UCLA

Unobserved confounders (UCs) are factors in a system that affect a treatment and its outcome, but whose states are unknown. When left uncontrolled, UCs present a major obstacle to inferring causal relations from statistical data, which can impede policy making and machine learning. Control of UCs has traditionally been accomplished by randomizing treatments, thus severing any causal influence of the UCs to the treatment assignment, and averaging their effects on the outcome in each randomized group. Although such interventional data can be used to appropriately inform population-level decisions, unit-level decisions are best informed by counterfactual quantities that provide information about the UCs relating to each unit. That said, arbitrary counterfactual computation can be performed in only certain scenarios, or in possession of a fully-specified causal model that requires knowledge of the distribution over UC states.

This work describes how additional information from a deciding agent can be utilized to empirically estimate certain counterfactuals, even in the presence of UCs and the absence of a fully-specified model of reality. The resulting technique yields strictly more information than standard randomization, and is specialized to personal decision-making. We first formalize this new strategy, called Intent-specific Decision-making (ISDM), in the context of the tools provided by causal inference. We then demonstrate its utility in online, reinforcement learning tasks with UCs, and support the efficacy of our technique in both human-subject and simulation experiments. We demonstrate how ISDM accommodates a fusion of observational, experimental, and counterfactual data, which can be used to accelerate policy learning. Finally, we extend ISDM to the offline experimental design domain, detailing its application toward improving the established randomized clinical trial.