Algorithmic Decision Making and the Cost of Fairness

Explore algorithmic decision-making and fairness in this Simons Institute symposium talk. Delve into the challenges of identifying bias in algorithmic decisions, focusing on a case study of pre-trial decision-making. Examine the limitations of benchmark tests and outcome tests, and understand the concept of infra-marginality. Investigate how to identify bias in human decisions and compare it to algorithmic decision-making. Analyze evidence from Broward County and discuss potential fairness concerns, including redlining and the insufficiency of calibration. Learn about sample bias, label bias, and subgroup validity. Evaluate the use of protected characteristics and statistical parity as measures of fairness. Understand the optimal rule for decision-making and the trade-offs between different fairness criteria. Draw analogies to tests for discrimination and explore the limitations of false positive rates. Gain insights into making fair decisions with algorithms and recognize the limitations of current approaches.

Intro
How do we identify bias in algorithmic decisions?
Case study: Pre-trial decision making
Problems with the benchmark test
The outcome test in Broward County
Risk distributions
The problem with the outcome test
The problem of infra-marginality
Identifying bias in human decisions
Making decisions with algorithms
Evidence from Broward County
Potential fairness concerns
Redlining
Why is calibration insufficient?
Sample bias
Label bias
Subgroup validity
Use of protected characteristics
Statistical parity as a measure of fairness
Where do these disparities come from?
The optimal rule is a single threshold
The fairness/fairness trade-off
Analogies to tests for discrimination
The problem with false positive rates
Making fair decisions with algorithms
Limitations