Causal Bandits Podcast Causality, Experimentation, and Marketplaces | Lawrence De Geest S2E10
Apr 1, 2026
Lawrence De Geest, economist and data scientist at Zoox (formerly Lyft and the NBA), riffs on marketplaces and experimentation. He explains why simple A/B tests fail in marketplaces. He walks through coarse randomization, synthetic-control quasi-experiments, and designs to handle interference. He also discusses supply elasticity, long-term re-equilibration, and practical trade-offs in real-world experiments.
AI Snips
Chapters
Books
Transcript
Episode notes
Use Coarser Randomization To Capture Indirect Effects
- When interference is expected, choose coarser randomization like time-split (switchback) or region-level experiments to capture direct and indirect effects.
- Balance bias and variance: coarser units reduce bias but increase noise, so pick design based on required precision.
Sweep Out Bias By Modeling Dispatch Mechanisms
- You can reduce interference bias by modeling dispatch or allocation mechanisms and 'sweeping out' marketplace-mediated effects.
- Lyft's MMV approach solves dispatch intervals via a linear program to adjust for marginal value and mitigate bias.
Launching In A Few Regions Acts Like An Experiment
- Lyft sometimes launches product in a subset of regions and treats it as an experiment, then evaluates with synthetic controls when randomization isn't possible.
- Synthetic controls work variably and require careful fiddling to match treated units and donor pools.
