Causal Inference: An Introduction

Qingyuan Zhao

Statistical Laboratory, University of Cambridge

4th March, 2020 @ Social Sciences Research Methods Programme (SSRMP),

University of Cambridge

Slides and more information are available at

http://www.statslab.cam.ac.uk/~qz280/.
About this lecture
About me
2019 – University Lecturer in the Statistical Laboratory (in Centre for
Mathematical Sciences, West Cambridge).
2016 – 2019 Postdoc: Wharton School, University of Pennsylvania.
2011 – 2016 PhD in Statistics: Stanford University.

Disclaimer
I am a statistician who work on causal inference, but not a social scientist.
Bad news: What’s in this lecture may not reflect the current practice of
causal inference in social sciences.
Good news (hopefully): What’s in this lecture will provide you an
up-to-date view on the design, methodology, and interpretation of causal
inference (especially observational studies).
I tried to make the materials as accessible as possible, but some amount of
maths seemed inevitable. Please bear with me and don’t hesitate to ask
questions.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 1 / 57

Growing interest in causal inference

●
United States ●
United Kingdom
100 ● ●

●
●
●
Interest (Google Trends)

● ●
● ●
75 ●
● ●
●
● ● ● ● ● ●
● ●
● ●
● ●
● ● ● ●● ●
● ●●
● ●
● ● ● ●
50 ●
●
●
●
●
● ●
●
● ● ●● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ● ●●
●
●
● ● ● ●
●● ● ● ● ●●
● ● ●
●● ● ● ●
● ●
● ● ● ● ● ● ●● ● ● ●
● ●
●
● ● ●● ● ● ● ● ● ● ●
● ● ● ● ● ● ●
●● ●
● ● ● ● ● ●
● ●
● ● ● ● ● ●
● ●●● ●●
25 ●
●
● ●
●●
● ● ●
●
● ● ● ● ● ● ●
●
● ● ● ●●● ● ● ●
● ●●● ● ● ●● ● ●
● ●● ● ● ● ● ● ●● ● ● ●
● ● ● ● ●
●● ● ● ●
● ● ● ● ●
● ● ●●●
● ● ● ● ● ● ● ● ●●●●● ● ●

0 ●● ● ● ●● ● ● ● ●

Jan 2010 Jan 2012 Jan 2014 Jan 2016 Jan 2018 Jan 2020
Time

Figure: Data from Google Trends.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 2 / 57

A diverse field

Causal inference is driven by applications and is at the core of statistics (the

science of using information discovered from collecting, organising, and studying
numbers—Cambridge Dictionary).

Many origins of causal inference

Biology and genetics;
Agriculture;
Epidemiology, public health, and medicine;
Economics, education, psychology, and other social sciences;
Artificial intelligence and computer science;
Management and business.

In the last decade, independent developments in these disciplines have been

merging into a single field called “Causal Inference”.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 3 / 57

Examples in social sciences

1 Economics: How does supply and demand (causally) depend on price?

2 Policy: Are job training programmes actually effective?
3 Education: Does learning “mindset” affect academic achievements?
4 Law: Is it justifiable to sue the factory over injuries due to poor working
conditions?
5 Psychology: What is the effect of family structure on children’s outcome?

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 4 / 57

Outline for this lecture
To study causal relationships, empirical studies can be categorised into

Randomised Experiments (Part I)

1 Completely randomised;
2 Stratified (pairs or blocks);
3 With regression adjustment (also called covariance adjustment)?
4 More sophisticated designs (e.g. sequential experiments).

↓↓ Question: How to define causality? (Part II) ↓↓

Observational Studies (Part III)
Also called quasi-experiments in social sciences (I think it’s a poor name).
1 Controlling for confounders;
2 Instrumental variables;
3 Regression discontinuity design;
4 Negative control (e.g. difference in differences).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 5 / 57

Part I: Randomised experiments
The breakthrough
The idea of randomised experiments dates back to the early development of
experimental psychology in the late 1800s by Charles Sanders Peirce
(American philosopher).
In 1920s, Sir Ronald Fisher established randomisation as a principled way for
causal inference in scientific research (The Design of Experiments, 1935).

Fundamental logic*
1 Suppose we let half of the participants to receive the treatment at random,
2 If significantly more treated participants have better outcome,
3 Then the treatment must be beneficial.

Randomisation (1) =⇒ a choice of statistical error (2) vs. causality (3).

(because there can be no other logical explanations)

*We will revisit this logic when moving to observational studies.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 7 / 57

Randomisation
Some notations
A is treatment (e.g. job training), for now let A be binary (0=control, 1=treated);
Y is outcome (e.g. employment status 6 months after job training).
X is a vector of covariates measured before the treatment (e.g. gender,
education, income, . . . ).
Subscript i = 1, . . . , n indexes the study participants.

Different designs of randomised experiments

Bernoulli trial: A1 , . . . , An independent and P(Ai = 1) = 0.2.
Completely randomised: !−1
n
P(A1 = a1 , . . . , An = an ) = if a1 + · · · + an = n/2.
n/2

Stratified: A1 , . . . , An independent, P(Ai = 1 | Xi ) = π(Xi ) where π(·) is a given

function. For example:
P(Ai = 1 | Xi1 = male) = 0.5 and P(Ai = 1 | Xi1 = female) = 0.75.
Blocked: Completely randomised within each block of participants similar in X .
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 8 / 57
Statistical inference: Approach 1

Randomisation inference (permutation test)

Test the hypothesis H0 : A ⊥
⊥ Y | X (or H0 : A ⊥
⊥ Y if randomisation does not
depend on X ).
1 Choose a test statistic T (X , A, Y ) (e.g. in a blocked experiment with
matched pairs, the average pairwise treated-minus-control difference in Y ).
2 Obtain the randomisation distribution of T (X , A, Y ) by permuting A,
according to how it was randomised.
3 Compute the p-value:
 
PA∼π T (X , A, Y ) ≥ T (X , Aobs , Y ) | X , Y .

Note that the randomisation inference treats X and Y as given and only
considers randomness in the treatment A ∼ π (which is exactly the
randomness introduced by the experimenter).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 9 / 57

Statistical inference: Approach 2
Regression analysis
Simplest form:
E[Y |A] = α + βA.
Regression adjustment (also called covariance adjustment):

E[Y |A, X ] = α + βA + γX + δAX .

More complex mixed-effect models, to account for heterogeneity of the

participants.

Interpretation of regression analysis

Slope coefficient β of the treatment A in these regression models is usually
interpreted as the average treatment effect, although this becomes difficult
to justify in complex designs/regression models.
To differentiate from structural equation models, regression models were
written in the form of E[Y |A] = α + βA instead of the “traditional” form
Y = α + βA + . We will explain their differences later.
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 10 / 57
Comparison of the two approaches
Randomisation inference
Advantages:
1 Only uses randomness in the design.
2 Distribution-free and exact finite-sample test.
Disadvantages:
1 Only gives a hypothesis test for “no treatment effect whatsoever” (can be
extended to constant treatment effect).

Regression analysis
Advantages:
1 Account for treatment effect heterogeneity.
2 Well-developed extensions: mixed-effect models, generalised linear models,
Cox proportional-hazards models, etc.
Disadvantages:
1 Inference usually relies on normality or large-sample approximations.
2 Causal interpretation is model-dependent!
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 11 / 57
Internal vs. external validity
Internal validity
Campbell and Stanley (1963): “Whether the experimental treatments make a
difference in this specific experimental instance”.
Exactly what randomisation inference tries to do.

External validity
Shadish, Cook and Campbell (2002): “Whether the cause-effect relationship
holds over variation in persons, settings, treatment variables, and
measurement variables”.

Related concepts
Another important concept in social sciences is construct validity: “the
validity if inferences about the higher order constructs that represent
sampling particulars”. See Shadish et al. (2002) for more discussion.
Perice’s three kinds of inferences: deduction, induction, abduction.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 12 / 57

How causal inference became irrelevant
The narrow-minded view of causality
“Correlation does not imply causation”
=⇒ Causality can only be established by randomised experiments
=⇒ Causal inference became absent in statistics until 1980s.
Example: “Use of Causal Language” in the author guidelines of JAMA:
Causal language (including use of terms such as effect and efficacy)
should be used only for randomised clinical trials. For all other study
designs, methods and results should be described in terms of association or
correlation and should avoid cause-and-effect wording.

Broken cycle of statistical research

Conjecture Data collection

X
Analysis Modelling

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 13 / 57

“Clouds” over randomised experiments
(Borrowing the metaphor from the famous 1900 speech by Kelvin.)

Smoking and Lung cancer (1950s)

Hill, Doll and others: Overwhelming association between smoking and lung
cancer, in many populations, and after conditioning on many variables.
Fisher and other statisticians: But correlation is not causation.

Infeasibility of randomised experiments

Ethical problems, high cost, and many other reasons.

Non-compliance
People may not comply with assigned treatment or drop out during the study.

=⇒ Need for causal inference from observational data.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 14 / 57

Part II: How to define causality?
Definition 0: Implicitly from randomisation
Recall the logic of randomised experiment:
1 Suppose we let half of the participants to receive the treatment at random,
2 If significantly more treated participants have better outcome,
3 Then the treatment must be beneficial (because there can be no other
logical explanation).

Randomisation (1) =⇒ a choice of statistical error (2) vs. causality (3).

(because there can be no other logical explanations)

For observational studies, we need a definition of causality that does not hinge
on (explicit) randomisation.
Pioneers in causal inference have come up with three definitions/languages:
1 Counterfactual (also called potential outcome);
2 Causal graphical model;
3 Structural equation model.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 16 / 57

Part II: How to define causality?
Definition 1: Counterfactuals (Neyman, 1923; Rubin, 1974)
Participants have two counterfactuals, Y (0) and Y (1).
We only observe one counterfactual (in any study, randomised or not),

Y (1), if A = 1,
Y = Y (A) =
Y (0), if A = 0.

i Yi (0) Yi (1) Ai Yi
1 -3.7 ? 0 -3.7
2 2.3 ? 0 2.3
3 ? 7.4 1 7.4
4 0.8 ? 0 0.8
.. .. .. .. ..
. . . . .
Rubin calls this the “science table” (I didn’t find this terminology useful).
The goal of causal inference is to infer the difference
Distribution of Y (0) vs. Distribution of Y (1).
Example: Average treatment effect is defined as E[Y (1) − Y (0)].
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 17 / 57
Part II: How to define causality?
Definition 1: Counterfactuals (Neyman, 1923; Rubin, 1974)
We would like to infer about the difference between

Distribution of Y (0) vs. Distribution of Y (1).

How is this possible? If we know A ⊥

⊥ Y (0) | X , then

P(Y (0) = y ) = E[P(Y (0) = y | X )]

= E[P(Y (0) = y | A = 0, X )]
= E[P(Y = y | A = 0, X )]

Remark 1: The above derivation is called causal identification.

Remark 2: In the literature, the key assumption A ⊥
⊥ Y (0) | X is called
“randomisation”, “ignorability”, or “no unmeasured confounders”.
Remark 3: An synonym for counterfactual is potential outcome. I like to
use potential outcome for randomised experiments (looking forward) and
counterfactual for observational studies (looking backward).
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 18 / 57
Part II: How to define causality?
Definition 2: Graphical models
X1 X2

A Y

Probabilistic graphical models/Bayesian networks (Pearl, 1985; Lauritzen,

1996): Joint distribution factorises according to the graph:
P(X1 = x, X2 = x, A = a, Y = y )
=P(X1 = x1 , X2 = x2 ) P(A = a | X1 = x1 , X2 = x2 ) P(Y = y | X2 = x2 , A = a).

We can obtain conditional independence between the variables by applying

the d-separation criterion (details omitted; imagine information flowing like
water).
Examples: Y ⊥
⊥ X1 | A; X1 ⊥
⊥ X2 but X1 6⊥
⊥ X2 | A (this is called collider
bias).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 19 / 57

How to define causality?

Definition 2: Graphical models

Causal graphical models (Robins, 1986; Spirtes et al., 1993; Pearl, 2000):
Joint distribution in interventional settings also described by the graph:
P(X1 = x1 , X2 = x2 , A = a, Y (a) = y )
=P(X1 = x1 , X2 = x2 ) P(A = a | X1 = x1 , X2 = x2 ) P(Y (a) = y | X2 = x2 ).

Remark: Computer scientists use the do notation introduced by Pearl:

P(Y = y | do(A = a)) = P(Y (a) = y ).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 20 / 57

How to define causality?
Definition 3: Structural equations (Wright, 1920s; Haavelmo, 1940s)
X1 X2

A Y

From the graph we may define a set of structural equations:

X1 = fX1 (X1 ),
X2 = fX2 (X2 ),
A = fA (X1 , X2 , A ),
Y = fY (A, X2 , Y ).
Parameters in the structural equations are causal effects. For example, if
fY (A, X2 , Y ) = βAY A + βXY X2 + Y , then βAY is the causal effect of A on Y .
Remark: Structural equations are different from regressions that only model
the conditional expectation E[Y | A, X ].

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 21 / 57

Unification of the definitions
Define counterfactual from graphs
Structural equations are structural instead of regression because they also
govern the interventional settings (Pearl, 2000):

Y (a) = FY (a, X , Y ).

That is, Y (0) = FY (0, X , Y ) and Y (1) = FY (1, X , Y ) share the

randomness in X and Y .

Single-world intervention graphs (Richardson and Robins, 2013)

Distribution of counterfactuals factorises according to an extended graph
(obtained by splitting and relabelling the nodes).

X1 X2

A a Y (a)

Apply the d-separation, we get Y (a) ⊥

⊥ A | X2 (and also Y (a) ⊥
⊥ A | X1 , X2 ).
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 22 / 57
Recap

“Equivalence” of the definitions of causality

Graphical models
→ Define structural equations
→ Define counterfactuals
→ Embed in extended graph.

Strengths of the different approaches

Graphical model: Good for understanding the scientific problems.
Structural equations: Good for fitting simultaneous models for the variables
(especially for abstract constructs in social sciences).
Counterfactuals: Good for articulating the inference for a small number of
causes and effects.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 23 / 57

Modern causal inference
Logic of randomised experiment
Randomisation (1) =⇒ a choice of statistical error (2) vs. causality (3).

Logic of observational studies

View randomisation as a breakable identification assumption.
I Examples: need to use pseudo-RNGs; non-compliance and missing data.
Causal inference from observational studies becomes a choice between
1 Identification and modelling assumptions being violated;
2 Statistical error;
3 True causality.
Causal inference is abductive (inference to the best explanation).
I Strength of causal inference = credibility of the assumptions.
Cycle of statistical research is restored:
Conjecture Data collection

Analysis Modelling
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 24 / 57
Part III: Designing observational studies

Conjecture Data collection

Analysis Modelling

Study design = How data are collected in a study.

This is slightly different from the traditional notion of experimental design

(often about how to minimise the statistical error in a regression analysis).
In modern causal inference, study design refers to how data are collected
to meet the identification assumption (independent of analysis).
I Common designs in observational studies: controlling for confounders,
instrumental variables, regression discontinuity, difference-in-differences.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 26 / 57

Design trumps analysis (Rubin, 2008)
Logic of observational studies
Causal inference from observational studies becomes a choice between
1 Identification and modelling assumptions being violated;
2 Statistical error;
3 True causality.

A decomposition of estimation error (Zhao, Keele, and Small, 2019)

Causal estimator − True causal effect
= Design bias + Modelling bias + Statistical noise.

The first term (Design bias) is fixed once we decide how to collect data.
The last two terms resemble the familiar bias-variance trade-off in statistics.
We can hope to make it small by using better statistical methods and or
having a large sample.
=⇒ Design  Modelling > Analysis.
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 27 / 57
Design 1: Controlling for confounders

X1 X2

A Y

Loosely speaking, confounders are common causal ancestors of the treatment

and the outcome (for example, X2 in the above graph).

Identifying assumption: No unmeasured confounders

In counterfactual terms: Y (0) ⊥
⊥ A | X and Y (1) ⊥
⊥ A | X for measured X .
In the above example, this would hold if X = X2 or X = (X1 , X2 ). It would
not hold if X = X2 and there is another U3 affecting both A and Y directly.
This can be checked using the single-world intervention graphs.
This assumption is also called ignorability, exogeneity, unconfoundedness,
selection on observables, etc.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 29 / 57

Which covariates should be controlled for?
Counterfactualists: Measuring pre-treatment covariate always helps
Rubin (2009), replying to Pearl and others:
I cannot think of a credible real-life situation where I would
intentionally allow substantially different observed distributions of a true
covariate in the treatment and control groups.
Logic: observational studies should try to mimic randomised experiments.

Graphists: Counterexample (M-bias)

U1 U2

A Y

X is measured, U1 and U2 are unmeasured, all temporally precede A.

Conditioning on X introduces spurious association between A and Y .

This debate is still ongoing. My take: measure as many covariates as

possible, but think about if any would introduce bias via the M-structure.
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 30 / 57
Statistical methods: Approach 1

Create a pseudo-population to mimic randomised experiment

Matching: Create pairs of treated and control participants with similar
pre-treatment characteristics (in terms of the covariates X ).
I Many algorithms: nearest-neighbour matching, Mahalanobis distance
matching, optimal matching, etc.
Propensity-score matching: Match on the (estimated) propensity score
π(X ) = P(A = 1 | X ) to reduce the dimensionality.
Stratification: Create strata/blocks in terms of X or π(X ). Treat
participants within a stratum/block as randomised.
Weighting: Weight the participants by the inverse of the probability of
receiving the observed treatment.
1 1
I That is, weight participant i by if Ai = 1 (treated) and by if
π(Xi ) 1 − π(Xi )
Ai = 0 (control).

Randomisation inference or regression analysis (for randomised experiments) can

then be applied to the pseudo-population.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 31 / 57

Statistical methods: Approach 2

Outcome regression (also called standardisation)

Recall that if A ⊥
⊥ Y (0) | X , then

E[Y (0)] = E[E(Y (0) | X )] = E[E[Y (0) | A = 0, X ]] = E[E[Y | A = 0, X ]].

Two steps to estimate E[Y (0)] (average counterfactual under control):

Estimate E[Y | A = 0, X ] by regression using control participants.
Average the predicted E[Y | A = 0, X ] over all participants.
We can do the same thing to estimate E[Y (1)] and take the difference to
estimate E[Y (1) − Y (0)] (average treatment effect).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 32 / 57

Statistical methods: Which one to use?
Both approaches are better than the “standard” regression (e.g.
Y = α + βA + γX + ), because interpreting the results of the “standard”
regression requires that we correctly specify the structural equation.
Both approaches are semiparametric in the sense that the “nuisance
parameters” π(X ) and E[Y | A = 0, X ] can be estimated nonparametrically.

More complicated methods

State-of-the-art: estimate π(X ) and E[Y | A = 0, X ] using machine
learning and then combine them in a “doubly robust” estimator.
What they are trying to do is to minimise the “Modelling bias”:

Causal estimator − True causal effect

= Design bias + Modelling bias + Statistical noise.

My take: Too much sophistication not really necessary in “normal”

applications. Save your time for study design and data collection. Choose the
method you are most comfortable with.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 33 / 57

Another key assumption

Overlap assumption (also called positivity)

A key assumption that was implicit in the above discussion is:

0 < π(x) = P(A = 1 | X = x) < 1, for all x.

This means that the treated participants and control participants have
overlapping X distributions.
In other words, any study participant have at least some chance of receiving
treatment (or control).

You should always check the overlap assumption and define your study
population accordingly (e.g. by comparing histograms).
Matching methods are helpful in this regard, because you can examine
whether the matched participants are indeed similar.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 34 / 57

Recap

Study designs discussed so far assume no unmeasured confounders

I Either by randomisation in randomised experiments;
I Or by treating it as an explcit assumption in observational studies.
Next: Other observational study designs that try to remove or reduce bias
due to unmeasured confounders.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 35 / 57

Design 2: Instrumental variables
U

Z A Y

Z is an instrumental variable (IV); U is unmeasured confounder.

Idea: use exogenous (or unconfounded) randomness in A.

Examples of IV
Draft lottery for Vietnam war (treatment: military service).
Distance to closest college (treatment: college education).
Favourable growing condition for crops (treatment: market price, outcome:
market demand).
Randomised cash incentive to quit smoking (treatment: quit smoking).
Randomised treatment assignment (treatment: actual treatment received,
could be different to the IV due to non-compliance).
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 37 / 57
Assumptions for instrumental variables

Z A Y

1 Z must affect A.
2 There is no unmeasured Z -Y confounders.
3 There is no direct effect from Z to Y .

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 38 / 57

Assumptions for instrumental variables

Z z A(z) a Y (a)

1 Z must affect A: A(z) depends on z.

2 There is no unmeasured Z -Y confounders: Y (a) ⊥
⊥ Z | X.
3 There is no direct effect from Z to Y : Y(a,z) = Y(a).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 39 / 57

Statistical methods for instrumental variables
U

Z A Y

Two-stage least squares (most widely used)

Stage 1: Regress A on Z and X .
Stage 2: Regress Y on predicted A from stage 1 and X .

Special case: when there is no X , this is equivalent to the Wald estimator:

Slope of Y ∼ Z regression
Slope of A ∼ Z regression
.
Remark: Can also use randomisation inference (Imbens and Rosenbaum, 2005).
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 40 / 57
How to interpret instrumental variable studies
Appropriateness of the assumptions
1 IV must affect treatment.
2 There is no unmeasured IV-outcome confounders.
3 There is no direct effect from IV to outcome.

Additional assumptions
Instrumental variable design often makes additional assumptions. Examples:
Homogeneity: Y (A = 1) − Y (A = 0) is constant.
Monotonicity: A(Z = 1) ≥ A(Z = 0) (e.g. IV is random encouragement).

Complier average treatment effect

Under monotonicity (and binary IV and treatment), it is well known that

The Wald estimator → E[Y (1) − Y (0) | A(1) = 1, A(0) = 0]

The condition {A(1) = 1, A(0) = 0} corresponds to the participants who would

comply with treatment encouragement.
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 41 / 57
Design 3: Regression discontinuity
Natural experiment: Sharp discontinuity
Covariate X : Test score.
Treatment A: Scholarship determined by test score A = I (X ≥ c).
Outcome Y : Future test score.

● Y(0)
Y(1)
Y

●
●● ●
● ●●
●● ●
● ● ●
● ●
● ●●● ●
●
●
● ● ●● ●
● ● ● ●● ●
● ●●
● ● ● ●
● ●● ●●
● ●
●
● ● ●
● ● ●
●
●
● ●● ● ●●●
●● ●
●
●
●

Regression discontinuity tries to estimate E[Y (1) − Y (0) | X = c].

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 43 / 57
Sharp regression discontinuity design
Assumptions
1 X has positive density around the discontinuity c.
2 E[Y (0) | X ] and E[Y (1) | X ] are continuous in x.

Remark: A = I (X ≥ c) satisfies the no unmeasured confounders assumption

Y (0) ⊥
⊥ A | X but not the overlap assumption 0 < P(A = 1 | X = x) < 1.

Statistical methods
Broken line regression: assume
(
α0 + γ0 x, if x < c,
E[Y | X ] =
α1 + γ1 x, if x ≥ c,

Jump can be estimated by (α̂1 − α̂0 ) + c(γ̂1 − γ̂0 ).

More robust: local linear regression using participants close to the
discontinuity.
Can also use randomisation inference (use randomness in X near c).
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 44 / 57
Extension
Fuzzy regression discontinuity design
A is not a deterministic function of X , but P(A = 1 | X = x) has a
discontinuity at x = c (jump size < 1).
● Y(0)
Y(1)

●
Y

● ● ●
● ●
●
●
● ● ● ●
●
● ● ●
● ●
● ●● ●
●
● ● ● ●
● ●
● ●
● ●●
● ● ●
● ● ● ●
● ● ●
● ●●● ● ● ● ●●
● ●
● ●
● ●

Can be similarly analysed (broken-line regression, local linear regression,

randomisation inference, . . . ).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 45 / 57

Design 4: Negative controls
Negative control is a general class of designs that utilise lack of direct
causal effect or association.
In other words, these designs utilise specificity of causal effect.
This approach is still under active development. It usually requires additional
assumptions beyond specificity.

Example: Instrumental variables

U

Z A Y

Key assumptions (specificity):

1 IV is independent of unmeasured confounder.
2 IV has no direct effect on outcome.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 47 / 57

Design 4: Negative control
Confirmatory factor analysis and latent variable models
X1 X4

βU
X2 U1 U2 X5

X3 X6

U1 and U2 : Latent abstract constructs (e.g. confidence, reading ability,

personality, . . . ).
X1 to X6 : Measurements of the latent variables.
Key assumption (specificity): lack of association between the
measurements (except those explained by the causal effect of U1 on U2 ).
Remark: Analysis of these designs usually relies on strong parametric
assumptions.
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 48 / 57
Design 4: Negative control

Example: Difference-in-differences (DID)

U

W A Y

W and Y are repeated measurements before and after the intervention.

Example: A is change in minimum wage. W and Y are unemployment rates
before and after the change.
Key assumption (specificity): Lack of direct effect of A on W .

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 49 / 57

Design 4: Negative control
Example: Difference-in-differences (DID)
DID requires an stronger assumption (than just specificity) called parallel
trends:
E[Y (0) − W | A = 1] = E[Y (0) − W | A = 0].

Estimator: “difference in differences” as illustrated in the figure.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 50 / 57

Summary
Part I: Randomised experiments
Randomisation =⇒ choose between 1. Statistical error and 2. Causality.
Statistical methods: randomisation inference and regression analysis.

Part II: How to define causality

1. Counterfactuals; 2. Graphical models; 3. Structural equations.
“Equivalence” of the definitions and their relative strengths.
Logic of observational studies: Choose between 1. False assumptions; 2.
Statistical error; 3. Causality.

Part III: Designing observational studies

Design 1: Controlling for confounders;
Design 2: Instrumental variables;
Design 3: Regression discontinuity;
Design 4: Negative controls.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 51 / 57

Principles of causal inference

Observation (seeing) is not intervention (doing).

Randomised experiment is the gold standard of causal inference.
Causal inference is abductive (inference to the best explanation).
Internal, external, and construct validities.
Design trumps analysis.
Cycle of statistical research.
Conjecture Data collection

Analysis Modelling

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 52 / 57

Further readings
Book-long treatments (from less mathematical to most mathematical):
Mackenzie and Pearl (2018) The Book of Why: The New Science of Cause
and Effect. [General]
Rosenbaum (2017) Observation and Experiment: An Introduction to Causal
Inference. [General]
Freedman (2009) Statistical Models: Theory and Practice. [Undergraduate]
Shadish, Cook, and Campbell (2002) Experimental and Quasi-Experimental
Designs. [Undergraduate/Postgraduate]
Angrist and Pischke (2008) Mostly Harmless Econometrics: An Empiricists
Companion. [Undergraduate/Postgraduate]
Hernán and Robins (2020) Causal Inference: What If. [Part I:
Undergraduate; Part II & III: Postgraduate]
Imbens and Rubin (2015) Causal Inference for Statistics, Social, and
Biomedical Sciences. [Postgraduate]
Pearl (2009) Causality: Models, Reasoning, and Inference. [Postgraduate]
Rosenbaum (2010) Design of Observational Studies. [Postgraduate]
Zhao (2019) Causal Inference Lecture Notes. [Postgraduate; unpublished and
available upon request].
Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 53 / 57
Further readings
Randomised experiments
Experimental design: Box (1978) Statistics for Experimenters: Design,
Innovation, and Discovery.
Randomisation inference: Rosenbaum (2002) Observational Studies.
Imbens and Rubin (2015, Chapter 5)
Regression adjustment: Imbens and Rubin (2015, Chapter 7).

Languages of causal inference

Counterfactuals: Imbens and Rubin (2015, Chapters 1–2); Hernán and
Robins (2020, Chapters 1–3).
Graphical models: Lauritzen (1996) Graphical Models [probabilistic
graphical models only]; Pearl (2009); Spirtes, Glymour, and Scheines (2000)
Causation, Prediction, and Search.
Structural equations: Bollen (1989) Structural Equations with Latent
Variables; Peters, Janzing, and Schölkopf (2017) Elements of Causal
Inference: Foundations and Learning Algorithms.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 54 / 57

Further readings
Observational studies
Controlling for confounders (randomisation inference): Rosenbaum
(2002, 2010);
Controlling for confounders (pseudo-population): Imbens and Rubin
(2015); Stuart (2010) Matching Methods for Causal Inference: A Review and
a Look Forward (in Statistical Science).
Controlling for confounders (regression and semiparametric inference):
Hernán and Robins (2020).
Instrumental variables: Angrist and Pischke (2008); Baiocchi, Cheng, Small
(2015) Tutorial in Biostatistics: Instrumental Variable Methods for Causal
Inference (in Statistics in Medicine).
Regression discontinuity: Shadish, Cook, and Campbell (2002); Imbens
and Lemieux (2008) Regression discontinuity designs: A guide to practice (in
Journal of Econometrics).
Structural equations with latent variables: Bollen (1989).
Difference in differences: Angrist and Pischke (2008).

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 55 / 57

Further readings
Topics not covered in this lecture
Sequentially randomised experiments: Multiple treatments at different time. See
Hernán and Robins (2020).
Effect modification (treatment effect heterogeneity): Estimate
E[Y (1) − Y (0) | X = x] as a function of x. See the results from a recent data
challenge in the journal Observational Studies.
Dynamic treatment regimes: How to optimally make sequential interventions?
See Kosorok and Laber (2019) Precision Medicine (in Annual Review of Statistics
and Its Application).
Sensitivity analysis: What if the identification assumptions are violated to a
limited degree? See Rosenbaum (2002, 2010).
Causal mediation analysis: Seperate direct and indirect causal effects. See
Vanderweele (2015) Explanation in Causal Inference: Methods for Mediation and
Interaction.
Corroboration of evidence (research synthesis): How to combine evidence from
different studies (possibily with different designs)? Often done in a qualitative way,
more quantitative developments needed. Classical book: Hedges and Olkin (1985)
Statistical Methods for Meta-Analysis.

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 56 / 57

Resources in Cambridge

The Statistical Laboratory has a free consulting service called Statistics Clinic
(http://www.talks.cam.ac.uk/show/index/21850).
I run a reading group in causal inference
(http://talks.cam.ac.uk/show/index/105688).
I run a Part III course in causal inference for maths students
(http://www.statslab.cam.ac.uk/~qz280/teaching/Causal_
Inference_2019.html).
There are several causal inference researchers in MRC Biostatistics Unit,
Cambridge social sciences and other subjects.
Best way to reach me: email me (qz280@cam) about my availability in the
Statistics Clinic.

That’s all! Questions?

Qingyuan Zhao (Stats Lab) Causal Inference: An Introduction SSRMP 57 / 57