Causal inference course note - Week 5

This is my note for the “A Crash Course in Causality: Inferring Causal Effects from Observational Data” course by Jason A. Roy on Coursera.



Table of Contents


Week 5. Instrumental Variables Methods

Introduction to instrumental variables

Confounding

Classic confounding situation: $X$ affects treatment $A$ and affects outcome Y: $X \rightarrow A \rightarrow Y \leftarrow X$. If $X$ is observed, we can analyze data using: matching, PS matching, IPTW. Even if there are risk factors, $V$, such that $V \rightarrow Y$, it’s valid to simply control for $X$.

Unmeasured confounding

Suppose there are unmeasured/unobserved variables, $U$, that affect $A$ and $Y$: $A \dashleftarrow U \dashrightarrow Y$. Then we have unmeasured confounding.

Instrumental variables

Instrumental variables (IV) is an alternative causal inference method that does not rely on the ignorability assumption. Here $Z$ is an IV:

\[Z \rightarrow A \rightarrow Y \leftarrow X \rightarrow A\]

It affects treatment, but does not (directly) affect the outcome. Think of $Z$ as encouragement.

Example

Concern: Could be unmeasured confounders.

Challenge: not ethical to randomly assign smoking to pregnant women.

Encouragement design

Sexton and Hebel, JAMA, 1984.

An intention-to-treat analysis would focus on the causal effect of encouragement:

\[E(Y^{Z=1}) - E(Y^{Z=0})\]

This is a valid causal effect and would likely be of some interest.

What can we say about the causal effect of smoking itself? This is the focus of IV methods.

Instrumental variables

Sometimes the IV is randomly assigned as part of the study. Other times, the IV is believed to be randomized in nature (natural experiment):

We will look at both situations in other lectures.

Randomized trials with noncompliance

Setup

Randomize trial:

DAG

\[Z \rightarrow A \rightarrow Y \leftarrow X \rightarrow A\]

Potential treatment

Causal effect of assignment on receipt

We can think of the average causal effect of treatment assignment on treatment received as:

\[E(A^1 - A^0)\]

Causal effect on assignment on outcome

We can think of the average causal effect of treatment assignment on the outcome as:

\[E(Y^{Z=1} - Y^{Z=0})\]

Causal effect of treatment

What about the causal effect of treatment received on the outcome?

Compliance classes

Potential values of treatment

We can classify people based on potential treatment.

$A^0$ $A^1$ Label
0 0 Never-takers
0 1 Compliers
1 0 Defiers
1 1 Always-takers

Angrist, Imbens, & Rubin, JASA, 1996.

Subpopulations

We can think of these as subpopulations of people:

Causal effects

A motivation for using IV methods in general is concern about possible unmeasured confounding.

IV methods do not focus on the average causal effect for the population.

Local average treatment effect

In IV methodcs, the target of inference is:

\[E(Y^{Z=1} \vert A^0 = 0, A^1 = 1) - E(Y^{Z=0} \vert A^0 = 0, A^1 = 1)\] \[= E(Y^{Z=1} - Y^{Z=0} \vert \text{compliers})\] \[= E(Y^{A=1} - Y^{A=0} \vert \text{compliers})\]

This is causal because it contrasts counterfactuals in a common population. Known as complier average causal effect (CACE).

Observed data

For each person we observe an $A$ and a $Z$, not ($A^0$, $A^1$).

$Z$ $A$ $A^0$ $A^1$ Class
0 0 0 ? Never-takers or compliers
0 1 1 ? Always-takers or defiers
1 0 ? 0 Never-takers or defiers
1 1 ? 1 Always-takers or compliers

Without additional assumptions, we cannot classify each subject into one of these categories. We can narrow it down to two options, however.

Identifiability

Compliance classes above are also known as principal strata. These are latent (not directly observable).

How can we estimate the complier average causal effect? What assumptions are needed? These are topics for other lectures.

Assumptions

Assumptions about IVs

A variable is an instrumental variable (IV) if:

  1. It is associated with the treatment.
  2. It affects the outcome only through its effect on treatment (known as the exclusion restriction).
\[Z \rightarrow A \rightarrow Y\] \[A \leftarrow X \rightarrow Y\]

If $Z$ is an IV, $Z$ must not directly affect $Y$ (exclusion restriction).

When there is unmeasured confounding,

\[Z \rightarrow A \rightarrow Y\] \[A \leftarrow X \rightarrow Y\] \[A \dashleftarrow U \dashrightarrow Y\]

$Z$ must not affect unmeasured confounders $U$. $Z$ cannot directly, or indirectly through its effect on $U$, affect the outcome $Y$.

Realistic?

IF $Z$ is random treatment assignment, are the IV assumptions met?

Causal effects

Assuming we have a valid IV, we are interested in using it to help us estimate the complier average causal effect:

\[E(Y^{A=1} - Y^{A=0} \vert \text{compliers})\]

The causal effect of treatment among subjects who only take treatment if they are randomized to $Z=1$.

Identification challenge

Challenge: We are interested in compliers, but we do not know who the compliers are.

Monotonicity assumption

The monotonicity assumption is there are no defiers.

With monotonicity,

$Z$ $A$ $A^0$ $A^1$ Class
0 0 0 ? Never-takers or compliers
0 1 1 ? 1 Always-takers or defiers
1 0 ? 0 0 Never-takers or defiers
1 1 ? 1 Always-takers or compliers

Causal effect identification and estimation

[TODO]

Comments