Andrew Hillis’s research while affiliated with Harvard University and other places

Research Summary We evaluate a pilot in an Inspections Department to explore the returns to a pair of algorithms that varied in their sophistication. We find that both algorithms provided substantial prediction gains, suggesting that even simple data may be helpful. However, these gains did not result in improved decisions. Inspectors often used their decision authority to override algorithmic recommendations, partly to consider other organizational objectives without improving outcomes. Interviews with 55 departments find that while some ran pilots seeking to prioritize inspections using data, all provided considerable decision authority to inspectors. These findings suggest that for algorithms to improve managerial decisions, organizations must consider both the returns to algorithms in the context and how decision authority is managed. Managerial Summary We evaluate a pilot in an Inspections Department to explore the returns to algorithms on decisions. We find that the greatest gains in this context come from integrating data into the decision process in the form of simple heuristics, rather than from increasing algorithmic sophistication or additional data. We also find that these improvements in prediction do not fully translate into improved decisions. Decision‐makers were less likely to follow data‐driven recommendations, partly in consideration of other organizational objectives, but without substantially improving on them overall. These findings suggest that organizations should consider the returns to technical sophistication in each context, and that the design and management of decision authority can be a key choice that impacts the value organizations can capture from using predictive analytics.

Economists have become increasingly interested in studying the nature of production functions in social policy applications, with the goal of improving productivity. Traditionally models have assumed workers are homogenous inputs. However, in practice, substantial variability in productivity means the marginal productivity of labor depends substantially on which new workers are hired--which requires not an estimate of a causal effect, but rather a prediction. We demonstrate that there can be large social welfare gains from using machine learning tools to predict worker productivity, using data from two important applications - police hiring and teacher tenure decisions.

Can open tournaments improve the quality of city services? The proliferation of big data makes it possible to use predictive analytics to better target services like hygiene inspections, but city governments rarely have the in-house talent needed for developing prediction algorithms. Cities could hire consultants, but a cheaper alternative is to crowdsource competence by making data public and offering a reward for the best algorithm. This paper provides a simple model suggesting that open tournaments dominate consulting contracts when cities have a reasonable tolerance for risk and when there is enough labor with low opportunity costs of time. We also illustrate how tournaments can be successful, by reporting on a Boston-based restaurant hygiene prediction tournament that we helped coordinate. The Boston tournament yielded algorithms—at low cost—that proved reasonably accurate when tested “out-of-sample” on hygiene inspections occurring after the algorithms were submitted. We draw upon our experience in working with Boston to provide practical suggestions for governments and other organizations seeking to run prediction tournaments in the future.