“Evaluation and Influence through Selective Learning of Attributes”
From buyers appraising complex products to policymakers evaluating novel program at pilot sites, important decisions rely on selective exploration of multi-attribute objects. This paper offers a model of learning with correlated attributes. Similarity across attributes is flexibly modeled through a Gaussian process. In the single-player benchmark, the optimal sample is (i) neutral and thorough, (ii) independent of prior attribute means or the sampling format, and (iii) consisting of the most central attributes in a corresponding inference graph, i.e. attaining the highest sample centrality. With separate authorities over sampling and adoption, the agent’s preference for a sample hinges both on its informativeness for the principal and its alignment of players’ interests – reflecting sampling’s dual purposes of influence and learning. We characterize distortions in sample size, content, and strategic delay. Our results have implications for purposive site selection and site selection bias in small-scale program evaluations.