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.
We view
and
as two treatment factors that can be manipulated, and define potential outcomes
for
and
. Based on these potential outcomes, we can define the controlled direct effect (CDE) below.
Define
![]() |
By definition,
is the average causal effect of the treatment if the intermediate variable is fixed at
.
The parameter
can capture the direct effect of the treatment holding the mediator at
. However, this formulation cannot capture the indirect effect. In particular, the parameter
only measures the effect of the mediator on the outcome holding the treatment at
This is not a meaningful definition of the indirect effect.
To identify
, we need the following assumption, which requires that
and
are jointly randomized given
.
I will focus on the case with a binary
and
.
Mathematically, we can just view this problem as an observational study with four treatment levels
![]() |
The following theorem extends the results for observational studies with a binary treatment, identifying
![]() |
based on outcome regression, IPW, and doubly robust estimation.
Define
![]() |
as the outcome mean conditional on the treatment, mediator, and covariates. Define
![]() |
![]() |
![]() |
||
![]() |
![]() |
as the probability of the joint value of
and
conditional on the covariates.
Under Assumption 28.1, we have
![]() |
Moreover, based on the working models
and
for
and
, respectively, we have the doubly robust formula
![]() |
which equals
if either
or
.
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
for
. Based on the treatment model, we can obtain
for
; based on the intermediate variable model, we can obtain
for
. We can then estimate
by outcome regression
![]() |
by IPW
![]() |
![]() |
![]() |
||
![]() |
![]() |
![]() |
or by augmented IPW
![]() |
We can then estimate
by
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.
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
can capture the direct effect of the treatment holding the mediator at
. However, this formulation cannot capture the indirect effect.
I summarize the causal frameworks for intermediate variables in Table 12.
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
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.
Show that under cross-world independence
for all
and
, the conditional controlled direct effect
and the conditional natural direct effect
have the following relationship:
![]() |
for a discrete
. Without the cross-world independence, does this relationship still hold in general?
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
, covariates
, and outcome
. Unit
has
potential outcomes
corresponding to the
treatment levels. In general, we can define causal effect in terms of contrasts of the potential outcomes:
![]() |
where
. The canonical choice of the pairwise comparison
![]() |
Therefore, the key is to identify and estimate the means of the potential outcomes
under the ignorability and overlap assumptions below based on IID data of
.
and
for 
Define the generalized propensity score as
![]() |
and define the conditional outcome mean as
![]() |
for
We have the following theorem.
Under Assumption 28.2, we have
![]() |
Moreover, based on the working models
and
for
and
, respectively, we have the doubly robust formula
![]() |
which equals
if either
or
.
Prove Theorem 28.2.
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。