An introduction to Causal Modelling

Gauranga Kumar Baishya

Instructor
Prof. M. R. Srinivasan
Professor, Chennai Mathematical Institute
mrsvasan@[Link]

May 14, 2025

2
 Acknowledgment

This work is based on a “Guided Project” where I read existing and latest literature in
the field of Causal Inference, modelling & learning.
My most sincere gratitude goes to Professor M.R. Srinivasan for his unwavering guidance
and unparalled support throughout time and course.

3
4
Contents

1 Introduction to Causal Inference 7

1.1 Potential Outcomes and Individual Causal Effects . . . . . . . . . . . . . 7
1.2 Average Causal Effects and Effect-Measure Scales . . . . . . . . . . . . . 8
1.3 Association versus Causation . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 RCT, Exchangeability, and Consistency . . . . . . . . . . . . . . . . . . . 11
1.5 Conditional Randomization, Stratification, and Blocking . . . . . . . . . 11
1.6 Crossover Experiments and Period Effects . . . . . . . . . . . . . . . . . 13
1.7 Inverse Probability Weighting (IPW) in Practice . . . . . . . . . . . . . . 13
1.8 Suggested Readings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2 Randomized Experiments 15
2.1 Selection Bias: Conceptual Foundations . . . . . . . . . . . . . . . . . . . 16
2.2 RCT: Assumptions and Theorem . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Sampling Framework for an RCT . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Point Estimation of Group Means . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Inference for the Average Treatment Effect . . . . . . . . . . . . . . . . . 18
2.6 The Delta Method and Relative Effectiveness . . . . . . . . . . . . . . . . 19
2.7 Covariate Adjustment and Treatment Effect . . . . . . . . . . . . . . . . 20
2.8 Limitations of RCTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.9 Conclusion & Suggested Readings . . . . . . . . . . . . . . . . . . . . . . 22

3 Double ML 23
3.1 Frisch–Waugh–Lovell (FWL) Partialling-Out . . . . . . . . . . . . . . . . 24
3.2 Neyman Orthogonality. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
√
3.3 Asymptotic Theory: n-Normality of α̂ . . . . . . . . . . . . . . . . . . 27
3.4 Conditional Convergence in Growth Economics . . . . . . . . . . . . . . 28
3.5 Beyond a Single Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.6 Alternative Orthogonal Methods . . . . . . . . . . . . . . . . . . . . . . . 29
3.7 Practical Checklist for Empirical Implementation . . . . . . . . . . . . . 31
3.8 Notebooks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.9 Suggested Readings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4 Metrics to Recover Causal Effects 33

4.1 Potential Outcomes Framework . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Linear Regression as a Conditioning Device . . . . . . . . . . . . . . . . . 35
4.3 The Horvitz–Thompson Method . . . . . . . . . . . . . . . . . . . . . . . 36
4.4 A Remark on Precision and Propensity Scores . . . . . . . . . . . . . . . 37
4.5 Covariate Balance Checks . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5
6 CONTENTS

4.6 Directed Acyclic Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.7 Putting the Pieces Together . . . . . . . . . . . . . . . . . . . . . . . . . 40

5 Graphical Models 41
5.1 General DAG and SEM via an Example . . . . . . . . . . . . . . . . . . 41
5.2 Conditional Ignorability and Exogeneity . . . . . . . . . . . . . . . . . . 43
5.3 DAGs, SEMs, and d-Separation . . . . . . . . . . . . . . . . . . . . . . . 45
5.4 Intervention, Counterfactual DAGs, and SWIGs . . . . . . . . . . . . . . 48
5.5 The Backdoor Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.6 Notebook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.7 Suggested Readings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Chapter 1

Introduction to Causal Inference

Why Causality Matters

Nearly every scientific or policy question that moves beyond description—from “Will
a carbon tax curb emissions?” to “Does aspirin prevent a second heart attack?”— is
inherently causal. Descriptive statistics can tell us what is, but acting wisely requires
knowing what would happen under different actions. Causal inference thus seeks to answer
counterfactual questions:
What would the outcome have been, had a different choice been made?
Historically, the formal study of causality lay scattered across philosophy, epidemiology,
statistics, and economics. Modern frameworks—notably the potential-outcomes model
and graphical DAG approaches—have unified these threads, giving us a precise language,
clear assumptions, and powerful tools for estimation and criticism. The remainder of this
single high-level chapter develops the essential building blocks of causal thinking, weaving
together intuitive stories, mathematical notation, and concrete examples.

1.1 Potential Outcomes and Individual Causal Effects

Historical roots. The potential-outcomes idea dates to Neyman(1923) in agricultural
field trials and was popularized by Rubin in the 1970s. Its genius is to treat each unit as
carrying a vector of latent responses—one for every treatment value—thereby separating
the science (how Nature assigns those latent responses) from the statistics (how the analyst
observes and aggregates them).

Notation and basic definitions. Let A ∈ {0, 1} be a binary treatment (with A = 1

for the active treatment and A = 0 for control) and let Y ∈ {0, 1} be a binary outcome
measured at the end of the study. For every individual i we postulate two potential
outcomes
Yi1 , Yi0 (equivalently written Yi (a) for a ∈ {0, 1}),
and collect them in the vector of potential outcomes
Yi = (Yi0 , Yi1 ).

1. These potential outcomes are well-defined only under the Stable Unit Treatment
Value Assumption (SUTVA):

7
8 CHAPTER 1. INTRODUCTION TO CAUSAL INFERENCE

(a) No multiple versions of treatment: each value of A corresponds to a unique,

well-specified intervention.
(b) No interference: an individual’s outcome depends only on their own treatment
assignment, not on the assignments of others.

2. The observed outcome is a reveal of exactly one coordinate:

Yi = YiAi .

The unobserved coordinate is called the counterfactual outcome.

The individual causal effect is defined as the contrast

∆i = Yi1 − Yi0 ∈ {−1, 0, 1},

indicating whether, for that person, treatment

1. harms (∆i = −1),

2. has no effect (∆i = 0), or

3. benefits (∆i = 1).

Because, for each individual, we can ever observe at most one of the two potential outcomes,
the pair (Yi0 , Yi1 ) is never jointly observed, making ∆i intrinsically unidentifiable. This is
not merely a statistical limitation but a logical one, often referred to as the fundamental
problem of causal inference.

Example. Zeus undergoes a heart transplant (A = 1) and dies on the fifth post-operative
day (Y = 1). Consequently we observe the treated potential outcome
1
YZeus = 1,
0
while the untreated potential outcome YZeus is fundamentally unobservable. Any statement
about what would have happened to Zeus without transplantation must rely on data
from other patients who did not receive a transplant, together with assumptions (e.g.
conditional ignorability) that make Zeus “exchangeable” with those patients. In this way,
information is borrowed from comparable untreated individuals to draw causal conclusions
about Zeus’s counterfactual outcome.

1.2 Average Causal Effects and Effect-Measure Scales

From individuals to populations. Recall that the individual causal effect ∆i =
Yi1 − Yi0 is never fully observed for any one person. Nevertheless, scientific inference and
policy decisions typically target average effects over a population or subpopulation. The
most fundamental estimand is the average treatment effect (ATE), defined by

ATE = E[Y 1 − Y 0 ] = E[Y 1 ] − E[Y 0 ].

In many applications one also considers two refined, subpopulation-specific estimands:

1.3. ASSOCIATION VERSUS CAUSATION 9

ATT Average Treatment Effect on the Treated:

ATT = E[Y 1 − Y 0 | A = 1],

which answers, “What is the average effect among those who actually received
treatment?”

ATC Average Treatment Effect on the Untreated (Controls):

ATC = E[Y 1 − Y 0 | A = 0],

useful for planning expansions of a program to currently untreated individuals.

Risk-based summary measures. When the outcome Y is binary, additional scale

choices help convey effect direction and public-health impact. Three common measures
are:

Risk Difference (RD):

RD = P (Y 1 = 1) − P (Y 0 = 1).

This absolute scale aligns with notions such as the “number needed to treat” (NNT =
1/RD).

Risk Ratio (RR):

P (Y 1 = 1)
RR = .
P (Y 0 = 1)
A value below 1 (e.g. RR = 0.6) can be read as “40% relative reduction.”

Odds Ratio (OR):

P (Y 1 = 1)/P (Y 1 = 0)
OR = .
P (Y 0 = 1)/P (Y 0 = 0)
Widely used in case–control studies and logistic regression. Note that for common
outcomes (prevalence > 10%), the OR can substantially overstate the RR.

Choosing a scale. The statistical efficiency, interpretability, and relevance of an effect-

measure scale depend on the scientific context (additive vs. multiplicative mechanism),
the target audience, and the study design. Always match the scale to both the underlying
causal model and the stakeholders’ needs.

1.3 Association versus Causation

Conditional versus Marginal Risk. Consider two distinct notions of risk:

P (Y = 1 | A = a) versus P (Y a = 1) .
| {z } | {z }
observed conditional risk counterfactual (marginal) risk

The first quantity, P (Y = 1 | A = a), is an observed conditional risk. It represents the

probability of the outcome given that treatment status A is observed to be a. Thus, this
measure is derived solely from individuals who actually received treatment a.
10 CHAPTER 1. INTRODUCTION TO CAUSAL INFERENCE

In contrast, the second quantity, P (Y a = 1), is a counterfactual (marginal) risk,

representing the probability that the outcome would be observed if the entire population
were universally assigned to treatment a. This counterfactual risk cannot typically be
observed directly, since it involves imagining the outcome under a hypothetical intervention,
regardless of individuals’ actual treatment assignments.
These two risks are conceptually distinct. The equality:

P (Y = 1 | A = a) = P (Y a = 1)

holds only under strong conditions, notably unconfoundedness, also known as condi-
tional exchangeability, where treatment assignment is independent of potential outcomes.
Such conditions are naturally satisfied in randomized trials but are rarely guaranteed
in observational studies, making direct equivalence of these risks uncommon in practice.
Careful methods of adjustment, conditioning, or weighting are typically required to make
valid causal inferences from observational data.

Confounding and Simpson’s Paradox. A variable L is said to confound the rela-

tionship between treatment A and outcome Y if it simultaneously influences both A and
Y and is not an intermediate variable on the causal pathway from A to Y . Confounding
thus generates a non-causal association, distorting our ability to discern the true causal
effect of treatment.
Simpson’s paradox vividly illustrates the perils of unadjusted analyses: it occurs when
the direction of the association between treatment and outcome observed in the overall
population reverses once we stratify (or condition) on the confounder L. This paradox
underscores the crucial need for proper adjustment or stratification in observational studies.
To systematically identify potential confounding and clarify causal structures, researchers
frequently rely on Directed Acyclic Graphs (DAGs). These graphical tools offer intuitive
and rigorous methods for pinpointing pathways that can introduce bias, ensuring that
analyses adjust for the right variables and preserve the integrity of causal inference.

Aspirin: Let’s take the example of aspirin treatment (A) and the chance of dying (Y )
among heart patients. Patients who are already at high risk are more likely to be given
aspirin. Because these patients are already more likely to die, we might see a higher death
rate among those who got aspirin: P (Y = 1 | A = 1) > P (Y = 1 | A = 0). At first glance,
this could wrongly make it seem like aspirin increases the risk of death. But that’s not
true—it’s just that sicker patients were more likely to get the drug. If we adjust for how
sick the patients were at the start (using something like a ”risk score”), or if we randomly
assign who gets aspirin and who doesn’t, we find the true story: aspirin actually helps.
After accounting for this confounding, we see that aspirin lowers the risk of death—it has
a protective, causal effect.

Thus, Raw associations in data are often misleading. Without appropriate analytical
methods or experimental designs that explicitly control for confounding, we risk drawing
incorrect conclusions. Robust causal inference requires thoughtfully severing all spurious
pathways and carefully distinguishing genuine causal effects from mere correlations.
1.4. RCT, EXCHANGEABILITY, AND CONSISTENCY 11

1.4 RCT, Exchangeability, and Consistency

Design Logic. Randomized controlled trials (RCTs) are the gold standard for causal
inference because randomization severs all backdoor paths from baseline prognostic
variables to treatment. This guarantees marginal exchangeability:

Y a ⊥⊥ A,

meaning that the potential outcomes are independent of treatment assignment. This
condition, combined with two other key assumptions—consistency,

Y = Y A,

which links the observed outcome to the potential outcome corresponding to the treatment
actually received, and positivity,

0 < P (A = a | L) < 1 for all strata L,

forms the standard identification triad. These assumptions allow us to estimate causal
effects from observed data in an RCT setting.

Intention-to-Treat (ITT) vs Per-Protocol: In practice, some participants may

not follow their assigned treatment (non-compliance). The Intention-to-Treat (ITT)
principle analyzes participants according to their original treatment assignment, regardless
of whether they adhered. This approach preserves the exchangeability provided by
randomization, avoiding bias due to post-randomization behavior. In contrast, a Per-
Protocol analysis compares participants based on the treatment actually received. While
this may reflect the causal efficacy of the treatment under ideal use, it requires strong
assumptions to adjust for any post-randomization confounders of adherence. Without
proper adjustment, the causal interpretation of per-protocol estimates may be invalid.

Blinding and Concealment: Blinding protects the study from differential behavior
or measurement that could arise if participants or researchers know the treatment assign-
ment. For example, caregivers might unconsciously deliver co-interventions differently, or
outcome assessors might introduce bias. Allocation concealment ensures that treatment
assignment is not known at the point of enrollment, preventing selection bias.

Thus, randomization is not just about flipping a fair coin—it involves a bundle of
practices designed to maintain the integrity of causal claims.

1.5 Conditional Randomization, Stratification, and

Blocking
In many randomized controlled trials, treatment is not assigned completely at random
across the entire study population. Instead, subjects are often grouped into strata (or
”blocks”) based on observed covariates such as age, sex, or disease severity. Within each
block, treatment is then randomized. This is known as a blocked design or stratified
[Link] improves the precision of causal estimates by ensuring that
treatment groups are balanced with respect to key baseline characteristics. In such designs,
12 CHAPTER 1. INTRODUCTION TO CAUSAL INFERENCE

treatment assignment is conditionally independent of the potential outcomes given the

strata, i.e.,
Y a ⊥⊥ A | L,
where L is the blocking (or stratifying) variable. However, marginal (unconditional)
independence Y a ⊥ A may no longer hold. Because treatment assignment is randomized
within strata but not across the entire population, the analysis must account for the
blocking. Ignoring this design structure can result in biased or inefficient estimates.
Standard approaches include stratified analyses such as the Mantel–Haenszel estimator or
regression models with interaction terms for each stratum.

Heart-Transplant Example in Depth. Consider a heart transplant trial where A = 1

indicates receiving a transplant and Y = 1 indicates death. Let L = 1 represent critically
ill patients and L = 0 represent stable patients. In this trial, suppose:

P (A = 1 | L = 1) = 0.75 and P (A = 1 | L = 0) = 0.50.

This indicates that sicker patients are more likely to be assigned to the treatment group,
making the overall (crude) comparison misleading. For instance, the crude mortality rates
are:
Treated: 7/13 vs Untreated: 3/7,
suggesting a higher risk of death in the treated group. However, when we stratify by L,
the mortality rates within each group are:

For L = 1 : 2/3 in both treated and control,

For L = 0 : 1/4 in both treated and control.

Within each stratum, the treatment effect disappears—mortality risk is identical. After
standardization (i.e., averaging the stratum-specific risks according to the distribution
of L), we find that the marginal (overall) risk ratio is actually 1. This illustrates how
conditioning on L can reveal the true causal story hidden beneath the crude associations.
1.6. CROSSOVER EXPERIMENTS AND PERIOD EFFECTS 13

Standard Error Implications. Ignoring the blocking structure in analysis can lead to:

1. Inflated standard error estimates, as the precision gains from stratification are lost;

2. Missed detection of effect modification, where the treatment effect might vary across
strata;

3. Potential confounding if the stratification variable is related to both treatment and

outcome.
To avoid these pitfalls, it is critical that statistical models reflect the study design.
At a minimum, the analysis should include indicator variables (dummy variables)
for each block, and, if warranted, interaction terms between treatment and blocks
to test for effect heterogeneity. Proper model specification ensures valid inference
and optimizes the use of stratified designs.

1.6 Crossover Experiments and Period Effects

Statistical model. Let t ∈ {0, 1} index periods, Ait the treatment at period t, and Yit the
response. Under assumptions of no carryover, equal period effects, and time-stable causal
effect αi , the within-subject difference Di = Yi1 − Yi0 identifies αi free of between-person
confounders.

Practical wrinkles. Wash-out periods mitigate biological carryover; period effects

(fatigue, learning) call for including a fixed effect for time. Missingness, often non-random,
erodes the self-controlled virtue and demands inverse-probability weighting in turn.

When infeasible. Irreversible treatments (organ transplant) or terminal outcomes

(death) preclude crossover schemes; alternative self-controlled designs (case-crossover,
N-of-1) may still apply when outcomes are transient.

1.7 Inverse Probability Weighting (IPW) in Practice

Estimand via Weighting. Inverse Probability Weighting (IPW) is a widely used
method for estimating average treatment effects in observational studies. The key idea
is to create a pseudo-population in which the distribution of observed covariates is
independent of treatment assignment—mimicking a randomized trial.
Let π(L) = P (A = 1 | L) denote the propensity score, i.e., the probability of
receiving treatment given baseline covariates L. The propensity score captures the
treatment selection mechanism. We then define the stabilized IPW weights for
individual i as:
Ai · P (A = 1) (1 − Ai ) · P (A = 0)
wi = + .
π(Li ) 1 − π(Li )
These weights aim to balance the covariate distributions across treatment groups. Stabi-
lized weights are preferred over naive weights like 1/π(Li ), as they help reduce variance
inflation and yield more stable estimators in finite samples.
14 CHAPTER 1. INTRODUCTION TO CAUSAL INFERENCE

Positivity and Truncation. The success of IPW depends crucially on the positivity
assumption, which requires that:

0 < π(Li ) < 1 for all i.

That is, every individual must have a non-zero probability of receiving either treatment
or control, regardless of their covariates L. In practice, some values of π(Li ) may be
very close to 0 or 1, leading to extremely large weights and inflated standard errors. To
address this, analysts often perform weight truncation or weight stabilization, for
example, by capping weights at the 1st and 99th percentiles. This introduces a small
amount of bias but substantially reduces the variance and leads to more reliable inference.
A more serious problem arises with structural positivity violations, where certain
strata of L have no treated or no control observations. In such cases, it is fundamentally
impossible to estimate treatment effects in those regions of the covariate space. The
solution is to redefine the causal question and restrict attention to the population where
treatment comparisons are possible—“where the data support the effect.” We will see
more in Chapter 4.

Sandwich Variance and Bootstrap. Since the propensity score π(L) is typically
estimated from data (e.g., via logistic regression or machine learning), the resulting
weights wi are random variables and introduce additional uncertainty. To account for this,
standard variance formulas must be adjusted. Two common methods for robust inference
are:

1. Sandwich variance estimators (also known as robust or heteroskedasticity-

consistent estimators), which correct the standard errors of IPW estimators by
accounting for the estimated nature of the weights.

2. Non-parametric bootstrap, which resamples the data to empirically estimate

the distribution of the IPW estimator. This approach is flexible and widely used,
especially when machine learning methods are used to estimate π̂(L).

1.8 Suggested Readings

1. James M. Robins and Miguel A. Hernán (2020). Chapter 1, 2, 3, What If ? The
Foundations of Causal Inference. Chapman and Hall/CRC. Also freely available at:
[Link]
Chapter 2

Randomized Experiments

Let D ∈ {0, 1} denote a binary treatment variable indicating whether an individual

smokes marijuana. Specifically, D = 1 denotes a smoker and D = 0 denotes a non-
smoker. Following the potential outcomes framework, we associate with each individual
two hypothetical (counterfactual) outcomes:

Y (1) and Y (0),

where Y (1) is the individual’s lifespan if they smoke, and Y (0) is the lifespan if they do not.
However, in practice, we only observe one of these outcomes per individual, depending on
their actual smoking status. To formalize this, we define a structural model for potential
outcomes:

Y (d) = η0 + η1 d (2.1)
Here, η0 is a person-specific intercept capturing their baseline health endowment (e.g.,
genetics, early childhood environment), and η1 reflects the biological effect of marijuana
smoking on longevity. For simplicity and didactic purposes, we assume η1 = 0, implying
that marijuana smoking has no causal effect on lifespan. In other words, the only differences
in observed lifespans will arise due to differences in η0 , not due to the treatment itself.

Now suppose that smoking behavior is not randomly assigned but arises from an
individual’s latent propensity or predisposition to smoke. This is modeled as:

D = 1{ν > 0}, (2.2)

where ν represents unobserved psychological factors such as risk tolerance, peer influence,
or personality traits. Crucially, we assume that this unobserved propensity is negatively
correlated with baseline health:

E[η0 ν] < 0. (2.3)

This means individuals who are more likely to smoke tend to start off with worse health
on average (due to for example, drinking). This assumption introduces selection bias, as
the treatment assignment (D) is no longer independent of potential outcomes. Under our
assumption that η1 = 0, the potential outcomes simplify to:

Y (1) = Y (0) = η0 ,
and hence the observed outcome is simply:

15
16 CHAPTER 2. RANDOMIZED EXPERIMENTS

Y = Y (D) = η0 .
However, in observational data, we do not observe η0 or ν directly. The only observed
data are the realized pairs (Y, D), where D is the observed treatment (smoking status),
and Y is the observed lifespan.

Apparent Health Gap. Even though smoking has no causal effect, selection into
smoking based on unobserved health characteristics distorts the observed comparison
of average lifespans between smokers and non-smokers. We can compute the expected
outcome (lifespan) for each group:

E[Y | D = 1] = E[η0 | ν > 0] < E[η0 ], (2.4)

E[Y | D = 0] = E[η0 | ν ≤ 0] > E[η0 ]. (2.5)

This leads to the observed inequality:

E[Y | D = 1] < E[Y | D = 0], (*)

suggesting that smokers have shorter lifespans than non-smokers. But this is an illusion.
Since the treatment D was not randomized and is negatively correlated with baseline
health (eq. (2.3)), the difference is entirely due to selection bias, not any biological harm
from marijuana use.

This toy example illustrates the fundamental problem of causal inference from obser-
vational data. The observed difference in means:

π = E[Y | D = 1] − E[Y | D = 0],

which we call the Average Predictive Effect (APE), does not equal the Average
Treatment Effect (ATE):

δ = E[Y (1)] − E[Y (0)],

whenever the treatment D is endogenously selected. That is, when treatment choice is
influenced by variables that also affect the outcome, simple comparisons of outcomes
between groups will confound correlation with causation. Without randomization or
adequate adjustment for confounding, we cannot identify causal effects reliably.

2.1 Selection Bias: Conceptual Foundations

Formally, bias arises whenever

E[Y (d) | D = 1] ̸= E[Y (d)], d ∈ {0, 1}.

The inequality reflects the non-independence of D and Y (d). Let ∆bias denote

∆bias = E[Y | D = 1] − E[Y (1)],

so that APE = δ + ∆bias . Strategies to neutralise ∆bias include:

2.2. RCT: ASSUMPTIONS AND THEOREM 17

1. Design: randomized controlled trials (RCTs), encouragement designs, cluster

randomization.

2. Re-weighting: inverse-probability weights, propensity-score matching, doubly

robust estimators.

3. Regression adjustment:—conditional exchangeability within covariate strata.

Among these, RCTs are the cleanest because they modify the data-generating process
itself rather than adjusting after the fact.

2.2 RCT: Assumptions and Theorem

Fundamental Assumption. In a well-executed randomized controlled trial (RCT), the
treatment assignment D is independent of all potential outcomes. That is, the potential
outcomes that would occur under each treatment level do not influence the probability of
receiving the treatment. Formally, this is expressed as:

D ⊥⊥ Y (1), Y (0) .
Additionally, the positivity assumption ensures that both treatment arms are repre-
sented in the data:

0 < P (D = 1) < 1.

Randomization Theorem. Under the independence assumption, the observed mean

outcomes in each treatment group can be interpreted as the mean potential outcomes.
Specifically, for d ∈ {0, 1},

E[Y | D = d] = E[Y (d) | D = d] = E[Y (d)],

which implies that:

π = E[Y | D = 1] − E[Y | D = 0] = δ,
where π is the observed average difference between treated and control groups, and δ is
the average treatment effect (ATE). Thus, randomization ensures that selection bias
is eliminated by design.

Note. The theoretical independence guaranteed by randomization holds only when

several practical conditions are met:

1. Allocation concealment: Treatment assignment must be hidden from participants

and investigators to prevent manipulation or selection bias.

2. No differential attrition: Dropout rates should not differ systematically between

treatment arms, as this can reintroduce bias.

3. Intention-to-treat (ITT) analysis: Participants should be analyzed based on

their original treatment assignment, regardless of compliance. This preserves the
benefits of randomization.
18 CHAPTER 2. RANDOMIZED EXPERIMENTS

2.3 Sampling Framework for an RCT

We observe
{(Yi , Di )}ni=1 ,
P
i.i.d. from the super-population distribution of (Y, D). Throughout we treat n1 = Di
and n0 = n − n1 as random binomial counts.

2.4 Point Estimation of Group Means

In the context of treatment effect estimation, one of the most fundamental tasks is to
estimate the expected outcome for each treatment group. For a binary treatment variable
D ∈ {0, 1}, we define the group-specific population means as:

θd = E[Y | D = d], for d = 0, 1.

These quantities represent the average outcomes we would observe within each treatment
arm if the entire population were observed under treatment status D = d.

Sample Estimators. Given data from a sample of n individuals indexed by i = 1, . . . , n,

a natural and unbiased estimator of θd is the empirical mean outcome within the subgroup
that received treatment status d. This is given by:
Pn
i=1 Yi · 1{Di = d} Ȳd
θ̂d = P n = ,
i=1 1{Di = d} p̂d
where:
1. 1{Di = d} is the indicator function that equals 1 when individual i received
treatment d, and 0 otherwise.
2. p̂d = nnd is the empirical
Pn proportion of individuals in the sample assigned to treatment
arm d, with nd = i=1 1{Di = d}.
3. Ȳd = n1d i:Di =d Yi is the sample mean outcome among individuals who received
P
treatment d.
These estimators θ̂0 and θ̂1 are simple group means and serve as foundational building
blocks for estimating causal effects, such as the average treatment effect (ATE), which
can be computed as θ̂1 − θ̂0 . They are unbiased estimators of the conditional expectations
E[Y | D = d], provided that sampling is random and no informative missingness or
selection bias is present.

2.5 Inference for the Average Treatment Effect

The average treatment effect (ATE) is defined as the difference in group means:

δ = θ1 − θ0 ,

where θd = E[Y | D = d] for d = 0, 1. The natural sample estimator is:

δ̂ = θ̂1 − θ̂0 ,
2.6. THE DELTA METHOD AND RELATIVE EFFECTIVENESS 19

where θ̂d is the sample mean of Y in treatment group d, as discussed previously. Assuming
finite variances within each treatment arm, the multivariate Central Limit Theorem (CLT)
applies:
√ θ̂0 − θ0 d
     2 
0 σ0 /p0 0
n −
→N , ,
θ̂1 − θ1 0 0 σ12 /p1
where:
1. σd2 = Var(Y | D = d) is the conditional variance,
2. pd = P(D = d) is the marginal probability of treatment arm d.
Applying the delta method (specifically for linear combinations), we derive the asymptotic
distribution of δ̂:
√ σ02 σ12
 
d
n(δ̂ − δ) −→ N 0, + .
p0 p1

Wald Confidence Interval. To construct confidence intervals for δ, we estimate the

variances σd2 using their sample analogues:
1 X
σ̂d2 = (Yi − Ȳd )2 .
nd − 1 i:D =d
i

A large-sample (asymptotic) 100(1 − α)% Wald confidence interval for δ is given by:
s
σ̂02 σ̂12
δ̂ ± zα/2 + ,
n0 n1
where zα/2 is the standard normal quantile corresponding to confidence level 1 − α, and
nd is the number of observations in group d.
This framework provides valid inference under random sampling and large-sample
approximations. In small samples or complex designs, bootstrap or robust methods may
be preferred.

2.6 The Delta Method and Relative Effectiveness

In many applied contexts, especially in health economics, epidemiology, and policy
evaluation, we are interested not just in the absolute effect of a treatment, but in its
relative effect. This is often used to express efficiency or cost-effectiveness. Define the
relative treatment effect as the proportionate change in the mean outcome:
θ1 − θ0
φ= ,
θ0
where θ1 = E[Y | D = 1] and θ0 = E[Y | D = 0] are the population means for the treated
and control groups, respectively. Let us define a function f : R2 → R that maps the
parameter vector θ = (θ0 , θ1 )⊤ to the relative effect:
θ1 − θ0
f (θ) = .
θ0
This simplifies to:
θ1
f (θ) = − 1.
θ0
20 CHAPTER 2. RANDOMIZED EXPERIMENTS

Delta Method Application. To obtain the large-sample distribution of the estimator

f (θ̂), we use the delta method. First, compute the gradient vector:
" # " #
∂f θ1
− 2
θ0
G = ∇f (θ) = ∂θ∂f
0
= 1 .
∂θ1 θ0

The asymptotic distribution of θ̂ = (θ̂0 , θ̂1 )⊤ ,

√ d
n(θ̂ − θ) −
→ N (0, V ) ,

where V is the asymptotic covariance matrix of θ̂, given by:

 2 
σ0 /p0 0
V = .
0 σ12 /p1

Then by the delta method, the relative effect estimator satisfies:

√  
d
n f (θ̂) − f (θ) − → N (0, G⊤ V G).

Use Cases: This approach allows us to construct standard errors and confidence intervals
for multiplicative effect measures such as:
1. Benefit-Cost Ratios in economics,

2. Vaccine Efficacy, typically computed as 1 − RR,

3. Rate Ratios or Relative Risks in epidemiological studies.

It is particularly useful when the absolute effect is less interpretable or when proportional
impacts provide more policy-relevant insights.

2.7 Covariate Adjustment and Treatment Effect

In many randomized experiments and observational studies, baseline covariates W — such
as age, gender, income, health status, or education — may influence both the outcome
and the treatment effect. To account for this, we define the Conditional Average
Treatment Effect (CATE):

δ(W ) = E[Y (1) − Y (0) | W ].

This quantity captures how the causal effect of treatment varies across individuals or
subgroups characterized by different values of W . In essence, it generalizes the average
treatment effect (ATE) to a conditional level, allowing for treatment effect heterogeneity.
Under the assumption of random treatment assignment (e.g., in a randomized controlled
trial), the treatment indicator D ∈ {0, 1} is independent of potential outcomes conditional
on W . That is,
Y (1), Y (0) ⊥⊥ D | W.
This implies the equality:

E[Y | D = d, W ] = E[Y (d) | W ], for d ∈ {0, 1}.

2.8. LIMITATIONS OF RCTS 21

Consequently, we can define the Conditional Average Predictive Effect:

π(W ) = E[Y | D = 1, W ] − E[Y | D = 0, W ],

and under ignorability, we obtain:

π(W ) = δ(W ).

That is, the difference in conditional means from observed data correctly identifies the
causal effect for covariate strata.

Efficiency gains through covariate adjustment. Even in randomized trials, adjust-

ing for baseline covariates can lead to efficiency gains. By modeling outcome variation
across covariates using methods such as ANCOVA (Analysis of Covariance), regression
adjustment, or modern machine learning methods like Targeted Minimum Loss Estimation
(TMLE) or double machine learning, one can:
1. Reduce residual variance,

2. Improve precision of treatment effect estimates,

3. Construct narrower confidence intervals,

4. Detect heterogeneity across subgroups.

These techniques leverage the predictive power of covariates W to ”de-noise” the treatment
effect estimation.
If covariates W contain missing values, simply removing such observations (complete-
case analysis) may introduce selection bias — especially if the missingness mechanism is not
completely at random, for instance, if it depends on both the treatment assignment D
and the outcome Y . In such settings, valid inference requires more sophisticated strategies:
1. Multiple Imputation: Fill in missing values using posterior draws from predictive
models.

2. Inverse Probability Weighting for Missingness: Weight complete cases by the

inverse of their response probabilities.

3. Doubly Robust Methods: Combine imputation and weighting to ensure consis-

tency even if one of the models is misspecified.

Covariate adjustment is a powerful tool for improving efficiency and exploring het-
erogeneity in treatment effects. However, proper handling of missing covariate data is
essential to maintain the validity of causal conclusions.

2.8 Limitations of RCTs

Violations of SUTVA. The Stable Unit Treatment Value Assumption has two clauses:
(i) no interference among units, (ii) no hidden versions of treatment. Herd-immunity
spill-overs, market-equilibrium effects, and varying implementation fidelity each breach
SUTVA, complicating causal interpretation.
22 CHAPTER 2. RANDOMIZED EXPERIMENTS

Ethical constraints. Randomly assigning harmful exposures (smoking initiation,

poverty) is unacceptable; equipoise, informed consent, and beneficence govern study
design. When RCTs are impossible, quasi-experimental designs—difference-in-differences,
instrumental variables, regression discontinuity—step in.

Cost and power. Field trials can be logistically daunting and expensive, especially for
rare outcomes requiring huge n for adequate power. Cluster randomization eases logistics
but inflates variance via intraclass correlation.

External validity. An RCT estimates the ATE for its sampling frame, yet effects may
attenuate or amplify when scaled up. Transportability analysis confronts such “all else
equal” fallacies.

2.9 Conclusion & Suggested Readings

Randomization underpins modern drug licensing (FDA), evidence-based development
economics (Banerjee, Duflo, Kremer), and behavioural interventions (Thaler’s nudge
agenda). While critics highlight cost and context specificity, the clarity of the causal
identification it affords keeps RCTs the gold standard against which observational methods
are benchmarked.

1. Victor Chernozhukov et al. (2023). Applied Causal Inference Powered by Machine

Learning and AI, Chapter 2. [Link]

2. Sylvain Chabé-Ferret (2024). Statistical Tools for Causal Inference, Chapter 3:

Randomized Controlled Trials. [Link]
html

3. Imbens & Rubin (2015). Causal Inference for Statistics, Social, and Biomedical
Sciences. Cambridge UP.
Chapter 3

Double Lasso, Neyman

Orthogonality, and Related Methods

In classical linear regression theory, the number of covariates p is assumed to be small

relative to the sample size n. Under this assumption, the Ordinary Least Squares (OLS)
estimator is well-behaved: it is unbiased, efficient, and asymptotically normal under
standard regularity conditions. This framework underlies much of traditional econometric
analysis. However, a few modern data applications often violate the low-dimensional
assumption. For example:
1. In genetics, DNA microarray data may involve p ∼ 104 gene expression levels for a
few hundred patients.
2. In digital marketing, clickstream data can generate p ∼ 105 binary indicators from
user behavior.
3. In cross-country development research, datasets may include dozens of institutional
indicators across fewer than only 200 countries.
When the number of covariates p is large relative to, or even exceeds, the sample size n,
traditional OLS estimation fails in several ways:
1. The matrix (X ⊤ X)−1 becomes ill-conditioned or undefined when p ≥ n, making the
OLS solution either unstable or non-existent.
2. Including all variables indiscriminately leads to inflated variance and overfitting,
degrading out-of-sample predictive performance.
3. Conversely, performing aggressive pre-screening of covariates risks omitting important
variables, introducing bias in coefficient estimates.
To address these challenges, modern regression analysis often relies on regularization
techniques, which penalize model complexity and encourage sparsity. Among the most
widely used is the Least Absolute Shrinkage and Selection Operator (Lasso).
Lasso estimation solves the optimization problem:
 
lasso 1 2
β̂ = arg min ∥Y − Xβ∥2 +λ∥β∥1 ,
β 2n
where λ is a tuning parameter controlling the degree of shrinkage. The Lasso estimator
offers key advantages in high-dimensional settings:

23
24 CHAPTER 3. DOUBLE ML

1. It performs automatic variable selection by shrinking some coefficients exactly to

zero.

2. It provides sparse solutions that improve interpretability and predictive performance.

However, Lasso introduces a critical trade-off. The shrinkage bias it induces pulls estimated
coefficients toward zero, which invalidates standard inferential tools such as t-tests and
confidence intervals derived from OLS theory. This creates a fundamental tension: how
can we retain the predictive power and sparsity benefits of Lasso while also conducting
valid statistical inference on key parameters—such as the effect of a treatment or policy
variable—amid high-dimensional confounding? Answering this question leads us toward
double selection methods, debiased Lasso, and double machine learning, which
build on Lasso’s strengths while correcting for its limitations.

3.1 Frisch–Waugh–Lovell (FWL) Partialling-Out

Consider the linear regression model:

Y = αD + β ⊤ W + ε, (3.1)

where:

1. Y is the outcome variable of interest (e.g., economic growth, wage, blood pressure,
etc.),

2. D ∈ R is the target regressor whose coefficient α represents the causal or predictive

effect we wish to estimate,

3. W ∈ Rp is a high-dimensional vector of covariates or controls,

4. ε is an idiosyncratic disturbance satisfying the orthogonality condition E[ε | D, W ] =

0.

Frisch–Waugh–Lovell (FWL) Theorem: In classical low-dimensional settings, α̂OLS

can be equivalently obtained by the following two-step residual regression procedure:

1. First, regress Y on W and obtain the residuals Ỹ = Y − π̂Y⊤ W .

⊤
2. Then, regress D on W and compute the residuals D̃ = D − π̂D W.

3. Finally, regress Ỹ on D̃ to obtain α̂ from the simple regression:

Ỹ = αD̃ + ε.

By removing the linear effects of W from both the outcome Y and the regressor of interest
D, we isolate the component of D that is uncorrelated with the confounders W , allowing
us to estimate its effect on Y without bias from omitted variables. When the number of
controls p becomes large relative to the sample size n, classical OLS techniques break down.
The matrix (W ⊤ W )−1 becomes ill-conditioned or non-invertible, and even if inversion
is numerically feasible, the variance of α̂ can explode due to overfitting. As a result,
standard OLS-based residualization becomes unreliable or impossible.
3.2. NEYMAN ORTHOGONALITY. 25

To overcome these limitations, modern econometrics uses regularized regression—specifically,

the Double Lasso or Partialling-Out Lasso method. Instead of projecting Y and D
onto W via OLS, we perform two separate Lasso regressions.

First, we run a Lasso of Y on W :

n
nX p o
X
⊤ 2
γ̂Y = arg minp (Yi − γ Wi ) + λ1 ψY j |γj | .
γ∈R
i=1 j=1

The penalty loadings ψY j (often ψY j = 1/σ̂j ) scale the ℓ1 constraint so that units of
different regressors are comparable. Independently we fit a Lasso of D on W :
nX X o
⊤ 2
γ̂D = arg min (Di − γ Wi ) + λ2 ψDj |γj | .
γ
i j

Both λ1 and λp
2 can be chosen via modified cross-validation or the theoretically motivated
formula λ ≍ σ 2 log p/n. We then compute

Y̌i = Yi − γ̂⊤Y Wi , Ďi = Di − γ̂⊤D Wi .

These are the pieces of Y and D orthogonal, up to Lasso error, to the entire W space.
Regressing the residualized outcome on residualized treatment gives
P
Ďi Y̌i
α̂ = Pi 2 = (Ď⊤ Ď)−1 (Ď⊤ Y̌ ).
i Ďi

Because (Ďi ) is a scalar, the matrix inverse above is merely a division.

Intuition. The union of variables selected in Step 1 or Step 2 acts like double selection.
Any regressor important for predicting either Y or D is implicitly adjusted for. Conse-
quently the omitted variable bias of missing “weak but critical” controls is dramatically
reduced compared with “single” Lasso selection that screens only on Y .

3.2 Neyman Orthogonality.

The success of Double Lasso relies on a powerful statistical property called Neyman
orthogonality. It ensures that small errors in the estimation of nuisance components (i.e.,
the Lasso fits for Y and D, discussed below) do not significantly bias the final estimate of
α. Specifically, the estimation procedure is first-order insensitive to small perturbations in
the nuisance parameters—an essential requirement for valid inference in high-dimensional
models. In contrast, methods that do not satisfy Neyman orthogonality—such as naive
single Lasso selection—allow first-stage bias in η̂ to leak directly into√ α̂ through linear
terms. This contamination prevents the estimator from achieving n-rate convergence
and undermines valid statistical inference. Thus, Neyman orthogonality is not merely a
technical requirement but a foundational design principle that enables double machine
learning procedures to perform reliable inference in high-dimensional settings where
traditional OLS methods fail. We prove it next. Let α(η) be the target parameter defined
implicitly by the moment condition
h i
M (α, η) := E (Ỹ (η) − αD̃(η))D̃(η) = 0,
26 CHAPTER 3. DOUBLE ML

where
Ỹ (η) = Y − η1⊤ W, D̃(η) = D − η2⊤ W.
Here, η = (η1 , η2 ) represents a collection of nuisance parameters, and η o denotes their true
values. The property of Neyman orthogonality requires that the estimator α(η) be
locally insensitive to small perturbations in η, i.e.,

∂η α(η o ) = 0.

In high-dimensional settings, nuisance parameters η are typically estimated (e.g., via

Lasso), introducing slight bias. Neyman orthogonality ensures that these small estimation
errors do not influence α to first order. Formally, for a small perturbation δ around the
true nuisance value η o ,

α(η o + δ) = α(η o ) + ∂η α(η o ) · δ +higher-order terms.

| {z }
=0

Proof. Since α(η) is implicitly defined via the moment equation M (α, η) = 0 and the
implicit function theorem gives:

∂η α(η o ) = − [∂α M (α, η o )]−1 ∂η M (α, η o ).

To show that ∂η α(η o ) = 0, it suffices to demonstrate:

∂η M (α, η o ) = 0.

Recall that h i
M (α, η) = E (Ỹ (η) − αD̃(η))D̃(η) ,
with
Ỹ (η) = Y − η1⊤ W, D̃(η) = D − η2⊤ W.
Differentiating with respect to η1 we get,
h i
∂η1 Ỹ (η) = −W ⇒ ∂η1 M (α, η o ) = E −W · D̃(η o ) .

Since D̃(η o ) is the residual from regressing D on W , it is orthogonal to W :

⊤
 
E W · (D − γD,W W ) = 0,

so
∂η1 M (α, η o ) = 0.
and the Derivative with respect to η2 gives,

∂η2 D̃(η) = −W.

So, the product rule gives

h    i
o o o o 2
∂η2 M (α, η ) = E ∂η2 Ỹ (η ) · D̃(η ) − α · ∂η2 D̃(η )
h i h i
= −E W · Ỹ (η o ) + 2α · E W · D̃(η o ) .
 √
3.3. ASYMPTOTIC THEORY: n-NORMALITY OF α̂ 27

Again, Ỹ (η o ) and D̃(η o ) are residuals from projecting Y and D on W , respectively.

Therefore:
E[W · Ỹ (η o )] = 0, E[W · D̃(η o )] = 0.
So,
∂η2 M (α, η o ) = 0.
Since both partial derivatives vanish, and it follows from the implicit function theorem
that:
∂η α(η o ) = − [∂α M (α, η o )]−1 · 0 = 0.

√
3.3 Asymptotic Theory: n-Normality of α̂
Under appropriate sparsity and moment assumptions—such as the condition that the
sparsity level s satisfies (s log p ≪ n—Belloni, Chernozhukov, and Hansen (2014)) a
central limit theorem for the double machine learning estimator α̂ can be established.

Let D̃ denote the residual from the population-level projection of D on W , and let ε
denote the structural error from the partially linear model

Y = αD + β ⊤ W + ε, E[ε | D, W ] = 0.

Then the debiased or orthogonalized estimator α̂ satisfies:

√ d
 
→ N 0, V := E[D̃2 ]−1 E[D̃2 ε2 ] E[D̃2 ]−1 .
n(α̂ − α) −

A consistent plug-in estimator V̂ replaces the population expectations with their sample
analogues:
n
!−2 n
!
1X 2 1X 2 2
V̂ = D̃ D̃ ε̂ ,
n i=1 i n i=1 i i

where ε̂i = Yi − α̂ · Di − β̂ ⊤ Wi is the estimated residual from the final stage. Using this
variance estimator, a Wald-type 95% confidence interval for α is given by:
s
V̂
α̂ ± 1.96
n

This interval has asymptotically correct coverage, with error rate converging to zero at
the standard O(n−1/2 ) rate.

A key strength of this double machine learning procedure is its robustness to imperfect
model selection. The validity of inference for α does not require perfect variable selection
in the high-dimensional regressions of Y and D on W . Instead, it suffices that the
Lasso-based nuisance estimators achieve a sufficiently accurate approximation of the best
s-sparse linear predictors. This flexibility is due to the Neyman orthogonality property,
which guarantees that small estimation errors in the first stage do not translate into
first-order biases in the second-stage estimate of the target parameter.
28 CHAPTER 3. DOUBLE ML

3.4 Conditional Convergence in Growth Economics

A central hypothesis in development economics is the theory of conditional convergence,
which posits that poorer countries tend to grow faster than richer ones, conditional on
similar structural characteristics. In other words, once we control for institutional quality,
educational attainment, trade openness, and other growth determinants, we expect
countries with lower initial income to exhibit higher subsequent growth rates. This
hypothesis translates into a regression framework where the dependent variable is the
economic growth rate and the main regressor of interest is initial GDP per capita:

Growthi = α InitialGDPi + β⊤ Wi + εi ,

where:

1. Growthi is the GDP growth rate for country i over a fixed period (e.g., 10 years),

2. InitialGDPi is the log of initial GDP per capita,

3. Wi is a high-dimensional vector of control variables, including proxies for human

capital, institutional quality, demographic structure, trade openness, and geographic
characteristics (here, p = 60),

4. and n = 90 denotes the number of countries in the sample.

In low-dimensional settings, ordinary least squares (OLS) would be the default estima-
tor. However, when the number of controls p is comparable to or exceeds the sample size n,
OLS becomes unstable and prone to overfitting. Moreover, the inclusion of many irrelevant
controls inflates variance, resulting in wide confidence intervals and reduced statistical
power. Double Lasso, on the other hand, reduces variance inflation and improves inference
validity in high-dimensional settings.

Method Estimate SE 95% CI

OLS −0.009 0.032 [−0.073, 0.054]
Double Lasso −0.045 0.018 [−0.080, −0.010]

The OLS estimate is close to zero and statistically insignificant, as its confidence interval
includes zero. This reflects the high noise-to-signal ratio in the presence of many controls
relative to sample size. In contrast, the Double Lasso estimate is larger in magnitude and
statistically significant at conventional levels. Its tighter confidence interval excludes zero,
offering empirical support for the hypothesis of conditional convergence.

The negative and significant estimate of α under Double Lasso suggests that, holding
constant various institutional and structural characteristics, countries with lower initial
GDP indeed tend to grow faster. This aligns with the predictions of the Solow growth
model.
3.5. BEYOND A SINGLE TREATMENT 29

3.5 Beyond a Single Treatment

Researchers often wish to infer a vector α = (α1 , . . . , αp1 ) attached to multiple policy
variables D1 , . . . , Dp1 . Double Lasso generalizes by residualizing each Dℓ on the control
set W and running separate low-dimensional regressions. Simultaneous confidence bands
may be constructed via the Gaussian multiplier bootstrap, controlling family-wise error
in the presence of many targets.
Applications include interaction effects (policy × industry), non-linear polynomial
expansions (D, D2 , D3 ), and heterogeneity by demographics.

3.6 Alternative Orthogonal Methods

An alternative to residualization in partialling-out is the Double Selection method. In
this approach, two separate Lasso regressions are performed:

1. Regress the outcome Y on the full set of covariates (D, W ) using Lasso. Let
SbY ⊆ {1, . . . , p} denote the selected variables (excluding D).

2. Regress the treatment variable D on the covariates W using Lasso. Let SbD ⊆
{1, . . . , p} denote the selected variables.

The final model includes the treatment D and the union of selected controls Sb = SbY ∪ SbD .
An ordinary least squares (OLS) regression of Y on D and the selected variables WSb is
then performed to estimate the treatment effect. This estimator:

1. Retains the Neyman orthogonality property,

2. Is robust to model selection errors in either the outcome or treatment model,

3. is typically more conservative—favoring inclusion of relevant covariates—than single-

selection or naive Lasso.

Another method is The Debiased Lasso, also known as the Desparsified Lasso,
addresses the regularization bias inherent in standard Lasso estimates. It constructs an
estimator with an asymptotic linear expansion by correcting the shrinkage in the Lasso
estimate. The procedure has been outlined below.

1. Run a Lasso regression of Y on D and W (with an appropriate penalty λ) to obtain

β̂.

2. Run a Lasso regression of D on W (with a suitable λ) to obtain γ̂.

3. Construct the residualized treatment:

D̃(γ̂) = D − γ̂ ′ W.

4. Obtain the estimator α̂ as the solution to the moment condition:

n
1 X ′

Yi − α̂Di − β̂ Wi D̃i (γ̂) = 0.
n i=1
30 CHAPTER 3. DOUBLE ML

This yields the explicit formula:

n
!−1 n
!
1X 1 X 
α̂ = Di D̃i (γ̂) Yi − β̂ ′ Wi D̃i (γ̂) .
n i=1 n i=1

The debiased estimator for a target coefficient α has the form:

n
1X ⊤  
α̂ = α̂Lasso + Θ̂ Xi Yi − Xi⊤ β̂Lasso ,
n i=1

where:
1. β̂Lasso is the Lasso estimate from regressing Y on the full covariate vector X,
2. Θ̂ is an estimate of the relevant row of the (pseudo-)inverse of the empirical Gram
matrix Σb = X ⊤ X/n.

This adjustment removes the bias from penalization and restores asymptotic normality:
√ d
→ N (0, σ 2 ),
n(α̂ − α) −
even in high-dimensional settings where p ≫ n. Confidence intervals and hypothesis tests
constructed from this estimator are asymptotically valid.

In the special case of a pure randomized controlled trial (RCT), the treatment D is
independent of the covariates W by design:
D ⊥⊥ W.
Therefore, the naive OLS regression of Y on D yields an unbiased estimate of the average
treatment effect (ATE), even without adjusting for W . While orthogonal methods like
Double Lasso or Debiased Lasso are not necessary to ensure unbiasedness in this setting,
they can still be employed to increase estimation efficiency (i.e., reduce variance) and
improve precision of inference when W explains a substantial amount of variation in Y .
While randomization solves the identification problem, orthogonal adjustments still play
a role in optimizing statistical performance.
1. Run a Lasso regression of Y on D and W (with an appropriate penalty λ) to obtain
β̂.
2. Run a Lasso regression of D on W (with a suitable λ) to obtain γ̂.
3. Construct the residualized treatment:
D̃(γ̂) = D − γ̂ ′ W.

4. Obtain the estimator α̂ as the solution to the moment condition:

n
1 X 
Yi − α̂Di − β̂ ′ Wi D̃i (γ̂) = 0.
n i=1

This yields the explicit formula:

n
!−1 n
!
1X 1 X ′

α̂ = Di D̃i (γ̂) Yi − β̂ Wi D̃i (γ̂) .
n i=1 n i=1
3.7. PRACTICAL CHECKLIST FOR EMPIRICAL IMPLEMENTATION 31

Equivalently, the debiased estimator for a target coefficient α has the form:
n
1X ⊤  ⊤

α̂ = α̂Lasso + Θ̂ Xi Yi − Xi β̂Lasso ,
n i=1

where:
1. β̂Lasso is the Lasso estimate from regressing Y on the full covariate vector X,

2. Θ̂ is an estimate of the relevant row of the (pseudo-)inverse of the empirical Gram

matrix Σb = X ⊤ X/n.

This correction term serves to ”undo” the bias introduced by Lasso’s penalization.

3.7 Practical Checklist for Empirical Implementation

When implementing orthogonal estimation methods such as Double Lasso or Debiased
Lasso, it is important to adhere to a set of practical best practices that ensure numerical
stability, valid inference, and robust results. Below is a detailed checklist to guide empirical
implementation:

1. Centre and Scale Regressors.

Always standardize the covariates in the high-dimensional matrix W . This involves
centering each column to have mean zero and scaling to unit variance. Standardiza-
tion:

(a) Improves numerical stability and convergence in optimization,

(b) Ensures that the Lasso penalty is applied uniformly across variables,
(c) Makes interpretation of coefficients more meaningful, especially when variables
are on different scales.

2. Penalty Selection (λ).

Choose the regularization parameter λ via one of two common methods:

(a) 10-fold Cross-Validation: Splits the data into 10 parts, trains on 9 and validates
on 1, cycling through all folds.
(b) Theory-Driven Plug-in Method: Based on theoretical guarantees that balance
false inclusion and exclusion probabilities (e.g., Belloni et al. plug-in).

Avoid under-penalization, which can lead to overfitting and poor generalization.

3. Inference: Robust Standard Errors.

Since weights and nuisance parameters are estimated, standard error estimation in
the second-stage regression must account for this:

(a) Use heteroskedasticity-robust sandwich estimators,

(b) Or apply the residual bootstrap to obtain valid confidence intervals in small
samples,
(c) If data are clustered (e.g., by region or school), use clustered standard errors
to account for intra-cluster correlation.
32 CHAPTER 3. DOUBLE ML

4. Diagnostic Plots.
Visual inspection of residuals and influential observations can prevent misinterpreta-
tion. Recommended diagnostics include:

(a) Histogram of Ďi : the partialled-out regressor, to verify variation,

(b) Leverage scores: to identify high-leverage points,
(c) Cook’s distance: to detect influential observations that could distort estimates.

5. Sensitivity Analysis.
Examine robustness of the estimated treatment effect α̂ by varying the penalty
parameter λ by ±25%. If results change dramatically, your model may be fragile or
overfit. Stable estimates across a penalty range provide greater empirical confidence.

3.8 Notebooks
The Jupyter notebooks for the referenced experimentation are available here in Github.

3.9 Suggested Readings

1. Victor Chernozhukov et al. (2023). Applied Causal Inference Powered by Machine
Learning and AI, Chapter 2. [Link]
Chapter 4

Conditioning and Propensity Scores

Recover Causal Effects

In observational studies, the treatment indicator D is almost never randomly assigned.

Patients choose therapies, students select majors, and firms adopt technologies according
to incentives, preferences, or constraints—factors that often also influence the outcome Y
we intend to study. The resulting selection bias makes naive comparisons unreliable:
average wage differences between trained and untrained workers reflect both training
benefits and pre-existing ability gaps. To address this challenge, two conceptual tools are
central:

1. Conditional ignorability (also known as unconfoundedness or selection on ob-

servables): This assumption posits that after conditioning on a sufficiently rich set
of covariates X, the treatment assignment D behaves as if it were random. That
is, any systematic differences in potential outcomes between treated and untreated
units can be fully explained by X.

2. Overlap (or positivity): This ensures that, for every covariate profile X, there
is a positive probability of observing both treated and untreated units. Formally,
0 < P (D = 1 | X) < 1 almost surely. Without this condition, comparisons across
treatment groups become ill-posed or undefined for certain strata.

When both of these conditions hold, researchers can recover the Average Treatment Effect
(ATE) using standard regression techniques or propensity-score reweighting—even without
the benefit of randomized experiments. What follows is a self-contained exposition of how
these tools work in practice.

4.1 Potential Outcomes Framework

In causal inference, the potential outcomes framework provides a rigorous way to define
and estimate causal effects. We first set up the core notations that enable the identification
of causal quantities from observational data and revisit a few assumptions.

Notation and Potential Outcomes.

1. D: A binary treatment indicator where D = 1 denotes treatment and D = 0 denotes
control.

33
34 CHAPTER 4. METRICS TO RECOVER CAUSAL EFFECTS

2. Y : The observed outcome variable.

3. Y (1), Y (0): Potential outcomes—Y (1) is the outcome that would be observed if
the unit were treated, and Y (0) is the outcome if the unit were not treated. These
together define the response function d 7→ Y (d) for d ∈ {0, 1}.

4. X: A vector of observed pre-treatment covariates (e.g., age, GPA, income, health

status, etc.).

We assume the Consistency axiom, which states:

Y = Y (D), almost surely,

meaning we observe the potential outcome corresponding to the realized treatment D.

Assumption of Conditional Ignorability: For each d ∈ {0, 1}, treatment assignment

is independent of the potential outcome given covariates:

D ⊥⊥ Y (d) | X.

To interpret, within covariate strata X = x, the treatment assignment behaves as if

randomized. That is, the treated and untreated subpopulations are statistically comparable
with respect to their potential outcomes. Any observed difference in outcomes between
treated and control units sharing the same covariate values can thus be attributed to the
effect of the treatment itself. For example, suppose two students have the same SAT score,
family income, and high school quality (i.e., same X). If one attends college (D = 1)
and the other does not (D = 0), conditional ignorability asserts that these students’
potential earnings Y (1), Y (0) are unrelated to the college attendance decision—once we
have conditioned on X:= SAT score, family income, and high school quality.

Assumption of Overlap (Common Support): Let p(X) = P (D = 1 | X) denote the

propensity score, the probability of receiving treatment given covariates. The overlap
condition requires:
0 < p(X) < 1 almost surely.
This ensures that every covariate profile X has a positive probability of being assigned
to both treatment and control groups. Without this assumption, there would be regions
of X where treated or untreated units are completely absent, making comparisons and
causal effect estimation ill-posed.

Under these assumptions of Conditional Ignorability and Overlap, we can eliminate

selection bias by conditioning on X:

E[Y | D = d, X] = E[Y (d) | D = d, X] (by Consistency)

= E[Y (d) | X] (by Ignorability).

Therefore, the Conditional Average Predictive Effect (CAPE) given by:

π(X) = E[Y | D = 1, X] − E[Y | D = 0, X],

4.2. LINEAR REGRESSION AS A CONDITIONING DEVICE 35

equals the Conditional Average Treatment Effect (CATE):

δ(X) = E[Y (1) − Y (0) | X].

and averaging over the covariate distribution yields the Average Treatment Effect
(ATE):
δ = E[δ(X)] = E[π(X)].
So under the assumptions of ignorability and overlap, causal effects can be identified
purely from observational data using regression or weighting methods that condition on
X, which we check next.

4.2 Linear Regression as a Conditioning Device

A standard approach to causal inference under the potential outcomes framework is to
use linear regression as a tool for conditioning on observed covariates. This approach
becomes valid under the assumption of conditional ignorability.

Suppose we posit the following linear model for the conditional expectation of the
outcome:
E[Y | D, X] = αD + β⊤ X, (4.1)
where

1. α represents the average causal effect of the treatment, D

2. β ∈ Rp is a vector of coefficients associated with the covariates X.

We then estimate the following linear regression model via ordinary least squares (OLS):

Y = αD + β⊤ X + ε,

where ε is a zero-mean error term satisfying E[ε | D, X] = 0. Under the assumptions of

linearity, ignorability, and overlap, the OLS estimator α̂ is a consistent estimator of the
Average Treatment Effect (ATE), denoted:

δ := E[Y (1) − Y (0)].

To account for heterogeneous treatment effects, we can extend the linear model to
allow the treatment effect to vary with covariates by introducing interaction terms between
D and a centered version of the covariates X. The modified model is:

E[Y | D, X] = α1 D + α⊤2 (X − X̄)D + β1 + β⊤2 (X − X̄),

where:

1. X̄ is the sample mean of the covariates,

2. α1 captures the average treatment effect across all covariate strata (ATE),

3. α2 allows the treatment effect to vary with characteristics.

36 CHAPTER 4. METRICS TO RECOVER CAUSAL EFFECTS

In this case, we interpret:

ATE = α1 , CATE(X) = α1 + α⊤2 (X − X̄).

The validity of this linear regression approach depends on the correctness of the
specified functional form. If the linear model in Equation (4.1) does not represent the
true conditional expectation function (CEF), then the resulting estimate α̂ may be biased.
When the functional form is in doubt or the relationship between Y , D, and X is highly
non-linear, it is advisable to turn to more flexible estimation techniques such as non-
parametric regression or modern machine learning methods (e.g., random forests, boosted
trees, neural networks). These approaches can better capture complex patterns in the
data and help avoid misspecification bias.

4.3 The Horvitz–Thompson Method

In observational studies, the identification of treatment effects often relies on conditioning
on observed covariates to model the outcome process, i.e., the conditional expectation
function
E[Y | D, X].
However, in many real-world applications, this function may be complex and difficult to
approximate accurately. As an alternative, we may focus on the treatment assignment
process, specifically the propensity score, which is defined as
p(X) = P (D = 1 | X).
The Horvitz–Thompson (HT) method leverages this score to reweight observed outcomes,
providing an alternative path to identification of potential outcome means. This approach
is especially effective when the covariate vector X is high-dimensional, and the propensity
score is either known (e.g., from design) or can be reliably estimated.
Theorem. Horvitz–Thompson Propensity Score Reweighting): Under the assumptions
of Conditional Ignorability (Y (d) ⊥ D | X) and Overlap (0 < P (D = d | X) < 1),
the conditional expectation of a properly reweighted observed outcome Y identifies the
conditional mean of the potential outcome Y (d):
 
Y · 1(D = d)
E X = E[Y (d) | X].
P (D = d | X)
Taking expectation over the distribution of X, we recover the marginal potential outcome
mean:  
Y · 1(D = d)
E = E[Y (d)].
P (D = d | X)
Proof. We apply the law of iterated expectations:
   
Y · 1(D = d) Y (d) · 1(D = d)
E X =E X (by consistency: Y = Y (d) if D = d)
P (D = d | X) P (D = d | X)
1
= E [Y (d) · 1(D = d) | X]
P (D = d | X)
1
= (E[Y (d) | X, D = d] · P (D = d | X))
P (D = d | X)
= E[Y (d) | X] (by ignorability: Y (d) ⊥ D | X).
4.4. A REMARK ON PRECISION AND PROPENSITY SCORES 37

Averaging both sides over X yields the identification of the average potential outcome
E[Y (d)].

As a direct application, we can define a transformed variable, the Horvitz–Thompson

Transform:
1{D = 1} 1{D = 0}
H= − ,
P (D = 1 | X) P (D = 0 | X)
and recover the Average Treatment Effect (ATE) via:

δ = E[Y · H].

This formulation ensures that each individual’s observed outcome Y is weighted inversely
by their probability of receiving the treatment they actually received, thus correcting for
the non-random assignment. Similarly, we define the Conditional Average Treatment
Effect (CATE) as
δ(X) = E[Y · H | X].
This formulation highlights how causal effects can be recovered both at the population
level and at the covariate level, solely via reweighting based on estimated or known
treatment probabilities.

4.4 A Remark on Precision and Propensity Scores

Suppose we are studying the effect of a job training program on wages. The propensity
score p(X) = P (D = 1 | X) may be estimated using various individual-level covariates
such as age, education, and prior work experience. Now imagine two individuals who
share the same estimated propensity score. This implies that they have an equal predicted
probability of receiving the training, given their observed covariates. However, this
similarity in the overall score may mask very different covariate profiles. For example:

1. One individual may have substantial work experience and little formal education.

2. The other may be highly educated but lack practical experience.

Even though their propensity scores are the same, their background differences might
influence the outcome variable (e.g., wages) in distinct ways. This situation highlights
a key limitation of relying solely on propensity scores for identification. Solution lies in
Double Machine Learning. By combining both:

1. Reweighting (based on the propensity score),

2. Regression adjustment (using the full covariate vector X),

we can more effectively de-noise the outcome variable. That is, we reduce residual variance
by accounting for detailed covariate patterns. This hybrid approach—central to double
machine learning—produces more efficient and precise estimates of the true treatment
effect by leveraging both the treatment selection mechanism and the outcome-generating
process.
38 CHAPTER 4. METRICS TO RECOVER CAUSAL EFFECTS

4.5 Covariate Balance Checks

To assess whether randomization (or reweighting via the propensity score) has succeeded
in balancing treatment and control groups, we perform a covariate balance check.
Under the assumption of conditional ignorability, it must be that:

E[H | X] = 0,

where H = 1{D=1}
p(X)
− 1{D=0}
1−p(X)
is the Horvitz–Thompson transform. If this expectation
does not hold—i.e., if some function of X is predictive of H—then systematic differences
remain between the treatment and control groups even after reweighting. This implies a
violation of ignorability, as treatment assignment remains confounded.

In a low-dimensional setting, we can test balance via regression. Specifically:

1. Construct a dictionary W = f (X), containing linear and non-linear transformations

of covariates.

2. Regress H on W .

3. Perform an F -test to assess whether W significantly predicts H.

If the regression shows predictive power (i.e., W significantly predicts H), then covariate
imbalance exists—suggesting that the treatment assignment was not successfully random-
ized or the propensity score model was misspecified. Covariate balance diagnostics are
essential validity checks for any causal analysis relying on ignorability. They are especially
critical in observational studies where treatment is not randomly assigned.

4.6 Directed Acyclic Graphs

To understand how treatment assignment and potential outcomes are connected, we use
Directed Acyclic Graphs (DAGs). These help us visualize the assumptions behind causal
inference, particularly the assumption of conditional ignorability.

D Y (d)

In this DAG:
1. The node X represents observed pre-treatment covariates (e.g., age, education).

2. The node D is the actual treatment assignment (e.g., whether the subject receives a
drug).

3. The node d denotes a hypothetical intervention assigning treatment status.

4.6. DIRECTED ACYCLIC GRAPHS 39

4. Y (d) is the potential outcome under treatment level d.

5. The edge X → D captures how baseline characteristics influence treatment choice.

6. The edge X → Y (d) encodes that these same characteristics may also affect the
outcome.

7. The edge d → Y (d) reflects that the potential outcome depends on the treatment
level.

For example in a healthcare context,

1. X: Patient’s characteristics (e.g., age, comorbidities).

2. d: A hypothetical assignment to treatment (d = 1: drug, d = 0: no drug).

3. Y (d): The patient’s blood pressure if we hypothetically assign them treatment d.

4. D: Actual treatment assignment observed in data.

5. X → D and X → Y (d): Characteristics influence both treatment decision and

health outcome.

6. d → Y (d): The potential outcome depends on which treatment is given.

7. Under ignorability, once we condition on X, treatment D behaves like a random

assignment: outcome differences between groups with the same X can be causally
attributed to treatment D.

We get the following as an Observed Outcome DAG: X → D, X → Y , D → Y .

D Y

This DAG depicts the factual world:

1. Covariates X are realized first.

2. Based on X, each subject is assigned a treatment D.

3. The outcome Y is then determined by both the assigned treatment D and baseline
covariates X.

4. Importantly, we only observe Y = Y (D), the outcome corresponding to the received

treatment, not both potential outcomes.
40 CHAPTER 4. METRICS TO RECOVER CAUSAL EFFECTS

4.7 Putting the Pieces Together

To implement causal inference in observational studies, based on the above discussion,
one must follow a structured approach combining both conceptual clarity and empirical
rigor. The first step is conceptual, wherein the researcher articulates the assumptions
of conditional ignorability and overlap within the context of the study and supports
this reasoning with a directed acyclic graph (DAG) that visualizes the assumed causal
structure. Next comes empirical step A, which involves estimating the treatment effect
parameter α by fitting a regression model of the outcome variable Y on the treatment
indicator D and covariates X. This regression can be linear or based on machine learning
algorithms, and should be validated through model diagnostics such as goodness-of-fit and
tests for functional form misspecification. Empirical step B focuses on the alternative
strategy of weighting: the propensity score p̂(X) is estimated, Horvitz–Thompson (HT)
weights H are computed, and covariate balance is evaluated to ensure the weights have
successfully removed systematic differences across treatment groups. Finally, in the
synthesis step, results from both regression and weighting methods are compared. If
the estimates align and covariate balance holds, one can credibly report estimates of the
Average Treatment Effect (ATE) and the Conditional Average Treatment Effect (CATE).
If not, the analyst should reconsider the underlying assumptions or explore alternative
identification strategies such as instrumental variable methods.

Recommended Readings
1. Rosenbaum, P. R., & Rubin, D. B. (1983). “The Central Role of the Propensity
Score in Observational Studies for Causal Effects.” Biometrika, 70(1), 41–55.

2. Imbens, G. W., & Rubin, D. B. (2015). Causal Inference for Statistics, Social, and
Biomedical Sciences: An Introduction. Cambridge University Press.

3. Hernán, M. A., & Robins, J. M. (2020). Causal Inference: What If ? Boca

Raton: Chapman & Hall/CRC. Available at: [Link]
whatifbook

4. Chernozhukov, V., Wüthrich, K., Kumar, M., Semenova, V., Yadlowsky, S., & Zhu,
Y. (2023). Applied Causal Inference Powered by Machine Learning and AI. Available
at: [Link]
Chapter 5

Structured Equations Modelling and

Graphical Models

Statistical data alone inform us about associations, but causal science asks what would
happen if we acted differently. Directed acyclic graphs (DAGs) encode qualitative
subject-matter knowledge—who can influence whom—while structural equation mod-
els (SEMs) supply quantitative functional relationships that generate the joint distribution.
Pearl (2000) showed that every (acyclic) SEM induces a DAG and, conversely, that the
graph together with independent disturbances suffices to reconstruct counterfactual out-
comes. We begin by exploring a fully linear and nonlinear, nonparametric formulation of
causal diagrams and their associated structural equation models (SEMs). These models
serve as powerful and flexible tools for uncovering the underlying structure necessary for
causal identification, enabling us to move beyond the confines of purely linear assumptions.
Within this framework, we define counterfactuals in a formal way—adhering to what
Judea Pearl describes as the ”First Law of Causal Inference”—which states that every
structural equation model naturally induces a system of counterfactual outcomes.

5.1 General DAG and SEM via an Example

D Y

X
F

We now illustrate a causal diagram and its associated structural equation model (SEM)

41
42 CHAPTER 5. GRAPHICAL MODELS

using a real-world scenario: the impact of 401(k) eligibility on financial wealth. The
directed acyclic graph (DAG) below encodes the causal relationships among various
observed and unobserved variables involved in this context.
In the United States, a 401(k) plan is an employer-sponsored, defined-contribution,
personal pension (savings) account, as defined in section 401(k) of the U.S. Internal
Revenue Code. This causal graph represents the possible channels through which 401(k)
eligibility (D) may affect an individual’s net financial wealth (Y ). The interpretation of
each node is as follows:

1. D: Binary indicator for 401(k) eligibility — the treatment variable.

2. Y : Net financial assets — the outcome of interest.

3. M : Employer’s matching contribution — a potential mediator between eligibility

and wealth.

4. X: Observed worker-level covariates such as age, income, and job tenure.

5. F : Unobserved firm-level covariates such as firm size or culture.

6. U : General latent factors, which may include unmeasured personality traits or risk
preferences.

The arrows reflect assumed causal relationships. For example, D → M and M → Y

represent the fact that eligibility for a 401(k) program may lead to a matching contribution,
which in turn could influence net wealth. The variable X affects many nodes: it influences
both treatment assignment (D) and outcome (Y ), as well as the mediator (M ). The
unobserved variables U and F introduce latent confounding: U may simultaneously affect
observed covariates X and firm characteristics F , while F affects both D and M .

This DAG can be translated into a structural system of equations where each node is
generated as a function of its parent nodes and an associated noise term.

X = fX (U, εX ),
F = fF (U, εF ),
D = fD (X, F, εD ),
M = fM (D, X, F, εM ),
Y = fY (D, M, X, εY ).
Each equation specifies how a variable is generated as a function of its direct causes
(or ”parents” in the DAG) and an associated idiosyncratic error term. For example,
the covariates X are influenced by a latent factor U , which could represent unobserved
personal attributes such as financial literacy or long-term planning ability, along with a
noise component εX . Similarly, the firm-level variable F is also influenced by U and its
own noise εF , indicating that some unobserved traits may jointly affect both worker- and
firm-level characteristics. The treatment assignment D, indicating 401(k) eligibility, is
determined by observed covariates X, firm-level factors F , and a residual εD capturing
individual-level randomness in eligibility. The employer’s matching contribution M
depends on D, X, and F , suggesting that employer contributions may vary based on both
employee and firm characteristics. Finally, the outcome of interest Y , such as financial
5.2. CONDITIONAL IGNORABILITY AND EXOGENEITY 43

wealth, is modeled as a function of the treatment D, mediator M , covariates X, and an

independent shock εY . Together, these equations define a recursive system that captures
the flow of causality in the model.

5.2 Conditional Ignorability and Exogeneity

A fundamental challenge in causal inference is to identify the conditions under which
we can estimate causal effects from observed data. One such key condition is known as
conditional ignorability or conditional exogeneity. This concept is central to modern causal
analysis, as it provides the bridge between the structural assumptions encoded in a causal
model and the statistical procedures used for estimation.

The connection between structural equation models (SEMs) and potential outcomes
is foundational in causal inference. The fact that an SEM implies the existence of
potential outcomes is sometimes called the First Law of Causal Inference. SEMs, or their
graphical representations as directed acyclic graphs (DAGs), encapsulate the contextual
and substantive knowledge about the causal relationships in a given problem. As a result,
they allow us to derive, rather than merely assume, important identification conditions
such as conditional ignorability.
For example, suppose we are interested in the effect of a binary treatment D on
an outcome Y , and we have observed covariates F and X. The SEM or DAG for the
problem may suggest that, after conditioning on F and X, the treatment assignment D is
independent of the potential outcome Y (d) for each value of d. Formally, this is written
as:
Y (d) ⊥⊥ D | F, X,
which implies the following equality of conditional expectations:

E[Y (d) | F, X] = E[Y | D = d, F, X].

This property allows us to identify average causal (or treatment) effects by adjusting for
(i.e., conditioning on) F and X. Conditional ignorability (or exogeneity) can be justified
using both functional (structural) arguments based on SEMs and graphical arguments
based on d-separation and the backdoor criterion.

To provide a functional argument, consider the structural equations that define the
data-generating process. In the counterfactual (or potential outcome) setting where we
fix D = d, the relevant structural equations are:

Y (d) = fY (d, M (d), X, ϵY ),

M (d) = fM (d, F, X, ϵM ),
where M is a mediator and ϵY , ϵM are exogenous noise terms. The actual treatment
assignment is generated by
D = fD (F, X, U, ϵD ),
where U and ϵD are additional exogenous variables. Once we condition on F and X, the
distribution of Y (d) is determined solely by d, M (d), X, and their associated noise terms.
The realized value of D is not relevant for the distribution of Y (d) once F and X are
44 CHAPTER 5. GRAPHICAL MODELS

given. In other words, knowing D provides no additional information about Y (d) beyond
what is already known from F and X:

Y (d) ⊥⊥ D | F, X.

This structural argument demonstrates how the ignorability condition can be justified by
the functional relationships in the SEM.

The same conclusion can be reached using graphical criteria. In the counterfactual
DAG, the node Y (d) receives inputs from M (d), X, and the fixed value d. The treatment
variable D is still present in the graph and is generated by its usual parents (F , X, and
U ), but there is no direct arrow from D to Y (d).
Any path from D to Y (d) must pass through either F or X. For instance, typical
paths include:

1. D ← X → Y (d),

2. D ← F → M (d) → Y (d),

3. D ← F ← U → X → Y (d).

By conditioning on F and X, we block all such paths. In graphical terms, this is known
as d-separation: conditioning on a node severs the flow of information along any path
passing through that node. The Global Markov property then tells us that d-separation
implies conditional independence, to conclude that

Y (d) ⊥⊥ D | F, X.

A related graphical concept is the backdoor criterion, which provides a practical

method for identifying a set of variables Z that, when conditioned on, blocks all non-
causal (backdoor) paths from D to Y . A set Z satisfies the backdoor criterion if:

1. No variable in Z is a descendant of D, and

2. Z blocks every backdoor path from D to Y (that is, every path that starts with an
arrow into D).

The first rule prevents us from blocking the causal effect of D on Y , while the second
ensures that all confounding paths are eliminated. In the context of the 401(k) example,
the relevant backdoor paths from D to Y run through F and X, such as:

1. D ← X → Y ,

2. D ← F → M → Y ,

3. D ← F ← U → X → Y .

By conditioning on both F and X, we block all such paths, ensuring that the observed
association between D and Y reflects only the causal effect of D on Y .
5.3. DAGS, SEMS, AND D-SEPARATION 45

5.3 DAGs, SEMs, and d-Separation

DAGs and their associated SEMs provide a precise mathematical language for representing
causal systems. These concepts are essential for both the design and analysis of causal
inference strategies. This section introduces the foundational concepts of directed acyclic
graphs (DAGs), their associated structural equation models (SEMs), and the graphical
criteria for conditional independence known as d-separation. These tools form the backbone
of modern causal inference, providing a rigorous language for representing, analyzing, and
reasoning about causal relationships.

A directed acyclic graph (DAG) is a graph G = (X, E) consisting of a set of nodes

X = {X1 , X2 , . . . , X|X| } and directed edges E, with the critical property that there are
no cycles—meaning it is impossible to start at any node and follow a sequence of directed
edges that eventually loops back to the starting node. Equivalently, the set of nodes V is
partially ordered by the edge structure E. Within a DAG, several relationships among
nodes are defined:

1. The parents of a node Xj , denoted P aj , are all nodes with directed edges pointing
into Xj :
P aj := {Xk : Xk → Xj }.

2. The children of Xj , denoted Chj , are all nodes that Xj points to:

Chj := {Xk : Xj → Xk }.

3. The ancestors of Xj , denoted Anj , are all nodes from which there exists a directed
path to Xj , including Xj itself:

Anj := {Xk : Xk < Xj } ∪ {Xj }.

4. The descendants of Xj , denoted Dsj , are all nodes that can be reached by a
directed path starting from Xj :

Dsj := {Xk : Xk > Xj }.

A DAG can be associated with an acyclic structural equation model (ASEM), which
formalizes how each variable is generated as a function of its parents and some exogenous
noise. For each node j ∈ V , the structural equation is:

Xj := fj (P aj , ϵj ),

where each ϵj is a random disturbance (exogenous variable), and the collection (ϵj )j∈V is
assumed to be jointly independent. A linear ASEM is a special case where the structural
equations are linear in the parents:

fj (P aj , ϵj ) := fj′ P aj + ϵj ,

with fj′ being a vector of coefficients. In linear ASEMs, the independence assumption
on the errors can be weakened to mere uncorrelatedness. The structural potential
46 CHAPTER 5. GRAPHICAL MODELS

response process for each variable describes how the value of Xj would respond to
arbitrary assignments of its parents:

Xj (paj ) := fj (paj , ϵj ),

viewed as a stochastic process indexed by the possible values of the parents.

To generate observable variables in an ASEM, one draws a realization of the exogenous

shocks {ϵj }j∈V and then solves the system of equations to obtain the endogenous variables
{Xj }j∈V . The exogenous variables are those not determined by the model (the ϵj ), while
the endogenous variables are those generated as outputs of the structural equations (the
Xj ).

A key property of ASEMs associated with DAGs is the Markov factorization, which
states that the joint distribution of all variables factorizes as:
Y
p({xℓ }ℓ∈V ) = p(xℓ | paℓ ).
ℓ∈V

Equivalently, each variable is independent of its non-descendants given its parents—this

is known as the local Markov property.
Understanding the flow of information and potential confounding in DAGs requires
precise definitions of paths and how they may be blocked. Now we define Paths, Blocked
Paths, and d-Separation in DAGs

1. A directed path is a sequence of nodes connected by edges all pointing in the same
direction:
Xv1 → Xv2 → · · · → Xvm .

2. A non-directed path allows some arrows to be reversed.

3. A node Xj is a collider on a path if the path includes a segment of the form:

→ Xj ← .

4. A backdoor path from Xl to Xk is a non-directed path that starts at Xl and ends

with an arrow into Xk .

A path, π is said to be blocked by a set of nodes S if either:

1. π contains a chain (i → m → j) or a fork (i ← m → j) with m ∈ S, or

2. π contains a collider (i → m ← j), and neither m nor any of its descendants are in
S.

If a path is not blocked by S, it is called open. Conditioning on a node in a chain or fork

blocks the path, while conditioning on a collider or its descendant opens a path that would
otherwise be blocked. For example, in the figures below, the backdoor path Y ← X → D
in (a) is blocked by conditioning on X. Conversely, a path like Y → C ← D in (b) is
blocked by default, but becomes open if we condition on the collider C.
5.3. DAGS, SEMS, AND D-SEPARATION 47

Z D Y Z D Y

X C

(a) (b)

The concept of d-separation provides a graphical criterion for determining conditional

independence in DAGs. Given a DAG G, a set of nodes S is said to d-separate nodes X
and Y if S blocks all paths between X and Y . This is denoted as:

(X ⊥⊥d Y |S)G .

By the foundational result of Pearl and Verma, d-separation implies conditional indepen-
dence in the probability distribution generated by the DAG:

X ⊥⊥ Y | S.

Intuitively, if all information flow between X and Y is interrupted by conditioning on

S, then knowing X provides no additional information about Y once S is known. The
formal proof of this equivalence is nontrivial, and the converse does not always hold.

Z X Y Z X Y

U U

(a) (b)

To make these concepts concrete, consider the following two graphical structures below.
In the first example, suppose the graph contains nodes Z, U , X, and Y , with edges such
that Z and U are parents of X, U is also a parent of Y , and there is a direct edge from Z
to Y . Here, the set S = {Z, U } d-separates X and Y , blocking all paths between them.
The Markov factorization yields:

p(y, x | u, z) = p(y | x, z, u) p(x | z, u) = p(y | u, z) p(x | z, u),

implying that X ⊥⊥ Y | Z, U . In the second example, suppose the structure is such that Z
points to X, X and Y both point to U , and Z also points to Y . In this case, conditioning
on Z alone d-separates X and Y , and the factorization becomes:

p(y, x | z) = p(y | z) p(x | z),

which implies X ⊥⊥ Y | Z.
48 CHAPTER 5. GRAPHICAL MODELS

5.4 Intervention, Counterfactual DAGs, and SWIGs

When analyzing causal effects, it is essential to understand how interventions modify the
structure of a causal system. This is formalized through the concepts of counterfactual
DAGs and Single World Intervention Graphs (SWIGs), which provide a unified graphical
framework for reasoning about interventions and potential outcomes.

Consider a causal system represented by a directed acyclic graph (DAG) and its
associated structural equation model (SEM). Each node in the DAG corresponds to a
variable, and the directed edges represent causal relationships. To analyze the effect of
an intervention—such as setting a treatment variable Xj to a specific value xj —we must
construct a new graphical object that reflects this hypothetical scenario. The intervention
fix(Xj = xj ) transforms the original DAG into a counterfactual DAG, known as a Single
World Intervention Graph (SWIG). The SWIG is constructed by a node-splitting operation:

1. The original node Xj is split into two distinct entities:

(a) Xj∗ , representing the natural (pre-intervention) value of Xj .

(b) A new deterministic node, denoted a (or Xa∗ ), set to the intervened value xj .

2. The intervention node Xa∗ inherits only the outgoing edges from Xj (i.e., eeai = eji
for all i) and has no incoming edges (e
eia = 0 for all i), reflecting that it is fixed by
the intervention.

3. The node Xj∗ inherits only the incoming edges from Xj (i.e., eeij = eij for all i) and
has no outgoing edges (eeji = 0 for all i), preserving its dependence on its original
causes.

4. All remaining edges are preserved: eeik = eik for all i and for all k ̸= j, k ̸= a, ensuring
the rest of the graph structure remains intact.

5. The counterfactual variables are assigned according to:

Xk∗ := fk (P a∗k , ϵk ), for k ̸= a,

where P a∗k denotes the parents of Xk∗ under E,

e adapting the structural equations to
the new graph.

The resulting SWIG, denoted G(xe j ), contains a set of counterfactual variables {X ∗ }k∈V ∪
k
{Xa∗ }, where each variable is now interpreted as a function of the intervention. The
counterfactual SEM (CF-ASEM) associated with the SWIG defines the structural equations
for the counterfactual variables:

Xk∗ := fk (P a∗k , ϵk ), for k ̸= a,

where P a∗k denotes the parents of Xk∗ in the modified edge set, and ϵk are the exogenous
noise terms. This construction ensures that the counterfactual variables are generated
consistently with the intervention. To illustrate, consider the following diagrams. The left
figure shows the original DAG with variables and their causal relationships. The right
figure shows the corresponding SWIG after intervening to set variable Xj to value xj :
5.4. INTERVENTION, COUNTERFACTUAL DAGS, AND SWIGS 49

Xj X6 Xj∗ X6∗

X3 X3∗

Xa∗ = xj

X4 X4
X1 X1

X5 X5

(a) Original DAG (b) SWIG after fix(Xj = xj )

A central result is that the SWIG encodes the conditional independence relations
among counterfactual variables under the intervention. Specifically, suppose we relabel the
treatment node as D, and let Y be any descendant of D. Construct the SWIG induced
by fix(D = d), and let S be any subset of nodes common to both the original DAG and
the SWIG such that Y (d) is d-separated from D by S in the SWIG. Then:

1. The conditional exogeneity condition holds:

Y (d) ⊥⊥ D | S.

2. The conditional average potential outcome is identified by:

E[g(Y (d)) | S = s] = E[g(Y ) | D = d, S = s],

for all s with p(s, d) > 0 and for all bounded functions g.

To further clarify, consider the following example inspired by Pearl’s work. The left
diagram shows the original DAG, and the right shows the SWIG after intervening to set
D = d:

d
Z2 X3 Y Z2 X3 Y (d)

X2 M X2
M (d)

Z1 X1 D
Z1 X1 D
(a) Original DAG
(b) SWIG after fix(D = d)
50 CHAPTER 5. GRAPHICAL MODELS

In this example, the goal is to estimate the causal effect of D on Y , i.e., the mapping
d 7→ Y (d). The set of variables

S = {{X1 , X2 }, {X2 , X3 }, {X2 , Z2 }, {X2 , Z1 }}

are valid adjustment sets that, when conditioned upon, block all backdoor paths between
D and Y (d) in the SWIG. Notably, conditioning on X2 alone is insufficient, as it opens a
path where X2 acts as a collider; adding X1 , X3 , Z1 , or Z2 blocks this path, ensuring the
required conditional ignorability.

An important and perhaps underappreciated advantage of the counterfactual DAG (or

SWIG) approach is its ability to clarify not only which adjustment sets are valid, but also
which are actually necessary and helpful for identifying causal effects. This is particularly
relevant when considering the efficiency of estimators and the potential pitfalls of including
superfluous control variables. Consider the simple causal DAG:

Z ← D → Y,

where D is the treatment, Y is the outcome, and Z is a variable influenced by D. In this

structure, there are no backdoor paths from D to Y that could introduce confounding.
The only paths from D to Y are direct, and all paths from D to Z are also direct. When we
construct the corresponding counterfactual DAG or SWIG, representing the intervention
fix(D = d), the graph becomes:

Z(d) ← d → Y (d).

Here, both Z(d) and Y (d) are potential outcomes under the intervention D = d, and d
is a fixed value. In this counterfactual graph, there are no open paths from d to either
Z(d) or Y (d) except the direct arrows, and crucially, there is no path from Z(d) to Y (d)
that could induce spurious association. The key insight is that, under this counterfactual
representation, no adjustment is required to identify the causal effect of D on Y . Formally,
the empty set is a valid adjustment set:

Y (d) ⊥⊥ D.

This means that the average causal effect of D on Y can be identified without controlling
for any other variables—adjustment is unnecessary because there is no confounding to
remove. However, it is also true that Z is a valid control variable in the sense that
adjusting for Z does not introduce bias. This can be seen by considering a ”cross-world”
DAG (below) that combines both factual and counterfactual variables from the respective
structural equation models.

ϵz Z D Y ϵy

Z(d) d Y (d)

In such a graph, Y (d) is d-separated from D by Z, so the conditional independence

Y (d) ⊥⊥ D | Z
5.5. THE BACKDOOR CRITERION 51

holds as well. Nevertheless, while Z is a valid control in the sense that it blocks no
necessary paths and does not introduce bias, it is also superfluous: adjusting for Z does
not improve identification and can in fact reduce the precision of our estimates. Including
unnecessary variables in the adjustment set can lead to less efficient estimators, as it
increases variance without reducing bias. This underscores the practical value of the
counterfactual DAG approach: it not only identifies all valid adjustment sets but also
helps to avoid overadjustment by highlighting when adjustment is unnecessary.

5.5 The Backdoor Criterion

A central question in causal inference is: under what conditions can we identify the causal
effect of a treatment variable on an outcome using observational data? The backdoor
criterion, formulated by Judea Pearl, provides a clear graphical answer to this question.
This section presents the theorem, illustrates it with figures, and walks through a canonical
example.

As discussed before, a backdoor path is any path from the treatment D to the
outcome Y that starts with an arrow into D, indicating a potential confounding influence.
The backdoor criterion provides a graphical condition for identifying a set of variables
(an adjustment set) that, when conditioned upon, ensures the identification of the causal
effect of D on Y .

Theorem. (Backdoor Criterion) Let G be a DAG representing an acyclic structural

equation model (ASEM). Relabel a treatment node Xj as D, and let Y be any descendant
of D. A set of variables S is a valid adjustment set—meaning it implies conditional
ignorability,
Y (d) ⊥⊥ D | S,
if the following two conditions hold:

1. No element of S is a descendant of D.

2. All backdoor paths from D to Y are blocked by S.

This criterion ensures that, by conditioning on S, all spurious (non-causal) associations

between D and Y are removed, leaving only the causal effect to be estimated. To visualize
the backdoor criterion, consider the following DAG, adapted from Pearl’s classic example:

(i) D ← X2 → Y
(ii) D ← X1 ← Z1 → X2 ← Z2 → X3 → Y

In this DAG, there are two notable backdoor paths from D to Y as shown above.
To apply the backdoor criterion, we must block all such paths by conditioning on an
appropriate set S. Conditioning on X2 alone blocks the inner backdoor path (i), since
X2 is a non-collider on this path and conditioning on it blocks the flow of association.
However, conditioning on X2 alone actually opens the outer backdoor path (ii), because X2
acts as a collider along that path. In general, conditioning on a collider (or its descendant)
opens a path that would otherwise be blocked, potentially introducing bias. Therefore, to
block the outer backdoor path as well, we must also condition on an additional variable
52 CHAPTER 5. GRAPHICAL MODELS

Z2 X3 Y

X2 M

Z1 X1 D

Figure 5.1: DAG illustrating backdoor paths from D to Y . Red: direct/inner backdoor
path. Blue: longer/outer backdoor path.

that lies on that path but is not a descendant of D—for example, X1 , X3 , Z1 , or Z2 . Thus,
valid adjustment sets include:

S1 = {X1 , X2 } or S2 = {X2 , X3 } or S3 = {X2 , Z1 } or S4 = {X2 , Z2 }

Each of these sets blocks all backdoor paths from D to Y and contains no descendants of
D.

It is important to note that conditioning on a descendant of D, such as M (an

intermediate outcome on the causal path from D to Y ), is not valid. Doing so can
introduce bias by blocking part of the direct causal effect or opening new spurious paths.
Additionally, The backdoor criterion systematically yields minimal adjustment sets for
identification. However, it may not capture every valid set. For example, consider the
simple DAG:
Z ← D → Y.
Here, conditioning on Z does not satisfy the backdoor criterion (since Z is a descendant
of D), yet Z is a valid control. In this case, D directly causes Y without confounding,
so there is no need to adjust for Z. Adjusting for Z may even lower the precision of
the estimated effect. This limitation is useful as it helps disregard controls that, while
valid, do not add any meaningful information for identifying the causal effect. The same
observation applies within the counterfactual approach.

5.6 Notebook
The Jupyter notebook is available here in Github.

5.7 Suggested Readings

1. Victor Chernozhukov et al. (2023). Applied Causal Inference Powered by Machine
Learning and AI, Chapter 2. [Link]