Suppose we have that $\left(Y(1), Y(0)\right)$ are potential outcomes with $X$ being the covariate and $Z$ the treatment assignment. Typically in causal inference, one will assume strong ignorability or unconfoundedness to equate these two:

$$ E[ Y(1) \mid Z=1, X ] = E[Y(1) \mid X] $$

What I am wondering is why we cannot just use iterated expectations to get:

\begin{align*} E\bigg( E\left( Y(1) \mid Z=1, X \right) \mid X\bigg) &= E[Y(1) \mid X] \end{align*}

since it appears to me that $X$ is a subcollection of $(Z=1, X)$?

  • 1
    $\begingroup$ Two things: if you are taking the expectation over $Z$, it can't be fixed to 1. Also, your goal is not to get rid of $Z$ (you observe $Z$) but to replace counterfactual quantities (like Y(1) which is unobserved) with observed quantities (like Y). $\endgroup$ Nov 13, 2018 at 3:59
  • $\begingroup$ What is the reason that we cannot fix $Z=1$ when taking the expectation over $Z$? $\endgroup$
    – user321627
    Oct 14, 2019 at 13:05

1 Answer 1


Ignorability assumption, a.k.a. unconfoundedness (Rosenbaum and Rubin, 1983) is based on the idea that we have measured ENOUGH features so that Y(1),Y(0) are independent of the treatment conditional to the covariates X. If you numerically test it with a toy example, using random variables in the mix, you will notice that you need a LOT of Y(1),Y(0) to get the assumption correct.


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.