Causal Inference Tutorial

Rahul Singh

Original: July 23, 2019; Updated: September 10, 2020

The goal of this tutorial is to introduce central concepts, algorithms, and techniques of causal inference for
a machine learning audience.

There are three sections.

1. Causal frameworks. I present the three most common languages for expressing causal assumptions:
potential outcomes, DAGs, and moment restrictions. These are three self-contained axiomatic frame-
works, with trade-offs in interpretability, expressiveness, and tractability. Roughly speaking, these
frameworks were developed by statisticians and epidemiologists; computer scientists; and economists,
respectively. A causal parameter is identified if it can be expressed as a function of data at the pop-
ulation level; it must be the case that two parameter values are not observationally equivalent. As a
running example, I present the relevant assumptions for identification of average treatment effect using
each of the three languages.
2. Reduced form toolkit. The previous section discusses causal assumptions at the population level.
How are these concepts brought to data? I introduce a toolkit for causal inference in observational
data that requires relatively few assumptions about the data generating process–specifically, it involves
assumptions about treatment assignment and adherence. This toolkit is called the reduced form toolkit
in econometrics because it deals with consequences of equilibrium rather than equilibrium itself. I refer
to some ways in which this toolkit is being reimagined with ML estimators.
3. De-biased machine learning. This section presents a different paradigm for combining ML and causal
inference: delegate prediction tasks to black-box ML estimators, and create an appropriate harness
around the ML estimators for valid causal inference. I present the DML algorithm, and I give references
for both its econometric theory and its statistical learning theory.

At the end of each section, I list further topics, canonical references, and relevant courses at MIT. Since
this tutorial summarizes a great deal of material across disciplines, I only refer to works that summarize a
topic (rather than all the papers in the intellectual history of that topic) or that present a machine learning
extension. Comprehensive citations will be found therein.

Contents

1 Causal frameworks 3

1.1 Potential outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.1.1 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.1.2 Treatment effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 DAGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1
 1.2.1 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.2 Treatment effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Moment restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.3.1 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.3.2 Treatment effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.4 Further topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.4.1 G-formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.4.2 Front door criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.4.3 Instrumental variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.4.4 Structural equation modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.4.5 Panel data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.4.6 Mediation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.4.7 Causal discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.5 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.5.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.5.2 Courses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Reduced form toolkit 9

2.1 Sub-classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3 Propensity score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.4 Further topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.4.1 Synthetic control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.4.2 Regression discontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.4.3 Two stage least squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.5 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.5.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.5.2 Courses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 De-biased machine learning 12

3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2 Influence adjustment and double robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2
 3.3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.4 Further topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.4.1 Econometric theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.4.2 Statistical learning theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.5 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.5.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.5.2 Courses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1 Causal frameworks

1.1 Potential outcomes

1.1.1 Framework

Let’s fix the following notation. Y is the outcome of interest, D is the treatment, and X are control variables.
For a given individual, we observe (Y, D, X). I suppress individual index i for readability. In the potential
outcomes framework, a given individual also possesses latent variables {Yd } where d ∈ D = supp(D). For
fixed treatment value d, Yd is the potential outcome an individual would experience had they received
treatment d. The outcome we actually observe, Y , is equal to the potential outcome Yd iff D = d.

Philosophically, in the potential outcomes framework we believe an individual contains many, latent selves.
For example, consider the treatment variable indicating attendance at this tutorial: D = 1 means attendance
and D = 0 means non-attendance. Let Y be research productivity. There exist two potential Anish’s, each
with a different productivity: Y1 is Anish’s productivity had he attended; and Y0 is Anish’s productivity had
he not attended. Since in reality D = 1 for Anish, his actual productivity Y will coincide with Y1 , and we
will never observe Anish’s Y0 . Formally, with binary treatment,

Y = DY1 + (1 − D)Y0

The individual treatment effect of the tutorial on Anish’s productivity is Y1 − Y0 . But we have a missing
data problem: we only ever observe Y1 or Y0 for a given individual. For this reason, we consider a population
average treatment effect instead:
θ0 = E[Y1 − Y0 ]
We now have a causal language to ask: for the average PhD, what is the productivity effect of attending the
tutorial vs. not?

1.1.2 Treatment effect

How can we formalize the above discussion into axioms that identify the average treatment effect? There are
three key axioms

1. Consistency: if D = d then Y = Yd . Recall our story about Anish’s productivity. This assumption
has more substance than it seems. It rules out interference across units: potential outcome of unit
i is unaffected by treatment of unit j. This condition is also called the stable unit treatment value
assumption. It would be violated if we were studying vaccines, for example. Or if Devavrat’s attendance
at the tutorial affects Anish’s potential productivity.

3
 2. Positivity: if P(X = x) > 0 then P(D = d|x) ∈ (0, 1). There are no PhD characteristics (# years,
subfield, etc.) that deterministically pin down treatment status. This condition is also called common
support.
3. Conditional exchangeability: {Yd } D|X. Here we formalize the notion that, conditional on controls

|=
X, treatment assignment D is as good as random; the observational data are quasi-experimental. This
condition is also called selection on observables. If we are interested in the effect of lab D (e.g. CSAIL
vs. LIDS) on research productivity Y , conditional exchangeability could be satisfied taking controls
X to be the set of labs to which an individual receives admission. Conditional on a choice set, it is
plausibly random which lab a PhD chooses.

Finally, we see our first identification argument. Under the three assumptions above

θ0 = E[Y1 − Y0 ]
Z
= {E[Y1 |x] − E[Y0 |x]}dP(x)
Z
= {E[Y1 |D = 1, x] − E[Y0 |D = 0, x]}dP(x)
Z
= {E[Y |D = 1, x] − E[Y |D = 0, x]}dP(x)

Note that we have expressed θ0 as a function of data at the population level; ATE is identified.

1.2 DAGs

1.2.1 Framework

I review the concept of directed acyclic graphs (DAGs), introduce causal DAGs, and finally present single
world intervention graphs (SWIGs) as a way to relate causal DAGs with potential outcomes.

A DAG is a graph whose vertices are random variables V = (V0 , ..., VJ ) with directed edges and no directed
cycles. Denote by pa(Vj ) the set of vertices from which there is a direct edge into Vj . A DAG represents the
joint density of V iff P(V ) = ΠJj=1 P(Vj |pa(Vj )). A given vertex Vj may or may not be observed.

Figure 1: DAG [23]

D-separation is a topological property of a DAG. If there are no active paths from A to B when C1 , ..., CK are
observed (shaded), then A d B|{C1 , ..., CK }. The rules governing whether a path is active are summarized
|=

by Bayes ball. In the figure below, an undirected path is active if a Bayes ball traveling along it never
encounters a stop symbol.

4
 Figure 2: The 10 rules of Bayes ball [21]

D-separation implies statistical independence. However, d-connection does not imply statistical dependence.
The faithfulness assumption patches this discrepancy.

A causal DAG is a DAG associated to a causal model. A lack of an arrow from Vj to Vk means the absence
of direct causal effect of Vj on Vk . Any variable is a cause of all its descendants. Temporally, a cause must
take place before an effect.

1.2.2 Treatment effect

Can we formalize the above discussion into axioms that identify the average treatment effect? There are two
key conditions

1. Faithfulness: d-connection implies statistical dependence. Bayes ball becomes sufficient for analysis of
independence.
2. Back-door criterion: outcome Y and treatment D are d-separated conditional on the measured controls
X, in a graph in which the arrows out of D are removed. The back-door criterion may be satisfied if

(a) No common causes of Y and D

(b) No unmeasured confounding. Though there are common causes of Y and D, there are sufficiently
many and well-placed controls X such that conditioning on X blocks all backdoor paths in the
modified graph described above

As stated, the backdoor criterion is difficult to verify at a glance. Single world intervention graphs (SWIGs)
at once make the backdoor criterion easier to verify and formalize the connection between causal DAGs and
potential outcomes. The procedure to convert a DAG to a SWIG is called node-splitting. The procedure is
as follows.

Topologically order the DAG from left to right. Consider the intervention of setting D = d (in Pearl’s
language, do(D = d)). Decompose the vertex corresponding to D into a two vertices: a left side, and a right
side. The left side, labeled D, inherits all the arrows that previously led into the original D. The right side,
labeled D = d, inherits all the arrows that previously led out of the original D. The left side and right side
are not connected. All children of the right side are now potential outcomes indexed by d.

5
 Figure 3: SWIG [23]

Figure 4: SWIG with confounding [23]

Figure 5: SWIG with confounding [23]

Figure 6: SWIG with multiple treatments [24]

Once we have converted a DAG to a SWIG, we can verify conditional exchangeability via d-separation.

6
1.3 Moment restrictions

1.3.1 Framework

We have seen the potential outcome and causal DAG frameworks, which are close to equivalent. The
former is quite interpretable, while the latter is quite expressive. Identification in both frameworks relies
on conditional independences, which are rather strong assumptions. Most of econometrics relies on moment
restrictions instead: assumptions about the error terms in a structural model. We will see how moment
restrictions are quite tractable; functional form assumptions permit us to relax conditional independence to
conditional mean independence and even just orthogonality of error terms.

A moment restriction is of the form

E[ψ(W, θ)] = 0 iff θ = θ0
where W is the concatenation of all variables in an observation, and θ0 is the causal parameter of interest.
ψ is called the moment function. Estimation is then possible via generalized method of moments (GMM),
which can be viewed as a unifying framework for method of moments, maximum likelihood, linear regression,
nonlinear regression, and more.

As a toy example, consider the linear regression model. W = (Y, D, X), and we assume
  
D
Y = θ0 D + X 0 β + , E  =0
X
Here the moment restriction is
     
D 0 D
0 = E[ψ(W, θ0 )] = E  = (Y − θ0 D − X β)
X X
We have traded in strong functional form assumptions to weaken our independence requirement to orthogo-
nality. θ0 is identified as the ATE. This is a popular empirical strategy in economics papers.

1.3.2 Treatment effect

It may be the case that moment function ψ depends on additional nuisance parameters. Indeed, this is the
case for nonparametric estimation of ATE. In the exposition below, I show how the potential outcome story
about ATE can be reformulated as a moment restriction. This approach previews de-biased machine learning
(DML).

Define the conditional mean function

γ0 (D, X) = E[Y |D, X]
We have seen that under consistency, positivity, and conditional exchangeability, ATE is identified and
Z
θ0 = {E[Y |D = 1, x] − E[Y |D = 0, x]}dP(x) = E[γ0 (D = 1, X) − γ0 (D = 0, X)]

In other words, we have moment restriction

0 = E[ψ(W, θ0 , γ0 )] = E[γ0 (D = 1, X) − γ0 (D = 0, X) − θ0 ]
where W = (Y, D, X) and the moment function ψ has nuisance parameter γ0 . DML will explain how to deal
with γ0 and other nuisance parameters in a sophisticated manner.

1.4 Further topics

1.4.1 G-formula

What if the intervention of interest is a regime of treatments that depend on patient status over time?

7
1.4.2 Front door criterion

What if treatment and outcome are confounded, but they are mediated by only one variable that is not
confounded?

1.4.3 Instrumental variables

What if there is imperfect compliance, so that (randomized) treatment assignment and actual treatment are
different variables?

1.4.4 Structural equation modeling

What if we know a great deal about how the data are generated and want to learn structural parameters
that allow us to simulate events that have never happened, e.g. a merger?

1.4.5 Panel data

What if we observe a cohort of individuals over time?

1.4.6 Mediation

Can we discern the mechanism through which treatment affects outcome?

1.4.7 Causal discovery

Can we actually learn the edges of the causal DAG from data?

1.5 Resources

1.5.1 Literature

Readable introductions to potential outcomes are in [7, 17]. The classic reference for causal DAGs is [22].
SWIGs are covered in [17] as well. The classic reference for the moment restriction framework is [20].

1.5.2 Courses

In 14.387a, and starting this year in 14.381, Josh Angrist teaches the potential outcomes framework. In
6.438, Guy Bresler teaches graphical models. In 6.244, Caroline Uhler gives a more advanced treatment with
applications to causal discovery. Harvard School of Public Health offers a one-week summer course (free
for students) covering SWIGs and other tools developed by epidemiologists. In 14.385, Alberto Abadie and
Whitney Newey teach the moment restriction framework.

8
2 Reduced form toolkit

2.1 Sub-classification

Recall that under consistency, positivity, and conditional exhangeability,

Z
θ0 = {E[Y |D = 1, x] − E[Y |D = 0, x]}dP(x)

The idea of sub-classification is to approximate each component of this expression.

Assume control variable X ∈ X = supp(X) = {x1 , ..., xK }. X may be discrete or discretized into K cells.
Define
n
X
nkd = 1Xi =xk ,Di =d
i=1
1 X
Ȳdk = Yi
nkd
i:Xi =xk ,Di =d

nkd is the number of observations in cell k with treatment value d. Ȳdk is the mean outcome among observations
in cell k with treatment value d. Then the subclassification estimator is the sample analogue to θ0 :
K
X nk0 + nk1
θ̂ = [Ȳ1k − Ȳ0k ]
n
k=1

Clearly this estimator is infeasible with even moderately dimensional control X, but it is a reasonable starting
point.

2.2 Matching

The central idea of matching is to impute the unobserved potential outcome for each observation using
nearest neighbors of the opposite treatment status. If observation i is treated, find the M nearest neighbors
j1 (i), ..., jM (i) to i among untreated observations. Likewise, if observation i is untreated, find the M nearest
neighbors j1 (i), ..., jM (i) to i among treated observations. Observations i and j are close if kXi − Xj k is
small, where the norm is often taken to be Mahanalobis distance. The matching estimator is then
n
( M
)
1X 1 X
θ̂ = [2Di − 1] Yi − Yj (i)
n i=1 M m=1 m

Further work includes a bias correction term.

2.3 Propensity score

The propensity score is defined as π(x) = P(D = 1|X = x). I present three important roles that the
propensity score plays in causal inference:

1. rephrasing conditional exchangeability

2. dimension reduction in matching

3. new class of estimators based on Riesz representation

9
The first role is classical. The second is practical. The third is a more modern interpretation, previewing
DML.

I now show that conditional exchangeability can be rephrased in terms of the propensity score. Formally,

{Yd } D|X =⇒ {Yd }

|=

|=

|=
D|π(X), D X|π(X)

Assume {Yd }
|=
D|X.

P(D = 1|Y1 , Y0 , π(X)) = E[D|Y1 , Y0 , π(X)]

= EX,Y1 ,Y0 [E[D|X, Y1 , Y0 ]|Y1 , Y0 , π(X)]
= EX,Y1 ,Y0 [E[D|X]|Y1 , Y0 , π(X)]
= EX,Y1 ,Y0 [π(X)|Y1 , Y0 , π(X)]
= π(X)

Similarly

P(D = 1|π(X)) = E[D|π(X)]

= EX [E[D|X]|π(X)]
= EX [π(X)|π(X)]
= π(X)

so P(D = 1|Y1 , Y0 , π(X)) = P(D = 1|π(X)), i.e. {Yd }

|=

D|π(X). To see the second result, write

P(D = 1|X, π(X)) = P(D = 1|X) = π(X)

|=

hence D X|π(X).

This result is practical in the sense that we can now perform sub-classification or matching using the scalar
propensity score π(X) instead of the possibly high dimensional control variable X. It is also practical in
giving us a way to test a consequence of conditional exchangeability: we expect the balancing property to
hold:
P(X|D = 1, π(X)) = P(X|D = 0, π(X))

Finally, I turn to another important class of reduced form estimators that make use of the propensity score.
Observe that
   
D 1
E Y X =E Y D = 1, X P(D = 1|X)
π(X) π(X)
 
1
=E Y D = 1, X π(X)
π(X)
= E [Y |D = 1, X]

and likewise  
1−D
E Y X = E [Y |D = 0, X]
1 − π(X)
In summary, we can write ATE as
Z
θ0 = {E[Y |D = 1, x] − E[Y |D = 0, x]}dP(x)
Z     
D 1−D
= E Y X −E Y X dP(x)
π(X) 1 − π(X)
   
D 1−D
=E Y −E Y
π(X) 1 − π(X)

This expression suggests, by the analogy principle, a Horvitz-Thompson estimator of the ATE.

10
Written in terms of γ0 (D, X) = E[Y |D, X],
  
D 1−D
θ0 = E[γ0 (D = 1, X) − γ0 (D = 0, X)] = E γ0 (D, X) −
π(X) 1 − π(X)

The middle expression formulates ATE as a functional–a scalar-valued summary–of an underlying regression
γ0 . The last expression formulates ATE as an inner product between γ0 and the an object we will call the
Riesz representer α0  
D 1−D
α0 (D, X) = −
π(X) 1 − π(X)
Our ability to write ATE in both ways is no coincidence; by Riesz representation theorem, we are guaranteed
that any continuous linear functional can be represented as an inner product with that functional’s unique
Riesz representer. Written more generally, for functional m and its unique Riesz representer α0

θ0 = E[m(W, γ0 )] = E[α0 (D, X)γ0 (D, X)]

and indeed
E[m(W, γ)] = E[α0 (D, X)γ(D, X)], ∀γ ∈ L2 i.e. E[γ(D, X)]2 < ∞

2.4 Further topics

2.4.1 Synthetic control

Can we estimate the treatment effect on a treated unit by estimating its potential outcome had it not received
treatment as a convex combination of untreated neighbors (w.r.t. X)?

2.4.2 Regression discontinuity

What if treatment is a deterministic and discontinuous function of a control variable, e.g. D = 1X≥x0 for
threshold value x0 ?

2.4.3 Two stage least squares

How can we bring instrumental variable identification (recall the imperfect compliance story) to data?

2.5 Resources

2.5.1 Literature

See [26] for sub-classification, [2] for matching, and [25] for propensity score. The canonical reference for
synthetic control is [1].

The theory of regression discontinuity design is [15] with popular user guide [19]. In [10], the authors propose
a Gaussian process regression generalization.

See [6] for a potential outcomes analysis of instrumental variables, and [5] for its linear implementation as
two-stage least squares (2SLS). In [16], the authors consider a neural net generalization. In [27], the authors
consider an RKHS generalization with finite sample excess risk guarantees.

There is a recent literature bringing the technology of random forests to some of the methods described above
[31, 9]. There is another literature bringing the technology of matrix completion to these methods [8, 4, 3].

11
2.5.2 Courses

In 14.385, Alberto Abadie and Whitney Newey teach the reduced form toolkit. I’m the TA for that course.

3 De-biased machine learning

3.1 Motivation

In causal inference, we are often interested in a parameter defined as a functional of some underlying regres-
sion. We have already seen how ATE is one such example. Other common examples are average treatment
effect on the treated, average policy effect, and average consumer surplus. In this discussion, I restrict atten-
tion to DML applied to functionals. This is the original setting of DML, called semi-parametric estimation.
DML has been extended to a wide variety of settings since.

I briefly recap the ATE formalism we have seen before in the context of a concrete example. Consider the
following simple design. The horizontal axis is X and the vertical axis is Y. Observations (grey dots) are
generated from γ0 (D = 1, X) = E[Y |D = 1, X] (the orange curve) if treated and from γ0 (D = 0, X) =
E[Y |D = 0, X] (the blue curve) if untreated. The fact that observations are scattered reflects sampling from
the respective curves with noise. Note that γ0 (D = 1, X) − γ0 (D = 0, X) = 1 everywhere, so in particular
the ATE is θ0 = 1 by construction. Given access only to the grey dots W = (Y, D, X) and not the data
generating curves, we wish to estimate the ATE with true value θ0 = 1.

1
y

-1

-2

-3
E[Y|D=0,z]
E[Y|D=1,z]
data
-4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
z

Figure 7: CEF for treated vs. untreated. ATE is θ0 = 1

Recall that in the ATE example,

θ0 = E[m(W, γ0 )] = E[γ0 (D = 1, X) − γ0 (D = 0, X)]

A naive approach, suggested earlier, is as follows.

1. learn CEF with some ML estimator γ̂

2. plug into the functional
n n
1X 1X
θ̂ = m(Wi , γ̂) = [γ̂(D = 1, Xi ) − γ̂(D = 0, Xi )]
n i=1 n i=1

12
This is equivalent to method of moments with moment function

ψ(w, θ, γ) = m(w, γ) − θ

and nuisance parameter γ. Let’s see how the naive approach performs in 100 simulations.

Figure 8: Naive approach θ̂

This phenomenon, in which ML estimation of nuisance parameters badly biases estimation of causal param-
eters, motivates a more sophisticated approach: de-biased machine learning.

3.2 Influence adjustment and double robustness

Previously we used
ψ(w, θ, γ) = m(w, γ) − θ
Now use
ψ(w, θ, γ, α) = m(w, γ) − θ+α(d, x)[y − γ(d, x)]
which has nuisance parameters (γ, α). As our prior discussion foreshadowed, the new nuisance parameter
is the Riesz representer. The adjustment term is the product of Riesz representer α and regression residual
y − γ(x).

In what sense is this new moment function more sophisticated? I prove the following properties:

1. E[ψ(W, θ0 , γ, α0 )] = 0, ∀γ ∈ L2 . The argument is as follows.

E[ψ(W, θ0 , γ, α0 )] = E[m(W, γ) − θ0 + α0 (D, X)[Y − γ(D, X)]]

= E[m(W, γ) − θ0 + α0 (D, X)[γ0 (D, X) − γ(D, X)]]
= E[m(W, γ) − θ0 + m(W, γ0 ) − m(W, γ)]]
= E[m(W, γ) − m(W, γ0 ) + m(W, γ0 ) − m(W, γ)]]

2. E[ψ(W, θ0 , γ0 , α)] = 0, ∀α ∈ L2 . The argument is as follows.

E[ψ(W, θ0 , γ0 , α)] = E[m(W, γ0 ) − θ0 + α(D, X)[Y − γ0 (D, X)]]

= E[m(W, γ0 ) − θ0 + α(D, X)[γ0 (D, X) − γ0 (D, X)]]
= E[m(W, γ0 ) − m(W, γ0 ) + α(D, X)[γ0 (D, X) − γ0 (D, X)]]

13
Our new ψ it is not only a valid moment function, but also doubly robust: the moment function remains valid
even for incorrect values of γ or α. Estimation of θ is robust to first stage estimation error of either nuisance
parameter. Double robustness is a strong property, and much DML analysis requires the weaker condition
of local robustness. For general moment function ψ(w, θ, η) with nuisance parameter η, local robustness is
the property that
∂η Eψ(W, θ0 , η)|η=η0 = 0
This condition is also called Neyman orthogonality.

It turns out there is a rich theory that delivers a formula for finding locally robust moment functions.
Astonishingly, the locally robust moment function for a semi-parametric parameter θ0 is the limit of the
Gateaux derivative of the parameter θ0 with respect to a smooth deviation away from the true distribution
P as the deviation approaches a point mass.

3.3 Algorithm

I finally present the DML algorithm. It involves the doubly robust moment function and sample splitting.

Partition observations into L distinct subsets {I` }L

`=1

1. learn CEF and Riesz representer

(a) estimate γ̂` from observations not in I`

(b) estimate α̂` from observations not in I`
• manually (plug-in)
• automatically (Lasso, Dantzig selector)
2. DR method of moments
L
1 XX
θ̂ = {m(Wi , γ̂` ) + α̂` (Wi )[Yi − γ̂` (Xi )]}
n
`=1 i∈I`

Any consistent, black-box ML estimator can be used for γ̂. Let me clarify the distinction between manual
and automatic estimation of α̂ in the context of ATE.

First, manual estimation of α̂. In ATE, we have analytically derived the functional form
 
D 1−D
α0 (D, X) = −
π(X) 1 − π(X)

An analyst, knowing the functional form, may estimate α̂ by estimating the propensity score π̂ and plugging it
into the formula above. We may be concerned about plugging-in an estimated probability in a denominator.
For more complicated parameters, we may be concerned about deriving the Riesz representer’s functional
form.

Second, automatic estimation of α̂. Recent research shows that α0 is directly identified from the data W ,
and so we can use machine learning–e.g. Lasso or Dantzig selector–to estimate α̂ directly. In other words, we
do not have to estimate its components, or even know the functional form of true α0 . We can automatically
de-bias ML with more ML!

Let’s return to the simple design and see how DML with automatic Riesz representer estimation performs in
100 simulations. First recall the set-up.

14
 5

y
0

-1

-2

-3
E[Y|D=0,z]
E[Y|D=1,z]
data
-4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
z

Figure 9: CEF for treated vs. untreated. ATE is θ0 = 1

Here is the performance of DML with automatic Riesz representer.

Figure 10: DML with automatic Riesz representer θ̂

Finally, a side-by-side comparison.

15
 Figure 11: Naive approach vs. DML with automatic Riesz representer

3.4 Further topics

3.4.1 Econometric theory

Is it possible to derive asymptotic confidence intervals for the estimator θ̂?

3.4.2 Statistical learning theory

Is it possible to derive finite sample bounds on excess risk of θ̂?

3.5 Resources

3.5.1 Literature

The core econometric theory of DML is in [11]. The automatic Riesz representer extension of DML is in
[12, 13]. The core statistical learning theory of DML is in [14]. See [18] for a nice review of influence function
theory in DML, and [30, Chapter 25] for a rigorous statistical background. Finally, see [28, 29] for extensions
with instrumental variables.

3.5.2 Courses

In 14.387b, Victor Chernozhukov introduces DML more thoroughly. In 14.386, Whitney Newey teaches some
influence function theory (which delivers locally robust moment functions).

16
Conclusion

Causal inference spans statistics, epidemiology, computer science, and economics. There are three languages
to express causal assumptions and conclusions: potential outcomes, causal DAGs, and moment restrictions.
Recent research has begun to reimagine the reduced form toolkit with ML techniques such as matrix comple-
tion, random forests, neural networks, and RKHS methods. Another paradigm for bridging causal inference
and ML is DML, a meta-algorithm around black-box ML estimators that guarantees valid causal inference.
My email address is [email protected] and I would love to talk more!

References
[1] Alberto Abadie, Alexis Diamond, and Jens Hainmueller. Synthetic control methods for comparative
case studies: Estimating the effect of California’s tobacco control program. Journal of the American
Statistical Association, 105(490):493–505, 2010.
[2] Alberto Abadie and Guido W Imbens. Large sample properties of matching estimators for average
treatment effects. Econometrica, 74(1):235–267, 2006.
[3] Anish Agarwal, Devavrat Shah, Dennis Shen, and Dogyoon Song. On robustness of principal component
regression. arXiv:1902.10920, 2019.
[4] Muhammad Amjad, Devavrat Shah, and Dennis Shen. Robust synthetic control. The Journal of Machine
Learning Research, 19(1):802–852, 2018.
[5] Joshua D Angrist and Guido W Imbens. Two-stage least squares estimation of average causal effects in
models with variable treatment intensity. Journal of the American Statistical Association, 90(430):431–
442, 1995.
[6] Joshua D Angrist, Guido W Imbens, and Donald B Rubin. Identification of causal effects using instru-
mental variables. Journal of the American Statistical Association, 91(434):444–455, 1996.
[7] Joshua D Angrist and Jörn-Steffen Pischke. Mostly Harmless Econometrics: An Empiricist’s Compan-
ion. Princeton University Press, 2008.
[8] Susan Athey, Mohsen Bayati, Nikolay Doudchenko, Guido Imbens, and Khashayar Khosravi. Matrix
completion methods for causal panel data models. Technical report, National Bureau of Economic
Research, 2018.
[9] Susan Athey, Julie Tibshirani, and Stefan Wager. Generalized random forests. The Annals of Statistics,
47(2):1148–1178, 2019.
[10] Zach Branson, Maxime Rischard, Luke Bornn, and Luke W Miratrix. A nonparametric Bayesian method-
ology for regression discontinuity designs. Journal of Statistical Planning and Inference, 202:14–30, 2019.
[11] Victor Chernozhukov, Juan Carlos Escanciano, Hidehiko Ichimura, Whitney K Newey, and James M
Robins. Locally robust semiparametric estimation. arXiv:1608.00033, 2016.
[12] Victor Chernozhukov, Whitney K Newey, and Rahul Singh. Double/de-biased machine learning of global
and local parameters using regularized Riesz representers. arXiv:1802.08667, 2018.
[13] Victor Chernozhukov, Whitney K Newey, and Rahul Singh. Learning L2 continuous regression func-
tionals via regularized Riesz representers. arXiv:1809.05224, 2018.
[14] Dylan J Foster and Vasilis Syrgkanis. Orthogonal statistical learning. arXiv:1901.09036, 2019.
[15] Jinyong Hahn, Petra Todd, and Wilbert Van der Klaauw. Identification and estimation of treatment
effects with a regression-discontinuity design. Econometrica, 69(1):201–209, 2001.
[16] Jason Hartford, Greg Lewis, Kevin Leyton-Brown, and Matt Taddy. Deep IV: A flexible approach for
counterfactual prediction. In International Conference on Machine Learning, pages 1414–1423, 2017.

17
[17] Miguel A Hernan and James M Robins. Causal Inference. CRC Press, 2010.
[18] Hidehiko Ichimura and Whitney K Newey. The influence function of semiparametric estimators.
arXiv:1508.01378, 2015.
[19] David S Lee and Thomas Lemieux. Regression discontinuity designs in economics. Journal of Economic
Literature, 48(2):281–355, 2010.
[20] Whitney K Newey and Daniel McFadden. Large sample estimation and hypothesis testing. Handbook
of Econometrics, 4:2111–2245, 1994.
[21] Mark Paskin. A short course on graphical models, 2003.

[22] Judea Pearl. Causality. Cambridge University Press, 2009.

[23] James Robins. Graphical models for causal inference, 2017.
[24] James Robins. Time-varying exposures and the g-formula, 2018.
[25] Paul R Rosenbaum and Donald B Rubin. The central role of the propensity score in observational
studies for causal effects. Biometrika, 70(1):41–55, 1983.
[26] Paul R Rosenbaum and Donald B Rubin. Reducing bias in observational studies using subclassification
on the propensity score. Journal of the American Statistical Association, 79(387):516–524, 1984.
[27] Rahul Singh, Maneesh Sahani, and Arthur Gretton. Kernel instrumental variable regression.
arXiv:1906.00232, 2019.

[28] Rahul Singh and Liyang Sun. De-biased machine learning for compliers. arXiv:1909.05244, 2019.
[29] Vasilis Syrgkanis, Victor Lei, Miruna Oprescu, Maggie Hei, Keith Battocchi, and Greg Lewis. Machine
learning estimation of heterogeneous treatment effects with instruments. arXiv:1905.10176, 2019.
[30] Aad W Van der Vaart. Asymptotic Statistics, volume 3. Cambridge University Press, 2000.

[31] Stefan Wager and Susan Athey. Estimation and inference of heterogeneous treatment effects using
random forests. Journal of the American Statistical Association, 113(523):1228–1242, 2018.

18