Attributes: Selective Learning and Influence
When different stages of the evaluation of a multi-attribute project rest with conflicting economic actors, which attributes are selectively explored and why? We provide a model of attribute sampling in which correlation across attributes is flexibly modeled through Gaussian processes. In the absence of conflict, the optimal sample of attributes maximizes informativeness by balancing out-of-sample extrapolation with correlation within the sample. It depends neither on the prior value of the project nor on the format of sampling. Agency conflict, in contrast, gives rise to distortions. Sampling serves a dual purpose of generating valuable information and influencing the co-player. When influence takes priority, optimal sampling either suppresses informativeness for both players or negatively correlates their interests. Casting site selection as an attribute problem, our framework provides a theoretical rationale for site selection bias in small-scale program evaluation.