How do counterfactuals relate to the expected reward or expected “utility” used in medical decision analysis?
To compute the counterfactual expectation, , where
is some binary treatment,
is continuous, and
contains all confounders, we write
In decision analysis we want to compute Recall that we have
as a binary treatment,
as a continuous post-treatment vector of covariates, and
as a continuous pre-treatment vector of covariates.
We write
Now suppose , and we care about the expected reward when fixing
In other words, we care about the expected reward under a policy
where we draw action
with probability 1.
We get then
Rename as
and
as
Another way: The expected counterfactual is
With the expected utility all we are doing is
Now what happens if Then
which is
Leave a comment