Bayesian causal inference:

a critical review
royalsocietypublishing.org/journal/rsta
Fan Li1 , Peng Ding2 and Fabrizia Mealli3
1 Duke University, Durham, NC, USA
2 University of California, Berkeley, CA, USA
Review 3 University of Florence and EUI, Florence, Italy

Cite this article: Li F, Ding P, Mealli F. 2023 FL, 0000-0002-0390-3673

Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

Bayesian causal inference: a critical review.

Phil. Trans. R. Soc. A 381: 20220153. This paper provides a critical review of the Bayesian
perspective of causal inference based on the
https://doi.org/10.1098/rsta.2022.0153
potential outcomes framework. We review the
causal estimands, assignment mechanism, the general
Received: 28 June 2022 structure of Bayesian inference of causal effects
Accepted: 23 October 2022 and sensitivity analysis. We highlight issues that are
unique to Bayesian causal inference, including the role
of the propensity score, the definition of identifiability,
One contribution of 16 to a theme issue the choice of priors in both low- and high-dimensional
‘Bayesian inference: challenges, perspectives, regimes. We point out the central role of covariate
and prospects’. overlap and more generally the design stage in
Bayesian causal inference. We extend the discussion
Subject Areas: to two complex assignment mechanisms: instrumental
statistics variable and time-varying treatments. We identify the
strengths and weaknesses of the Bayesian approach
Keywords: to causal inference. Throughout, we illustrate the key
concepts via examples.
causal inference, design, ignorability,
This article is part of the theme issue ‘Bayesian
potential outcomes, propensity score inference: challenges, perspectives, and prospects’.

Author for correspondence:

Fan Li
e-mail: [email protected]
1. Introduction
Causality has long been central to the human philoso-
phical debate and scientific pursuit. There are many
relevant questions, e.g. the philosophical meaning
of causation or deducing the causes of a given
phenomenon. Among these questions, statistics—which
concerns measurements—arguably can contribute the
most to the question of measuring the effects of
causes. Statistics infers associations between variables.
Even though the research questions in many statistics-
based studies are causal in nature, a first lesson
in elementary statistics is that association does not
imply causation. Distinguishing between causation and
spurious association between various events is a
challenging task in science. Broadly speaking, statistical
causal inference is about building a framework that (i)

2023 The Author(s) Published by the Royal Society. All rights reserved.
 defines causal effects under general scenarios, (ii) specifies assumptions under which one can
2
identify causation from association and (iii) assesses the sensitivity to the causal assumptions and
finds ways to mitigate.

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
A mainstream statistical framework for causal inference is the potential outcomes framework
[1–3]. Under this framework, following the dictum ‘no causation without manipulation’ [3], a
cause is a pre-specified treatment or intervention that is at least hypothetically manipulable. A
typical causal question is ‘would an individual have a better outcome had he taken treatment
A versus treatment B?’ Causal effects are defined as comparisons of potential outcomes, also
known as counterfactuals, under different treatment conditions for the same units. The main
hurdle to interpreting the association between the treatment and the outcome as a causal effect
is confounding, i.e. the presence of factors that are associated with both the treatment and the
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

outcome. For example, patients with worse health conditions may be more likely to obtain a
beneficial medical treatment; then directly comparing the outcomes of the treated and control
patients, without adjusting for the difference in their baseline health conditions, would bias
the causal comparisons and mistakenly conclude that the treatment is harmful. Randomized
experiments, known as A/B testing in industry or randomized controlled trials in medicine, are the
gold standard for casual inference by eliminating confounding via randomization. But modern
causal inference has increasingly relied on observational data. The potential outcomes framework
provides the basis for identifying and estimating causal effects—quantities defined based on
counterfactuals—from the factual data in the presence of confounding, using randomized or
observational data. This framework is applicable to a wide range of problems in many disciplines
and has been increasingly adopted in the area of machine learning. Other frameworks for causal
inference, including the causal diagram [4] and invariant prediction [5], are beyond the scope of
this review.
There are three primary inferential approaches within the potential outcomes framework [6]:
Fisherian randomization test, Neymanian repeated-sampling evaluation and Bayesian inference.
The first two approaches belong to the Frequentist paradigm and have been dominant, with many
popular tools such as propensity scores, matching and weighting. The Bayesian approach has
several established advantages for general statistical analysis, including automatic uncertainty
quantification, coherently incorporating prior knowledge, and offering a rich collection of
advanced models for complex data. As causal studies increasingly involve real-world big data,
there has been a recent surge of research in Bayesian inference of causal effects [7–11], but there
is no comprehensive appraisal of the current state of the research. This paper aims to fill this gap.
Due to the space limit, we do not intend to provide a catalogue of the existing research on this
topic, but rather discuss the big picture of why and how to conduct Bayesian causal inference
in general settings. We emphasize the unique questions, challenges and opportunities that the
Bayesian approach brings to causal inference. We hope this review can stimulate broader and
deeper cross-fertilization between causal inference and Bayesian analysis.
Section 2 introduces the preliminaries of the potential outcomes framework, and briefly
discusses several Frequentist methods to causal inference. Section 3 outlines the general structure
of Bayesian causal inference, focusing on ignorable treatment assignments at one time point.
Section 4 discusses model specification and implications in high-dimensional settings. Section
5 reviews the role and various uses of the propensity score in Bayesian causal inference. Section 6
outlines sensitivity analysis in observational studies. Section 7 describes two complex assignment
mechanisms: instrumental variable and time-varying treatments. Section 8 concludes.

2. Estimands, identification and frequentist estimation

To convey the main ideas, we focus on the case with a binary treatment at one time period, which
can be readily extended to multiple treatments and multiple time points. Consider a sample
of units drawn from a target population, indexed by i ∈ {1, . . . , N}. Each unit can potentially be
assigned to one of two treatment levels z, with z = 1 for the active treatment and z = 0 for the
control. Let Zi (= z) be the binary variable indicating unit i’s observed treatment status. For unit i,
 a vector of p covariates Xi are observed before the treatment, and an outcome Yi is observed after
3
the treatment. A confounder is a pre-treatment variable that is associated with both the treatment
and the outcome; it can be observed, as a subset of the covariates Xi , or unobserved. Below we use

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
covariates and confounders interchangeably. We use the A ⊥⊥ B | C notation to denote conditional
independence between two variables A and B given variable C [12]. We also use the bold font to
indicate a vector consisting of the corresponding variables for the N units, e.g. Z = (Z1 , . . . , ZN ) .
We maintain the standard stable unit treatment value assumption (SUTVA) [13], namely,
there is (i) no different version of a treatment, and (ii) no interference in the sense that one
unit’s potential outcomes are not affected by other units’ treatment assignment. Under SUTVA,
each unit i has two potential outcomes: Yi (1) and Yi (0). Causal effects are contrasts of potential
outcomes under different treatment conditions for the same set of units. The individual treatment
effect (ITE) for unit i is τi = Yi (1) − Yi (0). Averaging τi over a sample we obtain the sample average
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025


treatment effect (SATE): τ S ≡ N−1 N i=1 τi . Furthermore, the conditional average treatment effect
(CATE) is the average of the individual treatment effect of all units with the covariate value x:

τ (x) ≡ E{Yi (1) − Yi (0) | Xi = x} = μ1 (x) − μ0 (x), (2.1)

where μz (x) ≡ E{Yi (z) | Xi = x} for z = 0, 1. Averaging τi or τ (Xi ) over a target population gives the
population average treatment effect (PATE):

τ P ≡ E{Yi (1) − Yi (0)} = E{τ (Xi )}. (2.2)

The PATE is a function of the distribution of the potential outcomes in a population, whereas
the SATE is a function of the potential outcomes themselves. The subtle distinction in their
definitions leads to important differences in inferential and computational strategies, as will
be discussed later. Traditionally, the SATE is of interest in randomized experiments where the
target population is the specific sample, whereas the PATE is of interest in observational studies
where the target population is the population from which the sample is drawn. In general, the
choice of a causal estimand is determined by the scientific question in hand rather than statistical
considerations. Note that although both the ITE and CATE are important in characterizing
treatment effect heterogeneity, they are obviously different; however, these two estimands are
sometimes conflated in the literature.
The fundamental problem of causal inference [14] is that, for each unit only the potential
outcome corresponding to the actual treatment, Yiobs ≡ Yi = Yi (Zi ), is observed or factual, and
the other potential outcome, Yimis = Yi (1 − Zi ), is missing or counterfactual. Therefore, additional
assumptions are necessary to identify the causal effects. The key identifying assumptions concern
the assignment mechanism, i.e. the process that determines which units get what treatment and
hence which potential outcomes are observed or missing [15]. The vast majority of causal studies
assume certain versions of an ignorable assignment mechanism, where the treatment assignment is
independent of the potential outcomes conditional on some observed variables. Specifically, in
the simple setting of a binary treatment at one time, ignorability consists of two sub-assumptions
[15,16].
Assumption 2.1. (Ignorability). (a) Unconfoundedness. Pr{Zi | Yi (0), Yi (1), Xi } = Pr(Zi | Xi ), or
equivalently Zi ⊥⊥ {Yi (0), Yi (1)} | Xi . (b) Overlap. 0 < e(Xi ) < 1 for all i, where e(x) ≡ Pr(Zi = 1 | Xi = x)
is the propensity score [16].
The unconfoundedness assumption states that there is no unmeasured confounding, and the
overlap assumption states that each unit has non-zero probability of being assigned to each
treatment condition. These two assumptions together ensure that the conditional distribution of
the potential outcomes is identifiable from observed data as

μz (x) ≡ E{Yi (z) | Xi = x} = E(Yi | Zi = z, Xi = x), for all z, x. (2.3)

Therefore, the CATE is identified as τ (x) = μ1 (x) − μ0 (x), and the PATE is identified as τ P =
E{μ1 (Xi ) − μ0 (Xi )}. This underlines the estimation strategy of outcome modelling: we can specify
a model for the outcome function μz (x), and estimate the CATE by  τ (x) = 
μ1 (x) − 
μ0 (x), and the
 
PATE by  τ reg = N−1 N i=1 {
μ1 (Xi ) − μ0 (Xi )}, where  μz (x) is the estimated outcome model from the 4
observed data.
In randomized experiments, the treatment assignment is known and controlled by the

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
experimenters, and ignorability holds by design. In observational studies, the treatment
assignment is unknown and uncontrolled, and ignorability at best holds approximately. A key
concept in causal inference is overlap and balance, which refers to the similarity in the distributions
of the covariates between the comparison groups. In general, as the two groups become
more balanced, the causal estimates become less sensitive to the estimate strategy and model
specification. In the ideal case of a randomized experiment, all—measured and unmeasured—
covariates are balanced in expectation, i.e. they have the same multivariate distribution
in the two treatment arms. Consequently, the simple difference-in-means estimator,  τ=
 N N N
{ N
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

i=1 Y i Z i / i=1 Z i } − { i=1 Y i (1 − Z i )/ i=1 (1 − Z i )}, is unbiased for τ S

and τ P
; furthermore,
even a misspecified linear outcome model leads to a consistent estimate of τ S [17]. On the contrary,
in observational studies, the two groups are often imbalanced in many covariates, e.g. patients
receiving the treatment may be generally sicker and older than those receiving the control. In
such cases, directly comparing the difference in the outcomes between the two groups would
give biased estimates of the causal effects. Moreover, the fit of an outcome model would rely
on extrapolation in the regions where the two groups are poorly overlapped, and consequently
outcome-model-based estimators, such as  τ reg , are sensitive to the model specification. Therefore,
a main effort in causal inference with observational data is to ensure overlap and balance to mimic
a randomized experiment as closely as possible. This process does not involve the outcome and
is referred to as the design stage, in contrast to the analysis stage, which utilizes the outcome and
estimates causal effects given the design stage [18]. A causal analysis of an observational study
usually has both design and analysis stages, in parallel with those of a randomized controlled
experiment.
The propensity score plays a central role in causal inference with observational data, owing
to its two special properties [16]. First, the propensity score is a balancing score in the sense that
Zi ⊥⊥ e(Xi ) | Xi . This means that balancing the scalar propensity score balances the multivariate
distribution of the covariates. Second, if a treatment assignment is unconfounded given Xi , then
it is unconfounded given e(Xi ), that is, Zi ⊥⊥ {Yi (0), Yi (1)} | Xi implies Zi ⊥⊥ {Yi (0), Yi (1)} | e(Xi ). In
observational studies, e(Xi ) is usually unknown and needs to be estimated, e.g. via a logistic
regression model of the treatment on the covariates.
The propensity score is usually used with matching, weighting, or stratification to achieve
balance and estimate causal effects. Specifically, matching methods use a certain algorithm to
find pairs of units in the two groups with similar covariates according to a distance metric,
e.g. the propensity score or the Mahalanobis distance, and then calculate the difference in
the average observed outcome between the groups in the matched sample [19–21]. Weighting
methods assign a weight to each unit, so that the weighted distribution of the covariates in
the two groups are balanced [22], and then calculate the weighted difference in the outcomes
between the two groups. An important weighting scheme is inverse probability weighting (IPW),
based on the identification formula of the PATE: τ P = E{Zi Yi /e(Xi ) − (1 − Zi )Yi /(1 − e(Xi ))}. A

corresponding IPW estimator [23] is  τ ipw = N−1 N i=1 {Zi Yi /
e(Xi ) − (1 − Zi )Yi /(1 − e(Xi ))}, where

e(Xi ) denotes the estimated propensity score for unit i. One can further augment the IPW
estimator by an outcome model to obtain a semiparametric efficient estimator [24]:  τ dr = τ reg +
−1 N
N i=1 {Zi Ri /e(Xi ) − (1 − Zi )Ri /(1 − e(Xi ))}, where Ri = Yi −  μZi (Xi ) is the residual from the
outcome model. The IPW estimator  τ ipw is consistent for τ if the propensity score model is correct,
and the outcome-model estimator  τ reg is consistent if the outcome model is correct. Because the
bias of the estimator  τ is a product of the residual of the propensity score model and that of
dr

the outcome model,  τ dr is doubly robust in the sense that it is consistent if either the propensity
score or the outcome model, but not necessarily both, is correctly specified [25]. Despite the
seemingly different construction, matching estimators, with proper mathematical formulations,
can be viewed as non-parametric versions of  τ ipw , τ reg and τ dr based on nearest-neighbour
regressions [26]. These are the main Frequentist estimation strategies for τ P under ignorability.
 When the target estimand is the CATE, the primary estimation strategy is outcome modelling.
5
We will discuss how to specify the outcome model in §4(a).

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
3. General structure of Bayesian causal inference
(a) Basic factorization and versions of causal estimands
Because of the unavoidable missing potential outcomes, causal inference under the potential
outcomes framework is inherently a missing data problem [6,15]. The Bayesian paradigm offers
a unified framework for statistical inference with missing data and thus for causal inference [27].
Below we review the general structure of Bayesian causal inference that was first outlined in [15].
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

Four quantities are associated with each unit i, {Yi (0), Yi (1), Zi , Xi }, where {Zi , Xi , Yi (Zi )} are
observed but Yi (1 − Zi ) is missing. Bayesian inference views all these quantities as random
variables and centres around specifying a model for them. Based on the Bayesian model, we can
draw inference on causal estimands—functions of the model parameters, covariates and potential
outcomes—from the posterior predictive distributions of the parameters and the unobserved
potential outcomes. Specifically, we assume the joint distribution of these random variables of
all units is governed by a parameter θ = (θX , θZ , θY ), conditional on which the random variables
for each unit are i.i.d.. Then we can factorize the joint density Pr{Yi (0), Yi (1), Zi , Xi | θ} for each unit
i as
Pr{Zi | Yi (0), Yi (1), Xi ; θZ } · Pr{Yi (0), Yi (1) | Xi ; θY } · Pr(Xi ; θX ). (3.1)

The three terms in (3.1) represent the model for the assignment mechanism, potential outcomes,
and covariates, respectively. Under ignorability, the assignment mechanism further reduces to the
propensity score model Pr(Zi | Xi ; θZ ).
Before diving into the technical details, we first clarify the subtle but important difference
between the Bayesian estimation of the PATE and SATE estimands. For the PATE, we rewrite the

outcome-model-based identification formula in §2 as τ P = {μ1 (x; θY ) − μ0 (x; θY )}F(dx; θX ), which
depends only on the unknown parameters θX and θY . Therefore, Bayesian inference for the PATE
requires obtaining posterior distributions of (θX , θY ). By contrast, the SATE τ S is a function of
the potential outcomes {Yi (0), Yi (1)}Ni=1 , which involves both observed and missing quantities.
Bayesian inference for the SATE requires imputing the missing potential outcomes Yimis from
their posterior predictive distributions based on the outcome model, and consequently deriving
the posterior distribution of τ S .
However, in practice, we rarely model the possibly multi-dimensional covariates Xi , and
instead condition on the observed values of the covariates. This is equivalent to replacing F(x; θX )
with  FX , the empirical distribution of the covariates. Therefore, most Bayesian causal inference
(e.g. [9,28]) in fact focuses on the mixed average treatment effect (MATE) [6]
 
N
τ M ≡ τ (x; θY )
FX (dx) = N−1 τ (Xi ; θY ), (3.2)
i=1

where τ (x; θY ) = τ (x) highlights the dependence on the parameter θY . The MATE is a convenient
approximation of the PATE and is particularly natural under the Bayesian paradigm. The
difference between the MATE and SATE is subtle: the former equals the average of the CATE
whereas the latter equals the average of the ITEs over the finite sample. Based on the posterior
distributions, the PATE has the largest uncertainty, whereas the SATE has the smallest uncertainty.
The distinction between these estimands is illustrated in the following example.

Example 3.1. [Covariate adjustment in a randomized experiment] Consider a completely

randomized experiment with covariates X. Assume the true model for potential outcomes is
     
Yi (1) 2 2 β1 Xi σ12 ρσ1 σ0
| (Xi , β1 , β0 , σ1 , σ0 , ρ) ∼ N , , i = 1, . . . , N.
Yi (0) β0 Xi ρσ1 σ0 σ02
 This model implies two univariate normal marginal models: Yi (z) | Xi , βz , σz2 ∼ N (βz Xi , σz2 ) for
6
z = 0, 1. In this example, the CATE is τ (x) = (β1 − β0 ) x; the PATE, SATE and MATE are
⎫
τ P = (β1 − β0 ) E(Xi ), ⎪

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

⎪

...............................................................
⎪
⎪
⎪
⎪
N ⎬
−1
τ =N
S
{Yi (1) − Yi (0)}, (3.3)
⎪
⎪
i=1 ⎪
⎪
⎪
⎪
⎭
τ M = (β1 − β0 ) X̄,

respectively, where X̄ = N−1 N i=1 Xi is the sample mean of the covariates.

(b) Posterior inference of causal effects

Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

Regardless of the version of the target estimand, the following assumption is commonly adopted.

Assumption 3.2. (Prior independence). The parameters for the models of assignment mechanism θZ ,
outcome θY , and covariates θX are a priori distinct and independent.

Assumption 3.2 imposes independent prior distributions for parameters (θX , θZ , θY ). It is

unique to the Bayesian paradigm of causal inference. It is imposed primarily for modelling
and computational convenience and may appear innocuous. However, as elaborated in §4(b),
it may lead to unintended and undesirable implications in high-dimensional problems. Under
assumptions 2.1 and 3.2, the joint posterior distribution of θ = (θX , θZ , θY ) and the missing
potential outcomes is proportional to
N N N
Pr(θX ) Pr(Xi ; θX ) · Pr(θZ ) Pr(Zi | Xi ; θZ ) · Pr(θY ) Pr{Yi (1), Yi (0) | Xi ; θY }. (3.4)
i=1 i=1 i=1

From (3.4), the posterior distributions of θX and θY , and consequently of τ P , do not depend on the
second component corresponding to the propensity score. Therefore, the propensity score model
is ignorable in Bayesian inference of τ P . The same ignorability argument applies to other estimands
such as τ S , τ M and τ (x) [6,9,15]. Furthermore, inference of τ M does not depend on the covariate
model Pr(Xi ; θX ). Because of this, it is essential to specify the outcome model Pr{Yi (1), Yi (0) | Xi ; θY }
in Bayesian causal inference.
By definition, τ P = E{Yi (1)} − E{Yi (0)} does not depend on the association between Yi (0) and
Yi (1), denoted by the parameter ρ. Similarly, τ (x) does not depend on ρ, but τ S does. So in
the inference of τ P and τ (x), we usually directly specify the marginal models Pr{Yi (z) | Xi ; θY }
or equivalently Pr(Yi | Zi = z, Xi ; θY ) under ignorability [28]. The observed-data likelihood based
on (3.4) becomes i:Zi =1 Pr(Yi | Zi = 1, Xi ; θY ) i:Zi =0 Pr(Yi | Zi = 0, Xi ; θY ). Imposing a prior for θY ,
we can proceed to infer θY and subsequently τ P , τ M , or τ (x) using the usual Bayesian inferential
procedures.
Bayesian inference of τ S is more complex, because it depends on both Yi (0) and Yi (1) and
thus requires posterior sampling of both θY and Ymis . The most common sampling strategy is
through data augmentation: iteratively simulate θ and Ymis given each other and the observed
data, namely from Pr(θY | Ymis , Yobs , Z, X) and Pr(Ymis | Yobs , Z, X; θY ). The former, given the
observed data and the imputed Ymis , can be straightforwardly obtained by a complete-data
analysis based on Pr{θY | Y(1), Y(0), X} ∝ Pr(θY ) N i=1 Pr{Yi (1), Yi (0) | Xi ; θY }. The latter requires
more elaboration. Specifically, we can show that Pr(Ymis | Yobs , Z, X; θY ) is proportional to
i:Zi =1 Pr{Yi (0) | Yi (1), Xi ; θY } i:Zi =0 Pr{Yi (1) | Yi (0), Xi ; θY }. This renders that imputing the Y
mis

depends crucially on the joint distribution of {Yi (1), Yi (0)}. Because Yi (0) and Yi (1) are never jointly
observed, the data provide no information about their association ρ. Unless the specific marginal
model places constraints on ρ, the posterior distribution of ρ would be the same as its prior.
Consequently, the posterior distribution of τ S would be sensitive to the prior of ρ.
The above discussion prompts us to clarify the notion of identifiability in Bayesian inference.
Under the Frequentist paradigm, a parameter is identifiable if any of its two distinct values
 give two different distributions of the observed data. Under the Bayesian paradigm, there is
7
no consensus. For example, Lindley [29] argued that all parameters are identifiable in Bayesian
analysis because with proper prior distributions, posterior distributions are always proper. In

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
this sense, ρ is identifiable. However, due to the fundamental problem of causal inference, there
is no information in the data on ρ and it is reasonable to label it as non-identifiable. This is
distinct from the parameters that the data provide direct information on, e.g. those in the marginal
distributions of the outcomes in each arm, which are reasonable to label as identifiable. Lindley’s
perspective of identifiability blurs such distinction. A more informative perspective is provided
by Gustafson [30], who argued that a parameter is weakly or partially identifiable, if a substantial
region of its posterior distribution is flat, or its posterior distribution depends crucially on its
prior distribution even with large samples, such as ρ. Another example of a partially identifiable
parameter is Pr{Yi (1) > Yi (0)}, which depends on ρ [6,31]. In this perspective, identifiability in
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

Bayesian inference is no longer all-or-nothing; instead, it is a continuum in between. This issue

motivates the strategy of transparent parameterization, where one separates identifiable and non-
identifiable parameters, and treats the latter as sensitivity parameters in a sensitivity analysis
[32–36]. More discussion will be given in §6(b).
Example 1 revisited We now illustrate the posterior inference of the causal estimands in
example 3.1. Here, the parameters β’s and σ ’s are identifiable, but ρ is not in the Frequentist
sense. We fit a Bayesian linear regression model of Yi on Xi to each observed arm z, with
independent priors on (β1 , σ12 ) and (β0 , σ02 ). The observed likelihood factorizes into two parts:
the data in treatment group {(Xi , Yi ) : Zi = 1} and control group {(Xi , Yi ) : Zi = 0} contribute to
the likelihood of (β1 , σ12 ) and (β0 , σ02 ), respectively. For example, imposing the conventional
conjugate normal-inverse χ 2 priors, we can draw from the posterior distribution of β and σ ,
and thus that of the MATE by plugging the posterior draws into the closed-form of τ M in (3.3).
To obtain the PATE, we would have to specify a multivariate model for Pr(X; θ), and derive the
posterior distribution of θX and then plug it into the closed form of τ P in (3.3). This can also
be implemented, e.g. via a Bayesian bootstrap step without a model, as described in the next
paragraph. To obtain the SATE, we could specify a prior for ρ or fix it to a value. Given ρ and
each draw of (β1 , β0 , σ12 , σ02 ), we can impute Yimis as follows: for treated units, Yimis = Yi (0), and we
draw Yi (0) from N (β0 Xi + ρσ0 /σ1 · (Yi − β1 Xi ), σ02 (1 − ρ 2 )); for control units, Yimis = Yi (1), and we
draw Yi (1) from N (β1 Xi + ρσ1 /σ0 · (Yi − β0 Xi ), σ12 (1 − ρ 2 )). Plugging these posterior predictive
draws of Yimis and the observed outcomes into the definition of τ S , we obtain its posterior
distribution. We suggest varying the sensitivity parameter ρ from 0 to 1, which corresponds
to conditionally independent potential outcomes and perfectly correlated potential outcomes,
respectively.
An interesting alternative Bayesian strategy is through the Bayesian bootstrap [37], where the
units are re-weighted with weights drawn from a Dirichlet distribution. The Bayesian bootstrap
is a general strategy to simulate the posterior distribution of a parameter under a non-parametric
model, which can be viewed as the limit of the inference under the Dirichlet Process prior
[38]. This renders the Bayesian bootstrap relevant to causal inference in at least two ways.
First, one can generate posterior samples from the distribution of Pr(Xi ; θX ) without specifying
a parametric model. This is desirable in inferring the population estimands like the PATE and
the CATE [39]. However, how to integrate these samples of X into the inference of the target
causal estimand is case-dependent and generally adds complexity to the analysis compared to the
MATE. Second, the Bayesian bootstrap offers a general recipe for incorporating many standard
Frequentist procedures into Bayesian inference. For example, Taddy et al. [40] used it to quantify
the uncertainty in linear and tree-based methods for estimating the CATE. Chamberlain & Imbens
[41] used it in M-estimation with an application to the setting of instrumental variables (see §7(a)).
However, we view the Bayesian bootstrap approach as peripheral to Bayesian causal inference
because it does not capitalize on arguably the main strength of Bayesian inference, namely, a
unified inferential framework underpinned by the Bayes theorem with versatile choice of priors
and outcome models.
 4. Model specification 8

(a) Common specification of the outcome model

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
Section 3 shows that the central component in Bayesian inference of the CATE, PATE, and MATE
is to specify the outcome model μz (x) = E(Yi | Xi = x, Zi = z; θY ). We can either model the two
treatment groups jointly with a single function μz (x) = μ(z, x) or model each group separately
with two functions μ1 (x) and μ0 (x), known as S-learner or T-learner, respectively, in the literature
[42]. The most basic outcome model is a linear regression: μ(z, x) = x + z + xz, with Gaussian
error terms, where the treatment-covariate interaction term xz captures the treatment effect
heterogeneity. This model is equivalent to specifying a linear regression in each group. But the
equivalence does not hold for nonlinear models; in fact, S-learners and T-learners with the same
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

type of nonlinear models often lead to markedly different causal estimates.

Linear regressions are easy to implement and interpret, but they are often too restricted. In
real-world problems, it is crucial to specify μ(z, x) flexibly enough to approximate the possibly
complex underlying true data generating mechanism. This is particularly desirable as the recent
focus in causal inference has been moving toward heterogeneous treatment effects. Outcome
modelling is the most natural approach in these studies: one can simply specify an outcome
model and derive the CATE as a function of the model parameters. There has been a rapidly
increasing adoption of non-parametric and machine learning models for μ(z, x). One of the most
widely used such models is based on regression trees. At a high level, regression trees partition
the covariate space into non-overlapping regions and the prediction in each region is based solely
on the data that fall in that region. The parameters of a regression tree characterize where to split
the covariate space and how to predict the outcomes in a terminal node [43]. An ensemble of
regression trees—usually referred to as forests—are often combined to improve the prediction.
Within the Bayesian paradigm, the Bayesian Additive Regression Tree (BART) [44] has become
very popular for causal inference. BART places certain priors on the parameters of the regression
trees to control the depth of the tree and the degree of shrinkage of the mean parameters in
terminal nodes. Hill [9] first advocated using BART to specify the outcome model μ(z, x) in an
S-learner. One can also specify a T-learner with a separate BART model for each treatment group
μz (x). However, without any additional structure on the marginal models μz (x), T-learners often
result in large variance of the treatment effects. Hahn et al. [8] proposed the Bayesian Causal Forest
method based on an alternative reparametrization, μ(z, x) = g1 (x) + g2 (x)z, where g1 (x) models
the distribution of Y(0) and g2 (x) represents the heterogeneous treatment effect, with a separate
BART prior for g1 (x) and g2 (x). The BART models have a number of advantages, including fast
computation, good performance of default choice of hyperparameters and available software.
When a study has adequate covariate overlap, BART has been shown to outperform numerous
competing methods, including (the Frequentist) random forests, in many empirical applications,
e.g. [45,46]. Other Bayesian non-parametric models, such as Gaussian process [47], Dirichlet
process [48–51], have also been considered for causal inference. We refer interested readers to
[10] for a more detailed review of these methods.

(b) Challenges in high dimensions

Conducting statistical inference in high dimensions is challenging in general. We differentiate
between two high-dimensional settings: (i) an outcome model with an infinite or a large
number of parameters, regardless of the number of covariates, such as non-parametric and
semiparametric models, and (ii) a large number of covariates. Both settings are increasingly
common in causal inference, particularly in studies targeting the CATE. As discussed earlier,
outcome modelling is the primary method in these settings, and Bayesian non-parametric models
have become a mainstay of the model choice.
A straightforward application of the standard Bayesian non-parametric priors to outcome
modelling is sometimes inadequate for causal inference, even with low dimensional covariates.
 (a)
linear GP BART 9
10

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
0

outcome
–10

–20
(b)
25
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

20

CATE
15

10

25 50 75 25 50 75 25 50 75
X

Figure 1. Example 4.1: estimates of means of the potential outcomes (a) and the CATE (b) and corresponding uncertainty band
as a function of the single covariate by linear model, Gaussian Process, BART, respectively. Cross symbols: treated units; circles:
control units. (Online version in colour.)

An important consideration is that a desirable prior should accurately reflect uncertainty

according to the degree of covariate overlap because intuitively the uncertainty of causal
estimates should increase as the degree of overlap decreases. Below, we reproduce a simple
example in [52] (first constructed by Surya Tokdar) with a single covariate to illustrate this point.

Example 4.1. [Choice of priors in estimating the CATE] Consider a study with 250 treated and
250 control units. Each unit has a single covariate X that follows a Gamma distribution with
mean 60 and 35 in the control (Zi = 0) and treatment (Zi = 1) group, respectively, and with s.d. 8
in both groups. To convey the main message, we consider a true outcome model with constant
treatment effects: Yi (z) = 10 + 5z − 0.3Xi + i with i ∼ N (0, 1), where the CATE τ (x) = 5 for all x.
The scatterplots in the upper panel of figure 1 show that covariate overlap is good between the
groups in the middle of the range of X (around 40 to 50), but deteriorates towards the tails of X.
To estimate the CATE, we fit an outcome model separately in each group: μ(z, x) = fz (x) + i
with i ∼ N (0, σ 2 ). We choose three priors for fz (x): (i) a linear model fz (x) = αz + βz x with a
Gaussian prior for the coefficients; (ii) a BART prior similar to [9]; (iii) a Gaussian Process
prior [53] with the covariance function specified by a Gaussian kernel with signal-to-noise
ratio parameter ρ and inverse-bandwidth parameter λ: (fz (x1 ), fz (x2 ), . . . , fz (xN )) ∼ N (0, Σ) where
Σij = δ 2 ρ 2 exp{−λ2 ||xi − xj ||2 }). Figure 1 shows the posterior means of μz (X) and the CATE,
with corresponding uncertainty band as a function of X. Here, we focus on the uncertainty
quantification. In the region of good overlap, all three models lead to similar points and credible
interval estimates of the CATE, but a marked difference emerges in the region of poor overlap.
The linear model appears overconfident in estimating the CATE. The Gaussian process trades
potential bias with wider credible interval as overlap decreases and produces a more adaptive
uncertainty quantification. BART produces shorter error bars than the Gaussian Process (but
wider than linear models), but the width of the credible interval remains similar regardless of
the degree of overlap and thus is over confident in the presence of poor overlap.
 Example 4.1 illustrates that, even with low dimensional covariates, standard Bayesian
10
priors can have markedly different operating characteristics when the two groups are poorly
overlapped, and not all priors can adaptively capture the uncertainty according to the degree

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
of overlap. A primary reason for BART potentially underestimating uncertainty in poor overlap
is its lack of smoothness, in contrast to the Gaussian process. Nonetheless, such a problem can be
mitigated by soft decision trees as in [54].
When there are a large number of covariates, the Bayesian paradigm usually achieves
regularization through sparsity-inducing priors for the outcome model, such as the spike-and-
slab prior [55], the Bayesian LASSO [56], as well as the model averaging techniques [57–59].
The use of these methods in causal inference is surveyed in [10,50]. High-dimensional covariates
pose additional complications to Bayesian causal inference. Robins & Ritov [60] pointed out that
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

non-parametric estimates often have slow convergence rates in this regime, which translates into
poor finite-sample performance. A central challenge in causal inference is that covariate overlap
is rapidly diminishing as the covariates dimension increases, violating the overlap assumption
that underpins standard causal analysis [61]. Lack of overlap exacerbates the usual inferential
challenges—such as sparsity and slow convergence—in high-dimensional analysis. Even if we
assume linear outcome models, we must carefully specify the priors on the regression coefficients.
For example, Hahn et al. [7] showed that standard Bayesian regularization on the nuisance
parameters may indirectly regularize important causal parameters and thus induce bias, namely
the regularization induced confounding. This issue was rigorously investigated in [62]. Specifically,
Linero [62] defines the selection bias as δz = E(Yi | Zi = z) − E{Yi (z)}, and showed that, under the
seemingly innocuous prior independence assumption 3.2, many Bayesian regularization priors
would a priori induce the selection bias δz to sharply concentrate around zero as the number
of covariates, p, increases, to the extent that no amount of data would overcome such a bias.
This implies that assumption 3.2 effectively acts as a strongly informative prior as p increases.
Such a phenomenon is referred to as prior dogmaticism and is the Bayesian analogue of the
aforementioned problem in Ritov et al. [63]. This line of research highlighted the importance
of incorporating the propensity score in Bayesian causal inference [7,62,64,65], which echos the
insights from the Frequentist double machine learning method [66,67]. Specifically, the regularized
propensity score model or outcome model alone would not be sufficient for valid causal inference,
but combining the two would achieve desirable convergence rate and finite sample performance
in high-dimensional causal analysis.

5. The role of the propensity score

A major debate in Bayesian causal inference is the role of the propensity score, which characterizes
the assignment mechanism. On the one hand, as shown in §3, under assumptions 2.1 and 3.2,
the propensity score drops out from the likelihood and thus its value appears to be irrelevant
in Bayesian causal inference, which seemingly only involves the outcome model and thus the
analysis stage. On the other hand, §2 shows that the propensity score is ubiquitous in the
Frequentist approach to causal inference, e.g. in constructing weighting, matching and doubly-
robust estimators. Regardless of the mode of inference, the propensity score is essential in
ensuring overlap and balance in the design stage of an observational study, which consequently
reduces the sensitivity to the outcome model specification. Such sensitivity reduction is key to
robust Bayesian causal inference, which is primarily based on outcome modelling. The literature
has recognized the importance of incorporating the propensity scores into Bayesian causal
inference, either in the design or the analysis stage, but there is no consensus on how to proceed.
Below we review three existing strategies.

(a) Include the propensity score as a covariate in the outcome model

The propensity score was first proposed to be included as the only covariate in a Bayesian
outcome model under ignorability, which would reduce the model complexity [68]. However,
 as later pointed out by Zigler [59]: Pr{Y(z) | X, e(X)} = Pr{Y(z) | X} = Pr{Y(z) | e(X)}. So using the
11
propensity score as the single covariate in the outcome model would not lead to the target
outcome distribution Pr{Y(z) | X}, but using it as an additional covariate, i.e. specifying a model

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
μ(z, x) = μ(z, x, e(x)), would. This specification is effectively conducting an outcome regression at
each value of the propensity score, and thus can be viewed as a smoothed version of combining
propensity score stratification and outcome modelling. In a sense, this specification provides a
Bayesian analogue of the double robustness [52,69]. On the one hand, when the outcome model
is correctly specified, μ(Zi , Xi , e(Xi )) reduces to μ(Zi , Xi ) because e(Xi ) is a function of Xi and
thus is redundant regardless of its specification. On the other hand, when the outcome model is
misspecified but the propensity model is correctly specified, the results are robust to the outcome
model specification because the treatment and control groups are approximately balanced in
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

covariates within each stratum of the propensity score. Various reparametrizations have been
proposed. One example is to specify μ(z, x, e(x)) = g1 (x, e(x)) + g2 (z, x), with g1 (·) being a non-
parametric model and g2 (·) being a parametric model. Little [70] adopted a penalized spline
model of e(x) for g1 (·). In the aforementioned Bayesian Causal Forest, Hahn et al. [8] imposed a
separate BART model for g1 (·) and g2 (·), and demonstrated that adding the propensity score as an
additional predictor in g1 significantly improves the empirical estimation of the CATE.
This strategy is usually implemented in two stages: first estimate the propensity score as e(X)
and then plug it into the Bayesian outcome model μ(Z, X, e(X)). Such a two-stage procedure
is not dogmatically Bayesian, which generically refers to the procedure of specifying a model
with parameters and prior distributions of these parameters and then use the Bayes theorem to
obtain the posterior distributions of the parameters. A direct consequence is that this procedure
may not properly propagate the uncertainty of estimating the propensity score in the outcome
model [69]. A dogmatic Bayesian approach would jointly model e(X; θZ ) and μ(Z, X, e(X); θY ) and
draw posterior inference of θZ and θY simultaneously [71]. However, when the outcome model
is misspecified, the joint-modelling approach would introduce a feedback problem, that is, the
fit of the outcome model would inform the estimation of the propensity scores. This violates
the unconfoundedness assumption, distorts the balancing property of the propensity score, and
consequently leads to biased estimate of causal effects. A suggested remedy is to first fit a Bayesian
model for e and then plug the posterior predictive draws of e into the outcome [11]. Such a two-
stage procedure is still not dogmatically Bayesian, but provides more robust posterior inference
to model misspecification empirically.
However, adding the propensity score into the outcome model is controversial conceptually,
because the outcome model reflects the nature of the generating process of the potential outcomes,
which arguably should not depend on how the treatment is assigned [72].

(b) Dependent priors

The Bayesian causal inference outlined in §3 rests on the prior independence assumption 3.2,
without which the propensity score model cannot be ignored from the likelihood. But this
assumption is not always plausible in real applications. Various priors that do not rely on this
assumption have been proposed [47,63–65]. Below we show two simple examples.
The first example is due to [58] and is designed for simultaneous variable selection for the
propensity score and outcome models. Specifically, assume a logistic propensity score model
logit{Pr(Zi = 1 | Xi )} = α  Xi and a linear outcome model Yi | Zi , Xi ∼ N (τ Zi + β  Xi , σ 2 ). Assume
each of the jth components of the coefficients, αj , follow the spike-and-slab prior [73]: αj | γjα ∼
(1 − γjα )I0 + γjα N (0, σα2 ), where γjα is a latent indicator of whether Xj is included in the model
and I0 denotes the point mass at 0. A similar spike-and-slab prior is assumed for the coefficients
β β β β
of the outcome model with a latent inclusion indicator γj : βj | γj ∼ (1 − γj )I0 + γj N (0, σβ2 ).
β
Then assume the probability of the events {γjα = 0} and {γj = 0} are dependent a priori:
β β
Pr(γj = 1 | γjα = 1)/Pr(γj = 0 | γjα = 1) = ω, where ω ∈ [1, ∞) is a dependence hyperparameter
that controls the prior odds of including Xj into the outcome model when it is included in
 the propensity score model. Larger ω implies stronger prior dependence between the variable
12
selection in the two models.
The second example is due to [74]. Assume Yi (1) | Xi ∼ N (μ1 , σ12 e(Xi )) and Yi (0) | Xi ∼

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
N (μ0 , σ02 (1 − e(Xi ))), with flat priors on μ1 and μ0 . If the propensity scores are known, the
posterior mean of the PATE equals the Hajék version of the IPW estimator:
N N
Zi Yi /e(Xi ) i=1 (1 − Zi )Yi /(1 − e(Xi ))
τ hajek = i=1
 N
− N .
Z
i=1 i /e(X i ) i=1 (1 − Zi )/(1 − e(Xi ))
If the propensity scores are unknown, then the posterior mean of the PATE is closely related to

τ hajek averaged over the posterior predictive distribution of the propensity scores. This strategy
simply includes the propensity scores in the outcome model, but somewhat unusually in the
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

conditional variances, rather than the conditional means of the potential outcomes.
Carefully designed dependent priors often achieve desirable finite sample results and are more
reasonable in real world studies. However, specification of such priors is case-dependent, and
there is no general solution.

(c) Posterior predictive inference

A general, albeit not dogmatically Bayesian, strategy is to specify both a propensity score model
e(Xi ; θZ ) and an outcome model {μ1 (Xi ; θY ), μ0 (Xi ; θY )}, and obtain posterior draws of e(Xi ; θZ )
and {μ1 (Xi ; θY ), μ0 (Xi ; θY )} from their respective posterior predictive distributions, and then plug
these posterior draws into the doubly-robust estimator  τ dr [66,75]. A variance estimator of the
resulting estimator  τ is given in [66]. In the same vein, Ding & Guo [76] incorporated the
dr

propensity score in Bayesian posterior predictive p-value. For the model with the Fisher’s sharp
null hypothesis of no causal effect for any units whatsoever (i.e. Yi (1) = Yi (0) for all i), the
procedure in [76] is equivalent to the Fisher randomization test averaged over the posterior
predictive distribution of the propensity score. Simulations in [76] show the advantages of the
Bayesian p-value compared to the Frequentist analogue. This perspective offers a straightforward
and flexible strategy to integrate Bayesian modelling and common Frequentist procedures for
causal inference and enables proper uncertainty quantification.
Besides the above three strategies, another general approach is through the aforementioned
Bayesian bootstrap, which can be used to simulate the posterior distribution of any parameter that
can be formulated as M-estimation or estimating equation [41,77]. As special cases, because the
IPW estimator  τ ipw and the doubly robust estimator  τ dr —both involving the propensity scores—
are both solutions to estimating equations, they can be naturally combined with the Bayesian
bootstrap to devise a Bayesian version. However, such an approach may be guilty of ‘Bayesian for
the sake of being Bayesian’, and their methodological and practical value compared to competing
methods is unclear.

6. Sensitivity analysis in observational studies

Unconfoundedness is a central assumption for causal inference. It holds by design in
randomized experiments. However, its validity is fundamentally untestable in observational
studies. Therefore, it is of great importance to assess the sensitivity of the results with respect
to unmeasured confounding in any observational study. Such procedures are broadly called
sensitivity analysis. Different classes of sensitivity analysis methods are characterized by the
specific parametrization of confounding. Below we review the two most popular classes.

(a) Parametrization involving distributions with unmeasured confounders

The first parametrization used for sensitivity analysis is motivated by the intuition that a hidden
confounder may completely explain away the association between the treatment and the outcome
even after adjusting for observed covariates. In a historic debate, Fisher [78] hypothesized that
 the strong association between cigarette smoking and lung cancer might be due to a hidden
13
genetic factor as their ‘common cause’ or confounder. Cornfield et al. [79] derived an inequality
showing that to explain away the observed association, the association between the unmeasured

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
confounder and cigarette smoking must be larger than or equal to the association between
cigarette smoking and lung cancer. Their work helped to initiate the field of sensitivity analysis.
Let U denote an unmeasured confounder and assume that unconfoundedness holds
conditional on (X, U): Z ⊥⊥ {Y(0), Y(1)} | X, U. The joint distribution of all variables factorizes into

Pr{Y(1), Y(0), Z, X, U} = Pr{Y(1), Y(0) | X, U} · Pr(Z | X, U) · Pr(U | X) · Pr(X). (6.1)

Under the factorization (6.1), sensitivity analysis requires us to specify the models for
Pr{Y(1), Y(0) | X, U}, Pr(Z | X, U) and Pr(U | X). In the special case of a binary Z, a binary Y
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

and a discrete X (which can be thought as a stratified propensity score), Rosenbaum & Rubin
[80] assumed a logistic model for Y given (Z, X, U), a logistic model for Z given (X, U), and
a Bernoulli distribution for U, and treated the logistic regression coefficients of U and the
probability parameter of U as the sensitivity parameters. They integrated out U in the complete-
data likelihood and obtained the maximum likelihood estimates of τ P over a plausible range of
values of the sensitivity parameters. This method has been extended to more general settings
in the Frequentist fashion in [81,82]. The Bayesian analogue of [80] is straightforward and can
leverage the data augmentation algorithm to impute U to simplify the computation. Dorie et al.
[83] extended this method to impose a Bayesian semiparametric model with a BART component
for Pr{Y(1), Y(0) | X, U} to allow for model flexibility.
As an extension of [79], Ding & VanderWeele [84] treated the treatment-confounder (Z and
U) and outcome-confounder (Y and U) associations as two sensitivity parameters, and derived
analytical thresholds for them in order to explain away the observed treatment-outcome (Z and
Y) association. Based on that theory, VanderWeele & Ding [85] further simplified by assuming
the two associations to be the same and called the resulting threshold the E-value, as a measure
of robustness of the causal conclusions with respect to unmeasured confounding. The E-value
framework is model-free because it avoids modelling assumptions with U; it also avoids repeating
the analysis over a range of sensitivity parameters as in the competing methods and thus is simple
to calculate.

(b) Parametrization involving distributions of potential outcomes

The second parametrization is motivated by an alternative mathematical expression of the
unconfoundedness assumption: Pr{Y(z) | Z = 1, X} = Pr{Y(z) | Z = 0, X} for z = 0, 1, representing
the fact that the units in the two randomized arms are comparable in terms of potential outcomes.
This class of sensitivity analysis is based on sensitivity parameters that directly represent the
difference between the distributions Pr{Y(z) | Z = 1, X} and Pr{Y(z) | Z = 0, X} instead of modelling
the difference with an unobserved U. This is implemented in the context of time-varying
treatments (see §7(b)) and Frequentist semiparametric estimation [86]. Franks et al. [34] pointed
out the importance of distinguishing between model fit and sensitivity to unconfoundedness:
the former involves identifiable parameters (e.g. the parameters in the model of the marginal
distributions of the outcome Pr{Yi (z) | Zi = z, X}) whereas the latter involves unidentifiable
parameters (e.g. the association between Yi (1) and Yi (0)). The merit of this parametrization
is apparent in this perspective because it separates identifiable and unidentifiable parameters.
Franks et al. [34] proceeded under the Bayesian paradigm and used a copula—parameters of
which are the sensitivity parameters—to connect the two identifiable marginal distributions.
A related branch of sensitivity analysis is Rosenbaum’s bounds [87]. His original formulation
takes the association between Z and the potential outcomes conditional on observed X, denoted
by Γ , as the sole sensitivity parameter for quantifying unmeasured confounding. He has also
made connections to the parametrization in §6(a) [88]. Starting with a matched sample to mimic
a randomized matched-pairs experiment, one can then repeat the Fisher randomization test on
the sharp null hypothesis of no treatment effect given a range of Γ values, and find the threshold
 of Γ at which the p-value of the test changes from significant to insignificant. This approach was
14
later generalized to derive the bounds of a given estimator under different Γ values. Grounded
in Fisherian randomization inference, Rosenbaum’s framework does not have a natural Bayesian

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
analogue.
Besides the above two classes, there are numerous other approaches to sensitivity analysis
based on alternative parameterization of unmeasured confounding. However, a common
criticism of various approaches to sensitivity analysis is that, in order to assess the consequence
of the untestable unconfoundedness assumption, one has to make even more untestable
assumptions, e.g. specifying models involving U. Moreover, sensitivity analysis, after all, is a
secondary analysis in causal studies, and thus simple implementation and intuitive interpretation
is much desired. The considerations underpin the dominance of the E-value method over other
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

methods in practice, particularly in medicine and public health.

7. Complex assignment mechanisms

So far, we have discussed the simplest causal setting of an ignorable treatment at one time point.
The basic formulation can be extended to many more complex assignment mechanisms. There are
also many popular quasi-experimental designs that rely on identification strategies alternative
to ignorability, e.g. regression discontinuity designs, difference-in-differences, synthetic controls.
These designs are widely used in socioeconomic applications. Due to the space limit, below we
will only briefly review two important extensions and refer interested readers to [89] for a review
of quasi-experimental designs and related econometric methods.

(a) From instrumental variable to principal stratification

Instrumental variable (IV) is one of the most important techniques for causal inference in
economics and social sciences. IVs are used in settings where dependence of the assignment on
the potential outcomes cannot plausibly be ruled out, even conditional on observed covariates.
An IV is a variable that provides a source of exogenous (or unconfounded) variation that helps
identify causal effects. IV methods are based on a set of assumptions alternative to ignorability.
Specifically, an IV satisfies three conditions: (i) it occurs before a treatment; (ii) it is independent
of the treatment-outcome confounding; and (iii) it affects the outcome only through its (non-zero)
effects on the treatment assignment. Finding a valid IV is challenging in observational studies and
many clever natural experiments have been identified [89]. Given a valid IV, one can extract the
causal effects of the treatment on an outcome by a two-stage least-squares (2SLS) estimator: first,
fit a linear regression of the treatment on the IV; second, fit a linear regression of the outcome on
the fitted value of the treatment from the first stage, the coefficient of which is taken as the causal
effect of the treatment on the outcome. Covariates can be added in both stages.
The IV method has been developed within the structural equation model framework (see
[89] for a review), and the 2SLS IV estimator may not correspond to a causal effect within the
potential outcomes framework except for a few special cases. In a landmark paper, Angrist et al.
[90] connects IV to the potential outcomes framework in the setting of randomized experiments
with binary treatment and all-or-nothing compliance, with the initial random assignment playing
the role of an IV. But many questions remain on the connection between the IV method and the
potential outcomes framework in more general settings. Below we will describe the special setting
of Angrist et al. [90].
We introduce some new notation. For unit i, let Zi be the randomly assigned treatment
(1 for the treatment and 0 for the control), and Wi be the actual treatment status (1 for the
treatment and 0 for the control). When Zi = Wi , non-compliance arises. Because Wi occurs post-
assignment, it has two potential values, Wi (0) and Wi (1), with Wi = Wi (Zi ). As before, the outcome
Yi has two potential outcomes, Yi (0) and Yi (1). Based on their joint potential status of the actual
treatment Ui = (Wi (1), Wi (0)), the units fall into four compliance types: compliers Ui = (1, 0) = co,
never-takers Ui = (0, 0) = nt, always-takers Ui = (1, 1) = at and defiers Ui = (0, 1) = df [90]. A key
 property of Ui is that it is not affected by the treatment assignment, and thus can be regarded as
15
a pre-treatment characteristic. Therefore, comparisons of Yi (1) and Yi (0) within the stratum of Ui
have standard subgroup causal interpretations: τu = E{Yi (1) − Yi (0) | Ui = u}, for u = nt, co, at, df;

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
τu are later called principal causal effects. The conventional causal estimand in clinical trials
is the intention-to-treat effect that ignores the compliance information, which is the weighted

average of the four stratum-specific effects: E{Yi (1) − Yi (0)} = u πu τu , where πu = Pr(Ui = u) is
the proportion of the stratum u. The intention-to-treat effect measures the effect of the assignment
instead of the actual treatment.
Due to the fundamental problem of causal inference, individual compliance stratum
Ui is not observed. So the principal causal effects are non-identifiable without additional
assumptions. Besides randomization of Zi , Angrist et al. [90] make two additional assumptions:
(i) monotonicity: Wi (1) ≥ Wi (0), and (ii) exclusion restriction: Yi (1) = Yi (0) whenever Wi (1) =
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

Wi (0). Monotonicity rules out defiers, and exclusion restriction imposes that the assignment has
zero effects among never-takers and always-takers. Then the compiler average causal effect is
identified by
E(Y | Z = 1) − E(Y | Z = 0)
τco ≡ E{Y(1) − Y(0) | U = co} = ,
E(W | Z = 1) − E(W | Z = 0)
which is exactly the probability limit of the two-stage least-squares estimator [90]. Because
under monotonicity, only compilers’ actual treatments are affected by the assignment, τco can
be interpreted as the effect of the treatment.
We now describe the Bayesian inference of the IV set-up, first outlined in [91]. Without
additional assumptions, the observed cells of Z and W consist of a mixture of units from more
than one stratum. For example, the units who are assigned to the treatment arm and took
the treatment (Z = 1, W = 1) can be either always-takers or compliers. One must disentangle
the causal effects for different compliance types from observed data. Therefore, model-based
inference here resembles that of a mixture model. In Bayesian analysis, it is natural to impute the
missing label Ui under some model assumptions. Specifically, six quantities are now associated
with each unit, {Yi (1), Yi (0), Wi (1), Wi (0), Xi , Zi }, four of which, {Yiobs = Yi = Yi (Zi ), Wiobs = Wi =
Wi (Zi ), Zi , Xi }, are observed and the remaining two, {Yimis = Yi (1 − Zi ), Wimis = Wi (1 − Zi )}, are
unobserved. Assume the joint distribution of these random variables of all units is governed
by a parameter θ , conditional on which the random variables for each unit are iid. We assume
unconfoundedness Pr{Zi = 1 | Xi , Wi (1), Wi (0), Yi (1), Yi (0)} = Pr(Zi = 1 | Xi ), and impose a prior
distribution Pr(θ ). Then the joint posterior distribution of θ and the missing potential outcomes
are proportional to the complete-data likelihood as follows:
N
Pr(θ ) Pr{Yi (0), Yi (1) | Ui , Xi ; θY } Pr(Ui | Xi ; θU ) Pr(Xi | θX ). (7.1)
i=1

Without covariates, posterior inference of τu is straightforward because it is a function of θY (see

example 7.1 below). With covariates, we can condition on them and focus on a MATE estimand

τuM = N−1 N i=1 E{Yi (1) − Yi (0) | Ui = u, Xi }. The formula (7.1) suggests that we need to specify two
models for inferring τuM : (i) the compliance type model, Pr(Ui | Xi ; θU ), and (ii) the outcome model,
Pr{Yi (0), Yi (1) | Ui , Xi ; θY }. For example, we can specify a multinomial logistic regression for Ui and
a generalized linear model for Yi [91,92].
Using the same arguments as in §3, to infer population and mixed estimands, we only need
to specify two marginal outcome models for Yi (1) and Yi (0) instead of a joint model, and do
not need to impute the missing potential outcomes Ymis . But we do need to impute the latent
Ui , or, equivalently, the missing intermediate variable W mis . We can simulate the joint posterior
distribution Pr(θ , Wmis | Yobs , Wobs , Z, X) by iteratively imputing the missing U from Pr(Wmis |
Yobs , Wobs , Z, X, θ) and updating the posterior distribution of θ from Pr(θ | Yobs , Wobs , Wmis , Z, X).
Below we illustrate the Bayesian procedure via a simple example of the IV approach.
Example 7.1. [Randomized experiment with one-sided non-compliance] Consider a randomized
experiment with a binary outcome, where control units have no access to the treatment, i.e.
 Wi (0) = 0 for all units. Therefore, we only have two strata: Ui = co with Wi (1) = 1 and Ui = nt with
16
Wi (1) = 0, respectively, with πco + πnt = 1. Assume Yi (z) | Ui = co ∼ Bern(pco,z ) for z = 0, 1, and
Yi (1) = Yi (0) | Ui = nt ∼ Bern(pnt ). So τco = pco,1 − pco,0 . For simplicity, we impose conjugate priors

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
on the parameters: πco , pco,1 , pco,0 , pnt are iid Beta(1/2, 1/2). To sample the posterior distributions,
the key is to impute the missing Ui ’s given the observed data. If Zi = 1, then Wi = 1 implies Ui = co
and Wi = 0 implies Ui = nt, respectively. If Zi = 0, then Wi = 0 and Ui is latent. For units with
(Zi = 0, Wi = 0), we can impute Ui = co with probability

πco · pY
co,0 (1 − pco,0 )
i 1−Yi

πco · pY
co,0 (1 − pco,0 )
i 1−Yi + π · pYi (1 − p )1−Yi
nt nt nt

and Ui = nt with the rest probability. With the imputed Ui ’s, we can sample the parameters
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

 N
from standard Beta posteriors: (i) sample πco from Beta(1/2 + N i=1 1(U i = co), 1/2 + i=1 1(Ui =
 N
nt)) and obtain πnt = 1 − πco , (ii) sample pco,1 from Beta(1/2 + i=1 Zi 1(Ui = co)Yi , 1/2 +
N 
Zi 1(Ui = co)(1 − Yi )), (iii) sample pco,0 from Beta(1/2 + N − Zi )1(Ui = co)Yi , 1/2 +
i=1 (1 
i=1
N N
(1 − Z i )1(U i = co)(1 − Y i )), and (iv) sample pnt from Beta(1/2 + i=1 1(Ui = nt)Yi , 1/2 +
i=1
N
i=1 1(U i = nt)(1 − Y i )). We iterate until convergence and obtain the posterior distribution of
τco = pco,1 − pco,0 . Imbens & Rubin [91] provided more detailed discussions.

Frangakis & Rubin [93] generalized the IV approach to principal stratification, a unified
framework for causal inference with post-treatment confounding. In the simplest scenario, a post-
treatment confounded variable lies in the causal pathway between the treatment and the outcome;
it cannot be adjusted in the same fashion as a pre-treatment covariate in causal inference. A
principal stratification with respect to a post-treatment variable is the classification of units based
on the joint potential values of the post-treatment variable, and the stratum-specific effects are
called principal causal effects, of which τco is a special case. The post-treatment variable setting
includes a wide range of examples. For instance, in the non-compliance setting, the ‘treatment’
is the randomized treatment assignment, the ‘post-treatment’ variable is the actual treatment
received, and the compliance types are the principal strata [91,92,94,95]. Zeng et al. [96] connects
principal stratification to the local IV method with a continuous IV and binary treatment [97].
Other examples include censoring due to death [98], surrogate endpoints [99,100], regression
discontinuity designs [101], time-varying treatments [102], and many more. The choices of target
strata and thus estimands, interpretations, and identifying assumptions depend on specific
applications, details of which are omitted here.

(b) Time-varying treatment and confounding

In real-world situations, subjects often receive treatments sequentially at multiple time points, and
the treatment assignment at each time is affected by both baseline and time-varying confounders
as well as previous treatments [103,104]. Such settings are referred to as time-varying, or
sequential, or longitudinal treatments.
Consider a study where treatments are assigned at T time points. Let Zit denote the treatment
at time t for unit i (i = 1, . . . , N; t = 1, . . . , T). At baseline (t = 0), each unit i has time-invariant
covariates Li0 measured; then after the treatment assignment at time t − 1 and prior to the
assignment at time t, a set of time-varying confounders Li,t−1 are measured, which include the
intermediate measurements of the final outcome and the covariates that are affected by the
previous treatments. For example, in a cancer study, baseline covariates can include sex, age, race
and time-varying confounders can include intermediate cancer progression and other clinical
traits such as blood pressure measured prior to the next treatment. Denote the observed and
hypothetical treatment sequence of length t by Z̄it = (Zi1 , . . . , Zit ) and z̄t = (z1 , . . . , zt ), respectively,
and the sequence of time-varying confounders by L̄it = (Li0 , Li1 , . . . , Lit ). For each z̄T , there is a
potential outcome Y(z̄T ). The final observed outcome Yi = Yi (Z̄iT ), corresponding to the entire
observed treatment sequence, is measured after treatment assigned at T. A common causal
 estimand is the marginal effect comparing two pre-specified treatment sequences, z̄, z̄ ∈ {0, 1}T :
17
τz̄T ,z̄T = E{Yi (z̄T ) − Yi (z̄T )}. For simplicity, below we drop the subscript i.
The central question to causal inference with sequential treatments is the role of the time-

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
varying confounders Lt in the assignment mechanism. These variables are affected by the
previous treatments and also affect the future treatment assignment and outcome. Much of the
literature assumes a sequentially ignorable assignment mechanism [103], that is, the treatment at
each time is unconfounded conditional on the observed history, which consists of past treatments
Z̄t−1 and time-varying confounders L̄t−1 , as stated below.

Assumption 7.2. (Sequential Ignorability). Pr{Zt | Z̄t−1 , L̄t−1 , Y(z̄t ) for all z̄t } = Pr(Zt | Z̄t−1 , L̄t−1 )
for t = 1, . . . , T.
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

A full Bayesian approach to time-varying treatments [105] would specify a joint model for
treatment assignment Zt and time-varying confounders Lt at all time points as well as all the
potential outcomes Y(z̄T ), and then derive the posterior predictive distributions of the missing
potential outcomes and thus of the estimands. This procedure is a straightforward extension from
the structure introduced in §3. However, the joint modelling approach is rarely used because it
quickly becomes intractable as the time T and the number of time-varying confounders increases.
Instead, most of the Bayesian methods are grounded in the g-computation. Under sequential
ignorability, the average potential outcome E{Yi (z̄T )} is identified from the observed data via the
g-formula [103]:

E{Yi (z̄T )} = E(Y | Z̄T = z̄T , L̄T−1 ) · Pr(LT−1 | Z̄T−1 = z̄T−1, L̄T−2 )
L0 ,L1 ,...,LT−1

· · · Pr(L1 | Z1 = z1 , L0 ) · Pr(L0 ). (7.2)

To operationalize the g-formula, we can specify models for all the components of (7.2), including
an outcome model Pr(Y | Z̄T = z̄T , L̄T−1 ) and a model for the time-varying confounders Lt at each
time t, Pr(Lt | Z̄t = z̄t , L̄t−1 ). The g-formula is in essence an extension of the outcome regression
approach to time-varying treatments. The Bayesian version of the g-computation would specify a
Bayesian model for each component in the g-formula (7.2) and then combine the posterior draws
of the parameters to obtain the posterior distribution of the estimands. Below we present an
illustrative example of Bayesian g-computation due to [106].

Example 7.3. [Bayesian g-computation with two periods] Consider the simplest possible
scenario with two time periods, binary covariates and a binary outcome. Let L0 be a binary
baseline covariate, Z1 is a binary treatment at time 1, L1 is a binary time-varying covariate
between time 1 and 2, Z2 is a binary treatment, and Y is a binary outcome. To obtain the posterior
distribution of
 
E{Y(z1 , z2 )} = Pr(Y = 1 | Z1 = z1 , Z2 = z2 , L0 = l0 , L1 = l1 )
l0 =0,1 l1 =0,1

· Pr(L1 = l1 | Z1 = z1 , L0 = l0 ) · Pr(L0 = l0 ), (7.3)

it suffices to obtain the posterior distributions of the probabilities in (7.3). Assuming the standard
Beta(1/2, 1/2) conjugate priors. We can obtain the posterior of the probabilities as follows: (i)

sample Pr(Y = 1 | Z1 = z1 , Z2 = z2 , L0 = l0 , L1 = l1 ) from Beta(1/2 + N i=1 1(Zi1 = z1 , Zi2 = z2 , Li0 =
N
l0 , Li1 = l1 )Yi , 1/2 + i=1 1(Zi1 = z1 , Zi2 = z2 , Li0 = l0 , Li1 = l1 )(1 − Yi )), (ii) sample Pr(L1 = 1 | Z1 =
 N
z1 , L0 = l0 ) from Beta(1/2 + N i=1 1(Zi1 = z1 , Li0 = l0 )Li1 , 1/2 + i=1 1(Zi1 = z1 , Li0 = l0 )(1 − Li1 )),

and (iii) sample Pr(L0 = 1) from Beta(1/2 + N i=1 Li0 , 1/2 +
N
i=1 (1 − Li0 )). With these ingredients
and (7.3), we can obtain the posterior distributions of E{Y(z1 , z2 )}’s and their contrasts

z1 ,z2 c(z1 , z2 )E{Y(z1 , z2 )}.

G-computation quickly becomes complex as the number of time periods T and time-varying
confounders increases, which requires specifying a large number of models. Then it is necessary
to impose more structural restrictions on the data-generating process. However, Robins &
 Wasserman [107] showed that unsaturated models might rule out the null hypothesis of zero
18
causal effect a priori, a phenomenon termed the ‘g-null paradox’.
A popular alternative strategy is the marginal structural model [104], which generalizes IPW to

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
time-varying treatments. Saarela et al. [108] devised a Bayesian version of the marginal structural
model via the Bayesian bootstrap. Because the marginal structural model relies on IPW, a key
component in its implementation is to estimate the propensity scores and ensure overlap at each
time point. However, overlap between different treatment paths usually becomes limited as the
number of time periods increases, rendering the marginal structural model sensitive to extreme
weights.
The above discussion focuses on static treatment sequences. Another important class of time-
varying treatment is the dynamic treatment regime, which consists of a sequence of decision
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

rules, one per time point of intervention, that determines how to individualize treatments to
units based on evolving treatment and covariate history. Inferring optimal dynamic treatment
regimes requires combining causal inference and decision theory techniques and is closely related
to reinforcement learning. See [109] for a review. Due to the space limit, we omit the discussion
of the closely related topics of Bayesian multi-armed bandit [110] and Bayesian reinforcement
learning [111].

8. Discussion
This paper reviews the Bayesian approach to causal inference under the potential outcomes
framework. We discussed the causal estimands, identification strategies, the general structure
of Bayesian inference of causal effects, and sensitivity analysis. We highlight issues that are
unique to Bayesian causal inference, including the role of the propensity score, definition of
identifiability, the choice of priors in both low- and high-dimensional regimes. In particular, under
ignorability and prior independence, the propensity score is seemingly irrelevant for the posterior
distributions of the causal parameters. However, we pointed out that even in this setting, the
propensity score and more generally the design stage plays a central role in obtaining robust
Bayesian causal inference. Regardless of the mode of inference, a critical step in causal inference
with observational data is to ensure adequate covariate overlap and balance in the design or
analysis stages. In high dimensions, such a task is particularly challenging and what is the optimal
practice remains an open question.
The Bayesian approach offers several advantages for causal inference. First and most
importantly, by enabling imputation of all missing potential outcomes, the Bayesian paradigm
provides a unified inferential framework for any causal estimand. This is particularly appealing
for inferring complex estimands such as the conditional average treatment effects or individual
treatment effects as well as partially identifiable causal estimands such as the principal strata
causal effects. In contrast, the Frequentist approach to these problems needs to be customized for
each scenario, and the inference usually relies on bounds or asymptotic arguments, which are
often either non-informative or questionable in cases like individual treatment effects. Second,
the automatic uncertainty quantification of any estimand renders it straightforward to combine
causal inference and decision theory for dynamic decision-making, e.g. in personalized medicine.
Third, the Bayeian approach naturally incorporates prior knowledge into a causal analysis, e.g.
in evaluating spatially correlated treatments and/or outcomes. Fourth, there is rich collection
of Bayesian models for complex data with limited Frequentist counterparts. A few examples
are (i) spatial or temporal data, (ii) functional data, and (iii) interference, i.e. when the SUTVA
assumption is violated. In these settings, special care must be taken on issues key to causal
inference such as defining relevant estimands and ensuring overlap. Moreover, it is important
to ensure that the Bayesian models are coherent to the model-free identification assumptions
such as ignorability. For example, adding spatial random effects into an outcome model may
inadvertently bias the coefficient of the treatment variable as the estimate of a causal effect [112].
Research on Bayesian analysis of these topics has been rapidly increasing [10,112–115] and is
expected to continue to grow.
 Despite the above advantages, the theory and practice in causal inference has long been
19
dominated by non-Bayesian methods. One reason is that many popular Frequentist techniques,
such as matching and weighting, as well as Fisherian randomization test, do not require

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
specifying outcome models and prior distribution of parameters, and thus offer a perception
of ‘model-free’ or ‘objective’. This is appealing to many applied researchers. Another reason
is that the Bayesian approach requires more advanced computing and programming, which
may not be readily available to many practitioners. The Stan programming language [116]
mitigates some of these issues, but Bayesian computation remains inaccessible to most domain
scientists. To popularize Bayesian causal inference in practice, it is crucial to provide (i) more
examples of successful Bayesian applications with clear advantages over other inferential modes,
e.g. [45], (ii) accessible tutorials, ideally with generalizable computer code and illustrations
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

of important scientific problems and (iii) developing and disseminating user-friendly, general
purpose software packages.
We have occasionally commented on whether a method is dogmatically Bayesian in the
discussion. However, we do not regard the conceptual purity of being dogmatically Bayesian,
per se, as advantageous, nor should it be the motivating goal in real applications. When a
quasi-Bayesian method outperforms its dogmatically Bayesian counterpart (if available) with
methodological footing and empirical evidence, as the example of adding estimated propensity
score in an outcome model in §5(a), we would endorse the former over the latter. We also
doubt the value of devising a Bayesian version of an established Frequentist method without
clear theoretical or practical advantages. As a general view, we believe whether to choose a
Bayesian approach should be dictated by its practical utility in a specific context rather than an
unconditional commitment to the Bayesian doctrine. For causal inference and perhaps everything
in statistics, being Bayesian should be a tool, not a goal.
Data accessibility. This article has no additional data.
Authors’ contributions. F.L.: conceptualization, formal analysis, investigation, methodology, writing—original
draft, writing—review and editing; P.D.: formal analysis, investigation, methodology, writing—original draft,
writing—review and editing; F.M.: methodology, writing—review and editing.
All authors gave final approval for publication and agreed to be held accountable for the work performed
therein.
Conflict of interest declaration. We declare we have no competing interests.
Funding. P.D.’s research is partially funded by the US National Science Foundation grant no. 1945136. F.M.
thanks the Department of Excellence 2018–2022 funding provided by the Italian Ministry of University and
Research (MUR).
Acknowledgements. The authors are grateful to Joey Antonelli, Yuansi Chen, Sid Chib, Ruobin Gong, Guido
Imbens, Zhichao Jiang, Antonio Linero, Georgia Papadogeorgou, Donald Rubin, Surya Tokdar, Mike West,
Jason Xu, Cory Zigler, Anqi Zhao and two anonymous reviewers for discussions and suggestions.

References
1. Neyman J. 1923 On the application of probability theory to agricultural experiments: essay
on principles, §9. Masters Thesis. Portions translated into english by D. Dabrowska and T.
Speed (1990) in Stat. Sci., pp. 465–472.
2. Rubin DB. 1974 Estimating causal effects of treatments in randomized and nonrandomized
studies. J. Edu. Psychol. 66, 688–701. (doi:10.1037/h0037350)
3. Rubin DB. 1975 Bayesian inference for causality: the role of randomization. In Proc. of Social
Statistics Section of Am Stat. Assoc., pp. 233–239.
4. Pearl J. 2000 Causality: models, reasoning, and inference. New York, NY: Cambridge University
Press.
5. Peters J, Bühlmann P, Meinshausen N. 2016 Causal inference by using invariant prediction:
identification and confidence intervals. J. R. Stat. Soc. B 78, 947–1012. (doi:10.1111/rssb.12167)
6. Ding P, Li F. 2018 Causal inference: a missing data perspective. Stat. Sci. 33, 214–237.
(doi:10.1214/18-STS645)
 7. Hahn PR, Carvalho CM, Puelz F, He J. 2018 Regularization and confounding
20
in linear regression for treatment effect estimation. Bayesian Anal. 13, 163–182.
(doi:10.1214/16-BA1044)

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
8. Hahn PR, Murray JS, Carvalho CM. 2020 Bayesian regression tree models for causal
inference: regularization, confounding, and heterogeneous effects (with discussion). Bayesian
Anal. 15, 965–1056. (doi:10.1214/19-BA1195)
9. Hill JL. 2011 Bayesian nonparametric modeling for causal inference. J. Comput. Graph. Stat.
20, 217–240. (doi:10.1198/jcgs.2010.08162)
10. Linero AR, Antonelli JL. 2022 The how and why of Bayesian nonparametric causal inference.
Wiley Interdiscip. Rev.: Comput. Stat. 15, e1583.
11. Zigler CM, Dominici F. 2014 Uncertainty in propensity score estimation: Bayesian methods
for variable selection and model averaged causal effects. J. Am. Stat. Assoc. 109, 95–107.
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

(doi:10.1080/01621459.2013.869498)
12. Dawid AP. 1979 Conditional independence in statistical theory. J. R. Stat. Soc. B 41, 1–15.
13. Rubin DB. 1980 Comment on ‘Randomization analysis of experimental data: the Fisher
randomization test’ by D. Basu. J. Am. Stat. Assoc. 75, 591–593.
14. Holland PW. 1986 Statistics and causal inference. J. Am. Stat. Assoc. 81, 945–960.
(doi:10.1080/01621459.1986.10478354)
15. Rubin DB. 1978 Bayesian inference for causal effects: the role of randomization. Ann. Stat. 6,
34–58. (doi:10.1214/aos/1176344064)
16. Rosenbaum PR, Rubin DB. 1983 The central role of the propensity score in observational
studies for causal effects. Biometrika 70, 41–55. (doi:10.1093/biomet/70.1.41)
17. Lin W. 2013 Agnostic notes on regression adjustments to experimental data: reexamining
Freedman’s critique. Ann. Appl. Stat. 7, 295–318. (doi:10.1214/12-AOAS583)
18. Rubin DB. 2007 The design versus the analysis of observational studies for causal effects:
parallels with the design of randomized trials. Stat. Med. 26, 20–36. (doi:10.1002/sim.
2739)
19. Abadie A, Imbens GW. 2011 Bias corrected matching estimators for average treatment
effects. J. Bus. Econom. Stat. 29, 1–11. (doi:10.1198/jbes.2009.07333)
20. Abadie A, Imbens GW. 2006 Large sample properties of matching estimators for average
treatment effects. Econometrica 74, 235–267. (doi:10.1111/j.1468-0262.2006.00655.x)
21. Rubin DB. 2006 Matched sampling for causal effects. Cambridge, UK: Cambridge University
Press.
22. Li F, Morgan KL, Zaslavsky AM. 2018 Balancing covariates via propensity score weighting.
J. Am. Stat. Assoc. 113, 390–400. (doi:10.1080/01621459.2016.1260466)
23. Rosenbaum PR. 1987 Model-based direct adjustment. J. Am. Stat. Assoc. 82, 387–394.
(doi:10.1080/01621459.1987.10478441)
24. Robins JM, Rotnitzky A, Zhao LP. 1994 Estimation of regression coefficients when
some regressors are not always observed. J. Am. Stat. Assoc. 89, 846–866. (doi:10.1080/
01621459.1994.10476818)
25. Bang H, Robins JM. 2005 Doubly robust estimation in missing data and causal inference
models. Biometrics 61, 962–972. (doi:10.1111/j.1541-0420.2005.00377.x)
26. Lin Z, Ding P, Han F. 2021 Estimation based on nearest neighbor matching: from density
ratio to average treatment effect. Preprint (https://arxiv.org/abs/2112.13506).
27. Rubin DB. 1976 Inference and missing data. Biometrika 63, 581–592. (doi:10.1093/biomet/
63.3.581)
28. Chib S. 2007 Analysis of treatment response data without the joint distribution of potential
outcomes. J. Econom. 140, 401–412. (doi:10.1016/j.jeconom.2006.07.009)
29. Lindley DV. 1972 Bayesian statistics: a review. SIAM.
30. Gustafson P. 2015 Bayesian inference for partially identified models: exploring the limits of limited
data. New York, NY: CRC Press.
31. Lu J, Ding P, Dasgupta T. 2018 Treatment effects on ordinal outcomes: causal estimands and
sharp bounds. J. Educ. Behav. Stat. 43, 540–567. (doi:10.3102/1076998618776435)
32. Daniels MJ, Hogan JW. 2008 Missing data in longitudinal studies: strategies for Bayesian modeling
and sensitivity analysis. London, UK: Chapman and Hall/CRC.
33. Ding P, Dasgupta T. 2016 A potential tale of two-by-two tables from completely randomized
experiments. J. Am. Stat. Assoc. 111, 157–168. (doi:10.1080/01621459.2014.995796)
 34. Franks AM, D’Amour A, Feller A. 2020 Flexible sensitivity analysis for observational
21
studies without observable implications. J. Am. Stat. Assoc. 115, 1730–1746. (doi:10.1080/0162
1459.2019.1604369)

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
35. Gustafson P. 2009 What are the limits of posterior distributions arising from nonidentified
models, why should we care? J. Am. Stat. Assoc. 104, 1682–1695. (doi:10.1198/
jasa.2009.tm08603)
36. Richardson TS, Evans RJ, Robins JM. 2010 Transparent parameterizations of models for
potential outcomes. In Bayesian Statistics, vol. 9 (eds JM Bernardo, MJ Bayarri, JO Berger, AP
Dawid, D Heckerman, AFM Smith, M West), pp. 569–610. Oxford, UK: Oxford University
Press.
37. Rubin DB. 1981 The Bayesian bootstrap. Ann. Stat. 9, 130–134. (doi:10.1214/aos/1176345338)
38. Ferguson TS. 1973 A Bayesian analysis of some nonparametric problems. Ann. Stat. 1, 209–
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

230. (doi:10.1214/aos/1176342360)
39. Oganisian A, Mitra N, Roy JA. 2020 Hierarchical Bayesian bootstrap for heterogeneous
treatment effect estimation. (https://arxiv.org/abs/2009.10839).
40. Taddy M, Gardner M, Chen L, Draper D. 2016 A nonparametric Bayesian analysis of
heterogenous treatment effects in digital experimentation. J. Bus. Econ. Stat. 34, 661–672.
(doi:10.1080/07350015.2016.1172013)
41. Chamberlain G, Imbens GW. 2014 Nonparametric applications of Bayesian inference. J. Bus.
Econ. Stat. 21, 12–18. (doi:10.1198/073500102288618711)
42. Künzel SR, Sekhon JS, Bickel PJ, Yu B. 2019 Metalearners for estimating heterogeneous
treatment effects using machine learning. Proc. Natl Acad. Sci. 116, 4156–4165.
(doi:10.1073/pnas.1804597116)
43. Breiman L, Friedman JH, Olshen RA, Stone CJ. 2017 Classification and regression trees.
Routledge.
44. Chipman HA, George EI, McCulloch RE. 2010 BART: Bayesian additive regression trees. Ann.
Appl. Stat. 4, 266–298. (doi:10.1214/09-AOAS285)
45. Dorie V, Hill J, Shalit U, Scott M, Cervone D. 2019 Automated versus do-it-yourself methods
for causal inference: lessons learned from a data analysis competition. Stat. Sci. 4, 43–68.
(doi:10.1214/18-STS667)
46. Hu L, Ji J, Li F. 2021 Estimating heterogeneous survival treatment effect in observational data
using machine learning. Stat. Med. 40, 4691–4713. (doi:10.1002/sim.9090)
47. Ray K, van der Vaart A. 2020 Semiparametric Bayesian causal inference. Ann. Stat. 48, 2999–
3020. (doi:10.1214/19-AOS1919)
48. Chib S, Hamilton BH, 2002 Semiparametric Bayes analysis of longitudinal data treatment
models. J. Econom. 110, 67–89. (doi:10.1016/S0304-4076(02)00122-7)
49. Karabatsos G, Walker SG. 2012 A Bayesian nonparametric causal model. J. Stat. Plan. Inference
142, 925–934. (doi:10.1016/j.jspi.2011.10.013)
50. Oganisian A, Roy JA. 2021 A practical introduction to Bayesian estimation of causal effects:
parametric and nonparametric approaches. Stat. Med. 40, 518–551. (doi:10.1002/sim.8761)
51. Roy J, Lum KJ, Zeldow B, Dworkin JD, Re III VL, Daniels MJ. 2018 Bayesian nonparametric
generative models for causal inference with missing at random covariates. Biometrics 74,
1193–1202. (doi:10.1111/biom.12875)
52. Papadogeorgou G, Li F. 2020 Discussion for ‘Bayesian regression tree models for causal
inference: regularization, confounding, and heterogeneous effects’. Bayesian Anal. 15, 1007–
1013.
53. Williams CK, Rasmussen CE. 2006 Gaussian processes for machine learning. Cambridge, MA:
MIT Press.
54. Linero AR, Yang Y. 2018 Bayesian regression tree ensembles that adapt to smoothness and
sparsity. J. R. Stat. Soc. B 80, 1087–1110. (doi:10.1111/rssb.12293)
55. Antonelli JL, Parmigiani G, Dominici F. 2019 High-dimensional confounding adjustment
using continuous spike and slab priors. Bayesian Anal. 14, 805–828. (doi:10.1214/18-BA1131)
56. Park T, Casella G. 2008 The Bayesian lasso. J. Am. Stat. Assoc. 103, 681–686. (doi:10.1198/
016214508000000337)
57. Raftery AE, Madigan D, Hoeting JA. 1997 Bayesian model averaging for linear regression
models. J. Am. Stat. Assoc. 92, 179–191. (doi:10.1080/01621459.1997.10473615)
 58. Wang C, Parmigiani G, Dominici F. 2012 Bayesian effect estimation accounting for
22
adjustment uncertainty. Biometrics 68, 661–671. (doi:10.1111/j.1541-0420.2011.01731.x)
59. Zigler CM. 2016 The central role of Bayes’ theorem for joint estimation of causal effects and

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
propensity scores. Am. Stat. 70, 47–54. (doi:10.1080/00031305.2015.1111260)
60. Robins JM, Ritov Y. 1997 Toward a curse of dimensionality appropriate
(CODA) asymptotic theory for semi-parametric models. Stat. Med. 16, 285–319.
(doi:10.1002/(SICI)1097-0258(19970215)16:3<285::AID-SIM535>3.0.CO;2-%23)
61. D’Amour A, Ding P, Feller A, Lei L, Sekhon J. 2021 Overlap in observational studies with
high-dimensional covariates. J. Econom. 221, 644–654.
62. Linero AR. 2021 In nonparametric and high-dimensional models, Bayesian ignorability is an
informative prior. Preprint (https://arxiv.org/abs/2111.05137).
63. Ritov Y, Bickel BJ, Gamst AC, Kleijn BJ. 2014 The Bayesian analysis of complex, high-
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

dimensional models: can it be CODA. Stat. Sci. 29, 619–639. (doi:10.1214/14-STS483)

64. Harmeling S, Touissant M. 2007 Bayesian estimators for Robins-Titov’s problem. Technical
report, School of Informatics, University of Edinburgh. See https://argmin.lis.tu-berlin.de/
papers/07-harmeling-tr.pdf.
65. Sims CA. 2012 Robins-wasserman, round N. See http://sims.princeton.edu/yftp/
WassermanExmpl/WassermanR4a.pdf.
66. Antonelli JL, Papadogeorgou G, Dominici F. 2022 Causal inference in high dimensions: a
marriage between Bayesian modeling and good frequentist properties. Biometrics 78, 100–
114. (doi:10.1111/biom.13417)
67. Chernozhukov V, Chetverikov D, Demirer M, Duflo E, Hansen C, Newey W. 2017
Double/debiased/neyman machine learning of treatment effects. Am. Econom. Rev. 107,
261–265. (doi:10.1257/aer.p20171038)
68. Rubin DB. 1985 The use of propensity score in applied Bayesian inference. In Bayesian
statistics, vol. 2 (eds JM Bernardo, MH DeGroot, DV Lindley, AFM Smith), pp. 463–472.
North-Holland: Elsevier Science Publisher B.V.
69. Zigler CM, Watts K, Yeh RW, Wang Y, Coull BA, Dominici F. 2013 Model
feedback in Bayesian propensity score estimation. Biometrics 69, 263–273. (doi:10.1111/j.
1541-0420.2012.01830.x)
70. Little RJA, An H. 2004 Robust likelihood-based analysis of multivariate data with missing
values. Stat. Sin. 14, 949–968.
71. McCandless LC, Gustafson P, Austin PC. 2009 Bayesian propensity score analysis for
observational data. Stat. Med. 28, 94–112. (doi:10.1002/sim.3460)
72. Robins JM, Hernán MA, Wasserman L. 2015 Discussion of ‘Bayesian estimation of marginal
structural models’ by Saarela et al. Biometrics 71, 293–296. (doi:10.1111/biom.12273)
73. George EI, McCulloch RE. 1997 Approaches for Bayesian variable selection. Stat. Sin. 1, 339–
373.
74. Little RJA. 2004 To model or not to model? Competing modes of inference for finite
population sampling. J. Am. Stat. Assoc. 99, 546–556. (doi:10.1198/016214504000000467)
75. Saarela O, Belzile LR, Stephens DA. 2016 A Bayesian view of doubly robust causal inference.
Biometrika 103, 667–681. (doi:10.1093/biomet/asw025)
76. Ding P, Guo T. 2023 Posterior predictive propensity scores and p-values. Observational Studies
9, 3–18. (doi:10.1353/obs.2023.0015)
77. Lyddon SP, Holmes CC, Walker SG. 2019 General Bayesian updating and the loss-likelihood
bootstrap. Biometrika 106, 465–478. (doi:10.1093/biomet/asz006)
78. Ferguson TS. 1957 Dangers of cigarette-smoking. BMJ 2, 297–298.
79. Cornfield J, Haenszel W, Hammond EC, Lilienfeld AM, Shimkin MB, Wynder EL. 1959
Smoking and lung cancer: recent evidence and a discussion of some questions. J. Natl. Cancer
Inst. 22, 173–203.
80. Rosenbaum PR, Rubin DB. 1983 Assessing sensitivity to an unobserved binary covariate in
an observational study with binary outcome. J. R. Stat. Soc. B 45, 212–218.
81. Ichino A, Mealli F, Nannicini T. 2008 From temporary help jobs to permanent employment:
what can we learn from matching estimators and their sensitivity? J. Appl. Econ. 23, 305–327.
(doi:10.1002/jae.998)
82. Imbens GW. 2003 Sensitivity to exogeneity assumptions in program evaluation. Am. Econ.
Rev. 93, 126–132. (doi:10.1257/000282803321946921)
 83. Dorie V, Harada M, Carnegie NB, Hill J. 2016 A flexible, interpretable framework for
23
assessing sensitivity to unmeasured confounding. Stat. Med. 35, 3453–3470. (doi:10.1002/
sim.6973)

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
84. Ding P, VanderWeele TJ. 2016 Sensitivity analysis without assumptions. Epidemiology 27, 368–
377. (doi:10.1097/EDE.0000000000000457)
85. VanderWeele TJ, Ding P. 2017 Sensitivity analysis in observational research: introducing the
E-value. Ann. Intern. Med. 167, 268–274. (doi:10.7326/M16-2607)
86. Robins JM. 1999 Association, causation, and marginal structural models. Synthese 121, 151–
179. (doi:10.1023/A:1005285815569)
87. Rosenbaum PR. 2002 Observational studies. New York: Springer.
88. Rosenbaum PR, Silber JH. 2009 Amplification of sensitivity analysis in matched
observational studies. J. Am. Stat. Assoc. 104, 1398–1405. (doi:10.1198/jasa.2009.tm08470)
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

89. Angrist JD, Pischke J-S. 2009 Mostly harmless econometrics: an empiricist’s companion. Princeton
University Press.
90. Angrist JD, Imbens GW, Rubin DB. 1996 Identification of causal effects using instrumental
variables. J. Am. Stat. Assoc. 91, 444–455. (doi:10.1080/01621459.1996.10476902)
91. Imbens GW, Rubin DB. 1997 Bayesian inference for causal effects in randomized experiments
with noncompliance. Ann. Stat. 25, 305–327. (doi:10.1214/aos/1034276631)
92. Hirano K, Imbens GW, Rubin DB, Zhou XH. 2000 Assessing the effect of an influenza vaccine
in an encouragement design. Biostatistics 1, 69–88. (doi:10.1093/biostatistics/1.1.69)
93. Frangakis CE, Rubin DB. 2002 Principal stratification in causal inference. Biometrics 58, 21–29.
(doi:10.1111/j.0006-341X.2002.00021.x)
94. Frangakis CE, Rubin DB, Zhou XH. 2002 Clustered encouragement designs with individual
noncompliance: Bayesian inference with randomization, and application to advance
directive forms. Biostatistics 3, 147–164. (doi:10.1093/biostatistics/3.2.147)
95. Mealli F, Pacini B. 2013 Using secondary outcomes to sharpen inference in
randomized experiments with noncompliance. J. Am. Stat. Assoc. 108, 1120–1131.
(doi:10.1080/01621459.2013.802238)
96. Zeng S, Li F, Ding P. 2020 Is being an only child harmful to psychological health?: evidence
from an instrumental variable analysis of China’s one-child policy. J. R. Stat. Soc. A 15, 1615–
1635. (doi:10.1111/rssa.12595)
97. Heckman JJ, Vytlacil EJ. 1999 Local instrumental variables and latent variable models
for identifying and bounding treatment effects. Proc. Natl Acad. Sci. USA 96, 4730–4734.
(doi:10.1073/pnas.96.8.4730)
98. Zhang JL, Rubin DB, Mealli F. 2008 Evaluating the effects of job training programs on
wages through principal stratification. In Applied Bayesian modeling and causal inference from
incomplete-data perspectives, pp. 117–145. New York, NY: John Wiley & Sons.
99. Gilbert PB, Hudgens MG. 2008 Evaluating candidate principal surrogate endpoints.
Biometrics 64, 1146–1154. (doi:10.1111/j.1541-0420.2008.01014.x)
100. Jiang Z, Ding P, Geng Z. 2016 Principal causal effect identification and surrogate end point
evaluation by multiple trials. J. R. Stat. Soc. B 78, 829–848. (doi:10.1111/rssb.12135)
101. Li F, Mattei A, Mealli F. 2015 Evaluating the causal effect of university grants on student
dropout: evidence from a regression discontinuity design using principal stratification. Ann.
Appl. Stat. 9, 1906–1931. (doi:10.1214/15-AOAS881)
102. Ricciardi F, Mattei A, Mealli F. 2020 Bayesian inference for sequential treatments under
latent sequential ignorability. J. Am. Stat. Assoc. 115, 1498–1517. (doi:10.1080/01621459.
2019.1623039)
103. Robins JM. 1986 A new approach to causal inference in mortality studies with sustained
exposure periods—Application to control of the healthy worker survivor effect. Math. Modell.
7, 1393–1512. (doi:10.1016/0270-0255(86)90088-6)
104. Robins JM, Hernán MA, Brumback B. 2000 Marginal structural models and causal inference.
Epidemiology 11, 550–560. (doi:10.1097/00001648-200009000-00011)
105. Zajonc T. 2012 Bayesian inference for dynamic treatment regimes: mobility, equity,
and efficiency in student tracking. J. Am. Stat. Assoc. 107, 80–92. (doi:10.1080/01621459.
2011.643747)
106. Gustafson P. 2015 Discussion of ‘Bayesian estimation of marginal structural models’ by
Saarela et al. Biometrics 71, 291–293. (doi:10.1111/biom.12271)
 107. Robins JM, Wasserman L. 1997 Estimation of effects of sequential treatments by
24
reparameterizing directed acyclic graphs. In Proc. of the Thirteenth Conf. on Uncertainty in
Artificial Intelligence, pp. 409–420. Burlington, MA: Morgan Kaufmann Publishers Inc.

royalsocietypublishing.org/journal/rsta Phil. Trans. R. Soc. A 381: 20220153

...............................................................
108. Saarela O, Stephens DA, Moodie EEM, Klein MB. 2015 On Bayesian estimation of marginal
structural models. Biometrics 71, 279–301. (doi:10.1111/biom.12269)
109. Chakraborty B, Murphy SA. 2014 Dynamic treatment regimes. Ann. Rev. Stat. Appl. 1, 447–
464. (doi:10.1146/annurev-statistics-022513-115553)
110. Scott EL. 2010 A modern Bayesian look at the multi-armed bandit. Appl. Stoch. Models Bus.
Ind. 26, 639–658. (doi:10.1002/asmb.874)
111. Ghavamzadeh M, Mannor S, Pineau J, Tamar A. 2015 Bayesian reinforcement learning: a
survey. Found. Trends Mach. Learn. 8, 359–483. (doi:10.1561/2200000049)
112. Schnell PM, Papadogeorgou G. 2020 Mitigating unobserved spatial confounding when
Downloaded from https://royalsocietypublishing.org/ on 17 January 2025

estimating the effect of supermarket access on cardiovascular disease deaths. Ann. Appl. Stat.
14, 2069–2095. (doi:10.1214/20-AOAS1377)
113. Daniels MJ, Roy JA, Kim C, Hogan JW, Perri MG. 2012 Bayesian inference for the causal
effect of mediation. Biometrics 68, 1028–1036. (doi:10.1111/j.1541-0420.2012.01781.x)
114. Forastiere L, Mealli F, Wu A, Airoldi E. 2022 Estimating causal effects under interference
using Bayesian generalized propensity scores. J. Mach. Learn. Res. 23, 1–61.
115. Zeng S, Rosenbaum S, Alberts SC, Archie EA, Li F. 2021 Causal mediation analysis for sparse
and irregular longitudinal data. Ann. Appl. Stat. 15, 747–767. (doi:10.1214/20-AOAS1427)
116. Stan Development Team. 2022 RStan: the R interface to Stan. R package version 2.21.5.