Causal graphical model reading group 2022

Causal inference



\[\renewcommand{\var}{\operatorname{Var}} \renewcommand{\corr}{\operatorname{Corr}} \renewcommand{\dd}{\mathrm{d}} \renewcommand{\bb}[1]{\mathbb{#1}} \renewcommand{\vv}[1]{\boldsymbol{#1}} \renewcommand{\rv}[1]{\mathsf{#1}} \renewcommand{\vrv}[1]{\vv{\rv{#1}}} \renewcommand{\disteq}{\stackrel{d}{=}} \renewcommand{\gvn}{\mid} \renewcommand{\Ex}{\mathbb{E}} \renewcommand{P}{\mathbb{P}} \renewcommand{\indep}{\mathop{\perp\!\!\!\perp}}\]

My chunk (Chapter 3) of the internal reading group group covering the Brady Neal course.

Recap: potential outcomes

Last time we discussed the potential outcome framework, which answers the question: How do we calculate a treatment effect \(Y(t)\) for some treatment \(t\)? i.e. how do we calculate \[ \Ex[Y(1)-Y(0)]? \]

We used the following assumptions: \[\begin{aligned} (Y(1), Y(0))\indep T | X &\quad \text{Unconfoundedness}\\ Y = Y(t)&\quad \text{Consistency}\\ (Y(1), Y(0))\indep T&\quad \text{Overlap}\\ (Y(1), Y(0))\indep T&\quad \text{No interference}\\ \end{aligned}\]

Under those assumptions, we have the causal adjustment formula \[\Ex[Y(1) − Y(0)] = \Ex_{X}\left[ \Ex[Y | T{=}t, X] − \Ex[Y | T{=}s, X]\right]. \]

Aside: what is going on in positivity?

And now…

Graphical models for causation wrangling

We have a finite collections of random variables \(\mathbf{V}=\{X_1, X_2,\dots\}\).

For simplicity of exposition, each of the RVs will be discrete so that we may work with pmfs, and write \(P(X_i|X_j)\) for the pmf. I sometimes write \(P(x_i|x_j)\) to mean \(P(X_i=x_i|X_j=x_j)\).

More notation. We write \[ X \indep Y|Z \] to mean “\(X\) is independent of \(Y\) given \(Z\)”.

We can solve these questions via a graph formalism. That’s where the DAGs come in.

Directed Acyclic Graphs (DAGs)

A DAG is graph with directed edges, and no cycles. (you cannot return to the same starting node travelling only forward along the arrows.)

DAGs are defined by a set of vertexes and (directed) edges.

We show the directions of edges by writing them as arrows.

For nodes \(X,Y\in \mathbf{V}\) we write \(X \rightarrow Y\) to mean there is a directed edged joining them.

Bayesian networks

Local Markov assumption

Given its parents in the DAG, a node is independent of all its non-descendants.

With four variable example, the chain rule of probability tells us that we can factorize any \(P\) thus \[ P\left(x_{1}, x_{2}, x_{3}, x_{4}\right)=P\left(x_{1}\right) P\left(x_{2} \mid x_{1}\right) P\left(x_{3} \mid x_{2}, x_{1}\right) P\left(x_{4} \mid x_{3}, x_{2}, x_{1}\right) \text {. } \]

Abstract DAG

Figure 1: Abstract DAG

If \(P\) is Markov with respect to the above graph then we can simplify the last factor: \[ P\left(x_{1}, x_{2}, x_{3}, x_{4}\right)=P\left(x_{1}\right) P\left(x_{2} \mid x_{1}\right) P\left(x_{3} \mid x_{2}, x_{1}\right) P\left(x_{4} \mid x_{3}\right) . \]

If we further remove edges, removing \(X_{1} \rightarrow X_{2}\) and \(X_{2} \rightarrow X_{3}\) as the below figure,

Abstract DAG 2

Figure 2: Abstract DAG 2

we can further simplify the factorization of \(P\) : \[ P\left(x_{1}, x_{2}, x_{3}, x_{4}\right)=P\left(x_{1}\right) P\left(x_{2}\right) P\left(x_{3} \mid x_{1}\right) P\left(x_{4} \mid x_{3}\right) \]

Bayesian Network Factorization

Given a probability distribution \(P\) and a DAG \(G\), we say \(P\) factorizes according to \(G\) if \[ P\left(x_{1}, \ldots, x_{n}\right)=\prod_{i} P\left(x_{i} \mid \operatorname{parents}(X_i)\right). \]

Minimality

  1. Given its parents in the DAG, a node X is independent of all its non-descendants
  2. Adjacent nodes in the DAG are dependent.

Causal interpretation

Causal association from Neal (2020)

Causal Edges

In a directed graph, every parent is a direct cause of all its children.

\(Y\) “directly causing“ \(X\) means that a \(X=f(\operatorname{parent}(X),\omega)\) is a (stochastic) function of some parent set which includes \(Y,\) and some independent noise.

Causal Bayesian Networks

Causal Edges + Local Markov

Conditional independence in Bayesian networks

When we fix some nodes, which independences do we introduce?

Chains

B in a chain path

\[P(a,b,c) = P(a)P(b|a)P(c|b)\]

We assert that, conditional on B, A and C are independent: \[ A\indep C | B \\ \Leftrightarrow\\ P(a,c|b) = P(a|b)P(c|b) \]

In slow motion, \[\begin{aligned} P(a,b,c) &= P(a)P(b|a)P(c|b)\\ P(a,c|b) &=\frac{P(a)P(b|a)P(c|b)}{P(b)}\\ &=P(c|b)\frac{P(a)P(b|a)}{P(b)}\\ &=P(c|b)\frac{P(a,b)}{P(b)}\\ &=P(c|b)P(a|b) \end{aligned}\]

Forks

B in a fork path

\[P(a,b,c) = P(b)P(a|b)P(c|b)\]

We assert that, conditional on B, A and C are independent: \[ A\indep C | B \\ \Leftrightarrow\\ P(a,c|b) = P(a|b)P(c|b) \] In slow motion, \[\begin{aligned} P(a,b,c) &= P(b)P(a|b)P(c|b)\\ P(a,c|b) &=\frac{P(b)P(a|b)P(c|b)}{P(b)}\\ &=P(a|b)P(c|b) \end{aligned}\]

Immoralities

(Colliders when I grew up.)

B in a collider path

\[P(a,b,c) = P(b)P(c)P(a|b,c)\]

We assert that, conditional on B, A and C are not in general independent: \[ A \cancel{\indep} C | B \\ \Leftrightarrow\\ P(a,c|b) = P(a|b)P(c|b) \]

Proof that this never factorizes?

Blocked paths

A path between nodes \(X\) and \(Y\) is blocked by a (potentially empty) conditioning set \(Z\) if either of the following is true:

  1. Along the path, there is a chain \(\cdots \rightarrow W \rightarrow \cdots\) or a fork \(\cdots \leftarrow W \rightarrow \cdots\), where \(W\) is conditioned on \((W \in Z)\).
  2. There is a collider \(W\) on the path that is not conditioned on \((W \notin Z)\) and none of its descendants are conditioned on \((\operatorname{descendants}(W) \nsubseteq Z)\).

d-separation

Two (sets of) nodes \(\vv{X}\) and \(\vv{Y}\) are \(d\)-separated by a set of nodes \(\vv{Z}\) if all of the paths between (any node in) \(\vv{X}\) and (any node in) \(\vv{Y}\) are blocked by \(\vv{Z}\).

d-separation in Bayesian networks

We use the notation \(X \indep_{G} Y \mid Z\) to denote that \(X\) and \(Y\) are d-separated in the graph \(G\) when conditioning on \(Z\). Similarly, we use the notation \(X \indep_{P} Y \mid Z\) to denote that \(X\) and \(Y\) are independent in the distribution \(P\) when conditioning on \(Z\).

Given that \(P\) is Markov with respect to \(G\) if \(X\) and \(Y\) are d-separated in \(G\) conditioned on \(Z\), then \(X\) and \(Y\) are independent in \(P\) conditioned on \(Z\).

\[ X \indep_{G} Y |Z \Longrightarrow X \indep_{P} Y | Z. \]

d-separation implies Association is Causation

Chocolate and Nobel prizes, Messerli (2012)

Correlation tastes as good as causation.

References

Elwert, Felix. 2013. Graphical Causal Models.” In Handbook of Causal Analysis for Social Research, edited by Stephen L. Morgan, 245–73. Handbooks of Sociology and Social Research. Dordrecht: Springer Netherlands.
Messerli, Franz H. 2012. Chocolate Consumption, Cognitive Function, and Nobel Laureates.” New England Journal of Medicine 367 (16): 1562–64.
Neal, Brady. 2020. Introduction to Causal Inference from a Machine Learning Perspective.” Course Lecture Notes (Draft).
Pearl, Judea. 2018. Does Obesity Shorten Life? Or Is It the Soda? On Non-Manipulable Causes.” Journal of Causal Inference 6 (2): 20182001.

No comments yet. Why not leave one?

GitHub-flavored Markdown & a sane subset of HTML is supported.