Human in the loop : the role of individual and institutional behavior on predictive algorithms

Over the past decade, algorithmic decision systems (ADS)--applications of statistical or computational techniques designed to assist human-decision making processes--have moved from an obscure domain of statistics and computer science into the mainstream. The rapid decline in the cost of computer processing and ubiquity of digital data storage has created a dramatic rise in the adoption of ADS using applied machine learning algorithms, transforming various sectors of society from digital advertising to political campaigns, risk modeling for the banking sector, healthcare and beyond. Many agencies and practitioners in the public sector turn to ADS as a means to stretch limited public resources amidst growing public demands for equity and accountability. However, recent research from multiple fields has found that social and institutional biases, often reflected by input data used to generate predictions. The potential of perpetuated discrimination via input data is a particular concern in fields such as criminal justice where historical biases against minorities have the potential to exacerbate existing racial inequalities. In a series of three essays, this dissertation seeks to outline how institutional norms often shape algorithmic predictions, examine how ADSs alter the incentive structures for agents using the tools, and ultimately its impact on human decision-making.