Mathematical Economics Seminar: Algorithmic Decision-Making

Speaker: Ashesh Rambachan, MIT

Title: Human Decisions and Machine Predictions in Multistage Systems

Abstract: Many high-stakes decisions go through successive stages of human review before a final decision is made, complicating efforts to evaluate both the human decision-makers and algorithmic counterfactuals. We develop a framework for evaluating both humans and algorithms in such multistage systems and apply it to child protective services,where human screeners first decide whether to investigate an allegation and human investigators then decide whether to open a case and provide services. The quasi-random assignment of human decision-makers identifies the effects of deploying algo-rithms throughout the system, even though outcomes are selectively observed across stages. We find three main results when applying our framework to administrative data from Michigan’s CPS system. First, conventional single-stage comparisons show that human screeners significantly underperform an equally selective algorithm buthuman investigators perform roughly as well as the algorithm. Second, these conven-tional single-stage comparisons significantly overstate the gains of deploying algorithms throughout the system because of a double-counting problem, where a case left unad-dressed at both stages is credited to the algorithm twice even though maltreatment can be prevented only once. Third, there is meaningful heterogeneity across the human in-vestigators, with the highest performers gathering and acting on information collected during the investigation itself that is unavailable to the algorithm. These results point to a broader feature of multistage systems, where the scope for humans to complement algorithms grows as additional information becomes available.

Speaker: Kosuke Imai, Harvard University

Title: Triage Score: A Counterfactual Risk Assessment Instrument

Abstract: Risk assessment instruments, also known as “risk scores,” are widely used in high-stakes decision-making settings such as medicine and the criminal justice system. A risk score predicts the likelihood of an undesired outcome if no intervention is made. Thus, a sufficiently high score is often interpreted as a recommendation to intervene. However, risk scores fail to account for what would happen if a decision-maker does intervene. This failure is problematic because effective decision making requires consideration of both or multiple potential outcomes. We propose “triage scores,” which are based on additive counterfactual utilities and include risk scores as a special case. Unlike risk scores, triage scores can incorporate counterfactual outcomes under alternative decisions, enabling decision makers to incorporate a wide range of ethical and practical factors. We illustrate the use of triage scores with an application to our own randomized controlled trial evaluating a pretrial risk score. Our analysis demonstrates that triage scores are able to capture rich utility structures and yield substantively distinct results regarding policy evaluation and learning.