CAUSALITY - Discussion (Sjolander) Date: March 10, 2006
From: Arvid Sjolander, Dept. of Medical Epidemiology and Biostatistics, Karolinska Institutet
Subject: d-separation of counterfactuals

Question to author:
At the bottom of page 214 you mention that "any variable that is d-separated from Z* would also be d-separated from UZ" (and vice versa I guess?). You conclude that "if UZ obeys a certain independence relationship then Zx (more generally, ZpaZ) must obey that relationship as well" (and vice versa?). I guess that what you mean is

where V is any variable defined by v = f(>paV,uV), Q and P are any variables except V and UV, and denotes independence (hope the symbols look the same on your computer as on my).

On the top of page 215 you use the twin network to show that

and then refer to the previous argument to state that

I don't see why (3) follows from (1) and (2). How can (1) justify the replacement of UY with Yz and UZ with Zx in (2)? Certainly Yz is a function of UY and Zx a function of UZ, which motivates one to write (2) as

but how can you remove UY and UZ from (4)?

If I would examine if (3) is true I would naively construct the following triple network:

The left part corresponds to the world in which no intervention is imposed, the middle part in which do(X=x) is imposed, and in the right part do(Z=z) is imposed. In this network (3) does not hold since the path is open by conditioning on Y. Is there anything wrong with this use of the (generalized) twin network method?

Author's first reply (with Ilya Shpitser)
You are, indeed, correct.
Yx is not independent of X given Yz, Zx, Y.

However, for example, Yx is independent of X given Zx.

Your use of d-separation in the twin network to understand independencies between underlying counterfactual quantities is correct, and so is your generalization of the 'twin' network to more than two possible worlds. We called this the 'parallel worlds model' in the path-specific effects paper.

I have spoken to Judea about (3) last spring. Our conclusion was that (3) is not true in general, but is true if the functions from U to their children are one to one. I believe this is the case Judea had in mind when he wrote that section of the book.

Arvid's followup: 1) I see that a one-to-one correspondence between Uv and VpaV is a sufficient criteria for letting one of them serve as a proxy for the other. If paV={}, i.e. V doesn't have any parents except Uv, the criteria is automatically true since we can always choose a "non-redundant coding" of Uv such that each value V corresponds to exactly one value of Uv. If not however, it seems to me that this criteria implies quiet strong restrictions on the structural relationships in a causal model. For example it completely rules out the following simple relationship between dichotomous {z,uy} and y=f(z,u_z): y=1 if z=1 and uy=1
y=0 else

Is there ever a reason for assuming the criteria would hold true?

2) As a bonus your answer helped me with a similar question on the paper "Direct and Indirect Effects" (Pearl 2001), namely "why does (7) hold in the graph in Figure 1(a)?". If one draws a triple graph it is obvious that the path is open. If however U2 is completely determined by W, we block the path by conditioning on W, as in (7). But then again, do we any reason to believe in this one-to-one correspondance in this particular setting?