Chapter 28 Controlled Direct Effect

The formulation of mediation analysis in Chapter 27 relies on the nested potential outcomes, and fundamentally, some nested potential outcomes are not observable in any physical experiments. If we stick to the Popperian philosophy of science reviewed in Chapter 140.2, we should only define causal parameters in terms of quantities that are observable under some experiments. This chapter discusses an alternative view of causal inference with an intermediate variable. In this view, we only define the direct effect but can not define the indirect effect.

145 Definition of the controlled direct effect

We view Z and M as two treatment factors that can be manipulated, and define potential outcomes Y(z,m) for z=0,1 and m\in\mathcal{M}. Based on these potential outcomes, we can define the controlled direct effect (CDE) below.

Definition 28.1 (CDE)

Define

\textsc{cde}(m)=E\{Y(1,m)-Y(0,m)\}.

By definition, \textsc{cde}(m) is the average causal effect of the treatment if the intermediate variable is fixed at m. The parameter \textsc{cde}(m) can capture the direct effect of the treatment holding the mediator at m. However, this formulation cannot capture the indirect effect. In particular, the parameter E\{Y(z,1)-Y(z,0)\} only measures the effect of the mediator on the outcome holding the treatment at z. This is not a meaningful definition of the indirect effect.

146 Identification and estimation of the controlled direct effect

To identify \textsc{cde}(m), we need the following assumption, which requires that Z and M are jointly randomized given X.

Assumption 28.1

Sequential ignorability requires

Z\ind Y(z,m)\mid X,\quad M\ind Y(z,m)\mid(Z,X)

or, equivalently (see Problem 27.2),

(Z,M)\ind Y(z,m)\mid X.

I will focus on the case with a binary Z and M. Mathematically, we can just view this problem as an observational study with four treatment levels

(z,m)\in\{(0,0),(0,1),(1,0),(1,1)\}.

The following theorem extends the results for observational studies with a binary treatment, identifying

\mu_{zm}=E\{Y(z,m)\}

based on outcome regression, IPW, and doubly robust estimation.

Define

\mu_{zm}(x)=E(Y\mid Z=z,M=m,X=x)

as the outcome mean conditional on the treatment, mediator, and covariates. Define

\displaystyle e_{zm}(x) \displaystyle= \displaystyle\textup{pr}(Z=z,M=m\mid X=x)
\displaystyle= \displaystyle\textup{pr}(Z=z\mid X=x)\textup{pr}(M=m\mid Z=z,X=x)

as the probability of the joint value of Z and M conditional on the covariates.

Theorem 28.1

Under Assumption 28.1, we have

\mu_{zm}=E\{\mu_{zm}(X)\}=E\left\{\frac{I(Z=z,M=m)Y}{e_{zm}(X)}\right\}.

Moreover, based on the working models e_{zm}(X,\alpha) and \mu_{zm}(X,\beta) for e_{zm}(X) and \mu_{zm}(X), respectively, we have the doubly robust formula

\mu_{zm}^{\textup{dr}}=E\{\mu_{zm}(X,\beta)\}+E\left[\frac{I(Z=z,M=m)\{Y-\mu_{zm}(X,\beta)\}}{e_{zm}(X,\alpha)}\right],

which equals \mu_{zm} if either e_{zm}(X,\alpha)=e_{zm}(X) or \mu_{zm}(X,\beta)=\mu_{zm}(X).

The proof of Theorem 28.1 is similar to those for the standard unconfounded observational studies. Problem 28.2 gives a general result. Based on the outcome mean model, we can obtain \hat{\mu}_{zm}(x) for \mu_{zm}(x). Based on the treatment model, we can obtain \hat{e}_{z}(x) for \textup{pr}(Z=z\mid X=x); based on the intermediate variable model, we can obtain \hat{e}_{m}(z,x) for \textup{pr}(M=m\mid Z=z,X=x). We can then estimate \mu_{zm} by outcome regression

\hat{\mu}_{zm}^{\text{reg}}=n^{-1}\sum_{i=1}^{n}\hat{\mu}_{zm}(X_{i}),

by IPW

\displaystyle\hat{\mu}_{zm}^{\text{ht}} \displaystyle= \displaystyle n^{-1}\sum_{i=1}^{n}\frac{I(Z_{i}=z,M_{i}=m)Y_{i}}{\hat{e}_{z}(X_{i})\hat{e}_{m}(z,X_{i})},
\displaystyle\hat{\mu}_{zm}^{\text{haj}} \displaystyle= \displaystyle\sum_{i=1}^{n}\frac{I(Z_{i}=z,M_{i}=m)Y_{i}}{\hat{e}_{z}(X_{i})\hat{e}_{m}(z,X_{i})}\Big/\sum_{i=1}^{n}\frac{I(Z_{i}=z,M_{i}=m)}{\hat{e}_{z}(X_{i})\hat{e}_{m}(z,X_{i})},

or by augmented IPW

\hat{\mu}_{zm}^{\text{dr}}=\hat{\mu}_{zm}^{\text{reg}}+n^{-1}\sum_{i=1}^{n}\frac{I(Z_{i}=z,M_{i}=m)\{Y_{i}-\hat{\mu}_{zm}(X_{i})\}}{\hat{e}_{z}(X_{i})\hat{e}_{m}(z,X_{i})}.

We can then estimate \textsc{cde}(m) by \hat{\mu}_{1m}^{*}-\hat{\mu}_{0m}^{*}\ (*=\text{reg},\text{ht},\text{haj},\text{dr}) and use the bootstrap to approximate the standard error.

If we are willing to assume a linear outcome model, the controlled direct effect simplifies to the coefficient of the treatment. Example 28.1 below gives the details.

Example 28.1

Under Assumption 28.1 and a linear outcome model,

E(Y\mid Z,M,X)=\theta_{0}+\theta_{1}Z+\theta_{2}M+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X,

we can show that \textsc{cde}(m) equals the coefficient \theta_{1}, which coincides with the natural direct effect in the Baron–Kenny method. I relegate the proof to Problem 28.3.

147 Discussion

The formulation of the controlled direct effect does not involve nested or a priori counterfactual potential outcomes, and its identification does not require the cross-world counterfactual independence assumption. The parameter \textsc{cde}(m) can capture the direct effect of the treatment holding the mediator at m. However, this formulation cannot capture the indirect effect. I summarize the causal frameworks for intermediate variables in Table 12.

Table 12: Causal frameworks for intermediate variables
chapter framework direct effect indirect effect
26 principal stratification \tau(1,1),\ \tau(0,0) ?
27 mediation analysis nde nie
28 controlled direct effect \textsc{cde}(m) ?

The mediation analysis framework can decompose the total effect into natural direct and indirect effects, but it requires nested potential outcomes and cross-world independence. The principal stratification and controlled direct effect frameworks cannot define indirect effects but they do not involve nested potential outcomes and cross-world independence. Moreover, the principal stratification framework does not necessarily require that M lies on the causal pathway from the treatment to the outcome. However, its identification and estimation involves disentangling mixture distributions, which is a nontrivial task in statistics.

148 Homework problems

28.1 cde and nde

Show that under cross-world independence Y(z,m)\ind M(z^{\prime})\mid X for all z,z^{\prime} and m, the conditional controlled direct effect \textsc{cde}(m\mid x)=E\{Y(1,m)-Y(0,m)\mid X=x\} and the conditional natural direct effect \textsc{nde}(x)=E\{Y(1,M_{0})-Y(0,M_{0})\mid X=x\} have the following relationship:

\textsc{nde}(x)=\sum_{m}\textsc{cde}(m\mid x)\textup{pr}(M_{0}=m\mid X=x)

for a discrete M. Without the cross-world independence, does this relationship still hold in general?

28.2 Observational studies with a multi-valued treatment

Theorem 28.1 is a special case of the following theorem for unconfounded observational studies with multiple treatment levels (Imai and Van Dyk,, 2004; Cattaneo,, 2010). Below, I state the general problem and theorem.

Consider an observational study with a multi-valued treatment Z\in\{1,\ldots,K\}, covariates X, and outcome Y. Unit i has K potential outcomes Y_{i}(1),\ldots,Y_{i}(K) corresponding to the K treatment levels. In general, we can define causal effect in terms of contrasts of the potential outcomes:

\tau_{C}=\sum_{k=1}^{K}C_{k}E\{Y(k)\}

where \sum_{k=1}^{K}C_{k}=0. The canonical choice of the pairwise comparison

\tau_{k,k^{\prime}}=E\{Y(k)-Y(k^{\prime})\}.

Therefore, the key is to identify and estimate the means of the potential outcomes \mu_{k}=E\{Y(k)\} under the ignorability and overlap assumptions below based on IID data of (Z_{i},X_{i},Y_{i})_{i=1}^{n}.

Assumption 28.2

Z\ind \{Y(1),\ldots,Y(K)\}\mid X and \textup{pr}(Z=k\mid X)>0 for k=1,\ldots,K.

Define the generalized propensity score as

e_{k}(X)=\textup{pr}(Z=k\mid X),

and define the conditional outcome mean as

\mu_{k}(X)=E(Y\mid Z=k,X)

for k=1,\ldots,K. We have the following theorem.

Theorem 28.2

Under Assumption 28.2, we have

\mu_{k}=E\{\mu_{k}(X)\}=E\left\{\frac{I(Z=k)Y}{e_{k}(X)}\right\}.

Moreover, based on the working models e_{k}(X,\alpha) and \mu_{k}(X,\beta) for e_{k}(X) and \mu_{k}(X), respectively, we have the doubly robust formula

\mu_{k}^{\textup{dr}}=E\{\mu_{k}(X,\beta)\}+E\left[\frac{I(Z=k)\{Y-\mu_{k}(X,\beta)\}}{e_{k}(X,\alpha)}\right],

which equals \mu_{k} if either e_{k}(X,\alpha)=e_{k}(X) or \mu_{k}(X,\beta)=\mu_{k}(X).

Prove Theorem 28.2.

Remark: Theorem 28.1 is a special case of Theorem 28.2 if we view the (Z,M) in Theorem 28.1 as a treatment with four levels. The \textsc{cde}(m) is a special case of \tau_{C}.

28.3 cde in the linear outcome model

Show that under Assumption 28.1, if E(Y\mid Z,M,X)=\theta_{0}+\theta_{1}Z+\theta_{2}M+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X, then

\textsc{cde}(m)=\theta_{1}

for all m; if E(Y\mid Z,M,X)=\theta_{0}+\theta_{1}Z+\theta_{2}M+\theta_{3}ZM+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X, then

\textsc{cde}(m)=\theta_{1}+\theta_{3}m.
28.4 cde in the logistic outcome model

Show that for a binary outcome, under Assumption 28.1, if

\textup{logit}\{\textup{pr}(Y=1\mid Z,M,X)\}=\theta_{0}+\theta_{1}Z+\theta_{2}M+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X,

then

\textsc{cde}(m)=E\{\textup{expit}(\theta_{0}+\theta_{1}+\theta_{2}m+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X)-\textup{expit}(\theta_{0}+\theta_{2}m+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X)\};

if

\textup{logit}\{\textup{pr}(Y=1\mid Z,M,X)\}=\theta_{0}+\theta_{1}Z+\theta_{2}M+\theta_{3}ZM+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X,

then

\textsc{cde}(m)=E\{\textup{expit}(\theta_{0}+\theta_{1}+\theta_{2}m+\theta_{3}m+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X)-\textup{expit}(\theta_{0}+\theta_{2}m+\theta_{4}^{\mkern-1.5mu\mathsf{T}}X)\}.
28.5 Recommended reading

Nguyen et al., (2021) provided a friendly review of of the topics in Chapters 27 and 28.

本书:A First Course in Causal Inference,作者 Peng Ding。原文:arXiv:2305.18793v2,按 CC BY 4.0 许可发布。本站(yuanens.com)把 arXiv 网页版改成分章的纯 HTML,把公式渲染成图片,方便在 Kindle 等电子书阅读器上看,内容没有改动。转换难免有错,以原文为准。本站与作者和出版社无关。