Here is my problem:

There is a causal Markovian model as follows.

enter image description here

By the definition of interventional probability, since $\text{do}(x)$ makes no edges between $X$ and $Z_1, Z_2$, we have

$$ P(y\mid \text{do}(x)) =\sum_{z_1,z_2,z_3}P(z_1)P(z_2)P(z_3\mid z_1, z_2)P(y \mid z_2, z_3, x), > $$ where $\text{do}(\cdot)$ is so-called do-calculus (Judea pearl, 2010).

enter image description here

Then show that the summation over $z_2$ is $$ P(y\mid \text{do}(x)) =\sum_{z_1, z_3} P(z_1) P(z_3\mid z_1)P(y \mid z_1, z_3, x). $$

Here is what is tried.

To hold the equality of the above problem, we must have $$ \sum_{z_2} P(z_3 \mid z_1,z_2)P(y,z_2,z_3,x) $$ should be same with $$ P(y,z_1,z_3,x), $$ but I don't know how this can happen.


Reference: Judea Pearl, 2010, An Introduction to Causal Inference, The International Journal of Biostatistics, pp15-16


I just found that $z_1$ can be inserted in $P(y_1 \mid z_2, z_3, x)$ due to $z_2$ and $z_3$ are sufficient statistics.

$$ P(y_1 \mid z_2, z_3, x) = P(y_1 \mid z_1, z_2, z_3, x) $$

Therefore, the first equation becomes, $$ P(y_1 \mid \text{do}(x)) = \sum_{z_1,z_2,z_3}P(z_1)P(z_2)P(z_3\mid z_1,z_2)P(y_1 \mid z_1, z_2, z_3, x) \\ = \sum_{z_1,z_2,z_3}P(z_1, z_2, z_3)P(y_1 \mid z_1, z_2, z_3, x) \\ = \sum_{z_1,z_2,z_3}P(z_1, z_2, z_3)\frac{P(y_1, z_1, z_2, z_3, x)}{P(z_1, z_2, z_3, x)} \\ = \sum_{z_1,z_2,z_3}P(z_1, z_2, z_3)\frac{P(y_1, z_1, z_2, z_3, x)}{P(z_1, z_2, z_3)} \quad \text{(because X is forced to be constant x)} \\ = \sum_{z_1,z_2,z_3}P(y_1, z_1, z_2, z_3, x) \\ = \sum_{z_1,z_3}P(y_1, z_1, z_3, x) \\ = \sum_{z_1,z_3}P(z_1, z_3) \frac{P(y_1, z_1, z_3, x)}{P(z_1,z_3)} \\ = \sum_{z_1,z_3}P(z_1)P(z_1 \mid z_3)P(y_1\mid z_1, z_3, x) $$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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