Potential Outcomes Framework

Potential outcomes is a way to think about cause and effect by comparing what actually happened to what would have happened under different conditions. Every decision has multiple possible futures—think of it as a branching path where you can only walk down one, but we want to know what would've occurred down the other. Try this: The next time you see a headline about what "works," ask yourself what would have happened without it.

The potential outcomes framework defines causation as a comparison between two states of the world: what actually happened versus what would have happened under a different treatment or condition. Each person or unit has potential outcomes under each possible condition, but we only ever observe one. This creates the fundamental problem of causal inference—we can never directly observe the effect for any individual because we can't see both outcomes simultaneously.

The core insight is that causal effects are defined at the individual level as the difference between potential outcomes, even though we can never measure this directly.

The potential outcomes framework, also called the Rubin Causal Model, was formalized by Donald Rubin in the 1970s building on earlier work by Jerzy Neyman in 1923. It defines causal effects in terms of counterfactual quantities: for each unit, there exists a potential outcome under treatment (Y₁) and under control (Y₀), but only one is ever observed. The individual causal effect is Y₁ - Y₀, while the average treatment effect (ATE) is the expected value of this difference across the population. Holland (1986) dubbed this the "fundamental problem of causal inference"—we can never observe both potential outcomes for the same unit simultaneously [2].

The potential outcomes framework has several important limitations. First, it assumes consistency—the treatment version is well-defined and the same for all units—but real-world interventions often have multiple implementation versions, complicating causal claims (what exactly is "the training program"?). Second, it requires the stable unit treatment value assumption (SUTVA), meaning one unit's treatment doesn't affect another's outcome and no interference occurs between units; this fails in settings like network effects or herd immunity. Third, the framework struggles with dynamic decisions where treatment choices depend on past outcomes, requiring more complex modeling. Fourth, it focuses primarily on average effects across populations—individual-level causal effects remain fundamentally unobservable. Finally, the framework doesn't specify how causal effects operate mechanistically; it defines what they are but not why they occur, meaning it needs to be combined with subject-matter knowledge to be useful for explanation.