Proper Scoring Rules for Unobservable Outcomes: An Application to the Peer-Review Process

When eliciting private information from a group of experts, traditional devices used to promote honest reporting assume that there is an observable future outcome. In practice, this assumption is not always applicable. For example, usually there is no observable outcome, or a ground-truth review, in the peer-review process. With that in mind, we propose a scoring method built on strictly proper scoring rules to induce honest behavior without assuming observable outcomes. Our method provides scores based on pairwise comparisons of the reports made by each pair of experts in the group. For ease of exposition, we show how the scoring method can be applied to the peer-review process. We start by modeling this process using a Bayesian model where the uncertainty regarding the quality of the manuscript is taken into account. Thereafter, we introduce a scoring function to evaluate the reported reviews. Under the assumptions that reviewers are Bayesian decision-makers and that they cannot influence the reviews of other reviewers, we show that risk-neutral reviewers strictly maximize their expected scores by honestly disclosing their reviews. We also show how the group's scores can be used to find a consensual review. An implication of our model is that the distribution of the reported reviews converges to the probability distribution that represents the quality of the manuscript as the number of honest reviews increases. Experimental results show that encouraging honesty through the proposed scoring method creates more accurate reviews than the traditional peer-review process, thus corroborating our theoretical results.