Observation and intervention in causal models #
A partial assignment s : ∀ v, Flat (α v) can be read two ways in a causal model. As an
intervention it holds the variables it settles at their values (CausalModel.solve). As an
observation it picks out the contexts whose actual world agrees with it,
CausalModel.contexts, the conditioning step of Pearl's evaluation of a counterfactual.
The solution at a variable depends only on the intervention at that variable and its ancestors
(CausalModel.solve_congr), so an intervention leaves every variable it cannot reach as it was
(CausalModel.solve_update_of_not_reflTransGen).
Main definitions #
CausalModel.world: the actual world of a context, as a total partial assignmentCausalModel.contexts: the contexts in which an observation holds
Main results #
CausalModel.solve_congr: interventions agreeing on a variable's ancestors agree thereCausalModel.solve_update_of_not_reflTransGen: intervening oncleaves whatcdoes not reach unchangedCausalModel.solve_eq_solve_bot_of_mem_contexts: imposing an observation that holds changes nothing
References #
Interventions that agree on a variable and all its ancestors give it the same value.
Intervening on c leaves every variable that c is not an ancestor of unchanged.
M.world u is the actual world of the context u, the solution under no intervention with
every variable settled.
Instances For
M.contexts s is the set of contexts in which the observation s holds, those whose actual
world settles every variable s settles, at the same value.
Instances For
An observation that settles more holds in fewer contexts.
In a context where the observation holds, imposing it as an intervention changes nothing.
An observation extended by a fact of the context's actual world still holds there.