Is the model limited to randomized trials?

No; it also structures observational causal inference.

Can AI recover every missing potential outcome?

No; model predictions remain assumption-dependent.

Are prediction and causation interchangeable?

No; accurate forecasting under observed conditions does not establish effects of changing an intervention.

AI Glossary / Statistics and probability / Intermediate

Rubin causal model

Definition

The Rubin causal model is a framework for defining and studying causal effects using potential outcomes: the outcomes a unit would have under different interventions. A causal effect compares those outcomes for the same unit or population. Because only the outcome under the intervention actually received is observed, causal inference requires a study design and assumptions connecting observed data to the unobserved alternatives.

What it means

For a person, business, or device, imagine the outcome under a specified intervention and under a specified alternative. Their difference defines an individual causal effect, even though both outcomes cannot ordinarily be observed for that same unit at the same time.
The framework is often called the potential-outcomes framework or Neyman - Rubin model. It clarifies the question before choosing an estimator. It does not turn an ordinary predictive model or a raw difference between groups into causal evidence automatically.

How it works

Define the unit, intervention, comparator, outcome, and follow-up period. For example, offering an AI assistant and actually using it are different interventions. Ambiguous treatment versions can make the intended potential outcomes unclear.
For a binary treatment, write Y(1) for the potential outcome under treatment and Y(0) under control. The individual difference is Y(1)−Y(0). A population average treatment effect is E[Y(1)−Y(0)], while an effect among treated units is a different estimand.
Under consistency, the observed outcome equals the potential outcome corresponding to the treatment received. The familiar simple formulation also assumes a unit's outcome does not depend on other units' assignments and that treatment versions are adequately specified. Interference requires an expanded definition, not merely ignoring spillovers.
Study the assignment mechanism. Randomization supports comparison by making assignment independent of potential outcomes under the design. It provides balance in expectation rather than guaranteeing identical groups in every realized sample.
For observational identification through adjustment, a common requirement is conditional exchangeability: after conditioning on appropriate pre-treatment covariates, treatment assignment is independent of the relevant potential outcomes. This assumption can fail because of unmeasured confounding.
Positivity requires a chance of receiving each intervention being compared within the covariate strata relevant to the target population. If a treatment is never available for a group, estimating its effect there requires extrapolation or a different target question.
Choose an estimator consistent with the design and assumptions, such as a randomized group comparison or an appropriate adjustment procedure. Machine learning may help estimate nuisance relationships, but flexible prediction does not supply missing identification assumptions.
Report uncertainty, departures from the intended assignment, missing outcomes, and the population to which conclusions apply. Individual effects are generally not identified simply because an average effect can be estimated. Sensitivity analysis can examine how conclusions depend on assumptions that the observed data cannot establish alone.

Three practical examples

Product experimentation: a company defines the effect of offering a new interface on completion rates.
Workplace AI evaluation: researchers distinguish the effect of assigned access from the effect of actual assistant use.
Policy assessment: analysts specify counterfactual outcomes under alternative service rules before selecting an observational design.

Think of it this way

Think of two possible paths for the same business at the same starting point: one with a defined intervention and one without it. Reality reveals one path. Comparing different businesses can help estimate the missing contrast only when the design and assumptions justify treating them as informative about one another.

Worked example

Consider four fictional units with stipulated potential outcomes. Under control, Y(0) values are 10, 20, 30, and 40; under treatment, Y(1) values are 14, 22, 30, and 46. Their individual effects are 4, 2, 0, and 6, averaging 3 outcome units.
Suppose the first two units receive treatment and the last two receive control. The observed treated mean is (14+22)/2 = 18; the observed control mean is (30+40)/2 = 35. Their difference is −17, although the stipulated average causal effect is +3.
If exactly two units are selected uniformly at random for treatment, there are six possible assignments. Averaging the observed difference-in-means over all six gives +3. Thus randomization makes this estimator unbiased under the simple setup, but one realized estimate can differ greatly. In real data, the complete potential-outcome table is unavailable.

Illustrative case study

Fictional case: Willow Support tests access to an AI drafting assistant. The team defines its primary estimand as the effect of assigned access on ticket-resolution time over a fixed period. It randomizes eligible teams because workers within a team may share practices.
Analysts retain the assignment-based comparison and report uncertainty at the appropriate team level. They do not relabel an association among enthusiastic users as the causal effect of actual use without addressing the additional selection problem.

Common misconceptions and mistakes

Potential outcomes are not both observed merely because a predictive model generates two numbers.
Randomization does not guarantee a perfect estimate in every sample.
Adjusting for many variables is not automatically sufficient or safe; post-treatment variables can change or bias the intended comparison.
An average effect does not show that every unit benefits.
The framework defines causal questions, but credibility still depends on intervention clarity, design, data quality, and identifying assumptions.

Business use

Use this framework to turn vague impact claims into explicit comparisons. Before launching an AI project, specify what intervention is being evaluated, which outcome matters, and whether the target is an effect of offering, assigning, or using the tool. This makes experimental planning and interpretation more disciplined.

Check your understanding

Why does the worked example reveal both outcomes for each unit?

Answer: They are invented to demonstrate the concept, not real observations.

Does an unbiased estimator equal the causal effect in every experiment?

Answer: No; unbiasedness concerns its average over the assignment mechanism.

Related terms

Average treatment effect
Confounding
Consistency
Potential outcome
counterfactual
randomization
positivity
interference

Learn next

Write a potential-outcome question for one proposed business intervention. Identify the assignment mechanism and the assumptions needed before choosing a statistical or machine-learning estimator.

Machine learning

Sources

Rubin, Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies, Harvard repository: https://dash.harvard.edu/entities/publication/73120378-82c0-6bd4-e053-0100007fdf3b Hernán and Robins, Causal Inference: What If: https://miguelhernan.org/whatifbook The potential-outcome table and case are original fictional examples. Jonathan Wand, Stanford, Potential Outcomes and Causal Inference I: https://web.stanford.edu/class/polisci350c/classonly/causal.pdf

Previous: RST Next: Rule-based machine learning