Assumptions and Identifiability

Causal OPE relies on explicit assumptions. Each estimator lists its required assumptions in the API and documentation.

Core assumptions

• Sequential ignorability: no unmeasured confounding given observed history.
• Overlap: behavior policy supports target policy actions.
• Markov: state is sufficient for future evolution.
• Behavior policy known: propensities are known or correctly specified.
• Q-model realizability: value function lies in the chosen model class.
• Bridge identifiability: proximal bridge functions are well-posed.
• Bounded rewards: needed for high-confidence bounds.
• Bounded confounding: hidden bias is limited by a sensitivity parameter.

Practical guidance

• Diagnose overlap before trusting IS-based estimators.
• Prefer model-based or DR estimators when overlap is weak.
• Use sensitivity analysis when ignorability is questionable.