The SAT math gap- gender difference or selection bias
summary
A Bayesian model uses PreliZ for priors, PyMC for inference and ArviZ for validation to estimate latent efficacy, a logistic selection curve, and a Gaussian random walk that models score spikes. Hierarchical fitting and parameter recovery tests work well. Estimated latent means are 508 (male) and 502 (female), implying selection bias plausibly explains about 24 of the observed 30‑point gap (~80%). The approach quantifies a specific claim and demonstrates practical identifiability rather than asserting a topic.
AI-written summary · about
chapters (11)
- 0:00SAT gender gap debate
- 1:07Persistent SAT math gap
- 3:35Selection bias measurement
- 5:44Cheating baker fingerprints
- 9:39Modeling SAT score distributions
- 14:53Bayesian model for SAT data
- 19:362015 male test takers
- 24:29Hierarchical modeling years
- 26:57Selection bias plausibility
- 28:04Next steps state-level SAT
- 29:27Q&A
related talks
31mBias Meets Bayes: A Bayesian Perspective on Improving Model Fairness
18mConstantine
01h 31mHow To Measure And Mitigate Unfair Bias in Machine Learning Models
30mBuilding fairer models for finance
30mProbabilistic modelling with Scala
38mUsable A/B testing – A Bayesian approach
01h 20mBayesian Inference for Fun and Profit
33mHidden Biases In Data: Moving Beyond Black Box Models