Lecture 10: Interventions and Covariate Adjustments
This week we revisit the topic of causal inference, brought up in week 1.
Example network (slide 10-4):
Exposure variable (causal factor; exposition to this thing may have an effect on a particular disease) and outcome variable.
If we want to do causal inference, the other variables in the DAG may be confounding factors.
What are the numbers at the edges? They are literature references to justify the edges, i.e. they encode common scientific knowledge into the DAG.
Why is the idea that we need references for the edges misguided?
The testable assumptions in a DAG are in the edges that are not there. This means that we instead need references for the edges that are left out, instead of the ones that are currently in! So the logic "if we can't justify it, we leave it out" really doesn't hold here!
So if you start with an empty DAG and start by adding edges based on literature (which in itself is sensible), you should combine this with model testing to check for missing edges.
You cannot find references for edges that are currently not in here
Causility theory
- Structural causal models provide a language to encode causal relationships, which are the basis of causality theory.
- The usage of DAGs for causal inference really goes back to Judea Pearl.
- Causality theory helps decide when, and how, causation can be inferred from domain knowledge and data
Another example is Donald Rubin with his potential outcome framework.
This whole Bayesian networks course can be seen as a particular intro into causality theory.
Judea Pearl's claim in "Causility": "[] all approaches to causation are variants or abstractions of [...] structural theory". (Johannes Textor doesn't agree with this)
Do-operator: modelling interventions
Some philosophers argued that causation cannot be properly defined.
But you can define "causation" within the scope of Bayesian Network models. (And then we hope that BNs are a good model of reality).
A (model-implied) causal effect can be defined as the correlation ins a modified model where all influences stemming from "natural causes" of the exposure variable are removed.
If you observe P(cancer|coffee) in an observational regime, can we than also conclude that coffee causes cancer in an interventional regime P(cancer|do (coffee)).
So a causal effect is measured by intervening and measuring the effect of the intervention. (In practice it's hard to really ensure only the intervened variable is different; also enforcing the intervention is not always ethical).
In terms of a DAG, we build a mutilated graph (oh my god BN terminology...) that removes all edges going into your exposure.
Revisiting Simpson's Paradox
Recap paradox: if we look at the whole population the treatment was beneficial, but when selecting groups on gender not.
We saw this issue was due to the trial not being randomized. Because intuitively we knew gender was a confounding factor, so intuitively we should trust the results when split on group, showing that the treatment is not effective and in fact detrimental.
We first construct prob. tables based on "seeing", i.e. in the observational regime. We have gender as a confounder, so it may screw up our conclusions.
In the interventional regime we remove the edge between S and T (we set T), which leads to the interventional distribution P(c,s|do(t)). We do not have any parameters of treatment t because we set it.
We now find that the intervention of the treatment in fact has a negative causal effect.
What do we conclude? There is apparently a confounder (with a postitive sign) that overestimate the efficacy of our treatment.
Due to confounding we can both underestimate and overestimate the results, so it is good to think about the "direction" of this bias.
How do you find confounders? Draw out BNs and test and reason about them. You cannot do a "search" for confounders because the confounders may not be in the data. There's no shortcuts, you have to think!
Wrap up: seeing and doing
post hoc ergo propter hoc fallacy: just because it happens afterwards does not mean it was caused by that thing.
In the paradox example, we overestimated the effect of the treatment because:
- males were more likely to take the treatment
- males have a higher recovery rate
Covariate adjustments
What we did ... is average over man and women for ...
But we still used observational data, just differently by only taking the ones corresponding to a particular do operator as encoded in our mutilated graph!
When we work with observational data (cohort studies, case-control studies), causal inference requires making assumptions about the underlying process:
Nancy Cartwright: No causes in - no causes out.
DAGs are a way to encode those assumptions in a principled fashion and allows us to derive testable implications of our assumptions. In practice, this means we ask:
Given this DAG, is the causal effect of X on Y identifiable and if so, how?
10-17 despite the presence of an unobserved variable, I can still observe causal effects of some relations; we can get lucky if this is the one we are interested in.
We are thus not necessarily just fitting the whole network at once, if we are interested in a specific causal effect.
See in 10-17 how P(coffee|U) drops out in the mutilated network. This makes a critical assumption namely that U does not directly cause cancer. If there would be direct effect from a common unobserved cause, we can not determine a cancer relation.
I.e. we need to be able to make this assumption in other to be able to identify a causal effect. If not, too bad, we simply can't reach an unbiased conclusion.
(That requires us to be a bit more specific about what U may be; this is problematic in itself as we saw earlier. E.g. we cannot really estimate it when it has parents).
To estimate t he causal effect of coffee drinking on cancer, we adjust for the confounder "smoking". This is called covariate adjustment.
Graphical characterisation of adjustment set (Back-Door Criterion of Pearl 2009):
- Z does not contain any descendants of X
- Z d-separates all paths from X to Y starting with X <-
REVIEW FORMULA 10-18
Typical exam question would be: give an adjustment set for this DAG.
What is a confounder
A confounder is a common ancestor of X and Y in the graph where all arrows going out of X are deleted
Many examples from here.
Was very tired and a bit ill, so stopped making notes. Review later and do exercises from the lecture notes.
Necessity of causal analysis
The "more is better" misconception: heated debate between Dubin and Pearl.
Robust Covariate Adjustment
The adjustment set depends on the DAG, but the DAG may be wrong. E.g. a statistically equivalent DAG may be the true DAG. They imply very different adjustment sets.
Surprising result (of Textor's work himself): this is not always a critical problem. R example where the adjustment set for the whole equivalence class remains identical or near identical.
Covariate adjustment does not tend to be very sensitive to a few changed edge directions.