Decision-making selects an action or policy from alternatives with the aim of producing favorable consequences. Because actions cause outcomes, evaluating a decision requires asking what would happen under an intervention, not merely observing what happened when an action occurred. section 12.0 section 12.7
A policy maps information about a decision problem to an action. Bandit and reinforcement-learning systems search for policies that optimize outcomes over time, while probabilistic models represent randomness between an action and its consequences. section 8.1.5 section 12.0
Causal decision analysis distinguishes several related questions:
- Compare the observed outcome associated with an action, E(U(Y | X = x)), with the intervention outcome, E(U(YX=x)); the latter answers what would happen if the decision were imposed. section 12.7
- Choose the action that maximizes the expected utility of its consequences rather than relying only on predictive associations. section 12.7
- Represent sequential decisions with causal DAGs to expose confounder, mediator, and collider bias; policies behave like stochastic interventions while depending on past observations. section 12.7
- Use counterfactual simulations to compare alternative policies, estimate regret, and learn from conditions that did not occur naturally. section 8.1.5
Business and sequential decisions
A streaming service can ask whether subscribers would have left had a content deal failed, separating the deal’s causal contribution from the attraction of other content. An investing algorithm can backtest portfolio policies, while a game-playing system can replay saved games under alternative policies and use the resulting regret to improve. section 8.1.2 section 8.1.5
