Chapters 10–12 focused on the identification and estimation of the average causal effect
under the ignorability and overlap assumptions. Conceptually, it is straightforward to extend the discussion to the average causal effects on the treated and control units:
![]() |
![]() |
![]() |
||
![]() |
![]() |
![]() |
If
and
differ from
, then the average causal effects are heterogeneous across the treatment and control groups. Whether we should estimate
,
or
depends on the practical question of interest.
Because of the symmetry, this chapter focuses on
. Chapter 70 also discusses extensions to other estimands.
The average causal effect on the treated units equals
![]() |
where the first term
is directly identifiable from the data and the second term
is counterfactual. The key assumption to identify the second term is the following ignorability and overlap assumptions.
and
.
Because the key is to identify
, we only need the “one-sided” ignorability and overlap assumptions. Under Assumption 13.1, we have the following identification result for
.
Under Assumption 13.1, we have
![]() |
![]() |
![]() |
||
![]() |
![]() |
By Theorem 13.1 the counterfactual mean
equals the conditional mean of the observed outcomes under the control, averaged over the distribution of the covariates under the treatment.
It implies that
is nonparametrically identified by
![]() |
(70) |
With a discrete
, the identification formula in Theorem 13.1 reduces to
![]() |
motivating the following stratified estimator for
:
![]() |
where
is the proportion of category
of
among the treated units.
For continuous
, we need to fit an outcome model for
using the control units. If the fitted values for the control potential outcomes are
, then the outcome regression estimator is
![]() |
If we specify a linear model for all units
![]() |
then
![]() |
![]() |
![]() |
||
![]() |
![]() |
If we run OLS to obtain
, then we can use
to estimate
.
Section 56.2 shows that
is an estimator for
, and this example further shows that
is an estimator for
. This is not surprising because the linear model assumes constant causal effects across units.
The identification formula depends only on
, so we need only to specify a model for the control units. When this model is linear,
![]() |
we have
![]() |
![]() |
![]() |
||
![]() |
![]() |
If we run OLS with only the control units to obtain
, then the estimator is
![]() |
Using the property of the OLS (see 162), we have
![]() |
Therefore, the above estimator reduces to
![]() |
which is similar to () with a different coefficient for the difference in means of the covariates.
As an algebraic fact, we can show that this estimator equals the coefficient of
in the OLS fit of the outcome on the treatment, covariates, and their interactions, with the covariates centered as
. See Problem 13.2 for more details.

Proof of Theorem 13.2: The left-hand side of (71) equals
![]() |
![]() |
![]() |
||
![]() |
![]() |
|||
![]() |
![]() |
The right-hand side of (71) equals
![]() |
![]() |
![]() |
||
![]() |
![]() |
|||
![]() |
![]() |
|||
![]() |
![]() |
So (71) holds.

We have two IPW estimators
![]() |
and
![]() |
where
is the fitted odds of the treatment given covariates.
We also have a doubly robust estimator for
which combines the propensity score and the outcome models. Define
![]() |
(73) |
where 
Under Assumption 13.1, if either
or
, then 
Based on the population versions of
in (73), we can obtain its sample version to construct a doubly robust estimator for
.
)Based on the data
, we can obtain a doubly robust estimator for
by the following steps:
[(1)]
obtain the fitted values of the propensity scores
and then obtain the fitted values of the odds
;
obtain the fitted values of the outcome mean under control
;
construct the doubly robust estimator:
, where
![]() |
By Definition 13.1, we can rewrite
as
![]() |
or
![]() |
Similar to the discussion of
, we can estimate the variance of
via the bootstrap by resampling from
.
Hahn, (1998), Mercatanti and Li, (2014), Shinozaki and Matsuyama, (2015) and Yang and Ding, 2018b () are references on the estimation of
.
The following R function implements two outcome regression estimators, two IPW estimators, and the doubly robust estimator for
. To avoid extreme estimated propensity scores, we can also truncate them from the above.
The following R function further implements the bootstrap variance estimators.
Now we re-analyze the data in Example 10.3 to estimate
.
We obtain
without truncating the estimated propensity scores, and
by truncating the estimated propensity scores from the above at
. The HT estimator is sensitive to the truncation as expected. The regression estimator in Example 13.1 is quite different from other estimators. It imposes an unnecessary assumption that the regression functions in the treatment and control group share the same coefficient of
. The regression estimator in Example 13.2 is much closer to the Hajek and doubly robust estimators. The estimates above are slightly different from those in Section 64.3, suggesting the existence of treatment effect heterogeneity across
and
.
Li et al., 2018a () gave a unified discussion of the causal estimands in observational studies. Starting from the conditional average causal effect
, they proposed a general class of estimands
![]() |
indexed by a weighting function
with
. The normalization in the denominator is to ensure that a constant causal effect
averages to the same
.
Under ignorability,
![]() |
which motivates the outcome regression estimator
![]() |
Moreover, we can show that
has the following weighting form:
Under the ignorability and overlap assumption, we have
![]() |
The proof of Theorem 13.4 is similar to those of Theorems 11.2 and 13.2 which is relegated to Problem 13.9. Based on Theorem 13.4, we can construct the corresponding IPW estimator for
.
By Theorem 13.4, each unit is associated with the weight due to the definition of the estimand as well as the weight due to the inverse of the propensity score. Finally, the treated units are weighted by
and the control units are weighted by
Li et al., 2018a (, Table 1) summarized several estimands, and I present a part of it below:
| population | ![]() |
estimand | weights |
|---|---|---|---|
| combined | ![]() |
![]() |
and ![]() |
| treated | ![]() |
![]() |
1 and ![]() |
| control | ![]() |
![]() |
and 1 |
| overlap | ![]() |
![]() |
and ![]() |
The overlap population and the corresponding estimand
![]() |
is new to us. This estimand has the largest weight for units with
and down weights the units with extreme propensity scores.
A nice feature of this estimand is that its IPW estimator is stable without the possibly extremely small values of
and
in the denominator. If
including the special case of
, the parameter
reduces to
. In general, however, the estimand
may cause controversy because it changes the initial population and depends on the propensity score which may be misspecified in practice. Li et al., 2018a () and Li et al., (2019) gave some justifications and numerical evidence.
This estimand will appear again in Chapter 14.
We can also construct the doubly robust estimator for
. I relegate the details to Problem 13.10.
and 
Assume
. Recall
is the propensity score,
is the marginal probability of the treatment, and
is the CATE.
Show that
![]() |
Remark: The results above also imply that
. So the differences among
depend on the covariance between the propensity score and CATE.
When
and
are uncorrelated, their differences are 0.

This problem provides more details for Example 13.2.
Show that if we center the covariates by
for all units, then
equals the coefficient of
in the OLS fit of the outcome on the intercept, the treatment, covariates, and their interactions.
Chapter 12 ran some simulation studies for
. Run similar simulation studies for
with either correct or incorrect propensity score or outcome models.
You can choose different model parameters, a larger number of simulation settings, and a larger number of bootstrap replicates. Report your findings, including at least the bias, variance, and variance estimator via the bootstrap. You can also report other properties of the estimators, for example, the asymptotic Normality and the coverage rates of the confidence intervals.

Motivated by (73), we have an alternative form of doubly robust estimator for
:
![]() |
Show that under Assumption 13.1,
if either
or
. Give the sample analog of the doubly robust estimator for
.
Assume that
. The individual effect is
and the conditional average causal effect is
. Since we will discuss the individual effect, we do not drop the subscript
since
means the average causal effect, not the population version of
.
Under randomization with
and
, show that
![]() |
is an unbiased predictor of the individual effect in the sense that
![]() |
Further show that
for all
.
Under ignorability with
and
, show that
![]() |
is an unbiased predictor of the individual effect and the conditional average causal effect in the sense that
![]() |
Further show that
for all
.

Assume unconfoundedness.
Show that
if
, and
if
.

Show that
![]() |
Prove Theorem 13.4.
For a given
, we have the following formulas for constructing the doubly robust estimator for
:
![]() |
![]() |
![]() |
||
![]() |
![]() |
![]() |
Show that under ignorability and overlap,
if either
or
, then
;
if either
or
, then
;
if either
or
, then
![]() |
Remark:
Tao and Fu, (2019) proved the above results. However, they hold only for a given
. The most interesting cases of
and
all have weight depending on the propensity score
, which must be estimated in the first place. The above formulas do not apply to constructing the doubly robust estimators for
and
; there does not exist a doubly robust estimator for
.
Revisit Problem 12.5. Estimate
based on the methods introduced in this Chapter.
Shinozaki and Matsuyama, (2015) focused on
, and Li et al., 2018a () discussed general
.
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。