Background to question:
In statistics, machine learning, etc. there are algorithms for estimating the causal effect of an intervention X on an outcome Y.
In order to justify these estimates, they claim they have verified the following two assumptions.
- No unobserved confounders assumption (NUCA)
- Strong Ignorability/Positivity
The first is that all confounders Z that determine the assignment X, and that also determine the outcome Y, are in the model.
The second is that for all values z that a confounder Z takes on, the probability P(Z=z|X)>0. That is, the data has examples of all combinations of Z and X.
With these two assumptions satisfied, these algorithms can then adjust the estimate of the effect of X on Y according to the confounding effects of Z.
Question:
I do not know the source of the requirement of these two assumptions, but the source ought to be philosophical.
Can someone point me towards a review of philosophical literature that covers the requirements for inferring a counterfactual outcome?