Quantify persuasion against a downstream decision maker’s interests

Determine whether an experimenter can persuade a downstream decision maker to act against the decision maker’s own interests and, if so, quantify the extent of that influence within usage-aware experimental design.

Background

The paper considers experimental design when data are acted on by a downstream decision maker whose prior, loss function, or decision rule may differ from those of the experimenter. In this setting, the experimenter chooses the data-collection policy but cannot alter the likelihood generating the observations, and the downstream decision maker may observe all collected data.

The authors explain that a Bayesian downstream decision maker using their own Bayes act cannot be harmed in expectation by receiving data when the decision maker shares the likelihood and observes the data-collection process. They then identify two possible avenues for influence: selecting among informative experiments that favor the experimenter, and exploiting downstream rules that are not Bayes-optimal. The unresolved aspect is whether such influence can induce action contrary to the decision maker’s own interests and how large that effect can be.