A counterfactual is a statement or supposition about what would or might happen if circumstances differed from actuality. A typical example is: “If the switch had been pressed, the lamp would have lit.” In philosophy, counterfactuals raise questions about possibility, explanation, and causation; in empirical research, they help express outcomes under alternative interventions. Although their name suggests something contrary to fact, counterfactual grammatical forms do not invariably establish that their antecedents are false. (plato.stanford.edu)
Form and logical behavior
A counterfactual conditional has an antecedent, specifying the supposed circumstance, and a consequent, describing its proposed result. English commonly marks this construction through past-tense forms and modal expressions such as “would have.” These forms can signal distance from actuality rather than simply past time. An indicative conditional—“If the switch was pressed, the lamp lit”—usually presents a different conversational stance toward the antecedent. (plato.stanford.edu)
Counterfactuals differ from material implication in propositional logic. Material implication is true whenever its antecedent is false. That rule cannot distinguish a plausible counterfactual from an implausible one when both concern an event that never occurred. Counterfactual evaluation therefore requires more than the actual truth values of its component clauses. (plato.stanford.edu)
Their logic also differs from familiar material-conditional reasoning. Strengthening the antecedent is not generally valid: “If the switch had been pressed, the lamp would have lit” need not imply “If the switch had been pressed and the bulb had been broken, the lamp would have lit.” Likewise, contraposition and transitivity are not generally valid in standard counterfactual systems. These features connect counterfactual reasoning with nonmonotonic reasoning, in which additional information can defeat an earlier inference. (plato.stanford.edu)
Possible-world semantics
An influential approach to counterfactual semantics developed through Robert Stalnaker’s work in 1968 and David Lewis’s work in 1973. It evaluates counterfactuals using possible worlds: alternative ways reality could be. Roughly, “If A were the case, C would be the case” is true when C holds in the relevant A-worlds closest to actuality. This connects counterfactuals with modal logic, while distinguishing them from claims about everything that is possible. (plato.stanford.edu)
The theories differ in important details. Stalnaker’s selection-function approach selects a single closest antecedent-world. Lewis’s approach permits ties and does not require that a closest antecedent-world always exist. His more general formulation compares increasingly close alternatives rather than assuming a uniquely nearest one. (plato.stanford.edu)
Closeness is not necessarily geographical distance or a numerical measure. It concerns relevant similarity, including background circumstances and laws. Different judgments about what should remain fixed can change a counterfactual’s interpretation. Using possible worlds as a semantic framework does not itself require accepting Lewis’s metaphysical view that nonactual worlds are concrete realities. (plato.stanford.edu)
Causal explanation and structural models
Counterfactual theories analyze causal dependence through contrasts: if a cause had not occurred, its effect would not have occurred. Such dependence, however, does not straightforwardly capture every causal relationship. In preemption, an actual cause produces an effect while preventing a backup cause from doing so. The backup could have produced the effect had the actual cause been absent. Simple dependence on the actual cause may therefore fail, motivating more elaborate accounts. (plato.stanford.edu)
A structural causal model represents causal mechanisms through equations linking variables. An intervention replaces the equation governing a selected variable with a specified value, leaving other mechanisms intact. A directed acyclic graph can represent the dependencies of an acyclic model, but counterfactual evaluation also requires assumptions about the underlying mechanisms and background factors. (plato.stanford.edu)
A standard procedure has three stages:
- Abduction: use observed evidence to infer background circumstances.
- Action: modify the model to represent the hypothetical intervention.
- Prediction: calculate the outcome in the modified model using those inferred circumstances.
This distinguishes asking what an intervention would generally produce from asking what it would have produced for a particular observed case. Counterfactual probabilities can express uncertainty about that case’s unobserved background factors. (onlinelibrary.wiley.com)
Potential outcomes and empirical inference
In statistics and causal inference, the potential-outcomes framework assigns each unit an outcome under each specified intervention. For binary treatment, these are often written (Y_i(1)) and (Y_i(0)). Their difference defines an individual causal effect:
[ \tau_i=Y_i(1)-Y_i(0). ]
Only the outcome corresponding to the intervention actually received is observed. The other is counterfactual; comparing different individuals does not directly reveal both outcomes for the same individual. (hsph.harvard.edu)
Population effects, such as the expected value (E[Y(1)-Y(0)]), may nevertheless be identifiable under appropriate assumptions. Common conditions include consistency between observed and potential outcomes, exchangeability of comparison groups, and positivity of intervention assignment. Randomization supports exchangeability, whereas observational research must address confounding. Defining a counterfactual quantity is therefore distinct from establishing that available data identify it. (hsph.harvard.edu)
Thinking and computational explanations
In psychology, counterfactual thinking involves mentally constructing alternatives to past events. Upward counterfactuals imagine better outcomes; downward counterfactuals imagine worse ones. Research examines their relationships with regret, relief, judgments of responsibility, and subsequent behavior. Functional accounts describe how these thoughts can influence intentions and performance through specific lessons or through broader changes in motivation and emotion. (pmc.ncbi.nlm.nih.gov)
In machine learning, a counterfactual explanation identifies changes to an input that would produce a different model output. Within explainable artificial intelligence, this provides a contrastive explanation without necessarily describing the model’s internal operation. A changed prediction, however, is not automatically an established causal effect in the world: the explanation concerns the model’s response, and interpreting it as an intervention requires additional causal assumptions. (arxiv.org)