip function:
    [
    \gamma_t^g(a_t, l_t) = E[Y_{a_{t-1}, a_t, g_{t+1}} - Y_{a_{t-1}, g_t, g_{t+1}} | A_{t-1} = a_{t-1}, A_t = a_t, L_t = l_t]
    ]

  • Additive SNMMs allow for optimal regime identification (gopt=arg⁡max⁡gE[Yg]g_{\text{opt}} = \arg\max_g E[Y_g]).

总结精要（中文）

核心内容

  • 时间变化处理下传统混杂矫正方法失效，需要g方法（g-formula, IPW, g-estimation）
  • g方法依赖“序列交换性、正则性（positivity）、一致性”三大可识别条件
  • 各种方法可通过参数模型、插件法实现，需注意模型不正确时的偏倚
  • 可以构造双（多）重鲁棒法（即可兼容模型部分错配）
  • 删失（丢失访问）需看作也是时间变化的处理，g方法适用，只需权重或g公式加入删失概率
  • "big g-formula"说明理论上只要所有父节点（测量和未测量）都在g公式里，总是可识别，但实际没法访问未测量变量，是否仅用观测变量可识别由d分离判别

技术结论

  • 只要观测到所有混杂、满足可识别条件，g-formula、IPW和g-estimation结果应等价
  • g-estimation在正确模型下效率更高，IP加权较稳健，对模型偏差更不敏感
  • g公式组件可无单独因果意义，但整体在随机试验下有；IP权重直接描述逆概率下的伪人群
  • 复杂模型（高度维度）可用顺序回归、插件、机器学习等实现多重鲁棒
  • censoring（删失）看作是与A一样的处理变量，统一用g方法解决

参考公式列表

1. g-公式（静态/动态策略）：

[
E[Y_{\bar{a}}] = \sum_{\bar{l}} E[Y|\bar{A} = \bar{a}, \bar{L} = \bar{l}] \prod_{k=0}^K f(l_k | \bar{a}{k-1}, \bar{l})
]

2. IP权重：

[
W_{\bar{A}} = \prod_{k=0}^K \frac{1}{f(A_k|\bar{A}_{k-1}, \bar{L}_k)}
]

[
SW_{\bar{A}} = \prod_{k=0}^K \frac{f(A_k|\bar{A}{k-1})}{f(A_k|\bar{A}, \bar{L}_k)}
]

3. 双鲁棒估计量示例：

[
\hat{E}[Y^a] = \frac{1}{n}\sum_{i=1}^n \hat{E}[Y|A=a, L_i] + \frac{I(A_i = a)}{\hat{f}(a | L_i)} (Y_i - \hat{E}[Y|A=a, L_i])
]

4. SNMM & G-Estimation:

[
E[Y_{a_{k-1}, a_k, 0_{k+1}} - Y_{a_{k-1}, 0_k} | L_{a_{k-1},k} = l_k, A_{k-1} = a_{k-1}, A_k = a_k ] = a_k \gamma_k(a_{k-1}, l_k, \beta)
]

最后建议

  • 如估计结果有显著差异，应检查所有模型，防止系统性偏差。
  • 推荐多种方法并行实现，包含多种模型、鲁棒性加强。
  • 遇到删失数据，要系统地将其纳入g方法权重处理。

This summary preserves all crucial technical points, formulas, and logical structure from the chapter, with bilingual formula expression as requested. For specific implementation or program code (e.g. R or SAS packages), refer to indicated software.