2
$\begingroup$

I am trying to understand the practicalities of Pearl's front-door criteria to estimate causal effects under an unobserved unconfounder. First, to give context to the question we have a graph that looks like this:

Front-door criterion

Where $U$ is unobserved and, hence, no back-door adjustment formula can be used.

We can infer the causal effects on $Y$ of an intervention $do(X=x)$ by using the front-door adjustment formula Pearl (2009 p.81-83):

$$P(y | \textit{do}(X = x)) = \sum_z P(z | x) \sum_{x'} P(y|x', z)P(x').$$

My question is: What is the intuition of conditioning $Y$ on $x'$ ($P(y|x', z)$) if $Y$ is conditionally independent of $X$ given $Z$? Does that mean, in practical terms, that the way to estimate the front-door adjustment is by using the following:

$$P(y | \textit{do}(X = x)) = \sum_z P(z | x) \sum_{x'} P(y|z)P(x').$$

$\endgroup$
2
  • $\begingroup$ Or does it mean that I have to build an alternative estimator using both z and x as predictors? $\endgroup$
    – Sergio
    Commented Feb 5, 2020 at 15:10
  • 1
    $\begingroup$ If you haven't read it already, I would recommend Pearl's book The Book of Why, pages 224 to 231. He explains the intuition behind the front-door adjustment formula. $\endgroup$ Commented Feb 11, 2020 at 18:37

2 Answers 2

1
$\begingroup$

Y is independent of X given Z, but Z isn't independent of X. Which you're manipulating (or, well, simulating the manipulation, anyway).

An intervention on the value of X changes the probabilities of Zs you're going to observe. That's why we need the first part, summing/integrating over the post-intervention distribution $p(z|x)$.

However, both X and Y are influenced by the unobserved variable 'in the wild'. By intervening on X, we eliminate U's influence on X - but not on Y. It's still there, and we still need to account for it.

The $p(x')$ is the pre-intervention probability on values of X, so you cannot use your final equation. The term is actually derived using the independence assumptions to substitute for a $\sum_Up(y|z,u)p(u)$ term. Since U is not easily measurable while X' can be, we're inferring U from pre-intervention X.

$\endgroup$
0
$\begingroup$

To answer your question straight,

$$P(y | \textit{do}(X = x)) = \sum_z P(z | x) \sum_{x'} P(y|x', z)P(x').$$

can't be rewritten as

$$P(y | \textit{do}(X = x)) = \sum_z P(z | x) \sum_{x'} P(y|z)P(x')$$

because $P(y|x', z) \neq P(y|z)$ in the pre-intervention distribution given the presence of U, the confounder.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.