Learning Bayesian Statistics #154 Bayesian Causal Inference at Scale, with Thomas Pinder
Mar 25, 2026
Thomas Pinder, statistician and GPJax creator focused on Gaussian processes and Bayesian causal inference. He explains why GPJax was built for flexible, high-performance GP research. He discusses Bayesian approaches to causal inference, synthetic controls and SDID, and practical tradeoffs for scaling GPs and using hierarchical models in industry.
AI Snips
Chapters
Transcript
Episode notes
PhD GPs Led To Causal Work At Amazon
- Thomas Pinder moved from GP research in academia to applied experimentation at Amazon, where he worked on synthetic controls with Alberto Abadie.
- He described using GPs to accelerate Horvitz–Thompson estimators in supply chain internships, which convinced him industry could host deep statistical work.
GPJax Is A Researcher-Focused GP Toolbox
- GPJax was created to give researchers modular GP building blocks, not a guarded black-box GP solver.
- Thomas built it in JAX to leverage functional programming, autodiff, and JIT compilation for flexible, fast GP experimentation.
Bayes Gives Stakeholders Actionable Uncertainty
- Bayesian methods naturally fit causal questions because they deliver full posterior uncertainty rather than a binary p-value decision.
- Thomas stresses stakeholders benefit from probabilities and credible intervals to assess risk and make nuanced decisions.