Camelia
summary
A new threshold test claims to directly estimate search thresholds and risk distributions using a Bayesian hierarchical latent‑variable model applied to 4.5 million North Carolina stops and the Stanford Open Policing dataset, finding aggregated thresholds of about 23% for white drivers, 20.6% for black drivers and 14% for Hispanic drivers; inference swaps a beta for an LDA‑based discriminant family to make gradient computations fast and yields a 150× speed‑up and tractable statewide analysis.
AI-written summary · about
chapters (15)
- 0:00Start
- 1:45North Carolina Data Set
- 2:24Outcome Test
- 3:23Problem of infra-marginality
- 6:22Modeling a Traffic Stop
- 7:03Parameterizing the Risk Distribution
- 7:32Simplifying inference
- 8:23Graphical Model Representation
- 8:46Speeding up inference
- 9:12Mixture of Gaussians
- 10:27Discriminant distributions
- 11:28Key calculations in the threshold test
- 12:07A threshold test practical for broad use
- 13:28Posterior Predictive Checks
- 13:44Conclusions
related talks
40mEvaluating fairness in machine learning with PyMC3
15mExploring the math of thresholds
30mNon-parametric Causal Inference in Dynamic Thresholding Designs
46mThe Measure and Mismeasure of Fairness
28mUsing Changepoint and Bayesian Analysis to Drive Safety Improvements in Mining
38mTracking and Improving Information in the Service of Fairness
31mBias Meets Bayes: A Bayesian Perspective on Improving Model Fairness
45mBias in Artificial Intelligence