Fundamental Problem of Causal Inference

For each unit, at most one of the two potential outcomes can be observed:

$$Y_i(1) \quad \text{or} \quad Y_i(0),$$

but not both. This means causal effects are inherently missing-data problems at the unit level.

The entire logic of randomization, assignment mechanisms, and assumptions such as SUTVA is built around coping with this missing counterfactual information.